Add OpenAI compatible chat backend

.env.example

@@ -43,6 +43,18 @@ ONLY_FOR_CREATOR_MODE=false
 # Use user names in AI prompts
 USE_NAMES_IN_PROMPT=true
 
+# Disable all built-in local tools and keep only MCP tools
+DISABLE_LOCAL_TOOLS=false
+
+# Filter built-in local tools by name.
+# LOCAL_TOOL_ALLOWLIST lets through only the listed tools.
+# LOCAL_TOOL_DENYLIST removes the listed tools.
+# Examples:
+# LOCAL_TOOL_ALLOWLIST=get_datetime,web_search
+# LOCAL_TOOL_DENYLIST=shell_execute,python_interpreter
+LOCAL_TOOL_ALLOWLIST=
+LOCAL_TOOL_DENYLIST=
+
 # Custom system prompt for AI (or put it into data/SYSTEM_PROMPT.md)
 SYSTEM_PROMPT=
 
@@ -91,6 +103,10 @@ OLLAMA_MAX_CONCURRENT_REQUESTS=1
 # OpenAI
 OPENAI_API_KEY=
 OPENAI_BASE_URL=
+# Backend mode:
+# official = OpenAI responses API
+# compatible = OpenAI-compatible chat.completions servers like llama.cpp
+OPENAI_BACKEND=official
 OPENAI_MODEL=gpt-4.1-nano
 OPENAI_IMAGE_MODEL=gpt-image-1-mini
 OPENAI_TRANSCRIPTION_MODEL=gpt-4o-mini-transcribe
@@ -99,6 +115,14 @@ OPENAI_TTS_VOICE=alloy
 OPENAI_TTS_INSTRUCTIONS=
 OPENAI_MAX_CONCURRENT_REQUESTS=3
 
+# MCP servers
+# JSON array or {"mcpServers": {"name": {...}}}
+# Stdio example:
+# MCP_SERVERS=[{"name":"local-tools","transport":"stdio","command":"node","args":["./mcp-server.js"]}]
+# HTTP example:
+# MCP_SERVERS=[{"name":"remote-tools","transport":"http","url":"https://example.com/mcp"}]
+MCP_SERVERS=
+
 # Per-capability AI endpoint overrides
 # Pattern:
 # <PROVIDER>_<CAPABILITY>_MODEL=
@@ -113,6 +137,7 @@ OPENAI_MAX_CONCURRENT_REQUESTS=3
 # OLLAMA_ADDRESS or OLLAMA_BASE_URL.
 # Capability aliases are also supported: IMAGE, THINK, RAG, EMBEDDING,
 # TRANSCRIPTION, STT, TTS.
+# Backend override: OPENAI_BACKEND=official|compatible.
 #
 # Examples:
 # OPENAI_SPEECH_TO_TEXT_MODEL=gpt-4o-mini-transcribe

OPENAI_COMPATIBLE_TARGET_IMPLEMENTATION.md

@@ -0,0 +1,314 @@
+# OPENAI Compatible Target Implementation
+
+## Purpose
+
+Add a separate execution path for OpenAI-compatible backends such as `llama.cpp`, while keeping the current official OpenAI path unchanged.
+
+## Checklist
+
+- [x] Add explicit OpenAI backend mode in config
+- [x] Route OpenAI requests to separate official and compatible runners
+- [x] Keep official OpenAI on `responses.create(...)`
+- [x] Add compatible `chat.completions.create(...)` runner
+- [x] Add compatible tool-call extractors
+- [x] Add backend selection tests
+- [x] Add basic memory/config regression coverage
+- [x] Normalize compatible streaming tool-call assembly
+- [x] Preserve file upload behavior in compatible backend
+- [x] Guard unsupported OpenAI-only tools for compatible backend
+- [x] Add environment docs and example config entries
+- [x] Add real-server integration coverage for compatible backend
+- [x] Revisit shared orchestration extraction for further deduplication
+
+## Non-Goals
+
+1. Do not change the current official OpenAI `responses.create(...)` behavior.
+2. Do not auto-switch behavior only because `OPENAI_BASE_URL` is set.
+3. Do not merge compatible backend quirks into the official OpenAI runner.
+4. Do not remove or weaken existing tool ranking, memory, RAG, logging, or upload behavior.
+
+## Current State
+
+1. `src/ai/unified-ai-runner.openai.ts` currently uses the official `responses` API.
+2. `src/ai/provider-adapters.ts` already has provider-specific adapters and tool/result mapping.
+3. `src/ai/provider-adapter-contract.ts` already contains `responses`-style extractors.
+4. `src/ai/openai-chat-message.ts` currently models `responses`-style messages, not `chat.completions` tool messages.
+5. `src/ai/unified-ai-request-pipeline.ts` prepares chat context and runtime state before the model call.
+6. `src/ai/ai-runtime-target.ts` resolves provider targets, base URLs, models, and keys.
+7. `src/ai/unified-ai-runner.tool-ranker.ts` already uses a `chat.completions`-style call path, which is closer to compatible backends.
+
+## Target Architecture
+
+1. Official OpenAI backend stays on `responses.create(...)`.
+2. Compatible OpenAI backend uses `chat.completions.create(...)`.
+3. Backend selection is explicit through config, for example `OPENAI_BACKEND=official|compatible`.
+4. Shared preparation logic remains common.
+5. Transport-specific request formatting and response parsing are split.
+
+## Configuration Design
+
+1. Add a new config value `OPENAI_BACKEND`.
+2. Allowed values should be `official` and `compatible`.
+3. Default must be `official`.
+4. Keep `OPENAI_BASE_URL` as a transport setting only.
+5. `OPENAI_BASE_URL` must not imply compatible mode by itself.
+6. Extend environment schema and runtime config to expose this value.
+7. Update env docs and example env files.
+
+## Step 1: Config and Target Selection
+
+1. Update `src/common/environment.ts`.
+2. Add a new environment field for backend mode.
+3. Add setters if the codebase uses runtime env mutation in tests.
+4. Update the startup schema and runtime snapshot.
+5. Add tests for default value and explicit `compatible` selection.
+
+Expected result:
+- Official OpenAI stays unchanged by default.
+- Explicit `OPENAI_BACKEND=compatible` selects the new execution path.
+
+## Step 2: Split Runner Selection
+
+1. Update the unified AI execution entry point.
+2. Add a small backend selector for OpenAI targets.
+3. Route official mode to the current runner.
+4. Route compatible mode to a new compatible runner.
+5. Keep other providers untouched.
+
+Expected result:
+- One codepath for official OpenAI.
+- One codepath for OpenAI-compatible servers.
+
+## Step 3: Shared Orchestration Extraction
+
+1. Identify logic that is identical for both OpenAI branches.
+2. Extract common orchestration into a shared helper where possible.
+3. Keep these pieces shared:
+   - memory prompt injection
+   - tool ranking
+   - tool loop control
+   - logging and timing
+   - cancellation handling
+   - file upload post-processing
+   - document RAG preparation and cleanup
+4. Keep transport-specific pieces separate:
+   - request shape
+   - response parsing
+   - tool result message shape
+   - streaming event parsing
+
+Expected result:
+- Less duplicate logic.
+- Cleaner separation between official and compatible behavior.
+
+## Step 4: Compatible Message Model
+
+1. Update `src/ai/openai-chat-message.ts` or create a sibling type file for compatible chat messages.
+2. Model `system`, `user`, `assistant`, and `tool` roles explicitly.
+3. Support `tool_calls` on assistant messages.
+4. Support `tool_call_id` on tool result messages.
+5. Preserve support for text and multimodal user content where the backend supports it.
+6. Avoid forcing `responses` output types into `chat.completions`.
+
+Expected result:
+- Compatible runner can build valid `chat.completions` message arrays.
+
+## Step 5: Compatible Contract Extractors
+
+1. Extend `src/ai/provider-adapter-contract.ts`.
+2. Add extractors for `chat.completions` tool calls.
+3. Add extractors for `chat.completions` streaming tool call deltas.
+4. Keep existing `responses` extractors intact.
+5. Normalize tool call IDs, names, and argument text the same way as existing extractors.
+6. Ensure arguments are always represented as JSON text for the tool loop.
+
+Expected result:
+- Compatible runner can parse tool calls from both normal and streaming responses.
+
+## Step 6: Compatible Provider Adapter
+
+1. Update `src/ai/provider-adapters.ts`.
+2. Add a separate adapter or branch for OpenAI-compatible chat.completions behavior.
+3. Reuse existing tool ranking where safe.
+4. Make `appendToolResults(...)` emit `role: "tool"` messages with `tool_call_id`.
+5. Keep official OpenAI adapter outputting `function_call_output`.
+6. Keep Mistral and Ollama unchanged.
+
+Expected result:
+- Each backend uses the tool result shape it expects.
+
+## Step 7: Compatible Runner Implementation
+
+1. Create a new file such as `src/ai/unified-ai-runner.openai-compatible.ts`.
+2. Use `openai.chat.completions.create(...)`.
+3. Pass `messages`, `tools`, `model`, `stream`, and `signal`.
+4. Map system prompt and memory prompt into the `messages` array correctly.
+5. Keep the tool loop structure from the current runner.
+6. Append assistant tool-call messages and tool result messages between rounds.
+7. Continue until no tool calls remain or max rounds is reached.
+
+Expected result:
+- Compatible backends can complete multi-round tool flows.
+
+## Step 8: Tool Call Loop Semantics
+
+1. Preserve `MAX_TOOL_ROUNDS`.
+2. Preserve tool ranking before each round.
+3. Preserve memory tool selection.
+4. Preserve file search injection when document RAG is active.
+5. Preserve file upload post-processing.
+6. Preserve max-rounds warnings and continuation decisions.
+7. Keep the final text visible in the stream message exactly as today.
+
+Expected result:
+- Compatible backend behaves like the current runner from the user’s perspective.
+
+## Step 9: Streaming Behavior
+
+1. Implement streaming event handling for `chat.completions`.
+2. Parse text deltas and append them to `TelegramStreamMessage`.
+3. Parse `delta.tool_calls` and keep incremental tool-call state.
+4. Update status text when tool usage starts and ends.
+5. Keep image generation and file-search status handling if the backend emits compatible signals.
+6. Finalize the stream only after the terminal completion event.
+
+Expected result:
+- Streaming works without losing tool call state.
+
+## Step 10: Tool Result Handling
+
+1. After each tool execution round, append tool results using the compatible message format.
+2. Ensure each tool result keeps the correct `tool_call_id`.
+3. Preserve the existing file upload hook.
+4. If upload fails, convert the failure into a tool result error string.
+5. Preserve the same tool memory map behavior.
+
+Expected result:
+- The backend receives a valid message history for the next round.
+
+## Step 11: Prompt and Memory Injection
+
+1. Keep `buildSystemInstruction(...)` as the source of system prompt assembly.
+2. Keep `buildUserMemoryPrompt(...)` injected as a separate block.
+3. Preserve the explicit separation between assistant memory and user memory.
+4. Preserve the `user.md` and `system.md` memory layout.
+5. Ensure compatible backend receives the same semantic prompt content.
+
+Expected result:
+- Memory behavior stays identical across official and compatible backends.
+
+## Step 12: Tool Ranking Compatibility
+
+1. Review `src/ai/unified-ai-runner.tool-ranker.ts`.
+2. Verify whether the current JSON response handling is safe for compatible backends.
+3. If a backend cannot guarantee strict JSON mode, add a fallback parser.
+4. Keep ranking inputs and outputs consistent across both branches.
+5. Do not weaken tool selection heuristics.
+
+Expected result:
+- Tool ranking remains deterministic enough for both branches.
+
+## Step 13: File Search and RAG
+
+1. Keep document RAG preparation in the request pipeline.
+2. Keep vector store preparation for official OpenAI.
+3. Decide whether compatible backend supports file search or needs a no-op fallback.
+4. If unsupported, guard the tool list so the compatible backend never receives unsupported tools.
+5. Keep cleanup behavior for temporary artifacts.
+
+Expected result:
+- Compatible backend does not receive tools it cannot execute.
+
+## Step 14: Error Handling
+
+1. Preserve abort handling.
+2. Preserve response failure handling.
+3. Preserve stream error handling.
+4. Surface backend-specific incompatibilities as explicit errors.
+5. Do not silently fall back from compatible to official mode.
+6. Keep logs actionable.
+
+Expected result:
+- Failures are obvious and debuggable.
+
+## Step 15: Logging and Observability
+
+1. Keep the current AI logs and duration tracking.
+2. Add backend mode to log metadata.
+3. Log tool calls, tool outputs, and round transitions in both branches.
+4. Preserve existing observability hooks.
+5. Add explicit labels for official vs compatible runs.
+
+Expected result:
+- Debugging remains easy after the split.
+
+## Step 16: Tests
+
+1. Add unit tests for backend selection.
+2. Add unit tests for compatible message conversion.
+3. Add unit tests for compatible tool call extraction.
+4. Add integration tests for a tool-call round trip using mocked `chat.completions`.
+5. Add tests proving the official `responses` path is unchanged.
+6. Add tests for streaming tool call parsing if the backend supports it.
+7. Add tests for fallback behavior in the tool ranker if needed.
+
+Expected result:
+- Both branches are covered and regressions are visible quickly.
+
+## Step 17: Suggested File Changes
+
+1. `src/common/environment.ts`
+2. `src/ai/ai-runtime-target.ts`
+3. `src/ai/unified-ai-request-pipeline.ts`
+4. `src/ai/unified-ai-runner.openai.ts`
+5. `src/ai/unified-ai-runner.openai-compatible.ts`
+6. `src/ai/provider-adapter-contract.ts`
+7. `src/ai/provider-adapters.ts`
+8. `src/ai/openai-chat-message.ts`
+9. `src/ai/unified-ai-runner.tool-ranker.ts`
+10. `test/*.test.mjs`
+11. `.env.example`
+12. Documentation files for backend selection
+
+## Implementation Order
+
+1. [x] Add config flag and wire it through environment parsing.
+2. [x] Add backend selection logic.
+3. [x] Add compatible message and extractor support.
+4. [x] Create the compatible runner.
+5. [x] Reuse shared orchestration where possible.
+6. [x] Wire tests.
+7. [x] Verify official behavior is unchanged.
+8. [x] Verify compatible backend works with a real OpenAI-compatible server.
+
+## Verification Plan
+
+1. Run unit tests.
+2. Run integration tests.
+3. Verify official OpenAI path still uses `responses.create(...)`.
+4. Verify compatible path uses `chat.completions.create(...)`.
+5. Verify a `llama.cpp`-style server can complete a tool loop.
+6. Verify memory tools still work.
+7. Verify document RAG and file upload behavior do not regress.
+
+## Risks
+
+1. Some OpenAI-compatible servers do not support every official OpenAI feature.
+2. Streaming tool call deltas may differ across providers.
+3. JSON-mode assumptions in the ranker may not hold for all compatible servers.
+4. Tool schema filtering may need backend-specific allowlists.
+5. Message conversion mistakes can break tool loops silently if not tested.
+
+## Acceptance Criteria
+
+1. Official OpenAI behavior is unchanged.
+2. Compatible backend can run a full chat loop with tools.
+3. Tool calls are correctly extracted and executed.
+4. Tool results are appended in the correct format.
+5. Memory injection still works.
+6. Document RAG and file upload behavior remain functional or fail explicitly.
+7. Tests cover both branches.
+
+## Final Note
+
+The key design rule is simple: keep official OpenAI `responses` behavior intact, and introduce OpenAI-compatible `chat.completions` behavior as a separate backend mode with its own parsing and message shape.

PIPELINE_TODO.md

@@ -1,155 +0,0 @@
-# User Request Pipeline TODO
-
-Этот чеклист описывает оставшиеся задачи по доведению pipeline до чистой архитектуры. Текущее состояние уже рабочее: есть `UserRequestPipeline`, stage audit, `ai_requests`, internal artifacts, unified size gate, RAG/STT/final/error/tool-result artifacts и response pipeline. Ниже перечислены задачи, которые ещё нужно сделать, чтобы убрать оставшиеся архитектурные компромиссы.
-
-## 1. Нормализовать хранение attachments, artifacts и audit
-
-- [x] Создать отдельную таблицу `attachments`.
-- [x] Поля `attachments`: `id`, `messageChatId`, `messageId`, `direction`, `scope`, `kind`, `artifactKind`, `fileId`, `fileUniqueId`, `fileName`, `mimeType`, `cachePath`, `sizeBytes`, `sha256`, `metadata`, `createdAt`.
-- [x] Создать отдельную таблицу `artifacts`.
-- [x] Поля `artifacts`: `id`, `requestId`, `messageChatId`, `messageId`, `kind`, `stage`, `attachmentId`, `payload`, `createdAt`.
-- [x] Создать отдельную таблицу `request_audit`.
-- [x] Поля `request_audit`: `id`, `requestId`, `messageChatId`, `messageId`, `stage`, `status`, `startedAt`, `finishedAt`, `durationMs`, `provider`, `model`, `details`, `error`.
-- [x] Оставить обратную совместимость с текущими JSON-полями `messages.attachments` и `messages.pipelineAudit`.
-- [x] Добавить миграцию: переносить существующие `messages.attachments` в новую таблицу `attachments`.
-- [x] Добавить миграцию: переносить существующие `messages.pipelineAudit` в новую таблицу `request_audit`.
-- [x] Обновить backup/export/import, чтобы новые таблицы попадали в JSON и SQL dump.
-- [x] Добавить DAO/store слой: `AttachmentStore`, `ArtifactStore`, `RequestAuditStore`.
-- [x] Перевести новые записи на нормализованные таблицы.
-- [x] Оставить чтение legacy JSON только как fallback.
-
-## 2. Сделать единый ArtifactStore API
-
-- [x] Ввести `ArtifactStore.put(...)`.
-- [x] Ввести `ArtifactStore.getByRequestId(requestId)`.
-- [x] Ввести `ArtifactStore.getByMessage(chatId, messageId)`.
-- [x] Ввести `ArtifactStore.getLatestRagForReplyChain(chatId, messageId)`.
-- [x] Ввести `ArtifactStore.getTranscriptForMessage(chatId, messageId)`.
-- [x] Перевести `rag-artifact-store.ts` на `ArtifactStore`.
-- [x] Перевести `transcript-artifact-store.ts` на `ArtifactStore`.
-- [x] Перевести `final-response-artifact-store.ts` на `ArtifactStore`.
-- [x] Перевести `tool-result-artifact-store.ts` на `ArtifactStore`.
-- [x] Оставить физические JSON-файлы как storage backend для payload, но регистрировать их в БД.
-- [x] Добавить единый size gate для artifact payload до записи файла.
-- [x] Добавить cleanup policy для временных/устаревших artifact файлов.
-
-## 3. Расширить RAG artifact content
-
-- [x] Расширить общий тип `RagArtifact`.
-- [x] Для Ollama сохранять extracted documents.
-- [x] Для Ollama сохранять selected chunks.
-- [x] Для Ollama сохранять chunk scores.
-- [x] Для Ollama сохранять skipped documents и причины пропуска.
-- [x] Для Ollama сохранять embedding model, `topK`, `chunkSize`, `chunkOverlap`, `maxContextChars`.
-- [x] Для OpenAI сохранять `vectorStoreIds`.
-- [x] Для OpenAI сохранять source file mapping: local attachment -> uploaded/vector store file.
-- [x] Для Mistral сохранять `libraryId`.
-- [x] Для Mistral сохранять uploaded document ids.
-- [x] Для Mistral сохранять source file mapping: local attachment -> Mistral document id.
-- [ ] Добавить единый `providerState` schema для всех providers.
-- [ ] Добавить tests на сериализацию `RagArtifact`.
-- [ ] Добавить tests на то, что internal RAG artifacts не попадают обратно в user document context.
-
-## 4. Вынести provider runners в adapter layer
-
-- [ ] Ввести интерфейс `AiProviderAdapter`.
-- [ ] Методы adapter-а: `mapMessages`, `rankTools`, `callModel`, `extractTextDelta`, `extractToolCalls`, `appendToolResults`, `finalize`.
-- [ ] Реализовать `OpenAiProviderAdapter`.
-- [ ] Реализовать `MistralProviderAdapter`.
-- [ ] Реализовать `OllamaProviderAdapter`.
-- [ ] Перенести provider-specific tool schema mapping внутрь adapter-ов.
-- [ ] Перенести provider-specific streaming parsing внутрь adapter-ов.
-- [ ] Перенести provider-specific tool result append внутрь adapter-ов.
-- [ ] Упростить `runOpenAi`, `runMistral`, `runOllama` или заменить их adapter-driven runner-ом.
-- [ ] Оставить compatibility wrappers для текущих imports.
-- [ ] Добавить tests на adapter contract без реальных API.
-
-## 5. Сделать tool-ranker полноценным pipeline stage
-
-- [ ] Вынести вызов `ToolRanker.selectTools(...)` из provider runners.
-- [ ] Добавить stage `tool_rank`, который работает через provider adapter.
-- [ ] Добавить stage `filter_tools`, который фильтрует provider-specific tools по результату ranker.
-- [ ] Хранить `ToolRankDecision` в `UserRequestPipelineState.toolRankDecisions`.
-- [ ] Сохранять `ToolRankDecision` в `request_audit.details`.
-- [ ] Убрать дублирующий ручной `tool-rank-audit.ts`, если stage полностью заменит его.
-- [ ] Сохранить status UX: `🧩 Выбираю подходящие инструменты...`.
-- [ ] Гарантировать `clearStatus()` после ranker success/failure.
-- [ ] Добавить fallback через `PipelineFallbackExecutor`: main model, all tools, no tools.
-- [ ] Добавить tests на fallback ranker policy.
-
-## 6. Сделать model_call и tool_loop физически отдельными stages
-
-- [ ] Stage `model_call` должен делать только один model request.
-- [ ] Stage `model_call` должен возвращать normalized model output.
-- [ ] Stage `tool_loop` должен решать, есть ли tool calls.
-- [ ] Stage `tool_loop` должен выполнять tools через общий `executeToolBatch`.
-- [ ] Stage `tool_loop` должен добавлять tool results в provider adapter.
-- [ ] Stage `tool_loop` должен управлять max rounds.
-- [ ] Stage `tool_loop` должен сохранять tool result artifacts.
-- [ ] Stage `tool_loop` должен уметь завершаться без tools как `skipped`.
-- [ ] Убрать tool loop из `runOpenAi`.
-- [ ] Убрать tool loop из `runMistral`.
-- [ ] Убрать tool loop из `runOllama`.
-- [ ] Добавить tests на multi-round fake adapter.
-
-## 7. Довести fallback notifications до централизованного UX
-
-- [ ] Добавить `PipelineFallbackNotifier`.
-- [ ] Для `notify_user` отправлять пользователю понятное сообщение.
-- [ ] Для `continue_without_stage` писать короткий debug/audit без user notification.
-- [ ] Для `use_alternate_target` логировать исходный и alternate target.
-- [ ] Для `fail_request` завершать request через единый error path.
-- [ ] Добавить локализацию fallback messages.
-- [ ] Добавить отдельные тексты для RAG failure, STT failure, TTS failure, tool failure.
-- [ ] Не спамить пользователя несколькими fallback notifications за один request.
-- [ ] Сохранять fallback notification в `request_audit.details`.
-
-## 8. Улучшить поведение reply-chain с документами
-
-- [ ] Явно описать стратегию merge: current user attachments + reply-chain user attachments.
-- [ ] Исключать `scope: internal_artifact` всегда.
-- [ ] Исключать `scope: bot_output`, если это не user-provided file.
-- [ ] Если пользователь отвечает новым документом на ответ бота с предыдущим документом, использовать оба документа.
-- [ ] Если пользователь отвечает текстом на ответ бота, использовать документы из reply-chain.
-- [ ] Если пользователь явно говорит "этот файл", приоритет отдавать новому вложению.
-- [ ] Если несколько документов, добавлять их имена в prompt/RAG context.
-- [ ] Добавить tests на follow-up с новым документом.
-- [ ] Добавить tests на follow-up без нового документа.
-- [ ] Добавить tests на то, что RAG internal JSON не становится пользовательским документом.
-
-## 9. Интеграционные tests без реальных Telegram/AI API
-
-- [ ] Создать fake `TelegramStreamMessage`.
-- [ ] Создать fake provider adapter.
-- [ ] Создать fake message store или in-memory DB fixture.
-- [ ] Test: oversized input attachment rejected before download.
-- [ ] Test: document input creates RAG artifact.
-- [ ] Test: voice input creates transcript artifact.
-- [ ] Test: final answer creates final_text artifact.
-- [ ] Test: thrown error creates error artifact.
-- [ ] Test: tool call creates tool_result artifact.
-- [ ] Test: generated file creates generated_file artifact.
-- [ ] Test: TTS requested creates tts_audio artifact.
-- [ ] Test: fallback `continue_without_stage` continues request.
-- [ ] Test: fallback `fail_request` stops request.
-
-## 10. Operational cleanup and observability
-
-- [ ] Add retention policy for `data/cache/internal-artifacts`.
-- [ ] Add retention policy for stale RAG vector/library provider state.
-- [ ] Add command or admin view for recent `ai_requests`.
-- [ ] Add command or admin view for request audit by message id.
-- [ ] Add command to inspect artifacts for a message.
-- [ ] Add log correlation by `requestId` across AI logs, tool logs and DB audit.
-- [ ] Add metrics counters: requests, failures, fallbacks, tool calls, RAG runs, TTS runs.
-- [ ] Add startup migration logs for `ai_requests`, `attachments`, `artifacts`, `request_audit`.
-
-## Suggested order
-
-- [x] 1. Normalize DB tables: `attachments`, `artifacts`, `request_audit`.
-- [ ] 2. Build `ArtifactStore` and migrate current artifact helpers to it.
-- [ ] 3. Add fake integration tests for reply-chain documents and artifacts.
-- [ ] 4. Introduce provider adapter interface.
-- [ ] 5. Move `tool_rank` into pipeline stage.
-- [ ] 6. Split `model_call` and `tool_loop` physically.
-- [ ] 7. Add centralized fallback user notifications.

README.md

@@ -7,6 +7,7 @@ Bot for Telegram with a lot of commands and AI (Ollama/Mistral/OpenAI) written i
 ```bash
 cp .env.example .env
 # Edit .env: add BOT_TOKEN, CREATOR_ID and configure optional AI models (MISTRAL_API_KEY, OPENAI_API_KEY, OLLAMA_ADDRESS)
+# For OpenAI-compatible servers (llama.cpp, etc.), set OPENAI_BACKEND=compatible and OPENAI_BASE_URL.
 # Optional: set DATABASE_URL to postgres://... for PostgreSQL or :memory: for ephemeral SQLite.
 # Optional: set DATA_PATH if you want to override the default local storage directory.
 ```
@@ -27,6 +28,25 @@ The bot initializes and migrates its database schema automatically on startup.
 `/exportdb` sends the SQLite file when available, plus a `.sql` dump and a JSON backup.
 `/importdb` restores the database from the JSON backup format.
 
+MCP tool servers can be configured through `MCP_SERVERS` in `.env`. Use a JSON array with `stdio` or `http` transports. Example:
+
+```bash
+MCP_SERVERS=[{"name":"local-tools","transport":"stdio","command":"node","args":["./mcp-server.js"]}]
+```
+
+If you want to disable all built-in local tools and use only MCP tools, set:
+
+```bash
+DISABLE_LOCAL_TOOLS=true
+```
+
+If you want a partial filter instead, use tool names:
+
+```bash
+LOCAL_TOOL_ALLOWLIST=get_datetime,web_search
+LOCAL_TOOL_DENYLIST=shell_execute,python_interpreter
+```
+
 For local Ollama document RAG, install an embedding model locally and set it in `.env`:
 
 ```bash

bun.lock

@@ -14,8 +14,8 @@
         "emoji-regex": "^10.6.0",
         "fluent-ffmpeg": "^2.1.3",
         "ollama": "^0.6.3",
-        "openai": "^6.37.0",
-        "pg": "^8.20.0",
+        "openai": "^6.38.0",
+        "pg": "^8.21.0",
         "qrcode": "^1.5.4",
         "sharp": "^0.34.5",
         "systeminformation": "^5.31.6",
@@ -27,12 +27,12 @@
         "@eslint/js": "^9.39.4",
         "@types/bun": "^1.3.14",
         "@types/fluent-ffmpeg": "^2.1.28",
-        "@types/node": "^25.8.0",
+        "@types/node": "^25.9.1",
         "@types/pg": "^8.20.0",
         "@types/qrcode": "^1.5.6",
         "eslint": "^9.39.4",
         "typescript": "^6.0.3",
-        "typescript-eslint": "^8.59.3",
+        "typescript-eslint": "^8.59.4",
       },
     },
   },
@@ -179,7 +179,7 @@
 
     "@types/json-schema": ["@types/json-schema@7.0.15", "", {}, "sha512-5+fP8P8MFNC+AyZCDxrB2pkZFPGzqQWUzpSeuuVLvm8VMcorNYavBqoFcxK8bQz4Qsbn4oUEEem4wDLfcysGHA=="],
 
-    "@types/node": ["@types/node@25.8.0", "", { "dependencies": { "undici-types": ">=7.24.0 <7.24.7" } }, "sha512-TCFSk8IZh+iLX1xtksoBVtdmgL+1IX0fC9BeU4QqFSuNdN/K+HUlhqOzEmSYYpZUVsLYcPqc9KX+60iDuninSQ=="],
+    "@types/node": ["@types/node@25.9.1", "", { "dependencies": { "undici-types": ">=7.24.0 <7.24.7" } }, "sha512-xfrlY7UD5rMJk3ZVJP8BNzS28J36YJg+xp+LPXV1TdWxr8uMH5A860QNxYDGQe/ylDSgjxE52Q9VnO7p75tJxg=="],
 
     "@types/pg": ["@types/pg@8.20.0", "", { "dependencies": { "@types/node": "*", "pg-protocol": "*", "pg-types": "^2.2.0" } }, "sha512-bEPFOaMAHTEP1EzpvHTbmwR8UsFyHSKsRisLIHVMXnpNefSbGA1bD6CVy+qKjGSqmZqNqBDV2azOBo8TgkcVow=="],
 
@@ -187,25 +187,25 @@
 
     "@types/ws": ["@types/ws@8.18.1", "", { "dependencies": { "@types/node": "*" } }, "sha512-ThVF6DCVhA8kUGy+aazFQ4kXQ7E1Ty7A3ypFOe0IcJV8O/M511G99AW24irKrW56Wt44yG9+ij8FaqoBGkuBXg=="],
 
-    "@typescript-eslint/eslint-plugin": ["@typescript-eslint/eslint-plugin@8.59.3", "", { "dependencies": { "@eslint-community/regexpp": "^4.12.2", "@typescript-eslint/scope-manager": "8.59.3", "@typescript-eslint/type-utils": "8.59.3", "@typescript-eslint/utils": "8.59.3", "@typescript-eslint/visitor-keys": "8.59.3", "ignore": "^7.0.5", "natural-compare": "^1.4.0", "ts-api-utils": "^2.5.0" }, "peerDependencies": { "@typescript-eslint/parser": "^8.59.3", "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0", "typescript": ">=4.8.4 <6.1.0" } }, "sha512-PwFvSKsXGShKGW6n5bZOhGHEcCZXM8HofLK9fNsEwZXzFRjoY+XT1Vsf1zgyXdwTr0ZYz1/2tkZ0DBTT9jZjhw=="],
+    "@typescript-eslint/eslint-plugin": ["@typescript-eslint/eslint-plugin@8.59.4", "", { "dependencies": { "@eslint-community/regexpp": "^4.12.2", "@typescript-eslint/scope-manager": "8.59.4", "@typescript-eslint/type-utils": "8.59.4", "@typescript-eslint/utils": "8.59.4", "@typescript-eslint/visitor-keys": "8.59.4", "ignore": "^7.0.5", "natural-compare": "^1.4.0", "ts-api-utils": "^2.5.0" }, "peerDependencies": { "@typescript-eslint/parser": "^8.59.4", "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0", "typescript": ">=4.8.4 <6.1.0" } }, "sha512-PegsU+XfyJJNjd4+u/k6f9yTyp0lEXXiPopUNobZcIAUJFGICFLN+sP0Rb3JehVmiij1Ph0dFGYqODoRo/2+6A=="],
 
-    "@typescript-eslint/parser": ["@typescript-eslint/parser@8.59.3", "", { "dependencies": { "@typescript-eslint/scope-manager": "8.59.3", "@typescript-eslint/types": "8.59.3", "@typescript-eslint/typescript-estree": "8.59.3", "@typescript-eslint/visitor-keys": "8.59.3", "debug": "^4.4.3" }, "peerDependencies": { "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0", "typescript": ">=4.8.4 <6.1.0" } }, "sha512-HPwA+hVkfcriajbNvTmZv4VRauibay+cWArYUYq7u7W7PmGShMxbPxLvrwDme55a6d5alG3nrYfhyJ/G28XlLg=="],
+    "@typescript-eslint/parser": ["@typescript-eslint/parser@8.59.4", "", { "dependencies": { "@typescript-eslint/scope-manager": "8.59.4", "@typescript-eslint/types": "8.59.4", "@typescript-eslint/typescript-estree": "8.59.4", "@typescript-eslint/visitor-keys": "8.59.4", "debug": "^4.4.3" }, "peerDependencies": { "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0", "typescript": ">=4.8.4 <6.1.0" } }, "sha512-zORHqO/tuhxY1zWuTvMUqddRxpiFJ72xVfcNoWpqdLjs6lfPbuQBJuW4pk+49/uBMy7Ssr4bzgjiKmmDB1UbZQ=="],
 
-    "@typescript-eslint/project-service": ["@typescript-eslint/project-service@8.59.3", "", { "dependencies": { "@typescript-eslint/tsconfig-utils": "^8.59.3", "@typescript-eslint/types": "^8.59.3", "debug": "^4.4.3" }, "peerDependencies": { "typescript": ">=4.8.4 <6.1.0" } }, "sha512-ECiUWa/KYRGDFUqTNehaRgzDshnJfkTABJxVemHk4ko22gcr0ukloKjWvyQ64g8YCV/UI47kN1dbmjf/GaQYng=="],
+    "@typescript-eslint/project-service": ["@typescript-eslint/project-service@8.59.4", "", { "dependencies": { "@typescript-eslint/tsconfig-utils": "^8.59.4", "@typescript-eslint/types": "^8.59.4", "debug": "^4.4.3" }, "peerDependencies": { "typescript": ">=4.8.4 <6.1.0" } }, "sha512-Ly00Vu4oAacfDeHp2Zg85ioNG6l8HG+tN1D7J+xTHSxu9y0awYKJ2zH1rFBn8ZSfuGK+7FxK3Cgl3uAz0aZZLg=="],
 
-    "@typescript-eslint/scope-manager": ["@typescript-eslint/scope-manager@8.59.3", "", { "dependencies": { "@typescript-eslint/types": "8.59.3", "@typescript-eslint/visitor-keys": "8.59.3" } }, "sha512-t2LvZnoEfzKtnPjgeEu41xw5gxq9mQVfYy4OoZ4Vlt0sk3JwxmhCca/AR7DwOiHrjWgjAj6as4AhRLKSDfvZIA=="],
+    "@typescript-eslint/scope-manager": ["@typescript-eslint/scope-manager@8.59.4", "", { "dependencies": { "@typescript-eslint/types": "8.59.4", "@typescript-eslint/visitor-keys": "8.59.4" } }, "sha512-mUeR/3H1WrTAddJrwut8OoPjfauaztMQmRwV5fQTUyNVJCLiUXXe4lGEyYIL2oFDpP7UtgbGJXCt72wT0z2S3Q=="],
 
-    "@typescript-eslint/tsconfig-utils": ["@typescript-eslint/tsconfig-utils@8.59.3", "", { "peerDependencies": { "typescript": ">=4.8.4 <6.1.0" } }, "sha512-PcIJHjmaREXLgIAIzLnSY9VucEzz8FKXsRgFa1DmdGCK/5tJpW03TKJF01Q6VZd1lLdz2sIKPWaDUZN9dp//dw=="],
+    "@typescript-eslint/tsconfig-utils": ["@typescript-eslint/tsconfig-utils@8.59.4", "", { "peerDependencies": { "typescript": ">=4.8.4 <6.1.0" } }, "sha512-DLCpnKgD4alVxTBSKulK+gU1KCqOgUXfDRDXh2mZgzokQKa/70ax93I2uVO3m/LLvIAtWZIFoiifudmIqAxpMA=="],
 
-    "@typescript-eslint/type-utils": ["@typescript-eslint/type-utils@8.59.3", "", { "dependencies": { "@typescript-eslint/types": "8.59.3", "@typescript-eslint/typescript-estree": "8.59.3", "@typescript-eslint/utils": "8.59.3", "debug": "^4.4.3", "ts-api-utils": "^2.5.0" }, "peerDependencies": { "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0", "typescript": ">=4.8.4 <6.1.0" } }, "sha512-g71d8QD8UaiHGvrJwyIS1hCX5r63w6Jll+4VEYhEAHXTDIqX1JgxhTAbEHtKntL9kuc4jRo7/GWw5xfCepSccQ=="],
+    "@typescript-eslint/type-utils": ["@typescript-eslint/type-utils@8.59.4", "", { "dependencies": { "@typescript-eslint/types": "8.59.4", "@typescript-eslint/typescript-estree": "8.59.4", "@typescript-eslint/utils": "8.59.4", "debug": "^4.4.3", "ts-api-utils": "^2.5.0" }, "peerDependencies": { "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0", "typescript": ">=4.8.4 <6.1.0" } }, "sha512-uonTuPAAKr9XaBGqJ3LjYTh72zy5DyGesljO9gtmk/eFW0W1fRHjnwVYKB35Lm8d5Q5CluEW3gPHjTvZTmgrfA=="],
 
-    "@typescript-eslint/types": ["@typescript-eslint/types@8.59.3", "", {}, "sha512-ePFoH0g4ludssdRFqqDxQePCxU4WQyRa9+XVwjm7yLn0FKhMeoetC+qBEEI1Eyb1pGSDveTIT09Bvw2WhlGayg=="],
+    "@typescript-eslint/types": ["@typescript-eslint/types@8.59.4", "", {}, "sha512-F1o7WJcCq+bc8dwcO/YsSEOudAH8RDtaOhM6wcAQhcUsFhnWQl81JKy48q1hoxAU0qrzM89+31GYh1515Zde3Q=="],
 
-    "@typescript-eslint/typescript-estree": ["@typescript-eslint/typescript-estree@8.59.3", "", { "dependencies": { "@typescript-eslint/project-service": "8.59.3", "@typescript-eslint/tsconfig-utils": "8.59.3", "@typescript-eslint/types": "8.59.3", "@typescript-eslint/visitor-keys": "8.59.3", "debug": "^4.4.3", "minimatch": "^10.2.2", "semver": "^7.7.3", "tinyglobby": "^0.2.15", "ts-api-utils": "^2.5.0" }, "peerDependencies": { "typescript": ">=4.8.4 <6.1.0" } }, "sha512-CbRjVRAf7Lr9Kr8RopKcbY45p2VfmmHrm0ygOCYFi7oU8q19m0Fs/6iHS7kNOmwpp+ob07ZVcAqlxUod9lYdmg=="],
+    "@typescript-eslint/typescript-estree": ["@typescript-eslint/typescript-estree@8.59.4", "", { "dependencies": { "@typescript-eslint/project-service": "8.59.4", "@typescript-eslint/tsconfig-utils": "8.59.4", "@typescript-eslint/types": "8.59.4", "@typescript-eslint/visitor-keys": "8.59.4", "debug": "^4.4.3", "minimatch": "^10.2.2", "semver": "^7.7.3", "tinyglobby": "^0.2.15", "ts-api-utils": "^2.5.0" }, "peerDependencies": { "typescript": ">=4.8.4 <6.1.0" } }, "sha512-F+RuOmcDXo4+TPdfd/TCLS3m2nw8gE9XXyZLrA3JBfaA5tz9TtdkyD3YJFmPxulyc2cKbEok/CvFE3MgSLWnag=="],
 
-    "@typescript-eslint/utils": ["@typescript-eslint/utils@8.59.3", "", { "dependencies": { "@eslint-community/eslint-utils": "^4.9.1", "@typescript-eslint/scope-manager": "8.59.3", "@typescript-eslint/types": "8.59.3", "@typescript-eslint/typescript-estree": "8.59.3" }, "peerDependencies": { "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0", "typescript": ">=4.8.4 <6.1.0" } }, "sha512-JAvT14goBzRzzzZyqq3P9BLArIxTtQURUtFgQ/V7FO+eU+Gg6ES+5ymOPP1wRxXcxAYeivCk4uS3jCKWI1K8Zg=="],
+    "@typescript-eslint/utils": ["@typescript-eslint/utils@8.59.4", "", { "dependencies": { "@eslint-community/eslint-utils": "^4.9.1", "@typescript-eslint/scope-manager": "8.59.4", "@typescript-eslint/types": "8.59.4", "@typescript-eslint/typescript-estree": "8.59.4" }, "peerDependencies": { "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0", "typescript": ">=4.8.4 <6.1.0" } }, "sha512-cYXeNAUsG4lJo5dbc1FcKm+JwIWrj1/UpTORsC6tGMjEZ81DYcvIr9/ueikhMa/Y/gDQYGp+YX9/xQrXje5BJw=="],
 
-    "@typescript-eslint/visitor-keys": ["@typescript-eslint/visitor-keys@8.59.3", "", { "dependencies": { "@typescript-eslint/types": "8.59.3", "eslint-visitor-keys": "^5.0.0" } }, "sha512-f1UQF7ggd42YiwI5wGrRaPsa+P0CINBlrkLPmGfpq/u/I/oVtecoEIfFR9ag/oa1sLOsRNZ6xehf6qMZhQGBDg=="],
+    "@typescript-eslint/visitor-keys": ["@typescript-eslint/visitor-keys@8.59.4", "", { "dependencies": { "@typescript-eslint/types": "8.59.4", "eslint-visitor-keys": "^5.0.0" } }, "sha512-U3gxVaDVnuZKhSspW/MzMxE1kq7zOdc072FcSNoqA1I9p8HyKbBFfEHoWckBAMgNMph4MamwS5iTVzFmrnt8TQ=="],
 
     "acorn": ["acorn@8.16.0", "", { "bin": { "acorn": "bin/acorn" } }, "sha512-UVJyE9MttOsBQIDKw1skb9nAwQuR5wuGD3+82K6JgJlm/Y+KI92oNsMNGZCYdDsVtRHSak0pcV5Dno5+4jh9sw=="],
 
@@ -399,7 +399,7 @@
 
     "ollama": ["ollama@0.6.3", "", { "dependencies": { "whatwg-fetch": "^3.6.20" } }, "sha512-KEWEhIqE5wtfzEIZbDCLH51VFZ6Z3ZSa6sIOg/E/tBV8S51flyqBOXi+bRxlOYKDf8i327zG9eSTb8IJxvm3Zg=="],
 
-    "openai": ["openai@6.37.0", "", { "peerDependencies": { "ws": "^8.18.0", "zod": "^3.25 || ^4.0" }, "optionalPeers": ["ws", "zod"], "bin": { "openai": "bin/cli" } }, "sha512-0H5dEGFmmLv6KSd0W1w2nyL8WsLkX6yoLeQpU+dZAOuGcany5qkYQMmj35ZrKgb6yiyYqpUzFOpR8mZQkgqeEQ=="],
+    "openai": ["openai@6.38.0", "", { "peerDependencies": { "ws": "^8.18.0", "zod": "^3.25 || ^4.0" }, "optionalPeers": ["ws", "zod"], "bin": { "openai": "bin/cli" } }, "sha512-AoMplt2UalrpgUDMh3L09QWjNRlgJPipclQvA6sYAaeF6nHNBMgmikAZGmcYLn8on4d9sQY9Q8bOLfrBS7Lc8g=="],
 
     "optionator": ["optionator@0.9.4", "", { "dependencies": { "deep-is": "^0.1.3", "fast-levenshtein": "^2.0.6", "levn": "^0.4.1", "prelude-ls": "^1.2.1", "type-check": "^0.4.0", "word-wrap": "^1.2.5" } }, "sha512-6IpQ7mKUxRcZNLIObR0hz7lxsapSSIYNZJwXPGeF0mTVqGKFIXj1DQcMoT22S3ROcLyY/rz0PWaWZ9ayWmad9g=="],
 
@@ -415,15 +415,15 @@
 
     "path-key": ["path-key@3.1.1", "", {}, "sha512-ojmeN0qd+y0jszEtoY48r0Peq5dwMEkIlCOu6Q5f41lfkswXuKtYrhgoTpLnyIcHm24Uhqx+5Tqm2InSwLhE6Q=="],
 
-    "pg": ["pg@8.20.0", "", { "dependencies": { "pg-connection-string": "^2.12.0", "pg-pool": "^3.13.0", "pg-protocol": "^1.13.0", "pg-types": "2.2.0", "pgpass": "1.0.5" }, "optionalDependencies": { "pg-cloudflare": "^1.3.0" }, "peerDependencies": { "pg-native": ">=3.0.1" }, "optionalPeers": ["pg-native"] }, "sha512-ldhMxz2r8fl/6QkXnBD3CR9/xg694oT6DZQ2s6c/RI28OjtSOpxnPrUCGOBJ46RCUxcWdx3p6kw/xnDHjKvaRA=="],
+    "pg": ["pg@8.21.0", "", { "dependencies": { "pg-connection-string": "^2.13.0", "pg-pool": "^3.14.0", "pg-protocol": "^1.14.0", "pg-types": "2.2.0", "pgpass": "1.0.5" }, "optionalDependencies": { "pg-cloudflare": "^1.4.0" }, "peerDependencies": { "pg-native": ">=3.0.1" }, "optionalPeers": ["pg-native"] }, "sha512-AUP1EYJuHraQGsVoCQVIcM7TEJVGtDzxWtGFZd8rds9d+CCXlU5Js1rYgfLNvxy9iJrpHjGrRjoi/3BT9fRyiA=="],
 
-    "pg-cloudflare": ["pg-cloudflare@1.3.0", "", {}, "sha512-6lswVVSztmHiRtD6I8hw4qP/nDm1EJbKMRhf3HCYaqud7frGysPv7FYJ5noZQdhQtN2xJnimfMtvQq21pdbzyQ=="],
+    "pg-cloudflare": ["pg-cloudflare@1.4.0", "", {}, "sha512-Vo7z/6rrQYxpNRylp4Tlob2elzbh+N/MOQbxFVWCxS7oEx6jF53GTJFxK2WWpKuBRkmiin4Mt+xofFDjx09R0A=="],
 
-    "pg-connection-string": ["pg-connection-string@2.12.0", "", {}, "sha512-U7qg+bpswf3Cs5xLzRqbXbQl85ng0mfSV/J0nnA31MCLgvEaAo7CIhmeyrmJpOr7o+zm0rXK+hNnT5l9RHkCkQ=="],
+    "pg-connection-string": ["pg-connection-string@2.13.0", "", {}, "sha512-EMnU9E2fSULdsbErBbMaXJvFeD9B4+nPcM3f+4lsiCR0BHLPrLVjv3DbyM2hgQQviKJaTWIRRTjKjWlHg3p2ig=="],
 
     "pg-int8": ["pg-int8@1.0.1", "", {}, "sha512-WCtabS6t3c8SkpDBUlb1kjOs7l66xsGdKpIPZsg4wR+B3+u9UAum2odSsF9tnvxg80h4ZxLWMy4pRjOsFIqQpw=="],
 
-    "pg-pool": ["pg-pool@3.13.0", "", { "peerDependencies": { "pg": ">=8.0" } }, "sha512-gB+R+Xud1gLFuRD/QgOIgGOBE2KCQPaPwkzBBGC9oG69pHTkhQeIuejVIk3/cnDyX39av2AxomQiyPT13WKHQA=="],
+    "pg-pool": ["pg-pool@3.14.0", "", { "peerDependencies": { "pg": ">=8.0" } }, "sha512-gKtPkFdQPU3DksooVLi9LsjZxrsBUZIpa+7aVx+LV5pNh0KzP4Zleud2po+ConrxbuXGBJ6Hfer6hdgpIBpBaw=="],
 
     "pg-protocol": ["pg-protocol@1.13.0", "", {}, "sha512-zzdvXfS6v89r6v7OcFCHfHlyG/wvry1ALxZo4LqgUoy7W9xhBDMaqOuMiF3qEV45VqsN6rdlcehHrfDtlCPc8w=="],
 
@@ -495,7 +495,7 @@
 
     "typescript": ["typescript@6.0.3", "", { "bin": { "tsc": "bin/tsc", "tsserver": "bin/tsserver" } }, "sha512-y2TvuxSZPDyQakkFRPZHKFm+KKVqIisdg9/CZwm9ftvKXLP8NRWj38/ODjNbr43SsoXqNuAisEf1GdCxqWcdBw=="],
 
-    "typescript-eslint": ["typescript-eslint@8.59.3", "", { "dependencies": { "@typescript-eslint/eslint-plugin": "8.59.3", "@typescript-eslint/parser": "8.59.3", "@typescript-eslint/typescript-estree": "8.59.3", "@typescript-eslint/utils": "8.59.3" }, "peerDependencies": { "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0", "typescript": ">=4.8.4 <6.1.0" } }, "sha512-KgusgyDgG4LI8Ih/sWaCtZ06tckLAS5CvT5A4D1Q7bYVoAAyzwiZvE4BmwDHkhRVkvhRBepKeASoFzQetha7Fg=="],
+    "typescript-eslint": ["typescript-eslint@8.59.4", "", { "dependencies": { "@typescript-eslint/eslint-plugin": "8.59.4", "@typescript-eslint/parser": "8.59.4", "@typescript-eslint/typescript-estree": "8.59.4", "@typescript-eslint/utils": "8.59.4" }, "peerDependencies": { "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0", "typescript": ">=4.8.4 <6.1.0" } }, "sha512-Rw6+44QNFaXtgHSjPy+Kw8hrJniMYzR85E9yLmOLcfZ91/rz+JXQbDTCmc6ccxMPY6K6PgAq26f0JCBfR7LIPQ=="],
 
     "typescript-telegram-bot-api": ["typescript-telegram-bot-api@0.16.0", "", { "dependencies": { "axios": "^1.7.7", "form-data": "^4.0.1" } }, "sha512-NaAjXucQiZ87U8La/IMaWDOghbMlJzfMbU4rG8ppFpgPOvEwat/zfN5BM+J2QDvKVGN87qJ+1nELnkm3ctSLnQ=="],
 
@@ -551,12 +551,16 @@
 
     "@typescript-eslint/visitor-keys/eslint-visitor-keys": ["eslint-visitor-keys@5.0.1", "", {}, "sha512-tD40eHxA35h0PEIZNeIjkHoDR4YjjJp34biM0mDvplBe//mB+IHCqHDGV7pxF+7MklTvighcCPPZC7ynWyjdTA=="],
 
+    "bun-types/@types/node": ["@types/node@25.8.0", "", { "dependencies": { "undici-types": ">=7.24.0 <7.24.7" } }, "sha512-TCFSk8IZh+iLX1xtksoBVtdmgL+1IX0fC9BeU4QqFSuNdN/K+HUlhqOzEmSYYpZUVsLYcPqc9KX+60iDuninSQ=="],
+
     "cross-spawn/which": ["which@2.0.2", "", { "dependencies": { "isexe": "^2.0.0" }, "bin": { "node-which": "./bin/node-which" } }, "sha512-BLI3Tl1TW3Pvl70l3yq3Y64i+awpwXqsGBYWkkqMtnbXgrMD+yj7rhW0kuEDxzJaYXGjEW5ogapKNMEKNMjibA=="],
 
     "fs-extra/jsonfile": ["jsonfile@4.0.0", "", { "optionalDependencies": { "graceful-fs": "^4.1.6" } }, "sha512-m6F1R3z8jjlf2imQHS2Qez5sjKWQzbuuhuJ/FKYFRZvPE3PuHcSMVZzfsLhGVOkfd20obL5SWEBew5ShlquNxg=="],
 
     "libsql/detect-libc": ["detect-libc@2.0.2", "", {}, "sha512-UX6sGumvvqSaXgdKGUsgZWqcUyIXZ/vZTrlRT/iobiKhGL0zL4d3osHj3uqllWJK+i+sixDS/3COVEOFbupFyw=="],
 
+    "pg/pg-protocol": ["pg-protocol@1.14.0", "", {}, "sha512-n5taZ1kO3s9ngDTVxsEznOqCyToTgz0FLuPq0B33COy5pPpuWJpY3/2oRBVETuOgzdqRXfWpM9HIhp2LBBT1BA=="],
+
     "string-width/emoji-regex": ["emoji-regex@8.0.0", "", {}, "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A=="],
 
     "typescript-telegram-bot-api/axios": ["axios@1.16.0", "", { "dependencies": { "follow-redirects": "^1.16.0", "form-data": "^4.0.5", "proxy-from-env": "^2.1.0" } }, "sha512-6hp5CwvTPlN2A31g5dxnwAX0orzM7pmCRDLnZSX772mv8WDqICwFjowHuPs04Mc8deIld1+ejhtaMn5vp6b+1w=="],

locales/en.json

@@ -8,6 +8,13 @@
   },
   "providerChoice.default": "Default",
   "errorText": "⚠️ An error occurred.",
+  "pipelineFallback.generic": "⚠️ I had to skip part of the request, but I can continue.",
+  "pipelineFallback.notifyUser": "⚠️ I hit a problem and need to continue with a fallback.",
+  "pipelineFallback.failRequest": "⚠️ I could not finish this request.",
+  "pipelineFallback.documentRag": "⚠️ Document retrieval failed, so I will answer without RAG.",
+  "pipelineFallback.speechToText": "⚠️ Speech transcription failed, so I will continue without the audio transcript.",
+  "pipelineFallback.textToSpeech": "⚠️ Text-to-speech failed, so I will continue without audio output.",
+  "pipelineFallback.toolLoop": "⚠️ Tool execution failed, so I will continue without that tool.",
   "waitThinkText": "⏳ Let me think...",
   "analyzingPictureText": "🔍 Analyzing the image...",
   "analyzingPicturesText": "🔍 Analyzing the images...",
@@ -176,6 +183,9 @@
   "getWhenPluralUnitText": "{unit}s",
   "getWhenDurationText": "{prefix}{value} {unit}",
   "commandDescriptions": {
+    "aiAudit": "Inspect AI request audit and artifacts",
+    "aiMetrics": "Show AI observability counters",
+    "aiRequests": "Show recent AI requests",
     "ae": "evaluation",
     "adminsAdd": "Add user to admins",
     "adminsRemove": "Remove user from admins",

locales/ru.json

@@ -8,6 +8,13 @@
   },
   "providerChoice.default": "По умолчанию",
   "errorText": "⚠️ Произошла ошибка.",
+  "pipelineFallback.generic": "⚠️ Мне пришлось пропустить часть запроса, но я могу продолжить.",
+  "pipelineFallback.notifyUser": "⚠️ Возникла проблема, и я продолжу с запасным вариантом.",
+  "pipelineFallback.failRequest": "⚠️ Я не смог завершить этот запрос.",
+  "pipelineFallback.documentRag": "⚠️ Не удалось получить документы, поэтому я отвечу без RAG.",
+  "pipelineFallback.speechToText": "⚠️ Не удалось распознать речь, поэтому я продолжу без расшифровки аудио.",
+  "pipelineFallback.textToSpeech": "⚠️ Не удалось выполнить синтез речи, поэтому я продолжу без аудио.",
+  "pipelineFallback.toolLoop": "⚠️ Не удалось выполнить инструменты, поэтому я продолжу без них.",
   "waitThinkText": "⏳ Дайте-ка подумать...",
   "analyzingPictureText": "🔍 Анализирую изображение...",
   "analyzingPicturesText": "🔍 Анализирую изображения...",
@@ -202,6 +209,9 @@
   "getWhenPluralUnitText": "{unit}",
   "getWhenDurationText": "{prefix}{value} {unit}",
   "commandDescriptions": {
+    "aiRequests": "Показать последние AI-запросы",
+    "aiAudit": "Показать аудит AI-запроса и артефакты",
+    "aiMetrics": "Показать счётчики AI-обсервабилити",
     "ae": "вычисление",
     "adminsAdd": "Добавить пользователя в администраторы",
     "adminsRemove": "Удалить пользователя из администраторов",

locales/ua.json

@@ -8,6 +8,13 @@
   },
   "providerChoice.default": "За замовчуванням",
   "errorText": "⚠️ Сталася помилка.",
+  "pipelineFallback.generic": "⚠️ Мені довелося пропустити частину запиту, але я можу продовжити.",
+  "pipelineFallback.notifyUser": "⚠️ Виникла проблема, і я продовжу із запасним варіантом.",
+  "pipelineFallback.failRequest": "⚠️ Я не зміг завершити цей запит.",
+  "pipelineFallback.documentRag": "⚠️ Не вдалося отримати документи, тому я відповім без RAG.",
+  "pipelineFallback.speechToText": "⚠️ Не вдалося розпізнати мовлення, тому я продовжу без розшифровки аудіо.",
+  "pipelineFallback.textToSpeech": "⚠️ Не вдалося виконати синтез мовлення, тому я продовжу без аудіо.",
+  "pipelineFallback.toolLoop": "⚠️ Не вдалося виконати інструменти, тому я продовжу без них.",
   "waitThinkText": "⏳ Дайте-но подумати...",
   "analyzingPictureText": "🔍 Аналізую зображення...",
   "analyzingPicturesText": "🔍 Аналізую зображення...",
@@ -201,6 +208,9 @@
   "getWhenPluralUnitText": "{unit}",
   "getWhenDurationText": "{prefix}{value} {unit}",
   "commandDescriptions": {
+    "aiRequests": "Показати останні AI-запити",
+    "aiAudit": "Показати аудит AI-запиту та артефакти",
+    "aiMetrics": "Показати лічильники AI-спостережуваності",
     "help": "Показати список команд",
     "settings": "Налаштування користувача",
     "start": "Запустити бота",

package-lock.json

@@ -17,8 +17,8 @@
         "emoji-regex": "^10.6.0",
         "fluent-ffmpeg": "^2.1.3",
         "ollama": "^0.6.3",
-        "openai": "^6.37.0",
-        "pg": "^8.20.0",
+        "openai": "^6.38.0",
+        "pg": "^8.21.0",
         "qrcode": "^1.5.4",
         "sharp": "^0.34.5",
         "systeminformation": "^5.31.6",
@@ -30,12 +30,12 @@
         "@eslint/js": "^9.39.4",
         "@types/bun": "^1.3.14",
         "@types/fluent-ffmpeg": "^2.1.28",
-        "@types/node": "^25.8.0",
+        "@types/node": "^25.9.1",
         "@types/pg": "^8.20.0",
         "@types/qrcode": "^1.5.6",
         "eslint": "^9.39.4",
         "typescript": "^6.0.3",
-        "typescript-eslint": "^8.59.3"
+        "typescript-eslint": "^8.59.4"
       }
     },
     "node_modules/@emnapi/runtime": {
@@ -1246,9 +1246,9 @@
       "license": "MIT"
     },
     "node_modules/@types/node": {
-      "version": "25.8.0",
-      "resolved": "https://registry.npmjs.org/@types/node/-/node-25.8.0.tgz",
-      "integrity": "sha512-TCFSk8IZh+iLX1xtksoBVtdmgL+1IX0fC9BeU4QqFSuNdN/K+HUlhqOzEmSYYpZUVsLYcPqc9KX+60iDuninSQ==",
+      "version": "25.9.1",
+      "resolved": "https://registry.npmjs.org/@types/node/-/node-25.9.1.tgz",
+      "integrity": "sha512-xfrlY7UD5rMJk3ZVJP8BNzS28J36YJg+xp+LPXV1TdWxr8uMH5A860QNxYDGQe/ylDSgjxE52Q9VnO7p75tJxg==",
       "license": "MIT",
       "dependencies": {
         "undici-types": ">=7.24.0 <7.24.7"
@@ -1286,17 +1286,17 @@
       }
     },
     "node_modules/@typescript-eslint/eslint-plugin": {
-      "version": "8.59.3",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/eslint-plugin/-/eslint-plugin-8.59.3.tgz",
-      "integrity": "sha512-PwFvSKsXGShKGW6n5bZOhGHEcCZXM8HofLK9fNsEwZXzFRjoY+XT1Vsf1zgyXdwTr0ZYz1/2tkZ0DBTT9jZjhw==",
+      "version": "8.59.4",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/eslint-plugin/-/eslint-plugin-8.59.4.tgz",
+      "integrity": "sha512-PegsU+XfyJJNjd4+u/k6f9yTyp0lEXXiPopUNobZcIAUJFGICFLN+sP0Rb3JehVmiij1Ph0dFGYqODoRo/2+6A==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
         "@eslint-community/regexpp": "^4.12.2",
-        "@typescript-eslint/scope-manager": "8.59.3",
-        "@typescript-eslint/type-utils": "8.59.3",
-        "@typescript-eslint/utils": "8.59.3",
-        "@typescript-eslint/visitor-keys": "8.59.3",
+        "@typescript-eslint/scope-manager": "8.59.4",
+        "@typescript-eslint/type-utils": "8.59.4",
+        "@typescript-eslint/utils": "8.59.4",
+        "@typescript-eslint/visitor-keys": "8.59.4",
         "ignore": "^7.0.5",
         "natural-compare": "^1.4.0",
         "ts-api-utils": "^2.5.0"
@@ -1309,7 +1309,7 @@
         "url": "https://opencollective.com/typescript-eslint"
       },
       "peerDependencies": {
-        "@typescript-eslint/parser": "^8.59.3",
+        "@typescript-eslint/parser": "^8.59.4",
         "eslint": "^8.57.0 || ^9.0.0 || ^10.0.0",
         "typescript": ">=4.8.4 <6.1.0"
       }
@@ -1325,16 +1325,16 @@
       }
     },
     "node_modules/@typescript-eslint/parser": {
-      "version": "8.59.3",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/parser/-/parser-8.59.3.tgz",
-      "integrity": "sha512-HPwA+hVkfcriajbNvTmZv4VRauibay+cWArYUYq7u7W7PmGShMxbPxLvrwDme55a6d5alG3nrYfhyJ/G28XlLg==",
+      "version": "8.59.4",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/parser/-/parser-8.59.4.tgz",
+      "integrity": "sha512-zORHqO/tuhxY1zWuTvMUqddRxpiFJ72xVfcNoWpqdLjs6lfPbuQBJuW4pk+49/uBMy7Ssr4bzgjiKmmDB1UbZQ==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@typescript-eslint/scope-manager": "8.59.3",
-        "@typescript-eslint/types": "8.59.3",
-        "@typescript-eslint/typescript-estree": "8.59.3",
-        "@typescript-eslint/visitor-keys": "8.59.3",
+        "@typescript-eslint/scope-manager": "8.59.4",
+        "@typescript-eslint/types": "8.59.4",
+        "@typescript-eslint/typescript-estree": "8.59.4",
+        "@typescript-eslint/visitor-keys": "8.59.4",
         "debug": "^4.4.3"
       },
       "engines": {
@@ -1350,14 +1350,14 @@
       }
     },
     "node_modules/@typescript-eslint/project-service": {
-      "version": "8.59.3",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/project-service/-/project-service-8.59.3.tgz",
-      "integrity": "sha512-ECiUWa/KYRGDFUqTNehaRgzDshnJfkTABJxVemHk4ko22gcr0ukloKjWvyQ64g8YCV/UI47kN1dbmjf/GaQYng==",
+      "version": "8.59.4",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/project-service/-/project-service-8.59.4.tgz",
+      "integrity": "sha512-Ly00Vu4oAacfDeHp2Zg85ioNG6l8HG+tN1D7J+xTHSxu9y0awYKJ2zH1rFBn8ZSfuGK+7FxK3Cgl3uAz0aZZLg==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@typescript-eslint/tsconfig-utils": "^8.59.3",
-        "@typescript-eslint/types": "^8.59.3",
+        "@typescript-eslint/tsconfig-utils": "^8.59.4",
+        "@typescript-eslint/types": "^8.59.4",
         "debug": "^4.4.3"
       },
       "engines": {
@@ -1372,14 +1372,14 @@
       }
     },
     "node_modules/@typescript-eslint/scope-manager": {
-      "version": "8.59.3",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/scope-manager/-/scope-manager-8.59.3.tgz",
-      "integrity": "sha512-t2LvZnoEfzKtnPjgeEu41xw5gxq9mQVfYy4OoZ4Vlt0sk3JwxmhCca/AR7DwOiHrjWgjAj6as4AhRLKSDfvZIA==",
+      "version": "8.59.4",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/scope-manager/-/scope-manager-8.59.4.tgz",
+      "integrity": "sha512-mUeR/3H1WrTAddJrwut8OoPjfauaztMQmRwV5fQTUyNVJCLiUXXe4lGEyYIL2oFDpP7UtgbGJXCt72wT0z2S3Q==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@typescript-eslint/types": "8.59.3",
-        "@typescript-eslint/visitor-keys": "8.59.3"
+        "@typescript-eslint/types": "8.59.4",
+        "@typescript-eslint/visitor-keys": "8.59.4"
       },
       "engines": {
         "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
@@ -1390,9 +1390,9 @@
       }
     },
     "node_modules/@typescript-eslint/tsconfig-utils": {
-      "version": "8.59.3",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/tsconfig-utils/-/tsconfig-utils-8.59.3.tgz",
-      "integrity": "sha512-PcIJHjmaREXLgIAIzLnSY9VucEzz8FKXsRgFa1DmdGCK/5tJpW03TKJF01Q6VZd1lLdz2sIKPWaDUZN9dp//dw==",
+      "version": "8.59.4",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/tsconfig-utils/-/tsconfig-utils-8.59.4.tgz",
+      "integrity": "sha512-DLCpnKgD4alVxTBSKulK+gU1KCqOgUXfDRDXh2mZgzokQKa/70ax93I2uVO3m/LLvIAtWZIFoiifudmIqAxpMA==",
       "dev": true,
       "license": "MIT",
       "engines": {
@@ -1407,15 +1407,15 @@
       }
     },
     "node_modules/@typescript-eslint/type-utils": {
-      "version": "8.59.3",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/type-utils/-/type-utils-8.59.3.tgz",
-      "integrity": "sha512-g71d8QD8UaiHGvrJwyIS1hCX5r63w6Jll+4VEYhEAHXTDIqX1JgxhTAbEHtKntL9kuc4jRo7/GWw5xfCepSccQ==",
+      "version": "8.59.4",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/type-utils/-/type-utils-8.59.4.tgz",
+      "integrity": "sha512-uonTuPAAKr9XaBGqJ3LjYTh72zy5DyGesljO9gtmk/eFW0W1fRHjnwVYKB35Lm8d5Q5CluEW3gPHjTvZTmgrfA==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@typescript-eslint/types": "8.59.3",
-        "@typescript-eslint/typescript-estree": "8.59.3",
-        "@typescript-eslint/utils": "8.59.3",
+        "@typescript-eslint/types": "8.59.4",
+        "@typescript-eslint/typescript-estree": "8.59.4",
+        "@typescript-eslint/utils": "8.59.4",
         "debug": "^4.4.3",
         "ts-api-utils": "^2.5.0"
       },
@@ -1432,9 +1432,9 @@
       }
     },
     "node_modules/@typescript-eslint/types": {
-      "version": "8.59.3",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/types/-/types-8.59.3.tgz",
-      "integrity": "sha512-ePFoH0g4ludssdRFqqDxQePCxU4WQyRa9+XVwjm7yLn0FKhMeoetC+qBEEI1Eyb1pGSDveTIT09Bvw2WhlGayg==",
+      "version": "8.59.4",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/types/-/types-8.59.4.tgz",
+      "integrity": "sha512-F1o7WJcCq+bc8dwcO/YsSEOudAH8RDtaOhM6wcAQhcUsFhnWQl81JKy48q1hoxAU0qrzM89+31GYh1515Zde3Q==",
       "dev": true,
       "license": "MIT",
       "engines": {
@@ -1446,16 +1446,16 @@
       }
     },
     "node_modules/@typescript-eslint/typescript-estree": {
-      "version": "8.59.3",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/typescript-estree/-/typescript-estree-8.59.3.tgz",
-      "integrity": "sha512-CbRjVRAf7Lr9Kr8RopKcbY45p2VfmmHrm0ygOCYFi7oU8q19m0Fs/6iHS7kNOmwpp+ob07ZVcAqlxUod9lYdmg==",
+      "version": "8.59.4",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/typescript-estree/-/typescript-estree-8.59.4.tgz",
+      "integrity": "sha512-F+RuOmcDXo4+TPdfd/TCLS3m2nw8gE9XXyZLrA3JBfaA5tz9TtdkyD3YJFmPxulyc2cKbEok/CvFE3MgSLWnag==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@typescript-eslint/project-service": "8.59.3",
-        "@typescript-eslint/tsconfig-utils": "8.59.3",
-        "@typescript-eslint/types": "8.59.3",
-        "@typescript-eslint/visitor-keys": "8.59.3",
+        "@typescript-eslint/project-service": "8.59.4",
+        "@typescript-eslint/tsconfig-utils": "8.59.4",
+        "@typescript-eslint/types": "8.59.4",
+        "@typescript-eslint/visitor-keys": "8.59.4",
         "debug": "^4.4.3",
         "minimatch": "^10.2.2",
         "semver": "^7.7.3",
@@ -1513,16 +1513,16 @@
       }
     },
     "node_modules/@typescript-eslint/utils": {
-      "version": "8.59.3",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/utils/-/utils-8.59.3.tgz",
-      "integrity": "sha512-JAvT14goBzRzzzZyqq3P9BLArIxTtQURUtFgQ/V7FO+eU+Gg6ES+5ymOPP1wRxXcxAYeivCk4uS3jCKWI1K8Zg==",
+      "version": "8.59.4",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/utils/-/utils-8.59.4.tgz",
+      "integrity": "sha512-cYXeNAUsG4lJo5dbc1FcKm+JwIWrj1/UpTORsC6tGMjEZ81DYcvIr9/ueikhMa/Y/gDQYGp+YX9/xQrXje5BJw==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
         "@eslint-community/eslint-utils": "^4.9.1",
-        "@typescript-eslint/scope-manager": "8.59.3",
-        "@typescript-eslint/types": "8.59.3",
-        "@typescript-eslint/typescript-estree": "8.59.3"
+        "@typescript-eslint/scope-manager": "8.59.4",
+        "@typescript-eslint/types": "8.59.4",
+        "@typescript-eslint/typescript-estree": "8.59.4"
       },
       "engines": {
         "node": "^18.18.0 || ^20.9.0 || >=21.1.0"
@@ -1537,13 +1537,13 @@
       }
     },
     "node_modules/@typescript-eslint/visitor-keys": {
-      "version": "8.59.3",
-      "resolved": "https://registry.npmjs.org/@typescript-eslint/visitor-keys/-/visitor-keys-8.59.3.tgz",
-      "integrity": "sha512-f1UQF7ggd42YiwI5wGrRaPsa+P0CINBlrkLPmGfpq/u/I/oVtecoEIfFR9ag/oa1sLOsRNZ6xehf6qMZhQGBDg==",
+      "version": "8.59.4",
+      "resolved": "https://registry.npmjs.org/@typescript-eslint/visitor-keys/-/visitor-keys-8.59.4.tgz",
+      "integrity": "sha512-U3gxVaDVnuZKhSspW/MzMxE1kq7zOdc072FcSNoqA1I9p8HyKbBFfEHoWckBAMgNMph4MamwS5iTVzFmrnt8TQ==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@typescript-eslint/types": "8.59.3",
+        "@typescript-eslint/types": "8.59.4",
         "eslint-visitor-keys": "^5.0.0"
       },
       "engines": {
@@ -1590,6 +1590,18 @@
         "acorn": "^6.0.0 || ^7.0.0 || ^8.0.0"
       }
     },
+    "node_modules/agent-base": {
+      "version": "6.0.2",
+      "resolved": "https://registry.npmjs.org/agent-base/-/agent-base-6.0.2.tgz",
+      "integrity": "sha512-RZNwNclF7+MS/8bDg70amg32dyeZGZxiDuQmZxKLAlQjr3jGyLx+4Kkk58UO7D2QdgFIQCovuSuZESne6RG6XQ==",
+      "license": "MIT",
+      "dependencies": {
+        "debug": "4"
+      },
+      "engines": {
+        "node": ">= 6.0.0"
+      }
+    },
     "node_modules/ajv": {
       "version": "6.15.0",
       "resolved": "https://registry.npmjs.org/ajv/-/ajv-6.15.0.tgz",
@@ -1661,31 +1673,6 @@
         "proxy-from-env": "^2.1.0"
       }
     },
-    "node_modules/axios/node_modules/agent-base": {
-      "version": "6.0.2",
-      "resolved": "https://registry.npmjs.org/agent-base/-/agent-base-6.0.2.tgz",
-      "integrity": "sha512-RZNwNclF7+MS/8bDg70amg32dyeZGZxiDuQmZxKLAlQjr3jGyLx+4Kkk58UO7D2QdgFIQCovuSuZESne6RG6XQ==",
-      "license": "MIT",
-      "dependencies": {
-        "debug": "4"
-      },
-      "engines": {
-        "node": ">= 6.0.0"
-      }
-    },
-    "node_modules/axios/node_modules/https-proxy-agent": {
-      "version": "5.0.1",
-      "resolved": "https://registry.npmjs.org/https-proxy-agent/-/https-proxy-agent-5.0.1.tgz",
-      "integrity": "sha512-dFcAjpTQFgoLMzC2VwU+C/CbS7uRL0lWmxDITmqm7C+7F0Odmj6s9l6alZc6AELXhrnggM2CeWSXHGOdX2YtwA==",
-      "license": "MIT",
-      "dependencies": {
-        "agent-base": "6",
-        "debug": "4"
-      },
-      "engines": {
-        "node": ">= 6"
-      }
-    },
     "node_modules/balanced-match": {
       "version": "1.0.2",
       "resolved": "https://registry.npmjs.org/balanced-match/-/balanced-match-1.0.2.tgz",
@@ -2575,6 +2562,19 @@
         "node": ">= 0.4"
       }
     },
+    "node_modules/https-proxy-agent": {
+      "version": "5.0.1",
+      "resolved": "https://registry.npmjs.org/https-proxy-agent/-/https-proxy-agent-5.0.1.tgz",
+      "integrity": "sha512-dFcAjpTQFgoLMzC2VwU+C/CbS7uRL0lWmxDITmqm7C+7F0Odmj6s9l6alZc6AELXhrnggM2CeWSXHGOdX2YtwA==",
+      "license": "MIT",
+      "dependencies": {
+        "agent-base": "6",
+        "debug": "4"
+      },
+      "engines": {
+        "node": ">= 6"
+      }
+    },
     "node_modules/ignore": {
       "version": "5.3.2",
       "resolved": "https://registry.npmjs.org/ignore/-/ignore-5.3.2.tgz",
@@ -2847,9 +2847,9 @@
       }
     },
     "node_modules/openai": {
-      "version": "6.37.0",
-      "resolved": "https://registry.npmjs.org/openai/-/openai-6.37.0.tgz",
-      "integrity": "sha512-0H5dEGFmmLv6KSd0W1w2nyL8WsLkX6yoLeQpU+dZAOuGcany5qkYQMmj35ZrKgb6yiyYqpUzFOpR8mZQkgqeEQ==",
+      "version": "6.38.0",
+      "resolved": "https://registry.npmjs.org/openai/-/openai-6.38.0.tgz",
+      "integrity": "sha512-AoMplt2UalrpgUDMh3L09QWjNRlgJPipclQvA6sYAaeF6nHNBMgmikAZGmcYLn8on4d9sQY9Q8bOLfrBS7Lc8g==",
       "license": "Apache-2.0",
       "bin": {
         "openai": "bin/cli"
@@ -2959,14 +2959,14 @@
       }
     },
     "node_modules/pg": {
-      "version": "8.20.0",
-      "resolved": "https://registry.npmjs.org/pg/-/pg-8.20.0.tgz",
-      "integrity": "sha512-ldhMxz2r8fl/6QkXnBD3CR9/xg694oT6DZQ2s6c/RI28OjtSOpxnPrUCGOBJ46RCUxcWdx3p6kw/xnDHjKvaRA==",
+      "version": "8.21.0",
+      "resolved": "https://registry.npmjs.org/pg/-/pg-8.21.0.tgz",
+      "integrity": "sha512-AUP1EYJuHraQGsVoCQVIcM7TEJVGtDzxWtGFZd8rds9d+CCXlU5Js1rYgfLNvxy9iJrpHjGrRjoi/3BT9fRyiA==",
       "license": "MIT",
       "dependencies": {
-        "pg-connection-string": "^2.12.0",
-        "pg-pool": "^3.13.0",
-        "pg-protocol": "^1.13.0",
+        "pg-connection-string": "^2.13.0",
+        "pg-pool": "^3.14.0",
+        "pg-protocol": "^1.14.0",
         "pg-types": "2.2.0",
         "pgpass": "1.0.5"
       },
@@ -2974,7 +2974,7 @@
         "node": ">= 16.0.0"
       },
       "optionalDependencies": {
-        "pg-cloudflare": "^1.3.0"
+        "pg-cloudflare": "^1.4.0"
       },
       "peerDependencies": {
         "pg-native": ">=3.0.1"
@@ -2986,16 +2986,16 @@
       }
     },
     "node_modules/pg-cloudflare": {
-      "version": "1.3.0",
-      "resolved": "https://registry.npmjs.org/pg-cloudflare/-/pg-cloudflare-1.3.0.tgz",
-      "integrity": "sha512-6lswVVSztmHiRtD6I8hw4qP/nDm1EJbKMRhf3HCYaqud7frGysPv7FYJ5noZQdhQtN2xJnimfMtvQq21pdbzyQ==",
+      "version": "1.4.0",
+      "resolved": "https://registry.npmjs.org/pg-cloudflare/-/pg-cloudflare-1.4.0.tgz",
+      "integrity": "sha512-Vo7z/6rrQYxpNRylp4Tlob2elzbh+N/MOQbxFVWCxS7oEx6jF53GTJFxK2WWpKuBRkmiin4Mt+xofFDjx09R0A==",
       "license": "MIT",
       "optional": true
     },
     "node_modules/pg-connection-string": {
-      "version": "2.12.0",
-      "resolved": "https://registry.npmjs.org/pg-connection-string/-/pg-connection-string-2.12.0.tgz",
-      "integrity": "sha512-U7qg+bpswf3Cs5xLzRqbXbQl85ng0mfSV/J0nnA31MCLgvEaAo7CIhmeyrmJpOr7o+zm0rXK+hNnT5l9RHkCkQ==",
+      "version": "2.13.0",
+      "resolved": "https://registry.npmjs.org/pg-connection-string/-/pg-connection-string-2.13.0.tgz",
+      "integrity": "sha512-EMnU9E2fSULdsbErBbMaXJvFeD9B4+nPcM3f+4lsiCR0BHLPrLVjv3DbyM2hgQQviKJaTWIRRTjKjWlHg3p2ig==",
       "license": "MIT"
     },
     "node_modules/pg-int8": {
@@ -3008,18 +3008,18 @@
       }
     },
     "node_modules/pg-pool": {
-      "version": "3.13.0",
-      "resolved": "https://registry.npmjs.org/pg-pool/-/pg-pool-3.13.0.tgz",
-      "integrity": "sha512-gB+R+Xud1gLFuRD/QgOIgGOBE2KCQPaPwkzBBGC9oG69pHTkhQeIuejVIk3/cnDyX39av2AxomQiyPT13WKHQA==",
+      "version": "3.14.0",
+      "resolved": "https://registry.npmjs.org/pg-pool/-/pg-pool-3.14.0.tgz",
+      "integrity": "sha512-gKtPkFdQPU3DksooVLi9LsjZxrsBUZIpa+7aVx+LV5pNh0KzP4Zleud2po+ConrxbuXGBJ6Hfer6hdgpIBpBaw==",
       "license": "MIT",
       "peerDependencies": {
         "pg": ">=8.0"
       }
     },
     "node_modules/pg-protocol": {
-      "version": "1.13.0",
-      "resolved": "https://registry.npmjs.org/pg-protocol/-/pg-protocol-1.13.0.tgz",
-      "integrity": "sha512-zzdvXfS6v89r6v7OcFCHfHlyG/wvry1ALxZo4LqgUoy7W9xhBDMaqOuMiF3qEV45VqsN6rdlcehHrfDtlCPc8w==",
+      "version": "1.14.0",
+      "resolved": "https://registry.npmjs.org/pg-protocol/-/pg-protocol-1.14.0.tgz",
+      "integrity": "sha512-n5taZ1kO3s9ngDTVxsEznOqCyToTgz0FLuPq0B33COy5pPpuWJpY3/2oRBVETuOgzdqRXfWpM9HIhp2LBBT1BA==",
       "license": "MIT"
     },
     "node_modules/pg-types": {
@@ -3455,16 +3455,16 @@
       }
     },
     "node_modules/typescript-eslint": {
-      "version": "8.59.3",
-      "resolved": "https://registry.npmjs.org/typescript-eslint/-/typescript-eslint-8.59.3.tgz",
-      "integrity": "sha512-KgusgyDgG4LI8Ih/sWaCtZ06tckLAS5CvT5A4D1Q7bYVoAAyzwiZvE4BmwDHkhRVkvhRBepKeASoFzQetha7Fg==",
+      "version": "8.59.4",
+      "resolved": "https://registry.npmjs.org/typescript-eslint/-/typescript-eslint-8.59.4.tgz",
+      "integrity": "sha512-Rw6+44QNFaXtgHSjPy+Kw8hrJniMYzR85E9yLmOLcfZ91/rz+JXQbDTCmc6ccxMPY6K6PgAq26f0JCBfR7LIPQ==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
-        "@typescript-eslint/eslint-plugin": "8.59.3",
-        "@typescript-eslint/parser": "8.59.3",
-        "@typescript-eslint/typescript-estree": "8.59.3",
-        "@typescript-eslint/utils": "8.59.3"
+        "@typescript-eslint/eslint-plugin": "8.59.4",
+        "@typescript-eslint/parser": "8.59.4",
+        "@typescript-eslint/typescript-estree": "8.59.4",
+        "@typescript-eslint/utils": "8.59.4"
       },
       "engines": {
         "node": "^18.18.0 || ^20.9.0 || >=21.1.0"

package.json

@@ -22,8 +22,8 @@
     "emoji-regex": "^10.6.0",
     "fluent-ffmpeg": "^2.1.3",
     "ollama": "^0.6.3",
-    "openai": "^6.37.0",
-    "pg": "^8.20.0",
+    "openai": "^6.38.0",
+    "pg": "^8.21.0",
     "qrcode": "^1.5.4",
     "sharp": "^0.34.5",
     "systeminformation": "^5.31.6",
@@ -35,11 +35,11 @@
     "@eslint/js": "^9.39.4",
     "@types/bun": "^1.3.14",
     "@types/fluent-ffmpeg": "^2.1.28",
-    "@types/node": "^25.8.0",
+    "@types/node": "^25.9.1",
     "@types/pg": "^8.20.0",
     "@types/qrcode": "^1.5.6",
     "eslint": "^9.39.4",
     "typescript": "^6.0.3",
-    "typescript-eslint": "^8.59.3"
+    "typescript-eslint": "^8.59.4"
   }
 }

src/ai/ai-runtime-target.ts

@@ -1,12 +1,12 @@
 import {Mistral} from "@mistralai/mistralai";
 import {Ollama} from "ollama";
 import {OpenAI} from "openai";
-import {Environment} from "../common/environment";
-import {AiModelCapabilities} from "../model/ai-model-capabilities";
-import {AiProvider} from "../model/ai-provider";
+import {Environment} from "../common/environment.js";
+import {AiModelCapabilities} from "../model/ai-model-capabilities.js";
+import {AiProvider} from "../model/ai-provider.js";
 
 export type AiCapabilityName = keyof AiModelCapabilities;
-export type AiRuntimePurpose = AiCapabilityName | "chat";
+export type AiRuntimePurpose = AiCapabilityName | "chat" | "memoryCompress";
 
 export type AiRuntimeTarget = {
     provider: AiProvider;
@@ -24,6 +24,7 @@ const PURPOSE_SUFFIXES: Record<AiRuntimePurpose, string[]> = {
     thinking: ["THINKING", "THINK"],
     extendedThinking: ["EXTENDED_THINKING", "THINKING", "THINK"],
     tools: ["TOOLS", "CHAT"],
+    memoryCompress: ["MEMORY_COMPRESS"],
     toolRank: ["TOOL_RANK", "TOOL_RANKER"],
     audio: ["AUDIO"],
     documents: ["DOCUMENTS", "RAG", "EMBEDDING"],
@@ -155,6 +156,25 @@ export function resolveAiRuntimeTarget(
     return {provider, purpose, model, baseUrl, apiKey, systemPromptAdditions};
 }
 
+function hasExplicitTargetConfig(provider: AiProvider, purpose: AiRuntimePurpose): boolean {
+    const prefix = providerPrefix(provider);
+    return [
+        ...endpointEnvNames(provider, purpose),
+        ...apiKeyEnvNames(provider, purpose),
+        ...modelEnvNames(provider, purpose),
+        ...systemPromptEnvNames(provider, purpose),
+    ].some(name => !!env(name)) || !!env(`${prefix}_${PURPOSE_SUFFIXES[purpose][0]}_MODEL`);
+}
+
+export function resolveOptionalAiRuntimeTarget(
+    provider: AiProvider,
+    purpose: AiRuntimePurpose,
+    modelOverride?: string,
+): AiRuntimeTarget | undefined {
+    if (!hasExplicitTargetConfig(provider, purpose)) return undefined;
+    return resolveAiRuntimeTarget(provider, purpose, modelOverride);
+}
+
 export function sameRuntimeEndpoint(left: AiRuntimeTarget, right: AiRuntimeTarget): boolean {
     return left.provider === right.provider
         && (left.baseUrl ?? "") === (right.baseUrl ?? "")

src/ai/conversation-pipeline.ts

@@ -16,6 +16,7 @@ import type {AttachmentKind, AiRuntimeTarget, RuntimeConfigSnapshot} from "./uni
 import type {OpenAIChatMessage} from "./openai-chat-message";
 import type {MistralChatMessage} from "./mistral-chat-message";
 import type {OllamaChatMessage} from "./ollama-chat-message";
+import {buildUserMemoryPrompt} from "./tools/user-memory.js";
 
 export type ConversationAttachment = {
     kind: AttachmentKind;
@@ -32,6 +33,7 @@ export type ConversationTurn = {
     content: string;
     deletedByBotAt?: number | null;
     attachments: ConversationAttachment[];
+    documentNames?: string[];
 };
 
 export type ConversationSnapshot = {
@@ -123,6 +125,13 @@ function attachmentSummary(attachments: ConversationAttachment[]): string {
     return ["[attachments]:", ...lines].join("\n");
 }
 
+function namesSummary(kind: string, names: string[]): string {
+    const filtered = names.map(name => name.trim()).filter(Boolean);
+    if (!filtered.length) return "";
+
+    return [`[${kind}]:`, ...filtered.map(name => `- ${name}`)].join("\n");
+}
+
 function supportedAttachmentKinds(provider: AiProvider, bot: boolean): Set<AttachmentKind> {
     if (bot) return new Set<AttachmentKind>();
 
@@ -160,6 +169,10 @@ function renderContentText(
         parts.push("[message_state]: deleted_by_bot");
     }
 
+    if (turn.documentNames?.length) {
+        parts.push(namesSummary("documents", turn.documentNames));
+    }
+
     if (unsupported.length) {
         parts.push(attachmentSummary(unsupported));
     }
@@ -255,11 +268,13 @@ function buildSystemInstruction(
     responseLanguage: UserAiResponseLanguage,
     includePythonToolPrompt: boolean,
     additions?: string | null,
+    memoryInstruction?: string | null,
 ): string {
     return [
         config.useSystemPrompt ? getResponseLanguageInstruction(responseLanguage) : null,
         config.systemPrompt && config.useSystemPrompt ? config.systemPrompt : null,
         additions?.trim() ? additions.trim() : null,
+        memoryInstruction?.trim() ? memoryInstruction.trim() : null,
         includePythonToolPrompt ? pythonInterpreterToolPrompt : null,
     ].filter(Boolean).join("\n\n");
 }
@@ -291,17 +306,19 @@ export async function buildConversationSnapshot(
             content: part.content,
             deletedByBotAt: part.deletedByBotAt,
             attachments: buildConversationAttachments(part),
+            documentNames: part.documentNames,
         }));
 
     const imageCount = turns.reduce((sum, turn) => {
         if (turn.bot) return sum;
         return sum + turn.attachments.filter(attachment => attachment.kind === "image").length;
     }, 0);
+    const memoryInstruction = await buildUserMemoryPrompt(msg.from?.id);
 
     return {
         turns,
         imageCount,
-        systemInstruction: buildSystemInstruction(config, responseLanguage, includePythonToolPrompt, runtimeTarget.systemPromptAdditions),
+        systemInstruction: buildSystemInstruction(config, responseLanguage, includePythonToolPrompt, runtimeTarget.systemPromptAdditions, memoryInstruction),
     };
 }
 

src/ai/document-rag-pipeline.ts

@@ -42,6 +42,10 @@ export async function prepareDocumentRag(
     const documents = downloads.filter(download => download.kind === "document");
     if (!documents.length) return undefined;
 
+    if (provider === AiProvider.OPENAI && config.openAiBackend === "compatible") {
+        return undefined;
+    }
+
     switch (provider) {
         case AiProvider.OPENAI: {
             const openAi = createOpenAiClient(config.openAiChatTarget);

src/ai/mcp/mcp-client.ts

@@ -0,0 +1,421 @@
+import {spawn, type ChildProcessWithoutNullStreams} from "node:child_process";
+import type {BoundaryValue} from "../../common/boundary-types.js";
+import {toolsLogger} from "../tools/tool-logger.js";
+import type {McpServerConfig} from "./mcp-config.js";
+
+const logger = toolsLogger.child("mcp");
+const MCP_PROTOCOL_VERSION = "2025-06-18";
+const DEFAULT_REQUEST_TIMEOUT_MS = 30_000;
+
+export type McpToolDefinition = {
+    name: string;
+    description?: string;
+    inputSchema?: BoundaryValue;
+};
+
+type JsonRpcRequest = {
+    jsonrpc: "2.0";
+    id: number;
+    method: string;
+    params?: BoundaryValue;
+};
+
+type JsonRpcNotification = {
+    jsonrpc: "2.0";
+    method: string;
+    params?: BoundaryValue;
+};
+
+type JsonRpcResponse = {
+    jsonrpc?: "2.0";
+    id?: BoundaryValue;
+    result?: BoundaryValue;
+    error?: {
+        code?: number;
+        message?: string;
+        data?: BoundaryValue;
+    };
+};
+
+interface JsonRpcTransport {
+    request(method: string, params?: BoundaryValue): Promise<BoundaryValue>;
+    notify(method: string, params?: BoundaryValue): Promise<void>;
+    close(): Promise<void>;
+}
+
+function isRecord(value: BoundaryValue): value is Record<string, BoundaryValue> {
+    return value !== null && typeof value === "object" && !Array.isArray(value);
+}
+
+function toJsonRpcResponse(value: BoundaryValue): JsonRpcResponse | undefined {
+    if (!isRecord(value)) return undefined;
+    if (value.jsonrpc !== undefined && value.jsonrpc !== "2.0") return undefined;
+    return value as JsonRpcResponse;
+}
+
+function extractJsonRpcResult(response: BoundaryValue, expectedId?: number): BoundaryValue {
+    const parsed = toJsonRpcResponse(response);
+    if (!parsed) {
+        throw new Error("Invalid JSON-RPC response from MCP server.");
+    }
+
+    if (parsed.error) {
+        throw new Error(parsed.error.message || "MCP server returned an error.");
+    }
+
+    if (expectedId !== undefined && parsed.id !== undefined && parsed.id !== expectedId) {
+        throw new Error(`Unexpected JSON-RPC response id from MCP server. Expected ${expectedId}, got ${String(parsed.id)}.`);
+    }
+
+    return parsed.result ?? {};
+}
+
+function parseSsePayload(text: string): BoundaryValue[] {
+    const events: string[] = [];
+    let current: string[] = [];
+
+    for (const rawLine of text.split(/\r?\n/)) {
+        const line = rawLine.trimEnd();
+
+        if (!line) {
+            if (current.length) {
+                events.push(current.join("\n"));
+                current = [];
+            }
+            continue;
+        }
+
+        if (line.startsWith("data:")) {
+            current.push(line.slice(5).replace(/^ /, ""));
+        }
+    }
+
+    if (current.length) {
+        events.push(current.join("\n"));
+    }
+
+    return events.map(event => {
+        try {
+            return JSON.parse(event) as BoundaryValue;
+        } catch {
+            return undefined;
+        }
+    }).filter((event): event is BoundaryValue => event !== undefined);
+}
+
+function timeoutPromise<T>(promise: Promise<T>, timeoutMs: number, label: string): Promise<T> {
+    if (!Number.isFinite(timeoutMs) || timeoutMs <= 0) return promise;
+
+    let timeoutId: NodeJS.Timeout | undefined;
+    const timeout = new Promise<T>((_, reject) => {
+        timeoutId = setTimeout(() => {
+            reject(new Error(`${label} timed out after ${timeoutMs}ms`));
+        }, timeoutMs);
+    });
+
+    return Promise.race([promise, timeout]).finally(() => {
+        if (timeoutId) clearTimeout(timeoutId);
+    });
+}
+
+class StdioJsonRpcTransport implements JsonRpcTransport {
+    private readonly process: ChildProcessWithoutNullStreams;
+    private readonly pending = new Map<number, {resolve: (value: BoundaryValue) => void; reject: (error: Error) => void;}>();
+    private buffer = "";
+    private nextId = 1;
+
+    constructor(private readonly config: McpServerConfig) {
+        if (!config.command) {
+            throw new Error(`MCP stdio server '${config.name}' is missing command.`);
+        }
+
+        this.process = spawn(config.command, config.args ?? [], {
+            cwd: config.cwd,
+            env: {
+                ...process.env,
+                ...config.env,
+            },
+            stdio: ["pipe", "pipe", "pipe"],
+            windowsHide: true,
+        });
+
+        this.process.stdout.on("data", chunk => this.handleStdout(chunk));
+        this.process.stderr.on("data", chunk => {
+            const text = chunk.toString("utf8").trim();
+            if (text) logger.debug("stdio.stderr", {server: config.name, text});
+        });
+        this.process.on("error", error => this.failAll(error));
+        this.process.on("exit", code => this.failAll(new Error(`MCP stdio server '${config.name}' exited with code ${code ?? "unknown"}.`)));
+    }
+
+    private handleStdout(chunk: Buffer): void {
+        this.buffer += chunk.toString("utf8");
+
+        let newlineIndex = this.buffer.indexOf("\n");
+        while (newlineIndex !== -1) {
+            const line = this.buffer.slice(0, newlineIndex).trim();
+            this.buffer = this.buffer.slice(newlineIndex + 1);
+            newlineIndex = this.buffer.indexOf("\n");
+
+            if (!line) continue;
+
+            try {
+                const message = JSON.parse(line) as JsonRpcResponse | JsonRpcNotification;
+                if ("id" in message && message.id !== undefined) {
+                    const pending = this.pending.get(Number(message.id));
+                    if (pending) {
+                        this.pending.delete(Number(message.id));
+                        if ("error" in message && message.error) {
+                            pending.reject(new Error(message.error.message || "MCP stdio request failed."));
+                        } else {
+                            pending.resolve((message as JsonRpcResponse).result ?? {});
+                        }
+                    }
+                    continue;
+                }
+
+                if ("method" in message) {
+                    logger.debug("stdio.notification", {server: this.config.name, method: message.method});
+                }
+            } catch (error) {
+                logger.warn("stdio.parse_failed", {
+                    server: this.config.name,
+                    line: line.slice(0, 500),
+                    error: error instanceof Error ? error.message : String(error),
+                });
+            }
+        }
+    }
+
+    private failAll(error: Error): void {
+        for (const pending of this.pending.values()) {
+            pending.reject(error);
+        }
+        this.pending.clear();
+    }
+
+    async request(method: string, params?: BoundaryValue): Promise<BoundaryValue> {
+        if (this.process.exitCode !== null) {
+            throw new Error(`MCP stdio server '${this.config.name}' is not running.`);
+        }
+
+        const id = this.nextId++;
+        const request: JsonRpcRequest = {
+            jsonrpc: "2.0",
+            id,
+            method,
+            params,
+        };
+
+        const result = new Promise<BoundaryValue>((resolve, reject) => {
+            this.pending.set(id, {resolve, reject});
+        });
+
+        this.process.stdin.write(`${JSON.stringify(request)}\n`);
+        return timeoutPromise(result, this.config.timeoutMs ?? DEFAULT_REQUEST_TIMEOUT_MS, `${this.config.name}.${method}`);
+    }
+
+    async notify(method: string, params?: BoundaryValue): Promise<void> {
+        if (this.process.exitCode !== null) {
+            throw new Error(`MCP stdio server '${this.config.name}' is not running.`);
+        }
+
+        const notification: JsonRpcNotification = {
+            jsonrpc: "2.0",
+            method,
+            params,
+        };
+
+        this.process.stdin.write(`${JSON.stringify(notification)}\n`);
+    }
+
+    async close(): Promise<void> {
+        this.failAll(new Error(`MCP stdio server '${this.config.name}' closed.`));
+        if (!this.process.killed) {
+            this.process.kill();
+        }
+    }
+}
+
+class HttpJsonRpcTransport implements JsonRpcTransport {
+    private nextId = 1;
+    private sessionId?: string;
+
+    constructor(private readonly config: McpServerConfig) {
+        if (!config.url) {
+            throw new Error(`MCP HTTP server '${config.name}' is missing url.`);
+        }
+    }
+
+    private async post(body: BoundaryValue): Promise<Response> {
+        const controller = new AbortController();
+        const timeoutMs = this.config.timeoutMs ?? DEFAULT_REQUEST_TIMEOUT_MS;
+        const timeoutId = setTimeout(() => controller.abort(), timeoutMs);
+
+        try {
+            return await fetch(this.config.url!, {
+                method: "POST",
+                headers: {
+                    "Content-Type": "application/json",
+                    "Accept": "application/json, text/event-stream",
+                    ...(this.sessionId ? {"Mcp-Session-Id": this.sessionId} : {}),
+                    ...(this.config.headers ?? {}),
+                },
+                body: JSON.stringify(body),
+                signal: controller.signal,
+            }).finally(() => clearTimeout(timeoutId));
+        } catch (error) {
+            clearTimeout(timeoutId);
+            throw error;
+        }
+    }
+
+    async request(method: string, params?: BoundaryValue): Promise<BoundaryValue> {
+        const id = this.nextId++;
+        const request: JsonRpcRequest = {
+            jsonrpc: "2.0",
+            id,
+            method,
+            params,
+        };
+
+        const response = await this.post(request);
+        const sessionId = response.headers.get("Mcp-Session-Id");
+        if (sessionId) {
+            this.sessionId = sessionId;
+        }
+
+        if (!response.ok) {
+            const errorText = await response.text().catch(() => "");
+            throw new Error(`MCP HTTP server '${this.config.name}' returned ${response.status}: ${errorText || response.statusText}`);
+        }
+
+        const contentType = response.headers.get("content-type")?.toLowerCase() ?? "";
+        let payload: BoundaryValue;
+
+        if (contentType.includes("text/event-stream")) {
+            const text = await response.text();
+            const messages = parseSsePayload(text);
+            const responseMessage = messages.map(toJsonRpcResponse).find(message => message?.id === id && (message.result !== undefined || message.error));
+            payload = extractJsonRpcResult(responseMessage ?? messages[0] ?? {}, id);
+        } else {
+            payload = extractJsonRpcResult(await response.json() as BoundaryValue, id);
+        }
+
+        return payload;
+    }
+
+    async notify(method: string, params?: BoundaryValue): Promise<void> {
+        const response = await this.post({
+            jsonrpc: "2.0",
+            method,
+            params,
+        });
+
+        const sessionId = response.headers.get("Mcp-Session-Id");
+        if (sessionId) {
+            this.sessionId = sessionId;
+        }
+
+        if (!response.ok && response.status !== 202) {
+            const errorText = await response.text().catch(() => "");
+            throw new Error(`MCP HTTP notification failed for '${this.config.name}' with ${response.status}: ${errorText || response.statusText}`);
+        }
+    }
+
+    async close(): Promise<void> {
+        return;
+    }
+}
+
+function createTransport(config: McpServerConfig): JsonRpcTransport {
+    return config.transport === "stdio"
+        ? new StdioJsonRpcTransport(config)
+        : new HttpJsonRpcTransport(config);
+}
+
+function normalizeToolResultContent(content: BoundaryValue): string {
+    if (content === undefined || content === null) return "";
+    if (typeof content === "string") return content;
+    if (typeof content === "number" || typeof content === "boolean") return String(content);
+    if (Array.isArray(content)) return content.map(item => normalizeToolResultContent(item)).filter(Boolean).join("\n");
+    if (!isRecord(content)) return JSON.stringify(content);
+
+    if (content.type === "text" && typeof content.text === "string") return content.text;
+    if (content.type === "image") {
+        return `[image ${typeof content.mimeType === "string" ? content.mimeType : "unknown"}]`;
+    }
+    if (content.type === "resource" && isRecord(content.resource)) {
+        if (typeof content.resource.text === "string") return content.resource.text;
+        return JSON.stringify(content.resource);
+    }
+
+    return JSON.stringify(content);
+}
+
+export class McpClient {
+    private readonly transport: JsonRpcTransport;
+    private initialized = false;
+
+    constructor(readonly config: McpServerConfig) {
+        this.transport = createTransport(config);
+    }
+
+    async initialize(): Promise<void> {
+        if (this.initialized) return;
+
+        await this.transport.request("initialize", {
+            protocolVersion: MCP_PROTOCOL_VERSION,
+            clientInfo: {
+                name: "tg-chat-bot",
+                version: "1.0.0",
+            },
+            capabilities: {},
+        });
+
+        await this.transport.notify("notifications/initialized");
+        this.initialized = true;
+    }
+
+    async listTools(): Promise<McpToolDefinition[]> {
+        await this.initialize();
+        const result = await this.transport.request("tools/list");
+
+        if (!isRecord(result)) return [];
+
+        const tools = Array.isArray(result.tools) ? result.tools : [];
+        return tools.flatMap(tool => {
+            if (!isRecord(tool) || typeof tool.name !== "string") return [];
+            return [{
+                name: tool.name,
+                description: typeof tool.description === "string" ? tool.description : undefined,
+                inputSchema: tool.inputSchema,
+            }];
+        });
+    }
+
+    async callTool(name: string, args?: BoundaryValue): Promise<string> {
+        await this.initialize();
+        const result = await this.transport.request("tools/call", {
+            name,
+            arguments: args ?? {},
+        });
+
+        if (!isRecord(result)) {
+            return normalizeToolResultContent(result);
+        }
+
+        const content = Array.isArray(result.content) ? result.content : [];
+        const text = content.map(item => normalizeToolResultContent(item)).filter(Boolean).join("\n");
+
+        if (result.isError) {
+            return text ? `[MCP error] ${text}` : "[MCP error]";
+        }
+
+        return text || JSON.stringify(result);
+    }
+
+    async close(): Promise<void> {
+        await this.transport.close();
+    }
+}

src/ai/mcp/mcp-config.ts

@@ -0,0 +1,106 @@
+import type {BoundaryValue} from "../../common/boundary-types.js";
+
+export type McpTransport = "stdio" | "http";
+
+export type McpServerConfig = {
+    name: string;
+    transport: McpTransport;
+    command?: string;
+    args?: string[];
+    cwd?: string;
+    env?: Record<string, string>;
+    url?: string;
+    headers?: Record<string, string>;
+    timeoutMs?: number;
+};
+
+function isRecord(value: BoundaryValue): value is Record<string, BoundaryValue> {
+    return value !== null && typeof value === "object" && !Array.isArray(value);
+}
+
+function asString(value: BoundaryValue): string | undefined {
+    return typeof value === "string" && value.trim().length > 0 ? value.trim() : undefined;
+}
+
+function toStringRecord(value: BoundaryValue): Record<string, string> | undefined {
+    if (!isRecord(value)) return undefined;
+
+    const result: Record<string, string> = {};
+    for (const [key, entry] of Object.entries(value)) {
+        if (typeof entry === "string" || typeof entry === "number" || typeof entry === "boolean") {
+            result[key] = String(entry);
+        }
+    }
+
+    return Object.keys(result).length ? result : undefined;
+}
+
+function toStringArray(value: BoundaryValue): string[] | undefined {
+    if (!Array.isArray(value)) return undefined;
+    const items = value.filter((item): item is string => typeof item === "string" && item.trim().length > 0)
+        .map(item => item.trim());
+    return items.length ? items : undefined;
+}
+
+function toPositiveInt(value: BoundaryValue): number | undefined {
+    const n = typeof value === "number"
+        ? value
+        : typeof value === "string"
+            ? Number(value)
+            : NaN;
+
+    if (!Number.isFinite(n) || n <= 0) return undefined;
+    return Math.floor(n);
+}
+
+function normalizeServerConfig(value: BoundaryValue, fallbackName?: string): McpServerConfig | undefined {
+    if (!isRecord(value)) return undefined;
+
+    const name = asString(value.name) ?? fallbackName;
+    const transportRaw = asString(value.transport);
+    const transport = transportRaw === "http" || transportRaw === "stdio" ? transportRaw : undefined;
+
+    if (!name || !transport) return undefined;
+
+    return {
+        name,
+        transport,
+        command: asString(value.command),
+        args: toStringArray(value.args),
+        cwd: asString(value.cwd),
+        env: toStringRecord(value.env),
+        url: asString(value.url),
+        headers: toStringRecord(value.headers),
+        timeoutMs: toPositiveInt(value.timeoutMs),
+    };
+}
+
+export function parseMcpServerConfigs(raw: string | undefined): McpServerConfig[] {
+    if (!raw?.trim()) return [];
+
+    let parsed: BoundaryValue;
+    try {
+        parsed = JSON.parse(raw) as BoundaryValue;
+    } catch (error) {
+        throw new Error(`Invalid MCP_SERVERS JSON: ${error instanceof Error ? error.message : String(error)}`);
+    }
+
+    if (Array.isArray(parsed)) {
+        return parsed.flatMap((item, index) => normalizeServerConfig(item, `server-${index + 1}`) ? [normalizeServerConfig(item, `server-${index + 1}`)!] : []);
+    }
+
+    if (!isRecord(parsed)) {
+        return [];
+    }
+
+    if (Array.isArray(parsed.servers)) {
+        return parsed.servers.flatMap((item, index) => normalizeServerConfig(item, `server-${index + 1}`) ? [normalizeServerConfig(item, `server-${index + 1}`)!] : []);
+    }
+
+    if (isRecord(parsed.mcpServers)) {
+        return Object.entries(parsed.mcpServers).flatMap(([name, item]) => normalizeServerConfig(item, name) ? [normalizeServerConfig(item, name)!] : []);
+    }
+
+    const single = normalizeServerConfig(parsed);
+    return single ? [single] : [];
+}

src/ai/mcp/mcp-json-schema.ts

@@ -0,0 +1,123 @@
+import type {AiJsonValue, AiToolParameters} from "../tool-types.js";
+import type {BoundaryValue} from "../../common/boundary-types.js";
+
+type JsonSchemaRecord = Record<string, BoundaryValue>;
+
+function isRecord(value: BoundaryValue): value is JsonSchemaRecord {
+    return value !== null && typeof value === "object" && !Array.isArray(value);
+}
+
+function toAiJsonValue(value: BoundaryValue): AiJsonValue | undefined {
+    if (value === undefined) return undefined;
+    if (value === null) return null;
+    if (typeof value === "string" || typeof value === "number" || typeof value === "boolean") return value;
+
+    if (Array.isArray(value)) {
+        return value.map(item => toAiJsonValue(item) ?? null);
+    }
+
+    if (!isRecord(value)) return undefined;
+
+    const result: Record<string, AiJsonValue> = {};
+    for (const [key, entry] of Object.entries(value)) {
+        const normalized = toAiJsonValue(entry);
+        if (normalized !== undefined) {
+            result[key] = normalized;
+        }
+    }
+
+    return result;
+}
+
+function normalizeType(value: BoundaryValue): AiToolParameters["type"] | undefined {
+    const candidates = Array.isArray(value)
+        ? value.filter((item): item is string => typeof item === "string")
+        : typeof value === "string"
+            ? [value]
+            : [];
+
+    const prioritized = candidates.find(item => item !== "null") ?? candidates[0];
+    if (!prioritized) return undefined;
+
+    switch (prioritized) {
+        case "object":
+        case "string":
+        case "number":
+        case "integer":
+        case "boolean":
+        case "array":
+            return prioritized;
+        default:
+            return undefined;
+    }
+}
+
+export function convertJsonSchemaToToolParameters(schema: BoundaryValue): AiToolParameters | undefined {
+    if (!isRecord(schema)) return undefined;
+
+    const declaredType = normalizeType(schema.type);
+    const inferredType = declaredType
+        ?? (schema.properties !== undefined || schema.additionalProperties !== undefined ? "object" : undefined)
+        ?? (schema.items !== undefined ? "array" : undefined)
+        ?? "object";
+
+    const result: AiToolParameters = {
+        type: inferredType,
+    };
+
+    const description = typeof schema.description === "string" && schema.description.trim().length > 0
+        ? schema.description.trim()
+        : undefined;
+    if (description) result.description = description;
+
+    const defaultValue = toAiJsonValue(schema.default);
+    if (defaultValue !== undefined) result.default = defaultValue;
+
+    if (Array.isArray(schema.enum)) {
+        const enumValues = schema.enum
+            .filter((item): item is string => typeof item === "string" && item.length > 0);
+        if (enumValues.length) result.enum = enumValues;
+    }
+
+    if (typeof schema.minItems === "number") result.minItems = schema.minItems;
+    if (typeof schema.maxItems === "number") result.maxItems = schema.maxItems;
+    if (typeof schema.minimum === "number") result.minimum = schema.minimum;
+    if (typeof schema.maximum === "number") result.maximum = schema.maximum;
+
+    if (Array.isArray(schema.required)) {
+        const required = schema.required.filter((item): item is string => typeof item === "string" && item.trim().length > 0);
+        if (required.length) result.required = required;
+    }
+
+    if (inferredType === "object" || schema.properties !== undefined || schema.additionalProperties !== undefined) {
+        if (isRecord(schema.properties)) {
+            const properties: Record<string, AiToolParameters> = {};
+            for (const [key, value] of Object.entries(schema.properties)) {
+                const converted = convertJsonSchemaToToolParameters(value);
+                if (converted) properties[key] = converted;
+            }
+            if (Object.keys(properties).length) result.properties = properties;
+        }
+
+        if (schema.additionalProperties !== undefined) {
+            result.additionalProperties = typeof schema.additionalProperties === "boolean"
+                ? schema.additionalProperties
+                : convertJsonSchemaToToolParameters(schema.additionalProperties);
+        }
+    }
+
+    if (inferredType === "array" || schema.items !== undefined) {
+        if (Array.isArray(schema.items)) {
+            const firstItem = schema.items[0];
+            if (firstItem !== undefined) {
+                const converted = convertJsonSchemaToToolParameters(firstItem);
+                if (converted) result.items = converted;
+            }
+        } else {
+            const converted = convertJsonSchemaToToolParameters(schema.items);
+            if (converted) result.items = converted;
+        }
+    }
+
+    return result;
+}

src/ai/mcp/mcp-registry.ts

@@ -0,0 +1,165 @@
+import {Environment} from "../../common/environment.js";
+import type {AiTool} from "../tool-types.js";
+import type {ToolHandler} from "../tools/types.js";
+import {normalizeToolArguments} from "../tools/utils.js";
+import {toolsLogger} from "../tools/tool-logger.js";
+import {convertJsonSchemaToToolParameters} from "./mcp-json-schema.js";
+import {McpClient, type McpToolDefinition} from "./mcp-client.js";
+import {parseMcpServerConfigs, type McpServerConfig} from "./mcp-config.js";
+
+const logger = toolsLogger.child("mcp-registry");
+
+type McpToolBinding = {
+    server: McpServerConfig;
+    client: McpClient;
+    remoteToolName: string;
+    localToolName: string;
+    tool: AiTool;
+};
+
+type McpInitSummary = {
+    servers: number;
+    loadedServers: number;
+    tools: number;
+    failedServers: string[];
+};
+
+const toolBindings = new Map<string, McpToolBinding>();
+const clients = new Map<string, McpClient>();
+let initPromise: Promise<McpInitSummary> | undefined;
+
+function sanitizeSegment(value: string): string {
+    return value
+        .trim()
+        .replace(/[^a-zA-Z0-9_]+/g, "_")
+        .replace(/_+/g, "_")
+        .replace(/^_+|_+$/g, "") || "tool";
+}
+
+function buildLocalToolName(serverName: string, toolName: string): string {
+    return `mcp__${sanitizeSegment(serverName)}__${sanitizeSegment(toolName)}`;
+}
+
+function buildTool(serverName: string, tool: McpToolDefinition): AiTool {
+    const localName = buildLocalToolName(serverName, tool.name);
+    const description = tool.description?.trim()
+        ? `[MCP ${serverName}] ${tool.description.trim()}`
+        : `[MCP ${serverName}] ${tool.name}`;
+
+    return {
+        type: "function",
+        function: {
+            name: localName,
+            description,
+            parameters: convertJsonSchemaToToolParameters(tool.inputSchema),
+        },
+    };
+}
+
+async function loadServer(config: McpServerConfig): Promise<{loaded: boolean; tools: number}> {
+    const client = new McpClient(config);
+    clients.set(config.name, client);
+
+    try {
+        const remoteTools = await client.listTools();
+        let loaded = 0;
+
+        for (const remoteTool of remoteTools) {
+            const localName = buildLocalToolName(config.name, remoteTool.name);
+            if (toolBindings.has(localName)) {
+                logger.warn("tool.duplicate", {
+                    server: config.name,
+                    tool: remoteTool.name,
+                    localName,
+                });
+                continue;
+            }
+
+            const binding: McpToolBinding = {
+                server: config,
+                client,
+                remoteToolName: remoteTool.name,
+                localToolName: localName,
+                tool: buildTool(config.name, remoteTool),
+            };
+
+            toolBindings.set(localName, binding);
+            loaded += 1;
+        }
+
+        logger.info("server.loaded", {
+            server: config.name,
+            transport: config.transport,
+            tools: loaded,
+        });
+        return {loaded: true, tools: loaded};
+    } catch (error) {
+        logger.error("server.failed", {
+            server: config.name,
+            transport: config.transport,
+            error: error instanceof Error ? error.message : String(error),
+        });
+        await client.close().catch(() => undefined);
+        clients.delete(config.name);
+        return {loaded: false, tools: 0};
+    }
+}
+
+export async function initializeMcpTools(): Promise<McpInitSummary> {
+    if (initPromise) return initPromise;
+
+    initPromise = (async () => {
+        toolBindings.clear();
+        await Promise.all([...clients.values()].map(client => client.close().catch(() => undefined)));
+        clients.clear();
+
+        const configs = parseMcpServerConfigs(Environment.MCP_SERVERS);
+        const results = await Promise.all(configs.map(config => loadServer(config)));
+
+        return {
+            servers: configs.length,
+            loadedServers: results.filter(result => result.loaded).length,
+            tools: [...results].reduce((sum, result) => sum + result.tools, 0),
+            failedServers: configs.filter((_, index) => !results[index]?.loaded).map(config => config.name),
+        };
+    })();
+
+    try {
+        const summary = await initPromise;
+        logger.info("init.done", summary);
+        return summary;
+    } catch (error) {
+        initPromise = undefined;
+        logger.error("init.failed", {error: error instanceof Error ? error.message : String(error)});
+        throw error;
+    }
+}
+
+export function getMcpTools(): AiTool[] {
+    return [...toolBindings.values()].map(binding => binding.tool);
+}
+
+export function getMcpToolHandlers(): Record<string, ToolHandler> {
+    const handlers: Record<string, ToolHandler> = {};
+
+    for (const binding of toolBindings.values()) {
+        handlers[binding.localToolName] = async args => {
+            const normalized = normalizeToolArguments(args, undefined);
+            return binding.client.callTool(binding.remoteToolName, normalized);
+        };
+    }
+
+    return handlers;
+}
+
+export function getMcpToolPrompts(_toolNames: string[]): string[] {
+    return [];
+}
+
+export async function shutdownMcpTools(): Promise<void> {
+    initPromise = undefined;
+    toolBindings.clear();
+
+    await Promise.all([...clients.values()].map(client => client.close().catch(() => undefined)));
+    clients.clear();
+}

src/ai/model-call-stage.ts

@@ -0,0 +1,5 @@
+export async function runSingleModelRequest<T>(params: {
+    execute: () => Promise<T>;
+}): Promise<T> {
+    return await params.execute();
+}

src/ai/openai-chat-completions.ts

@@ -0,0 +1,66 @@
+import {isRecord} from "./unified-ai-runner.shared.js";
+import type {OpenAIChatMessage, OpenAICompatibleChatMessage} from "./openai-chat-message.js";
+import type {ToolCallData} from "./unified-ai-runner.shared.js";
+
+export function responseContentToText(content: unknown): string {
+    if (typeof content === "string") return content;
+    if (!Array.isArray(content)) return "";
+
+    return content
+        .map(part => isRecord(part) && typeof part.text === "string" ? part.text : "")
+        .join("");
+}
+
+export function openAiResponseMessagesToChatCompletions(messages: OpenAIChatMessage[]): OpenAICompatibleChatMessage[] {
+    return messages.map((message): OpenAICompatibleChatMessage => {
+        if (message.role === "system") {
+            return {role: "system", content: responseContentToText(message.content)};
+        }
+
+        if (message.role === "assistant") {
+            const text = responseContentToText(message.content);
+            return text.length
+                ? {role: "assistant", content: text}
+                : {role: "assistant", content: null};
+        }
+
+        const content = Array.isArray(message.content)
+            ? (() => {
+                const parts = message.content.map((part): {type: "text"; text: string} | {type: "image_url"; image_url: {url: string}} => {
+                    if (isRecord(part) && part.type === "input_image") {
+                        return {
+                            type: "image_url",
+                            image_url: {url: String(part.image_url ?? "")},
+                        };
+                    }
+
+                    return {
+                        type: "text",
+                        text: isRecord(part) && typeof part.text === "string" ? part.text : "",
+                    };
+                });
+
+                return parts.every(part => part.type === "text")
+                    ? parts.map(part => part.text).join("")
+                    : parts;
+            })()
+            : message.content;
+
+        return {role: "user", content};
+    });
+}
+
+export function buildAssistantToolMessage(calls: ToolCallData[], text: string): OpenAICompatibleChatMessage {
+    return {
+        role: "assistant",
+        content: text,
+        tool_calls: calls.map(call => ({
+            id: call.id,
+            type: "function",
+            function: {
+                name: call.name,
+                arguments: call.argumentsText,
+            },
+        })),
+    };
+}

src/ai/openai-chat-message.ts

@@ -2,6 +2,7 @@ import type {
     ResponseInputMessageContentList,
     ResponseOutputMessage,
 } from "openai/resources/responses/responses";
+import type {ChatCompletionMessageParam} from "openai/resources/chat/completions";
 
 type OpenAIInputChatMessage = {
     type: "message";
@@ -17,3 +18,5 @@ type OpenAIOutputChatMessage = {
 } & Pick<ResponseOutputMessage, "id" | "status">;
 
 export type OpenAIChatMessage = OpenAIInputChatMessage | OpenAIOutputChatMessage;
+
+export type OpenAICompatibleChatMessage = ChatCompletionMessageParam;

src/ai/openai-upload-files.ts

@@ -0,0 +1,74 @@
+import {Message} from "typescript-telegram-bot-api";
+import fs from "node:fs";
+import path from "node:path";
+import {bot} from "../index.js";
+import {Environment} from "../common/environment.js";
+import {logError} from "../util/utils.js";
+import {errorMessage} from "./unified-ai-runner.shared.js";
+import {SendFileAttachmentResult, SendFileAttachmentResultSchema} from "./tools/files.js";
+
+export async function tryToUploadFiles(
+    msg: Message,
+    toolResults: string[]
+): Promise<
+    | { found: false }
+    | { found: true, uploaded: true }
+    | { found: boolean, uploaded: false, error: string, toolIndex: number }
+> {
+    let sendFileAttachment: {
+        result: SendFileAttachmentResult & { success: true },
+        toolIndex: number
+    } | null = null;
+
+    let found = false;
+
+    try {
+        for (const [index, toolResult] of toolResults.entries()) {
+            const raw = JSON.parse(toolResult);
+            const res = SendFileAttachmentResultSchema.safeParse(raw);
+
+            if (res.success) {
+                found = true;
+
+                if (res.data.success) {
+                    sendFileAttachment = {result: res.data, toolIndex: index};
+                }
+            }
+        }
+
+        if (!found) {
+            return {found: false};
+        }
+
+        const attachmentRoot = Environment.FILE_TOOLS_ROOT_DIR;
+        const attachmentPath = attachmentRoot
+            ? path.join(
+                attachmentRoot,
+                String(msg.from?.id),
+                sendFileAttachment?.result?.attachment?.relativePath ?? "",
+            )
+            : "";
+
+        if (!fs.existsSync(attachmentPath)) {
+            throw new Error(`Attachment file does not exist: ${attachmentPath}`);
+        }
+
+        await bot.sendDocument({
+            chat_id: msg.chat.id,
+            reply_parameters: {
+                message_id: msg.message_id,
+            },
+            document: fs.createReadStream(attachmentPath),
+        });
+
+        return {found: true, uploaded: true};
+    } catch (e) {
+        logError(e instanceof Error ? e : String(e));
+        return {
+            found: found,
+            uploaded: false,
+            error: errorMessage(e instanceof Error ? e : String(e)),
+            toolIndex: sendFileAttachment?.toolIndex ?? -1
+        };
+    }
+}

src/ai/provider-adapter-contract.ts

@@ -0,0 +1,198 @@
+import type {ToolCallData} from "./unified-ai-runner.shared.js";
+import type {ResponseStreamEvent} from "openai/resources/responses/responses";
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+    return !!value && typeof value === "object" && !Array.isArray(value);
+}
+
+function normalizeToolCallId(value: unknown, fallback: string): string {
+    return typeof value === "string" && value.trim().length > 0 ? value : fallback;
+}
+
+function normalizeToolArguments(value: unknown): string {
+    if (typeof value === "string") return value;
+    return JSON.stringify(value ?? {});
+}
+
+function normalizeToolArgumentsChunk(value: unknown): string {
+    if (typeof value === "string") return value;
+    if (value === undefined || value === null) return "";
+    return JSON.stringify(value);
+}
+
+export function extractOpenAiToolCalls(response: unknown): ToolCallData[] {
+    const output = isRecord(response) && Array.isArray(response.output) ? response.output : [];
+
+    return output
+        .filter(item => isRecord(item) && item.type === "function_call" && (typeof item.call_id === "string" || typeof item.name === "string"))
+        .map((item, index) => ({
+            id: normalizeToolCallId(item.call_id, `openai_${index}`),
+            name: typeof item.name === "string" ? item.name : "",
+            argumentsText: normalizeToolArguments(item.arguments),
+        }))
+        .filter(call => call.name.length > 0);
+}
+
+export function extractOpenAiTextDelta(input: unknown): string {
+    const event = input as ResponseStreamEvent | undefined;
+    return event?.type === "response.output_text.delta" ? event.delta ?? "" : "";
+}
+
+export function extractOpenAiChatTextDelta(input: unknown): string {
+    const event = isRecord(input) ? input : undefined;
+    const choice = event && Array.isArray(event.choices) && isRecord(event.choices[0]) ? event.choices[0] : undefined;
+    const delta = isRecord(choice?.delta) ? choice.delta : undefined;
+    const content = delta && typeof delta.content === "string" ? delta.content : "";
+    return content;
+}
+
+export function normalizeStreamingTextDelta(existingText: string, deltaText: string): string {
+    if (!deltaText) return "";
+    if (!existingText) return deltaText;
+
+    if (deltaText.startsWith(existingText)) {
+        return deltaText.slice(existingText.length);
+    }
+
+    return deltaText;
+}
+
+export function extractOpenAiChatToolCalls(response: unknown): ToolCallData[] {
+    const record = isRecord(response) ? response : undefined;
+    const choice = record && Array.isArray(record.choices) && isRecord(record.choices[0]) ? record.choices[0] : undefined;
+    const message = isRecord(choice?.message) ? choice.message : undefined;
+    const toolCalls = message && Array.isArray(message.tool_calls) ? message.tool_calls : [];
+
+    return toolCalls
+        .filter((item, index) => isRecord(item) && ((typeof item.id === "string") || typeof item.index === "number" || index >= 0))
+        .map((item, index) => {
+            const call = isRecord(item) ? item : {};
+            const fn = isRecord(call.function) ? call.function : undefined;
+            const name = typeof fn?.name === "string" ? fn.name : typeof call.name === "string" ? call.name : "";
+            return {
+                id: normalizeToolCallId(call.id, `openai_chat_${typeof call.index === "number" ? call.index : index}`),
+                name,
+                argumentsText: normalizeToolArguments(fn?.arguments ?? call.arguments),
+            };
+        })
+        .filter(call => call.name.length > 0);
+}
+
+export function extractOpenAiChatStreamingToolCalls(input: unknown): ToolCallData[] {
+    const event = isRecord(input) ? input : undefined;
+    const choice = event && Array.isArray(event.choices) && isRecord(event.choices[0]) ? event.choices[0] : undefined;
+    const delta = isRecord(choice?.delta) ? choice.delta : undefined;
+    const toolCalls = Array.isArray(delta?.tool_calls) ? delta.tool_calls : [];
+
+    return toolCalls
+        .map((item, index) => {
+            const call = isRecord(item) ? item : {};
+            const fn = isRecord(call.function) ? call.function : undefined;
+            const name = typeof fn?.name === "string" ? fn.name : typeof call.name === "string" ? call.name : "";
+            return {
+                id: normalizeToolCallId(call.id, `openai_chat_${typeof call.index === "number" ? call.index : index}`),
+                name,
+                argumentsText: normalizeToolArgumentsChunk(fn?.arguments ?? call.arguments),
+            };
+        })
+        .filter(call => call.id.length > 0);
+}
+
+export function mergeToolCallChunks(existing: ToolCallData[], chunks: ToolCallData[]): ToolCallData[] {
+    const merged = new Map<string, ToolCallData>(existing.map(call => [call.id, {...call}]));
+
+    for (const chunk of chunks) {
+        const current = merged.get(chunk.id);
+        if (!current) {
+            merged.set(chunk.id, {...chunk});
+            continue;
+        }
+
+        merged.set(chunk.id, {
+            id: current.id,
+            name: current.name || chunk.name,
+            argumentsText: current.argumentsText + (chunk.argumentsText ?? ""),
+        });
+    }
+
+    return [...merged.values()];
+}
+
+export function extractOpenAiStreamingToolCalls(input: unknown): ToolCallData[] {
+    const event = input as ResponseStreamEvent | undefined;
+    if (event?.type === "response.output_item.added" && isRecord(event.item) && event.item.type === "function_call") {
+        return extractOpenAiToolCalls({
+            output: [{
+                type: "function_call",
+                call_id: event.item.call_id ?? event.item.id,
+                name: event.item.name,
+                arguments: event.item.arguments,
+            }],
+        });
+    }
+
+    return [];
+}
+
+export function extractMistralToolCalls(calls: unknown): ToolCallData[] {
+    const normalized = Array.isArray(calls)
+        ? calls
+        : isRecord(calls) && (Array.isArray(calls.toolCalls) || Array.isArray(calls.tool_calls))
+            ? (calls.toolCalls ?? calls.tool_calls)
+            : [];
+
+    if (!Array.isArray(normalized)) return [];
+
+    return normalized
+        .map((item, index) => {
+            const call = isRecord(item) ? item : {};
+            const fn = isRecord(call.function) ? call.function : undefined;
+            const name = typeof fn?.name === "string" ? fn.name : typeof call.name === "string" ? call.name : "";
+            return {
+                id: normalizeToolCallId(call.id, `mistral_${index}`),
+                name,
+                argumentsText: normalizeToolArguments(fn?.arguments ?? call.arguments),
+            };
+        })
+        .filter(call => call.name.length > 0);
+}
+
+export function extractMistralTextDelta(input: unknown): string {
+    const delta = isRecord(input) ? input : {};
+    const content = delta.content;
+    if (typeof content === "string") return content;
+    if (Array.isArray(content)) {
+        return content
+            .map(part => isRecord(part) && typeof part.text === "string" ? part.text : "")
+            .join("");
+    }
+    return "";
+}
+
+export function extractOllamaToolCalls(calls: unknown): ToolCallData[] {
+    const normalized = Array.isArray(calls)
+        ? calls
+        : isRecord(calls) && Array.isArray(calls.tool_calls)
+            ? calls.tool_calls
+            : [];
+
+    if (!Array.isArray(normalized)) return [];
+
+    return normalized
+        .map((item, index) => {
+            const call = isRecord(item) ? item : {};
+            const fn = isRecord(call.function) ? call.function : undefined;
+            const name = typeof fn?.name === "string" ? fn.name : typeof call.name === "string" ? call.name : "";
+            return {
+                id: normalizeToolCallId(call.id, `ollama_${index}`),
+                name,
+                argumentsText: normalizeToolArguments(fn?.arguments ?? call.arguments),
+            };
+        })
+        .filter(call => call.name.length > 0);
+}
+
+export function extractOllamaTextDelta(input: unknown): string {
+    const chunk = isRecord(input) ? input.message : undefined;
+    return isRecord(chunk) && typeof chunk.content === "string" ? chunk.content : "";
+}

src/ai/provider-adapters.ts

@@ -0,0 +1,196 @@
+import {AiProvider} from "../model/ai-provider.js";
+import type {BoundaryValue} from "../common/boundary-types.js";
+import type {RuntimeConfigSnapshot, ToolCallData} from "./unified-ai-runner.shared.js";
+import {getMistralTools, getOllamaTools, getOpenAIResponsesTools, getOpenAICodeInterpreterTool} from "./tool-mappers.js";
+import type {MistralChatMessage as MistralMessageType} from "./mistral-chat-message.js";
+import type {OpenAIChatMessage as OpenAiMessageType} from "./openai-chat-message.js";
+import type {Message as OllamaMessage} from "ollama";
+import {
+    extractMistralTextDelta,
+    extractMistralToolCalls,
+    extractOllamaTextDelta,
+    extractOllamaToolCalls,
+    extractOpenAiTextDelta,
+    extractOpenAiStreamingToolCalls,
+    extractOpenAiToolCalls,
+} from "./provider-adapter-contract.js";
+
+export type ProviderRankToolOptions = {
+    forCreator?: boolean;
+    vectorStoreIds?: string[];
+};
+
+export interface AiProviderAdapter {
+    readonly provider: AiProvider;
+    mapMessages(messages: readonly unknown[]): unknown[];
+    rankTools(config: RuntimeConfigSnapshot, options?: ProviderRankToolOptions): readonly BoundaryValue[];
+    callModel<T>(request: unknown, execute: () => Promise<T>): Promise<T>;
+    extractTextDelta(input: unknown): string;
+    extractToolCalls(input: unknown): ToolCallData[];
+    extractStreamingToolCalls(input: unknown): ToolCallData[];
+    appendToolResults(messages: unknown[], calls: ToolCallData[], results: string[]): void;
+    finalize(): Promise<void>;
+}
+
+function appendOllamaToolResults(messages: unknown[], calls: ToolCallData[], results: string[]): void {
+    for (const [index, call] of calls.entries()) {
+        messages.push({
+            role: "tool",
+            content: results[index] ?? "",
+            tool_name: call.name,
+        });
+    }
+}
+
+class OpenAiProviderAdapter implements AiProviderAdapter {
+    readonly provider = AiProvider.OPENAI;
+
+    mapMessages(messages: readonly unknown[]): unknown[] {
+        return messages as OpenAiMessageType[];
+    }
+
+    rankTools(config: RuntimeConfigSnapshot, options?: ProviderRankToolOptions): readonly BoundaryValue[] {
+        const tools: BoundaryValue[] = [
+            ...getOpenAIResponsesTools(options?.forCreator) as BoundaryValue[],
+            getOpenAICodeInterpreterTool() as BoundaryValue,
+            {
+                type: "image_generation",
+                model: config.openAiImageTarget.model,
+                size: "auto",
+                moderation: "low",
+                output_format: "png",
+                partial_images: 3,
+            },
+            {type: "web_search"},
+        ];
+
+        if (options?.vectorStoreIds?.length) {
+            tools.unshift({
+                type: "file_search",
+                vector_store_ids: options.vectorStoreIds,
+            });
+        }
+
+        return tools;
+    }
+
+    async callModel<T>(_request: unknown, execute: () => Promise<T>): Promise<T> {
+        return execute();
+    }
+
+    extractTextDelta(input: unknown): string {
+        return extractOpenAiTextDelta(input);
+    }
+
+    extractToolCalls(input: unknown): ToolCallData[] {
+        return extractOpenAiToolCalls(input);
+    }
+
+    extractStreamingToolCalls(input: unknown): ToolCallData[] {
+        return extractOpenAiStreamingToolCalls(input);
+    }
+
+    appendToolResults(messages: unknown[], calls: ToolCallData[], results: string[]): void {
+        for (const [index, call] of calls.entries()) {
+            messages.push({
+                type: "function_call_output",
+                call_id: call.id,
+                output: results[index] ?? "",
+            });
+        }
+    }
+
+    async finalize(): Promise<void> {
+        return;
+    }
+}
+
+class MistralProviderAdapter implements AiProviderAdapter {
+    readonly provider = AiProvider.MISTRAL;
+
+    mapMessages(messages: readonly unknown[]): unknown[] {
+        return messages as MistralMessageType[];
+    }
+
+    rankTools(_config: RuntimeConfigSnapshot, options?: ProviderRankToolOptions): readonly BoundaryValue[] {
+        return getMistralTools(options?.forCreator) as BoundaryValue[];
+    }
+
+    async callModel<T>(_request: unknown, execute: () => Promise<T>): Promise<T> {
+        return execute();
+    }
+
+    extractTextDelta(input: unknown): string {
+        return extractMistralTextDelta(input);
+    }
+
+    extractToolCalls(input: unknown): ToolCallData[] {
+        return extractMistralToolCalls(input);
+    }
+
+    extractStreamingToolCalls(input: unknown): ToolCallData[] {
+        return this.extractToolCalls(input);
+    }
+
+    appendToolResults(messages: unknown[], calls: ToolCallData[], results: string[]): void {
+        for (const [index, call] of calls.entries()) {
+            messages.push({
+                role: "tool",
+                name: call.name,
+                toolCallId: call.id,
+                content: results[index] ?? "",
+            });
+        }
+    }
+
+    async finalize(): Promise<void> {
+        return;
+    }
+}
+
+class OllamaProviderAdapter implements AiProviderAdapter {
+    readonly provider = AiProvider.OLLAMA;
+
+    mapMessages(messages: readonly unknown[]): unknown[] {
+        return messages as OllamaMessage[];
+    }
+
+    rankTools(_config: RuntimeConfigSnapshot, options?: ProviderRankToolOptions): readonly BoundaryValue[] {
+        return getOllamaTools(options?.forCreator) as BoundaryValue[];
+    }
+
+    async callModel<T>(_request: unknown, execute: () => Promise<T>): Promise<T> {
+        return execute();
+    }
+
+    extractTextDelta(input: unknown): string {
+        return extractOllamaTextDelta(input);
+    }
+
+    extractToolCalls(input: unknown): ToolCallData[] {
+        return extractOllamaToolCalls(input);
+    }
+
+    extractStreamingToolCalls(input: unknown): ToolCallData[] {
+        return this.extractToolCalls(input);
+    }
+
+    appendToolResults(messages: unknown[], calls: ToolCallData[], results: string[]): void {
+        appendOllamaToolResults(messages, calls, results);
+    }
+
+    async finalize(): Promise<void> {
+        return;
+    }
+}
+
+export function getProviderAdapter(provider: AiProvider): AiProviderAdapter {
+    switch (provider) {
+        case AiProvider.OPENAI:
+            return new OpenAiProviderAdapter();
+        case AiProvider.MISTRAL:
+            return new MistralProviderAdapter();
+        case AiProvider.OLLAMA:
+            return new OllamaProviderAdapter();
+    }
+}

src/ai/provider-model-runtime.ts

@@ -196,7 +196,8 @@ export async function getRuntimeCapabilities(
     target?: AiRuntimeTarget
 ): Promise<AiModelCapabilities> {
     const runtimeTarget = target ?? resolveAiRuntimeTarget(provider, "chat", model ?? getRuntimeModel(provider));
-        const result = await getModelCapabilities(provider, runtimeTarget.model, target?.purpose ?? "chat") ?? buildCapabilities({});
+    const targetPurpose = target?.purpose && target.purpose !== "memoryCompress" ? target.purpose : "chat";
+    const result = await getModelCapabilities(provider, runtimeTarget.model, targetPurpose) ?? buildCapabilities({});
 
     for (const capabilityName of CAPABILITY_NAMES) {
         if (provider === AiProvider.OPENAI && (capabilityName === "vision" || capabilityName === "ocr")) {

src/ai/rag-artifact-payload.ts

@@ -0,0 +1,77 @@
+import type {AiProvider} from "../model/ai-provider";
+
+export type RagArtifactSource = {
+    fileId: string;
+    fileName: string;
+    mimeType?: string;
+    sizeBytes?: number;
+    sha256?: string;
+    uploadedFileId?: string;
+    documentId?: string;
+};
+
+export type RagArtifactPayload = {
+    artifactKind: "rag";
+    provider: AiProvider;
+    createdAt: string;
+    sources: RagArtifactSource[];
+    providerState:
+        | {
+            provider: AiProvider.OPENAI;
+            vectorStoreIds: string[];
+            uploadedFileIds: string[];
+        }
+        | {
+            provider: AiProvider.MISTRAL;
+            libraryId?: string;
+            documentCount: number;
+        }
+        | {
+            provider: AiProvider.OLLAMA;
+            prepared: boolean;
+            embeddingModel?: string;
+            topK?: number;
+            chunkSize?: number;
+            chunkOverlap?: number;
+            maxContextChars?: number;
+            extractedDocuments: Array<{
+                documentIndex: number;
+                fileName: string;
+                textChars: number;
+            }>;
+            selectedChunks: Array<{
+                sourceId: string;
+                documentIndex: number;
+                documentName: string;
+                chunkIndex: number;
+                chunkCount: number;
+                textChars: number;
+                score?: number;
+            }>;
+            skippedDocuments: Array<{
+                documentIndex: number;
+                fileName: string;
+                reason: string;
+            }>;
+            query: string;
+            minScore: number;
+            maxArchiveFiles: number;
+            maxArchiveBytes: number;
+            maxArchiveDepth: number;
+        };
+};
+
+export function buildRagArtifactPayload(params: {
+    provider: AiProvider;
+    createdAt?: string;
+    sources: RagArtifactSource[];
+    providerState: RagArtifactPayload["providerState"];
+}): RagArtifactPayload {
+    return {
+        artifactKind: "rag",
+        provider: params.provider,
+        createdAt: params.createdAt ?? new Date().toISOString(),
+        sources: params.sources,
+        providerState: params.providerState,
+    };
+}

src/ai/rag-artifact-store.ts

@@ -4,75 +4,39 @@ import type {AiDownloadedFile} from "./telegram-attachments";
 import type {PreparedDocumentRag} from "./document-rag-pipeline";
 import type {OllamaRagArtifactDetails} from "./ollama-rag";
 import {persistInternalJsonArtifactAttachment} from "./internal-artifact-store";
-
-type RagArtifactPayload = {
-    artifactKind: "rag";
-    provider: AiProvider;
-    createdAt: string;
-    sources: Array<{
-        fileId: string;
-        fileName: string;
-        mimeType?: string;
-        sizeBytes?: number;
-        sha256?: string;
-        uploadedFileId?: string;
-        documentId?: string;
-    }>;
-    providerState: {
-        vectorStoreIds?: string[];
-        libraryId?: string;
-        documentCount?: number;
-        prepared?: boolean;
-        uploadedFileIds?: string[];
-        embeddingModel?: string;
-        topK?: number;
-        chunkSize?: number;
-        chunkOverlap?: number;
-        maxContextChars?: number;
-        extractedDocuments?: Array<{
-            documentIndex: number;
-            fileName: string;
-            textChars: number;
-        }>;
-        selectedChunks?: Array<{
-            sourceId: string;
-            documentIndex: number;
-            documentName: string;
-            chunkIndex: number;
-            chunkCount: number;
-            textChars: number;
-            score?: number;
-        }>;
-        skippedDocuments?: Array<{
-            documentIndex: number;
-            fileName: string;
-            reason: string;
-        }>;
-        query?: string;
-        ollama?: OllamaRagArtifactDetails["providerState"];
-    };
-};
+import {buildRagArtifactPayload, type RagArtifactPayload} from "./rag-artifact-payload";
 
 function providerState(prepared: PreparedDocumentRag, details?: NonNullable<Parameters<typeof persistRagArtifactAttachment>[0]["details"]>): RagArtifactPayload["providerState"] {
     switch (prepared.provider) {
         case AiProvider.OPENAI:
             return {
+                provider: AiProvider.OPENAI,
                 vectorStoreIds: prepared.vectorStoreIds,
                 uploadedFileIds: prepared.uploadedFileIds,
             };
         case AiProvider.MISTRAL:
             return {
+                provider: AiProvider.MISTRAL,
                 libraryId: prepared.libraryId,
                 documentCount: prepared.documents.length,
             };
         case AiProvider.OLLAMA:
             return {
+                provider: AiProvider.OLLAMA,
                 prepared: prepared.prepared,
                 embeddingModel: details?.embeddingModel,
                 topK: details?.topK,
                 chunkSize: details?.chunkSize,
                 chunkOverlap: details?.chunkOverlap,
                 maxContextChars: details?.maxContextChars,
+                extractedDocuments: details?.artifact?.extractedDocuments ?? [],
+                selectedChunks: details?.artifact?.selectedChunks ?? [],
+                skippedDocuments: details?.artifact?.skippedDocuments ?? [],
+                query: details?.artifact?.query ?? "",
+                minScore: details?.artifact?.providerState?.minScore ?? 0,
+                maxArchiveFiles: details?.artifact?.providerState?.maxArchiveFiles ?? 0,
+                maxArchiveBytes: details?.artifact?.providerState?.maxArchiveBytes ?? 0,
+                maxArchiveDepth: details?.artifact?.providerState?.maxArchiveDepth ?? 0,
             };
     }
 }
@@ -117,22 +81,11 @@ export async function persistRagArtifactAttachment(params: {
 
     if (!sources.length) return Promise.resolve(undefined);
 
-    const payload: RagArtifactPayload = {
-        artifactKind: "rag",
+    const payload = buildRagArtifactPayload({
         provider: params.provider,
-        createdAt: new Date().toISOString(),
         sources,
-        providerState: {
-            ...providerState(params.prepared, params.details),
-            ...(params.details?.artifact ? {
-                extractedDocuments: params.details.artifact.extractedDocuments,
-                selectedChunks: params.details.artifact.selectedChunks,
-                skippedDocuments: params.details.artifact.skippedDocuments,
-                query: params.details.artifact.query,
-                ollama: params.details.artifact.providerState,
-            } : {}),
-        },
-    };
+        providerState: providerState(params.prepared, params.details),
+    });
     return await persistInternalJsonArtifactAttachment({
         artifactKind: "rag",
         fileNamePrefix: "rag",
@@ -140,14 +93,8 @@ export async function persistRagArtifactAttachment(params: {
         messageId: params.messageId,
         payload,
         metadata: {
-            provider: params.provider,
             sourceFileNames: sources.map(source => source.fileName),
             ...payload.providerState,
-            embeddingModel: params.details?.embeddingModel,
-            topK: params.details?.topK,
-            chunkSize: params.details?.chunkSize,
-            chunkOverlap: params.details?.chunkOverlap,
-            maxContextChars: params.details?.maxContextChars,
         },
     });
 }

src/ai/rag-retention-planner.ts

@@ -0,0 +1,75 @@
+import type {RagArtifactPayload} from "./rag-artifact-payload";
+
+export type ArtifactLike = {
+    id: string;
+    createdAt: string;
+    payload: string;
+};
+
+export type RagCleanupTarget = {
+    artifactId: string;
+    createdAt: string;
+    provider: RagArtifactPayload["providerState"]["provider"];
+    vectorStoreIds?: string[];
+    uploadedFileIds?: string[];
+    libraryId?: string;
+};
+
+export type RagCleanupPlan = {
+    cutoffAt: string;
+    targets: RagCleanupTarget[];
+};
+
+function parseRagArtifactPayload(payload: string): RagArtifactPayload | null {
+    try {
+        const parsed = JSON.parse(payload) as Partial<RagArtifactPayload>;
+        if (!parsed || parsed.artifactKind !== "rag" || !parsed.providerState) return null;
+        return parsed as RagArtifactPayload;
+    } catch {
+        return null;
+    }
+}
+
+export function buildStaleRagCleanupPlan(
+    artifacts: ArtifactLike[],
+    retentionDays = 14,
+    now = new Date(),
+): RagCleanupPlan {
+    const cutoffAt = new Date(now.getTime() - retentionDays * 24 * 60 * 60 * 1000).toISOString();
+    const targets: RagCleanupTarget[] = [];
+
+    for (const artifact of artifacts) {
+        if (artifact.createdAt > cutoffAt) continue;
+
+        const payload = parseRagArtifactPayload(artifact.payload);
+        if (!payload || payload.artifactKind !== "rag") continue;
+
+        switch (payload.providerState.provider) {
+            case "OPENAI":
+                if (payload.providerState.vectorStoreIds.length || payload.providerState.uploadedFileIds.length) {
+                    targets.push({
+                        artifactId: artifact.id,
+                        createdAt: artifact.createdAt,
+                        provider: payload.providerState.provider,
+                        vectorStoreIds: [...payload.providerState.vectorStoreIds],
+                        uploadedFileIds: [...payload.providerState.uploadedFileIds],
+                    });
+                }
+                break;
+            case "MISTRAL":
+                if (payload.providerState.libraryId) {
+                    targets.push({
+                        artifactId: artifact.id,
+                        createdAt: artifact.createdAt,
+                        provider: payload.providerState.provider,
+                        libraryId: payload.providerState.libraryId,
+                    });
+                }
+                break;
+            case "OLLAMA":
+                break;
+        }
+    }
+
+    return {cutoffAt, targets};
+}

src/ai/rag-retention.ts

@@ -0,0 +1,117 @@
+import {appLogger} from "../logging/logger.js";
+import {DatabaseManager} from "../db/database-manager.js";
+import {AiProvider} from "../model/ai-provider.js";
+import {createOpenAiClient, resolveAiRuntimeTarget} from "./ai-runtime-target.js";
+import {deleteMistralLibrary} from "./unified-ai-runner.shared.js";
+import {buildStaleRagCleanupPlan} from "./rag-retention-planner.js";
+
+const logger = appLogger.child("rag-retention");
+
+function unique(values: string[]): string[] {
+    return [...new Set(values.filter(Boolean))];
+}
+
+async function cleanupOpenAiRag(vectorStoreIds: string[], uploadedFileIds: string[]): Promise<void> {
+    const target = resolveAiRuntimeTarget(AiProvider.OPENAI, "documents");
+    const client = createOpenAiClient(target);
+
+    for (const vectorStoreId of unique(vectorStoreIds)) {
+        const startedAt = Date.now();
+        logger.info("openai.vector_store.cleanup.start", {vectorStoreId});
+        try {
+            await client.vectorStores.delete(vectorStoreId);
+            logger.success("openai.vector_store.cleanup.done", {vectorStoreId, duration: `${Date.now() - startedAt}ms`});
+        } catch (error) {
+            logger.warn("openai.vector_store.cleanup.failed", {
+                vectorStoreId,
+                duration: `${Date.now() - startedAt}ms`,
+                error: error instanceof Error ? error : String(error),
+            });
+        }
+    }
+
+    for (const fileId of unique(uploadedFileIds)) {
+        const startedAt = Date.now();
+        logger.info("openai.file.cleanup.start", {fileId});
+        try {
+            await client.files.delete(fileId);
+            logger.success("openai.file.cleanup.done", {fileId, duration: `${Date.now() - startedAt}ms`});
+        } catch (error) {
+            logger.warn("openai.file.cleanup.failed", {
+                fileId,
+                duration: `${Date.now() - startedAt}ms`,
+                error: error instanceof Error ? error : String(error),
+            });
+        }
+    }
+}
+
+async function cleanupMistralRag(libraryId: string): Promise<void> {
+    const target = resolveAiRuntimeTarget(AiProvider.MISTRAL, "documents");
+    const startedAt = Date.now();
+    logger.info("mistral.library.cleanup.start", {libraryId});
+    try {
+        await deleteMistralLibrary(libraryId, target);
+        logger.success("mistral.library.cleanup.done", {libraryId, duration: `${Date.now() - startedAt}ms`});
+    } catch (error) {
+        logger.warn("mistral.library.cleanup.failed", {
+            libraryId,
+            duration: `${Date.now() - startedAt}ms`,
+            error: error instanceof Error ? error : String(error),
+        });
+    }
+}
+
+export async function cleanupStaleRagProviderState(retentionDays = 14): Promise<{
+    scannedArtifacts: number;
+    cleanupTargets: number;
+    openaiTargets: number;
+    mistralTargets: number;
+}> {
+    const startedAt = Date.now();
+    const artifacts = await DatabaseManager.getAllArtifacts().catch(() => []);
+    const plan = buildStaleRagCleanupPlan(artifacts, retentionDays);
+
+    logger.info("cleanup.start", {
+        retentionDays,
+        scannedArtifacts: artifacts.length,
+        cleanupTargets: plan.targets.length,
+        cutoffAt: plan.cutoffAt,
+    });
+
+    let openaiTargets = 0;
+    let mistralTargets = 0;
+
+    for (const target of plan.targets) {
+        switch (target.provider) {
+            case "OPENAI":
+                openaiTargets += 1;
+                await cleanupOpenAiRag(target.vectorStoreIds ?? [], target.uploadedFileIds ?? []);
+                break;
+            case "MISTRAL":
+                mistralTargets += 1;
+                if (target.libraryId) {
+                    await cleanupMistralRag(target.libraryId);
+                }
+                break;
+            case "OLLAMA":
+                break;
+        }
+    }
+
+    logger.success("cleanup.done", {
+        retentionDays,
+        scannedArtifacts: artifacts.length,
+        cleanupTargets: plan.targets.length,
+        openaiTargets,
+        mistralTargets,
+        duration: `${Date.now() - startedAt}ms`,
+    });
+
+    return {
+        scannedArtifacts: artifacts.length,
+        cleanupTargets: plan.targets.length,
+        openaiTargets,
+        mistralTargets,
+    };
+}

src/ai/reply-chain-downloads.ts

@@ -0,0 +1,39 @@
+import type {AiDownloadedFile} from "./telegram-attachments.js";
+
+function downloadKey(download: AiDownloadedFile): string {
+    return [
+        download.kind,
+        download.fileId,
+        download.sha256 ?? "",
+        download.fileName,
+    ].join(":");
+}
+
+export function mergeReplyChainDownloads(
+    currentDownloads: readonly AiDownloadedFile[],
+    replyChainDownloads: readonly AiDownloadedFile[],
+): AiDownloadedFile[] {
+    const result: AiDownloadedFile[] = [];
+    const seen = new Set<string>();
+
+    for (const download of [...currentDownloads, ...replyChainDownloads]) {
+        const key = downloadKey(download);
+        if (seen.has(key)) continue;
+        seen.add(key);
+        result.push(download);
+    }
+
+    return result;
+}
+
+export function shouldPreferCurrentDownloads(text: string, currentDownloads: readonly AiDownloadedFile[]): boolean {
+    if (!currentDownloads.length) return false;
+
+    const normalized = text.trim().toLowerCase();
+    if (!normalized) return false;
+
+    return normalized.includes("this file")
+        || normalized.includes("this document")
+        || normalized.includes("этот файл")
+        || normalized.includes("этот документ");
+}

src/ai/response-model-output.ts

@@ -0,0 +1,19 @@
+import type {TelegramOutputAttachmentRecord, TelegramToolExecutionRecord} from "./telegram-stream-message.js";
+
+export type NormalizedModelOutput = {
+    text: string;
+    toolExecutions: TelegramToolExecutionRecord[];
+    outputAttachments: TelegramOutputAttachmentRecord[];
+};
+
+export function summarizeModelOutput(params: {
+    text: string;
+    toolExecutions: readonly TelegramToolExecutionRecord[];
+    outputAttachments: readonly TelegramOutputAttachmentRecord[];
+}): NormalizedModelOutput {
+    return {
+        text: params.text.trim(),
+        toolExecutions: [...params.toolExecutions],
+        outputAttachments: [...params.outputAttachments],
+    };
+}

src/ai/telegram-stream-message.ts

@@ -5,6 +5,7 @@ import {Environment} from "../common/environment";
 import {MessageStore} from "../common/message-store";
 import {createQueuedFunction} from "../util/async-lock";
 import {enqueueTelegramApiCall} from "../util/telegram-api-queue";
+import {appLogger} from "../logging/logger";
 import fs from "node:fs";
 import path from "node:path";
 import {StoredAttachment, StoredAttachmentKind} from "../model/stored-attachment";
@@ -13,11 +14,13 @@ import {prepareTelegramMarkdownV2} from "../util/markdown-v2-renderer";
 import {AiProvider} from "../model/ai-provider";
 import {AI_IMAGE_OUTPUT_MODE_DOCUMENT, UserAiImageOutputMode} from "../common/user-ai-settings";
 import {PIPELINE_ATTACHMENT_LIMIT_BYTES} from "./user-request-pipeline";
+import {recordToolCall} from "../common/ai-observability.js";
 
 const TELEGRAM_LIMIT = 4096;
 const TELEGRAM_CAPTION_LIMIT = 1024;
 const TELEGRAM_PHOTO_LIMIT_BYTES = 10 * 1024 * 1024;
 const EDIT_INTERVAL_MS = 4500;
+const logger = appLogger.child("telegram-stream-message");
 
 export type TelegramArtifactFile = {
     kind: "image" | "file";
@@ -238,6 +241,13 @@ export class TelegramStreamMessage {
 
     recordToolExecution(record: TelegramToolExecutionRecord): void {
         this.toolExecutions.push(record);
+        recordToolCall();
+        logger.debug("tool.execution.recorded", {
+            requestId: this.cancelRequestId,
+            toolName: record.toolName,
+            callId: record.callId,
+            resultChars: record.resultChars,
+        });
     }
 
     getToolExecutions(): TelegramToolExecutionRecord[] {
@@ -246,6 +256,13 @@ export class TelegramStreamMessage {
 
     recordOutputAttachment(record: TelegramOutputAttachmentRecord): void {
         this.outputAttachments.push(record);
+        logger.debug("output_attachment.recorded", {
+            requestId: this.cancelRequestId,
+            artifactKind: record.artifactKind,
+            fileName: record.fileName,
+            sizeBytes: record.sizeBytes,
+            messageId: record.messageId,
+        });
     }
 
     getOutputAttachments(): TelegramOutputAttachmentRecord[] {

src/ai/tool-batch-runner.ts

@@ -0,0 +1,28 @@
+import type {AiProviderAdapter} from "./provider-adapters.js";
+import {executeToolBatch, type ToolCallData, type ToolExecutionMemory} from "./unified-ai-runner.shared.js";
+import type {TelegramStreamMessage} from "./telegram-stream-message.js";
+import type {ToolRuntimeContext} from "./tools/runtime.js";
+
+export async function executeToolBatchWithAdapter(params: {
+    userId: number | undefined | null;
+    toolCalls: ToolCallData[];
+    streamMessage: TelegramStreamMessage;
+    toolContext: ToolRuntimeContext;
+    toolMemory: ToolExecutionMemory;
+    adapter: AiProviderAdapter;
+    appendTargets?: unknown[][];
+}): Promise<string[]> {
+    const results = await executeToolBatch(
+        params.userId,
+        params.toolCalls,
+        params.streamMessage,
+        params.toolContext,
+        params.toolMemory,
+    );
+
+    for (const target of params.appendTargets ?? []) {
+        params.adapter.appendToolResults(target, params.toolCalls, results);
+    }
+
+    return results;
+}

src/ai/tool-loop-artifact-store.ts

@@ -0,0 +1,39 @@
+import type {StoredAttachment} from "../model/stored-attachment";
+import type {TelegramOutputAttachmentRecord, TelegramToolExecutionRecord} from "./telegram-stream-message.js";
+import {persistInternalJsonArtifactAttachment} from "./internal-artifact-store";
+
+export async function persistToolLoopSummaryArtifactAttachment(params: {
+    chatId: number;
+    messageId: number;
+    text: string;
+    executions: readonly TelegramToolExecutionRecord[];
+    outputAttachments: readonly TelegramOutputAttachmentRecord[];
+}): Promise<StoredAttachment | undefined> {
+    if (!params.executions.length) return undefined;
+
+    return await persistInternalJsonArtifactAttachment({
+        artifactKind: "tool_result",
+        fileNamePrefix: "tool-loop-summary",
+        chatId: params.chatId,
+        messageId: params.messageId,
+        payload: {
+            stage: "tool_loop",
+            text: params.text.trim(),
+            executions: params.executions.map(execution => ({
+                toolName: execution.toolName,
+                callId: execution.callId,
+                argumentsText: execution.argumentsText,
+                resultChars: execution.resultChars,
+                startedAt: execution.startedAt,
+                finishedAt: execution.finishedAt,
+            })),
+            outputAttachments: params.outputAttachments,
+        },
+        metadata: {
+            stage: "tool_loop",
+            toolExecutions: params.executions.length,
+            outputAttachments: params.outputAttachments.length,
+            textChars: params.text.trim().length,
+        },
+    });
+}

src/ai/tool-loop-control.ts

@@ -0,0 +1,38 @@
+import type {ToolCallData} from "./unified-ai-runner.shared.js";
+
+export type ToolLoopStopReason = "no_tool_calls" | "max_rounds_reached";
+
+export type ToolLoopContinuation = {
+    continue: boolean;
+    reason?: ToolLoopStopReason;
+    remainingRounds: number;
+};
+
+export function decideToolLoopContinuation(params: {
+    round: number;
+    maxRounds: number;
+    toolCalls: readonly ToolCallData[];
+}): ToolLoopContinuation {
+    const remainingRounds = Math.max(params.maxRounds - params.round - 1, 0);
+
+    if (!params.toolCalls.length) {
+        return {
+            continue: false,
+            reason: "no_tool_calls",
+            remainingRounds,
+        };
+    }
+
+    if (remainingRounds === 0) {
+        return {
+            continue: false,
+            reason: "max_rounds_reached",
+            remainingRounds,
+        };
+    }
+
+    return {
+        continue: true,
+        remainingRounds,
+    };
+}

src/ai/tool-loop-runner.ts

@@ -0,0 +1,22 @@
+export type ToolLoopRoundOutcome = {
+    shouldContinue: boolean;
+    maxRoundsReached?: boolean;
+};
+
+export async function runToolLoopRounds(params: {
+    maxRounds: number;
+    onRound: (round: number) => Promise<ToolLoopRoundOutcome>;
+    onMaxRoundsReached?: (round: number) => Promise<void> | void;
+}): Promise<void> {
+    for (let round = 0; round < params.maxRounds; round++) {
+        const outcome = await params.onRound(round);
+        if (!outcome.shouldContinue) {
+            if (outcome.maxRoundsReached) {
+                await params.onMaxRoundsReached?.(round);
+            }
+            return;
+        }
+    }
+
+    await params.onMaxRoundsReached?.(params.maxRounds - 1);
+}

src/ai/tool-loop-summary.ts

@@ -0,0 +1,56 @@
+import type {PipelineArtifact} from "./user-request-pipeline/types.js";
+import type {TelegramOutputAttachmentRecord, TelegramToolExecutionRecord} from "./telegram-stream-message.js";
+import {summarizeModelOutput} from "./response-model-output.js";
+
+export type ToolLoopSummary = {
+    status: "succeeded" | "skipped";
+    fallbackAction?: "continue_without_stage";
+    details: {
+        modelOutput: ReturnType<typeof summarizeModelOutput>;
+        count: number;
+        tools: Array<{
+            toolName: string;
+            callId: string;
+            resultChars: number;
+        }>;
+    };
+    artifacts?: PipelineArtifact[];
+};
+
+export function summarizeToolLoop(params: {
+    text: string;
+    executions: readonly TelegramToolExecutionRecord[];
+    outputAttachments: readonly TelegramOutputAttachmentRecord[];
+}): ToolLoopSummary {
+    const count = params.executions.length;
+    const tools = params.executions.map(execution => ({
+        toolName: execution.toolName,
+        callId: execution.callId,
+        resultChars: execution.resultChars,
+    }));
+
+    return {
+        status: count ? "succeeded" : "skipped",
+        fallbackAction: count ? undefined : "continue_without_stage",
+        details: {
+            modelOutput: summarizeModelOutput({
+                text: params.text,
+                toolExecutions: params.executions,
+                outputAttachments: params.outputAttachments,
+            }),
+            count,
+            tools,
+        },
+        artifacts: count ? [{
+            kind: "tool_result",
+            stage: "tool_loop",
+            createdAt: new Date().toISOString(),
+            toolName: "summary",
+            callId: "tool_loop_summary",
+            resultText: JSON.stringify({
+                count,
+                tools,
+            }),
+        }] : undefined,
+    };
+}

src/ai/tool-mappers.ts

@@ -1,8 +1,9 @@
 import {AiTool} from "./tool-types";
-import {AiProvider} from "../model/ai-provider";
-import {getTools} from "./tools/registry";
-import {WEB_SEARCH_TOOL_NAME} from "./tools/web-search";
-import {PYTHON_INTERPRETER_TOOL_NAME} from "./tools/python-interpretator";
+import {AiProvider} from "../model/ai-provider.js";
+import {getTools} from "./tools/registry.js";
+import {WEB_SEARCH_TOOL_NAME} from "./tools/web-search.js";
+import {PYTHON_INTERPRETER_TOOL_NAME} from "./tools/python-interpretator.js";
+import {toolSchemaNames} from "./tool-schema-utils.js";
 
 export type AiProviderName = "ollama" | "openai" | "mistral";
 
@@ -26,6 +27,11 @@ export function getOpenAITools(forCreator?: boolean): AiTool[] {
     }));
 }
 
+export function getOpenAICompatibleTools(forCreator?: boolean): AiTool[] {
+    // The compatible chat.completions backend only accepts plain function tools.
+    return getOpenAITools(forCreator);
+}
+
 export type OpenAiResponseTool = {
     type: "function";
     name: string;
@@ -79,3 +85,20 @@ export function getProviderTools(provider: AiProvider, forCreator?: boolean): Ai
             return getOpenAITools(forCreator);
     }
 }
+
+export function ensureToolsSelected<T>(availableTools: readonly T[], selectedTools: readonly T[], toolNames: readonly string[]): T[] {
+    const selected = [...selectedTools];
+    const selectedNames = new Set(selected.flatMap(tool => toolSchemaNames(tool as never)));
+
+    for (const toolName of toolNames) {
+        if (selectedNames.has(toolName)) continue;
+
+        const extraTool = availableTools.find(tool => toolSchemaNames(tool as never).includes(toolName));
+        if (extraTool) {
+            selected.unshift(extraTool);
+            selectedNames.add(toolName);
+        }
+    }
+
+    return selected;
+}

src/ai/tool-rank-audit.ts

@@ -1,32 +0,0 @@
-import {AiProvider} from "../model/ai-provider";
-import type {TelegramStreamMessage} from "./telegram-stream-message";
-import type {PipelineAuditEvent} from "./user-request-pipeline";
-import {logError} from "../util/utils";
-
-export async function storeToolRankAudit(params: {
-    streamMessage: TelegramStreamMessage;
-    provider: AiProvider;
-    model: string;
-    round: number;
-    startedAt: number;
-    startedAtIso: string;
-    selectedTools?: string[];
-    error?: unknown;
-}): Promise<void> {
-    const event: PipelineAuditEvent = {
-        stage: "tool_rank",
-        status: params.error ? "failed" : "succeeded",
-        startedAt: params.startedAtIso,
-        finishedAt: new Date().toISOString(),
-        durationMs: Date.now() - params.startedAt,
-        provider: params.provider,
-        model: params.model,
-        details: {
-            round: params.round,
-            selectedTools: params.selectedTools ?? [],
-        },
-        error: params.error instanceof Error ? params.error.message : params.error ? String(params.error) : undefined,
-    };
-
-    await params.streamMessage.storePipelineAudit([event]).catch(logError);
-}

src/ai/tool-rank-stage.ts

@@ -0,0 +1,146 @@
+import {AiProvider} from "../model/ai-provider.js";
+import type {BoundaryValue} from "../common/boundary-types.js";
+import type {TelegramStreamMessage} from "./telegram-stream-message.js";
+import type {RuntimeConfigSnapshot} from "./unified-ai-runner.shared.js";
+import {allToolSchemaNames, toolSchemaNames} from "./tool-schema-utils.js";
+import type {ToolRanker} from "./unified-ai-runner.tool-ranker.js";
+import type {PipelineAuditEvent} from "./user-request-pipeline/types.js";
+
+function latestUserText(messages: readonly { role?: string; content?: unknown }[]): string {
+    for (let i = messages.length - 1; i >= 0; i--) {
+        const message = messages[i];
+        if (message?.role !== "user") continue;
+        if (typeof message.content === "string") return message.content;
+        if (Array.isArray(message.content)) {
+            return message.content
+                .map(part => typeof part === "object" && part !== null && "text" in part && typeof (part as { text?: unknown }).text === "string"
+                    ? (part as { text: string }).text
+                    : "")
+                .filter(Boolean)
+                .join("\n");
+        }
+    }
+
+    return "";
+}
+
+export async function runToolRankStage(params: {
+    provider: AiProvider;
+    model: string;
+    round: number;
+    config: RuntimeConfigSnapshot;
+    availableTools: readonly BoundaryValue[];
+    messages: readonly { role?: string; content?: unknown }[];
+    streamMessage: TelegramStreamMessage;
+    signal: AbortSignal;
+    toolRanker?: ToolRanker;
+    storeAudit?: (params: {
+        streamMessage: TelegramStreamMessage;
+        provider: AiProvider;
+        model: string;
+        round: number;
+        startedAt: number;
+        startedAtIso: string;
+        availableTools: string[];
+        selectedTools?: string[];
+        usedRanker?: boolean;
+        error?: unknown;
+    }) => Promise<void>;
+}): Promise<{
+    filteredTools: BoundaryValue[];
+    selectedToolNames: string[];
+    usedRanker: boolean;
+}> {
+    const toolRanker = params.toolRanker ?? new (await import("./unified-ai-runner.tool-ranker.js")).ToolRanker(params.config);
+    const startedAt = Date.now();
+    const startedAtIso = new Date().toISOString();
+    const filterSelectedTools = (selectedToolNames: readonly string[]): BoundaryValue[] => {
+        const selected = new Set(selectedToolNames);
+        return params.availableTools.filter(tool => toolSchemaNames(tool).some(name => selected.has(name)));
+    };
+    const storeAudit = params.storeAudit ?? (async (auditParams: {
+        streamMessage: TelegramStreamMessage;
+        provider: AiProvider;
+        model: string;
+        round: number;
+        startedAt: number;
+        startedAtIso: string;
+        availableTools: string[];
+        selectedTools?: string[];
+        usedRanker?: boolean;
+        error?: unknown;
+    }) => {
+        const event: PipelineAuditEvent = {
+            stage: "tool_rank",
+            status: auditParams.error ? "failed" : "succeeded",
+            startedAt: auditParams.startedAtIso,
+            finishedAt: new Date().toISOString(),
+            durationMs: Date.now() - auditParams.startedAt,
+            provider: auditParams.provider,
+            model: auditParams.model,
+            details: {
+                round: auditParams.round,
+                availableTools: auditParams.availableTools,
+                selectedTools: auditParams.selectedTools ?? [],
+                usedRanker: auditParams.usedRanker ?? false,
+                toolRankDecision: {
+                    provider: auditParams.provider,
+                    round: auditParams.round,
+                    availableTools: auditParams.availableTools,
+                    selectedTools: auditParams.selectedTools ?? [],
+                    usedRanker: auditParams.usedRanker ?? false,
+                },
+            },
+            error: auditParams.error instanceof Error ? auditParams.error.message : auditParams.error ? String(auditParams.error) : undefined,
+        };
+
+        await auditParams.streamMessage.storePipelineAudit([event]);
+    });
+
+    params.streamMessage.setStatus("🧩 Выбираю подходящие инструменты...");
+    await params.streamMessage.flush();
+
+    try {
+        const selection = await toolRanker.selectTools({
+            provider: params.provider,
+            userQuery: latestUserText(params.messages),
+            availableTools: params.availableTools,
+            round: params.round,
+            signal: params.signal,
+        });
+
+        params.streamMessage.clearStatus();
+        await params.streamMessage.flush();
+        await storeAudit({
+            streamMessage: params.streamMessage,
+            provider: params.provider,
+            model: params.model,
+            round: params.round,
+            startedAt,
+            startedAtIso,
+            availableTools: allToolSchemaNames(params.availableTools),
+            selectedTools: selection.toolNames,
+            usedRanker: selection.usedRanker,
+        });
+
+        return {
+            filteredTools: filterSelectedTools(selection.toolNames),
+            selectedToolNames: selection.toolNames,
+            usedRanker: selection.usedRanker,
+        };
+    } catch (error) {
+        params.streamMessage.clearStatus();
+        await params.streamMessage.flush();
+        await storeAudit({
+            streamMessage: params.streamMessage,
+            provider: params.provider,
+            model: params.model,
+            round: params.round,
+            startedAt,
+            startedAtIso,
+            availableTools: allToolSchemaNames(params.availableTools),
+            error,
+        });
+        throw error;
+    }
+}

src/ai/tool-ranker-fallback.ts

@@ -0,0 +1,56 @@
+import {ToolRankerFallbackPolicy} from "../common/policies.js";
+import {decidePipelineFallback, type PipelineFallbackDecision} from "./user-request-pipeline/fallback-executor.js";
+
+export type ToolRankerFallbackSelection = {
+    toolNames: string[];
+    usedRanker: boolean;
+};
+
+export type ToolRankerFallbackDecision = PipelineFallbackDecision & ToolRankerFallbackSelection;
+
+function fallbackActionForPolicy(policy: ToolRankerFallbackPolicy) {
+    return policy === ToolRankerFallbackPolicy.MAIN_MODEL
+        ? "use_alternate_target"
+        : "continue_without_stage";
+}
+
+export function decideToolRankerFallback(params: {
+    fallbackPolicy: ToolRankerFallbackPolicy;
+    availableToolNames: readonly string[];
+    reason: "unavailable" | "failed";
+}): ToolRankerFallbackDecision {
+    const action = fallbackActionForPolicy(params.fallbackPolicy);
+    const decision = decidePipelineFallback({
+        stage: "tool_rank",
+        reason: params.reason,
+        policies: [{
+            stage: "tool_rank",
+            onUnavailable: action,
+            onFailed: action,
+        }],
+    });
+
+    return {
+        ...decision,
+        toolNames: params.fallbackPolicy === ToolRankerFallbackPolicy.NO_TOOLS
+            ? []
+            : [...params.availableToolNames],
+        usedRanker: false,
+    };
+}
+
+export function resolveToolRankerFallbackSelection(params: {
+    fallbackPolicy: ToolRankerFallbackPolicy;
+    availableToolNames: readonly string[];
+}): ToolRankerFallbackSelection {
+    const decision = decideToolRankerFallback({
+        fallbackPolicy: params.fallbackPolicy,
+        availableToolNames: params.availableToolNames,
+        reason: "failed",
+    });
+
+    return {
+        toolNames: decision.toolNames,
+        usedRanker: decision.usedRanker,
+    };
+}

src/ai/tool-ranker-metadata.ts

@@ -102,6 +102,100 @@ export const TOOL_RANKER_TOOL_INFOS = {
             example("где определён BotService?", ["search_files"]),
         ],
     ),
+    read_user_info: tool(
+        "read_user_info",
+        "Read persistent user memory from user.md.",
+        "Use before editing or when the user asks what you remember about them.",
+        [
+            example("что ты помнишь обо мне?", ["read_user_info"]),
+            example("покажи мою память", ["read_user_info"]),
+        ],
+    ),
+    read_system_info: tool(
+        "read_system_info",
+        "Read persistent assistant memory from system.md.",
+        "Use before editing or when the user asks what instructions you remember about yourself.",
+        [
+            example("что ты помнишь о себе?", ["read_system_info"]),
+            example("покажи память о тебе", ["read_system_info"]),
+        ],
+    ),
+    add_user_info: tool(
+        "add_user_info",
+        "Append a durable fact about the user to persistent memory.",
+        "Use when the user asks to remember a new fact, preference, identity detail, or profile information about themselves.",
+        [
+            example("запомни, что меня зовут Иван", ["add_user_info"]),
+            example("запомни, что я люблю чай", ["add_user_info"]),
+            example("remember that I like short answers", ["add_user_info"]),
+        ],
+    ),
+    add_system_info: tool(
+        "add_system_info",
+        "Append a durable instruction about the assistant to persistent memory.",
+        "Use when the user asks to remember a new assistant identity, style, or behavior instruction.",
+        [
+            example("тебя зовут Евлампий", ["add_system_info"]),
+            example("ты ИИ помощник", ["add_system_info"]),
+            example("remember you are a concise assistant", ["add_system_info"]),
+        ],
+    ),
+    remove_user_info: tool(
+        "remove_user_info",
+        "Remove a specific user fact from persistent memory.",
+        "Use when the user asks to forget, delete, or remove a specific fact about themselves.",
+        [
+            example("забудь, что я люблю кофе", ["remove_user_info"]),
+            example("удали из памяти, что я живу в Москве", ["remove_user_info"]),
+            example("forget that I work at ACME", ["remove_user_info"]),
+        ],
+    ),
+    remove_system_info: tool(
+        "remove_system_info",
+        "Remove a specific assistant instruction from persistent memory.",
+        "Use when the user asks to forget or remove a specific instruction about the assistant.",
+        [
+            example("забудь, что тебя зовут Евлампий", ["remove_system_info"]),
+            example("убери правило отвечать коротко", ["remove_system_info"]),
+            example("forget that you are a concise assistant", ["remove_system_info"]),
+        ],
+    ),
+    replace_user_info: tool(
+        "replace_user_info",
+        "Replace the full user memory with a new compact version.",
+        "Use when the user wants to overwrite all remembered user info, for example when they say to forget everything and keep only the new fact.",
+        [
+            example("забудь всё обо мне и запиши только это: меня зовут Иван", ["replace_user_info"]),
+            example("замени всю память обо мне на: люблю чай и короткие ответы", ["replace_user_info"]),
+        ],
+    ),
+    replace_system_info: tool(
+        "replace_system_info",
+        "Replace the full assistant memory with a new compact version.",
+        "Use when the user wants to overwrite all remembered assistant info or instructions.",
+        [
+            example("забудь всё о себе и запиши только это: тебя зовут Евлампий", ["replace_system_info"]),
+            example("замени инструкцию о себе на: ты краткий ИИ помощник", ["replace_system_info"]),
+        ],
+    ),
+    delete_user_info: tool(
+        "delete_user_info",
+        "Delete user.md entirely.",
+        "Use when the user explicitly asks to delete all remembered user info, not just a fragment.",
+        [
+            example("удали всю память обо мне", ["delete_user_info"]),
+            example("forget all user memory", ["delete_user_info"]),
+        ],
+    ),
+    delete_system_info: tool(
+        "delete_system_info",
+        "Delete system.md entirely.",
+        "Use when the user explicitly asks to delete all remembered assistant info, not just a fragment.",
+        [
+            example("удали всю память о себе", ["delete_system_info"]),
+            example("forget all assistant memory", ["delete_system_info"]),
+        ],
+    ),
     create_file: tool(
         "create_file",
         "Create a new small file.",
@@ -352,6 +446,20 @@ function toolNamesFromTool(tool: BoundaryValue): string[] {
     return name ? [name] : [];
 }
 
+function fallbackToolInfoFromTool(toolValue: BoundaryValue, name: string): ToolRankerToolInfo | undefined {
+    if (!isRecord(toolValue)) return undefined;
+
+    const fn = isRecord(toolValue.function) ? toolValue.function : undefined;
+    const description = asOptionalString(fn?.description ?? toolValue.description)
+        ?? `Tool ${name}.`;
+
+    return tool(
+        name,
+        description,
+        "Use when the tool description matches the user's request.",
+    );
+}
+
 export function getToolRankerToolInfo(name: string): ToolRankerToolInfo | undefined {
     return TOOL_RANKER_TOOL_INFOS[name as ToolRankerToolName];
 }
@@ -363,10 +471,25 @@ export function getToolRankerToolInfos(names: readonly string[]): ToolRankerTool
 }
 
 export function getToolRankerAvailableToolInfos(availableTools: readonly BoundaryValue[]): ToolRankerToolInfo[] {
-    return getToolRankerToolInfos([
-        "no_tool",
-        ...availableTools.flatMap(toolNamesFromTool),
-    ]);
+    const infos = new Map<string, ToolRankerToolInfo>();
+
+    infos.set("no_tool", TOOL_RANKER_TOOL_INFOS.no_tool);
+
+    for (const tool of availableTools) {
+        for (const name of toolNamesFromTool(tool)) {
+            if (infos.has(name)) continue;
+
+            const known = getToolRankerToolInfo(name);
+            const fallback = fallbackToolInfoFromTool(tool, name);
+            if (known) {
+                infos.set(name, known);
+            } else if (fallback) {
+                infos.set(name, fallback);
+            }
+        }
+    }
+
+    return [...infos.values()];
 }
 
 function renderToolLine(tool: ToolRankerToolInfo, compact: boolean): string {
@@ -414,6 +537,16 @@ function buildPriorityLines(tools: readonly ToolRankerToolInfo[]): string[] {
     pushIfAvailable("read_file", "known local file path -> read_file");
     pushIfAvailable("list_directory", "project structure or directory listing -> list_directory");
     pushIfAvailable("search_files", "local file/content search or unknown file path -> search_files");
+    pushIfAvailable("read_user_info", "inspect remembered user info -> read_user_info");
+    pushIfAvailable("read_system_info", "inspect remembered assistant info -> read_system_info");
+    pushIfAvailable("add_user_info", "remember a new user fact -> add_user_info");
+    pushIfAvailable("add_system_info", "remember a new assistant instruction -> add_system_info");
+    pushIfAvailable("remove_user_info", "forget a user fact -> remove_user_info");
+    pushIfAvailable("remove_system_info", "forget an assistant instruction -> remove_system_info");
+    pushIfAvailable("replace_user_info", "overwrite all user memory -> replace_user_info");
+    pushIfAvailable("replace_system_info", "overwrite all assistant memory -> replace_system_info");
+    pushIfAvailable("delete_user_info", "delete all user memory -> delete_user_info");
+    pushIfAvailable("delete_system_info", "delete all assistant memory -> delete_system_info");
     pushIfAvailable("edit_file_patch", "targeted existing file edit -> edit_file_patch");
     pushIfAvailable("update_file", "full existing file replacement -> update_file");
     pushIfAvailable("create_file", "small new file -> create_file");
@@ -471,7 +604,7 @@ export function buildToolRankerSystemPrompt(params: {
     const includeExamples = params.includeExamples ?? false;
     const maxExamplesPerTool = Math.max(0, params.maxExamplesPerTool ?? 1);
     const compact = params.compact ?? true;
-    const availableTools = getToolRankerToolInfos(params.availableTools.map(tool => tool.name));
+    const availableTools = params.availableTools;
     const availableToolNames = availableTools.map(tool => tool.name);
 
     const sections: string[] = [

src/ai/tool-ranker-pipeline.ts

@@ -1,12 +1,12 @@
-import type {BoundaryValue} from "../common/boundary-types";
-import type {AiRuntimeTarget} from "./ai-runtime-target";
-import {AiProvider} from "../model/ai-provider";
-import {RuntimeConfigSnapshot, toolSchemaNames} from "./unified-ai-runner.shared";
+import type {BoundaryValue} from "../common/boundary-types.js";
+import type {AiRuntimeTarget} from "./ai-runtime-target.js";
+import {AiProvider} from "../model/ai-provider.js";
+import {RuntimeConfigSnapshot, toolSchemaNames} from "./unified-ai-runner.shared.js";
 import {
     buildToolRankerSystemPrompt,
     getToolRankerAvailableToolInfos,
     type ToolRankerToolInfo,
-} from "./tool-ranker-metadata";
+} from "./tool-ranker-metadata.js";
 
 export type ToolRankerMessage = {
     role?: string;

src/ai/tool-schema-utils.ts

@@ -0,0 +1,33 @@
+import type {BoundaryValue} from "../common/boundary-types.js";
+
+function isRecord(value: BoundaryValue): value is Record<string, BoundaryValue> {
+    return value !== null && typeof value === "object" && !Array.isArray(value);
+}
+
+function asOptionalString(value: BoundaryValue): string | undefined {
+    return typeof value === "string" && value.trim().length > 0 ? value.trim() : undefined;
+}
+
+export function toolSchemaName(tool: BoundaryValue): string | undefined {
+    if (!isRecord(tool)) return undefined;
+    const fn = isRecord(tool.function) ? tool.function : undefined;
+    const directName = fn?.name ?? tool.name ?? (typeof tool.type === "string" && tool.type !== "function" ? tool.type : undefined);
+    return asOptionalString(directName);
+}
+
+export function toolSchemaNames(tool: BoundaryValue): string[] {
+    if (!isRecord(tool)) return [];
+
+    if (Array.isArray(tool.functionDeclarations)) {
+        return tool.functionDeclarations
+            .map(declaration => isRecord(declaration) ? asOptionalString(declaration.name) : undefined)
+            .filter((name): name is string => !!name);
+    }
+
+    const name = toolSchemaName(tool);
+    return name ? [name] : [];
+}
+
+export function allToolSchemaNames(tools: readonly BoundaryValue[]): string[] {
+    return [...new Set(tools.flatMap(toolSchemaNames))];
+}

src/ai/tools/create-note.ts

@@ -1,11 +1,11 @@
-import {AiTool} from "../tool-types";
+import {AiTool} from "../tool-types.js";
 import path from "node:path";
 import {readFile, writeFile} from "node:fs/promises";
-import {NOTES_HEADER, notesDir, notesRootFile} from "../../index";
-import {asNonEmptyString} from "./utils";
+import {NOTES_HEADER, notesDir, notesRootFile} from "../../index.js";
+import {asNonEmptyString} from "./utils.js";
 import fs from "node:fs";
-import {toolsLogger} from "./tool-logger";
-import {AiJsonObject} from "../tool-types";
+import {toolsLogger} from "./tool-logger.js";
+import {AiJsonObject} from "../tool-types.js";
 
 const logger = toolsLogger.child("create-note");
 

src/ai/tools/datetime.ts

@@ -1,5 +1,5 @@
 import {AiTool} from "../tool-types";
-import {asNonEmptyString} from "./utils";
+import {asNonEmptyString} from "./utils.js";
 import {AiJsonObject} from "../tool-types";
 
 export const getCurrentDateTimeTool = {

src/ai/tools/files.ts

@@ -3,8 +3,8 @@ import fs from "node:fs";
 import path from "node:path";
 import {z} from "zod";
 
-import {Environment} from "../../common/environment";
-import {AiJsonObject, AiJsonValue, AiTool} from "../tool-types";
+import {Environment} from "../../common/environment.js";
+import {AiJsonObject, AiJsonValue, AiTool} from "../tool-types.js";
 import {
     MAX_COPY_ENTRIES,
     MAX_COPY_TOTAL_BYTES,
@@ -23,8 +23,8 @@ import {
     MAX_PATCH_SEARCH_BYTES,
     MAX_STREAM_WRITE_IDLE_MS,
     MAX_STREAM_WRITE_SESSIONS,
-} from "./limits";
-import {asBoolean, asNonEmptyString, asPositiveInt, asString} from "./utils";
+} from "./limits.js";
+import {asBoolean, asNonEmptyString, asPositiveInt, asString} from "./utils.js";
 
 // =============================================================================
 // Public types and schemas

src/ai/tools/market-rates.ts

@@ -1,7 +1,7 @@
-import {AiTool} from "../tool-types";
+import {AiTool} from "../tool-types.js";
 import axios from "axios";
-import {toolsLogger} from "./tool-logger";
-import {AiJsonObject} from "../tool-types";
+import {toolsLogger} from "./tool-logger.js";
+import {AiJsonObject} from "../tool-types.js";
 
 const logger = toolsLogger.child("market-rates");
 

src/ai/tools/notes.ts

@@ -1,11 +1,11 @@
-import {AiTool} from "../tool-types";
+import {AiTool} from "../tool-types.js";
 import path from "node:path";
 import {readdir, readFile, stat, unlink, writeFile} from "node:fs/promises";
-import {notesDir, notesRootFile} from "../../index";
-import {asNonEmptyString} from "./utils";
-import {toolsLogger} from "./tool-logger";
+import {notesDir, notesRootFile} from "../../index.js";
+import {asNonEmptyString} from "./utils.js";
+import {toolsLogger} from "./tool-logger.js";
 import {z} from "zod";
-import {AiJsonObject} from "../tool-types";
+import {AiJsonObject} from "../tool-types.js";
 
 const logger = toolsLogger.child("notes");
 

src/ai/tools/python-interpretator.ts

@@ -2,11 +2,11 @@ import {spawn} from "node:child_process";
 import {copyFile, lstat, mkdir, readdir, rm, writeFile} from "node:fs/promises";
 import os from "node:os";
 import path from "node:path";
-import {AiTool} from "../tool-types";
-import {Environment} from "../../common/environment";
-import {toolsLogger} from "./tool-logger";
+import {AiTool} from "../tool-types.js";
+import {Environment} from "../../common/environment.js";
+import {toolsLogger} from "./tool-logger.js";
 import {randomUUID} from "node:crypto";
-import {AiJsonObject} from "../tool-types";
+import {AiJsonObject} from "../tool-types.js";
 
 const logger = toolsLogger.child("python-interpreter");
 

src/ai/tools/registry.ts

@@ -1,17 +1,17 @@
-import {Environment} from "../../common/environment";
-import {AiTool} from "../tool-types";
-import {WEB_SEARCH_TOOL_NAME, webSearch, webSearchTool, webSearchToolPrompt} from "./web-search";
-import {getCurrentDateTime, getCurrentDateTimeTool} from "./datetime";
-import {shellExecute, shellExecuteTool} from "./shell";
-import {ToolHandler} from "./types";
-import {getWeather, getWeatherTool} from "./weather";
+import {Environment} from "../../common/environment.js";
+import {AiTool} from "../tool-types.js";
+import {WEB_SEARCH_TOOL_NAME, webSearch, webSearchTool, webSearchToolPrompt} from "./web-search.js";
+import {getCurrentDateTime, getCurrentDateTimeTool} from "./datetime.js";
+import {shellExecute, shellExecuteTool} from "./shell.js";
+import {ToolHandler} from "./types.js";
+import {getWeather, getWeatherTool} from "./weather.js";
 import {
     GET_FINANCIAL_MARKET_DATA_TOOL_NAME,
     getFinancialMarketData,
     getFinancialMarketDataToolPrompt,
     getMarketRates
-} from "./market-rates";
-import {pythonInterpreterTool, runPythonInterpreter} from "./python-interpretator";
+} from "./market-rates.js";
+import {pythonInterpreterTool, runPythonInterpreter} from "./python-interpretator.js";
 import {
     beginFileWrite,
     beginFileWriteTool,
@@ -44,11 +44,14 @@ import {
     updateFileTool,
     writeFileChunk,
     writeFileChunkTool
-} from "./files";
+} from "./files.js";
+import {executeMemoryTool, memoryToolPrompt, memoryTools, type MemoryToolName} from "./user-memory.js";
+import {getMcpToolHandlers, getMcpToolPrompts, getMcpTools} from "../mcp/mcp-registry.js";
 
 export const defaultTools: AiTool[] = [
     getCurrentDateTimeTool,
     getFinancialMarketData,
+    ...memoryTools,
 ];
 
 export const fileTools = [
@@ -72,44 +75,67 @@ export const fileTools = [
     deletePathTool,
 ] satisfies AiTool[];
 
-// export const notesFileTools: AiTool[] = [
-//     createNoteTool,
-//     listNotesTool,
-//     getNoteContentTool,
-//     updateNoteContentTool,
-//     deleteNoteTool,
-//     sendNoteAsFileTool,
-//     searchNotesTool
-// ]
+function parseToolNameSet(raw: string | undefined): Set<string> | undefined {
+    if (!raw?.trim()) return undefined;
+
+    const names = raw
+        .split(",")
+        .map(item => item.trim().toLowerCase())
+        .filter(Boolean);
+
+    return names.length ? new Set(names) : undefined;
+}
+
+function isLocalToolEnabled(toolName: string): boolean {
+    if (Environment.DISABLE_LOCAL_TOOLS) return false;
+
+    const allowlist = parseToolNameSet(Environment.LOCAL_TOOL_ALLOWLIST);
+    if (allowlist && !allowlist.has(toolName.toLowerCase())) return false;
+
+    const denylist = parseToolNameSet(Environment.LOCAL_TOOL_DENYLIST);
+    if (denylist && denylist.has(toolName.toLowerCase())) return false;
+
+    return true;
+}
+
+function filterEnabledTools(tools: AiTool[]): AiTool[] {
+    return tools.filter(tool => isLocalToolEnabled(tool.function.name));
+}
 
 export const getTools = (forCreator?: boolean) => {
-    const tools: AiTool[] = [
-        ...defaultTools,
-        // ...notesFileTools
-    ];
+    const tools: AiTool[] = [];
+
+    if (Environment.DISABLE_LOCAL_TOOLS) {
+        tools.push(...getMcpTools());
+        return tools;
+    }
+
+    tools.push(...filterEnabledTools(defaultTools));
 
     if (Environment.BRAVE_SEARCH_API_KEY) {
-        tools.push(webSearchTool);
+        tools.push(...filterEnabledTools([webSearchTool]));
     }
 
     if (Environment.OPEN_WEATHER_MAP_API_KEY) {
-        tools.push(getWeatherTool);
+        tools.push(...filterEnabledTools([getWeatherTool]));
     }
 
     if (Environment.FILE_TOOLS_ROOT_DIR && Environment.ENABLE_FS_TOOLS) {
-        tools.push(...fileTools);
+        tools.push(...filterEnabledTools(fileTools));
     }
 
     if (forCreator) {
         if (Environment.ENABLE_PYTHON_INTERPRETER) {
-            tools.push(pythonInterpreterTool);
+            tools.push(...filterEnabledTools([pythonInterpreterTool]));
         }
 
         if (Environment.ENABLE_UNSAFE_EVAL) {
-            tools.push(shellExecuteTool);
+            tools.push(...filterEnabledTools([shellExecuteTool]));
         }
     }
 
+    tools.push(...getMcpTools());
+
     return tools;
 };
 
@@ -135,44 +161,83 @@ export const fileToolHandlers = {
 };
 
 export const getToolHandlers = () => {
-    let handlers: Record<string, ToolHandler> = {
-        get_datetime: getCurrentDateTime,
-        get_financial_market_data: getMarketRates,
-
-        // create_note: createNote,
-        // list_notes: listNotes,
-        // get_note_content: getNoteContent,
-        // update_note_content: updateNoteContent,
-        // delete_note: deleteNote,
-        // send_note_as_file: sendNoteAsFile,
-        // search_notes: searchNotes,
-
-        ...fileToolHandlers,
-
-
-        python_interpreter: runPythonInterpreter,
-
-        shell_execute: shellExecute,
-
-        web_search: webSearch,
-
-        get_weather: getWeather,
-
+    const handlers: Record<string, ToolHandler> = {
+        ...getMcpToolHandlers(),
     };
 
+    if (Environment.DISABLE_LOCAL_TOOLS) {
+        return handlers;
+    }
+
+    if (isLocalToolEnabled("get_datetime")) handlers.get_datetime = getCurrentDateTime;
+    if (isLocalToolEnabled("get_financial_market_data")) handlers.get_financial_market_data = getMarketRates;
+    for (const tool of memoryTools) {
+        if (!isLocalToolEnabled(tool.function.name)) continue;
+        handlers[tool.function.name] = async (args, context) => {
+            const userId = typeof args?.userId === "number" ? args.userId : undefined;
+            if (!userId) {
+                return {success: false, error: "Missing userId"};
+            }
+
+            return executeMemoryTool(tool.function.name as MemoryToolName, {
+                userId,
+                content: typeof args?.content === "string" ? args.content : undefined,
+            }, context);
+        };
+    }
+
+    if (isLocalToolEnabled("read_file")) handlers.read_file = readFile;
+    if (isLocalToolEnabled("list_directory")) handlers.list_directory = listDirectory;
+    if (isLocalToolEnabled("search_files")) handlers.search_files = searchFiles;
+    if (isLocalToolEnabled("create_file")) handlers.create_file = createFile;
+    if (isLocalToolEnabled("begin_file_write")) handlers.begin_file_write = beginFileWrite;
+    if (isLocalToolEnabled("write_file_chunk")) handlers.write_file_chunk = writeFileChunk;
+    if (isLocalToolEnabled("finish_file_write")) handlers.finish_file_write = finishFileWrite;
+    if (isLocalToolEnabled("cancel_file_write")) handlers.cancel_file_write = cancelFileWrite;
+    if (isLocalToolEnabled("send_file_as_attachment")) handlers.send_file_as_attachment = sendFileAsAttachment;
+    if (isLocalToolEnabled("create_directory")) handlers.create_directory = createDirectory;
+    if (isLocalToolEnabled("copy_path")) handlers.copy_path = copyPath;
+    if (isLocalToolEnabled("update_file")) handlers.update_file = updateFile;
+    if (isLocalToolEnabled("edit_file_patch")) handlers.edit_file_patch = editFilePatch;
+    if (isLocalToolEnabled("rename_path")) handlers.rename_path = renamePath;
+    if (isLocalToolEnabled("delete_path")) handlers.delete_path = deletePath;
+
+    if (isLocalToolEnabled("python_interpreter")) handlers.python_interpreter = (args, _context) => runPythonInterpreter(args);
+    if (isLocalToolEnabled("shell_execute")) handlers.shell_execute = shellExecute;
+    if (isLocalToolEnabled("web_search")) handlers.web_search = webSearch;
+    if (isLocalToolEnabled("get_weather")) handlers.get_weather = getWeather;
+
     return handlers;
 };
 
 export function getToolPrompts(toolNames: string[]): string[] {
+    if (Environment.DISABLE_LOCAL_TOOLS) {
+        return getMcpToolPrompts(toolNames);
+    }
+
     const prompts: string[] = [];
+    const memoryToolNames = new Set(memoryTools.map(tool => tool.function.name));
+    let memoryPromptAdded = false;
 
     for (const toolName of toolNames) {
+        if (!isLocalToolEnabled(toolName)) {
+            continue;
+        }
+
         if (!prompts.includes(fileToolsToolPrompt) &&
             fileTools.map(t => t.function.name).includes(toolName)) {
             prompts.push(fileToolsToolPrompt);
             continue;
         }
 
+        if (memoryToolNames.has(toolName)) {
+            if (!memoryPromptAdded) {
+                prompts.push(memoryToolPrompt);
+                memoryPromptAdded = true;
+            }
+            continue;
+        }
+
         switch (toolName) {
             case GET_FINANCIAL_MARKET_DATA_TOOL_NAME:
                 prompts.push(getFinancialMarketDataToolPrompt);
@@ -185,5 +250,6 @@ export function getToolPrompts(toolNames: string[]): string[] {
         }
     }
 
+    prompts.push(...getMcpToolPrompts(toolNames));
     return prompts;
 }

src/ai/tools/runtime.ts

@@ -1,14 +1,19 @@
-import {getToolHandlers} from "./registry";
-import {normalizeToolArguments} from "./utils";
-import {PYTHON_INTERPRETER_TOOL_NAME, PythonInterpreterInputFile, runPythonInterpreter} from "./python-interpretator";
-import {toolsLogger} from "./tool-logger";
-import {AiJsonObject, AiJsonValue} from "../tool-types";
+import {getToolHandlers} from "./registry.js";
+import {normalizeToolArguments} from "./utils.js";
+import {PYTHON_INTERPRETER_TOOL_NAME, PythonInterpreterInputFile, runPythonInterpreter} from "./python-interpretator.js";
+import {toolsLogger} from "./tool-logger.js";
+import {AiJsonObject, AiJsonValue} from "../tool-types.js";
+import type {MemoryRuntimeContext} from "./user-memory.js";
+import type {AiRuntimeTarget} from "../ai-runtime-target.js";
+import type {AiProvider} from "../../model/ai-provider.js";
 
 const logger = toolsLogger.child("runtime");
 
 export type ToolRuntimeContext = {
     pythonInputFiles?: PythonInterpreterInputFile[];
-};
+    provider?: AiProvider;
+    runtimeTarget?: AiRuntimeTarget;
+} & MemoryRuntimeContext;
 
 function stringifyToolResult(result: AiJsonValue): string {
     if (typeof result === "string") return result;
@@ -48,7 +53,7 @@ export async function executeToolCall(
         }
 
         const arguments1 = normalizeToolArguments(args, userId);
-        const result = await handler(arguments1);
+        const result = await handler(arguments1, context);
         const s = stringifyToolResult(result);
         logger.debug("execute.done", {name, chars: s.length, duration: logger.duration(startedAt)});
         return s;

src/ai/tools/search-notes.ts

@@ -1,10 +1,10 @@
-import {AiTool} from "../tool-types";
+import {AiTool} from "../tool-types.js";
 import path from "node:path";
 import {readdir, readFile} from "node:fs/promises";
-import {notesDir, notesRootFile} from "../../index";
-import {asNonEmptyString} from "./utils";
-import {toolsLogger} from "./tool-logger";
-import {AiJsonObject, AiJsonValue} from "../tool-types";
+import {notesDir, notesRootFile} from "../../index.js";
+import {asNonEmptyString} from "./utils.js";
+import {toolsLogger} from "./tool-logger.js";
+import {AiJsonObject, AiJsonValue} from "../tool-types.js";
 
 const logger = toolsLogger.child("search-notes");
 

src/ai/tools/shell.ts

@@ -1,6 +1,6 @@
 import {AiTool} from "../tool-types";
-import {runCommand} from "../../util/utils";
-import {asNonEmptyString} from "./utils";
+import {runCommand} from "../../util/utils.js";
+import {asNonEmptyString} from "./utils.js";
 import {AiJsonObject} from "../tool-types";
 
 export const shellExecuteTool = {

src/ai/tools/tool-logger.ts

@@ -1,3 +1,3 @@
-import {appLogger} from "../../logging/logger";
+import {appLogger} from "../../logging/logger.js";
 
 export const toolsLogger = appLogger.child("ai-tools");

src/ai/tools/types.ts

@@ -1,3 +1,4 @@
 import {AiJsonObject, AiJsonValue} from "../tool-types";
+import type {ToolRuntimeContext} from "./runtime.js";
 
-export type ToolHandler = (args?: AiJsonObject) => Promise<AiJsonValue | string | null | undefined> | AiJsonValue | string | null | undefined;
+export type ToolHandler = (args?: AiJsonObject, context?: ToolRuntimeContext) => Promise<AiJsonValue | string | null | undefined> | AiJsonValue | string | null | undefined;

src/ai/tools/user-memory.ts

@@ -0,0 +1,582 @@
+import path from "node:path";
+import {readFile, rename, writeFile, mkdir, rm} from "node:fs/promises";
+import {AiProvider} from "../../model/ai-provider.js";
+import {Environment} from "../../common/environment.js";
+import {createMistralClient, createOllamaClient, createOpenAiClient, resolveOptionalAiRuntimeTarget, type AiRuntimeTarget} from "../ai-runtime-target.js";
+import {AiTool} from "../tool-types.js";
+import {toolsLogger} from "./tool-logger.js";
+import {asNonEmptyString} from "./utils.js";
+
+const logger = toolsLogger.child("user-memory");
+
+function memoryDir(): string {
+    return path.join(Environment.DATA_PATH, "memory");
+}
+
+export const USER_MEMORY_MAX_CHARS = 1000;
+
+export type MemoryScope = "user" | "system";
+export type MemoryAction = "add" | "replace" | "remove";
+
+export type MemoryRuntimeContext = {
+    provider?: AiProvider;
+    runtimeTarget?: AiRuntimeTarget;
+};
+
+export type MemoryOperationResult =
+    | {success: true; scope: MemoryScope; filePath: string; content: string; chars: number; compressed: boolean}
+    | {success: false; scope: MemoryScope; error: string};
+
+type CompressionRunResult = {
+    content: string;
+};
+
+export type MemoryCompressionRunner = (params: {
+    target: AiRuntimeTarget;
+    scope: MemoryScope;
+    currentText: string;
+    limit: number;
+}) => Promise<string>;
+
+function extractMistralText(content: unknown): string {
+    if (typeof content === "string") return content;
+    if (!Array.isArray(content)) return "";
+
+    return content
+        .map(part => {
+            if (typeof part === "string") return part;
+            if (part && typeof part === "object" && "text" in part && typeof (part as {text?: unknown}).text === "string") {
+                return (part as {text: string}).text;
+            }
+            return "";
+        })
+        .join("");
+}
+
+export type MemoryToolName =
+    | "read_user_info"
+    | "read_system_info"
+    | "add_user_info"
+    | "add_system_info"
+    | "remove_user_info"
+    | "remove_system_info"
+    | "replace_user_info"
+    | "replace_system_info"
+    | "delete_user_info"
+    | "delete_system_info";
+
+export const MEMORY_TOOL_NAMES: MemoryToolName[] = [
+    "read_user_info",
+    "read_system_info",
+    "add_user_info",
+    "add_system_info",
+    "remove_user_info",
+    "remove_system_info",
+    "replace_user_info",
+    "replace_system_info",
+    "delete_user_info",
+    "delete_system_info",
+];
+
+type MemoryToolSpec = {
+    name: MemoryToolName;
+    scope: MemoryScope;
+    kind: "read" | "write" | "delete";
+    action?: MemoryAction;
+    description: string;
+    prompt: string;
+};
+
+const MEMORY_TOOL_SPECS: MemoryToolSpec[] = [
+    {
+        name: "read_user_info",
+        scope: "user",
+        kind: "read",
+        description: "Read persistent user memory from user.md.",
+        prompt: `Use when you need to inspect remembered user facts before editing or answering.`,
+    },
+    {
+        name: "read_system_info",
+        scope: "system",
+        kind: "read",
+        description: "Read persistent assistant memory from system.md.",
+        prompt: `Use when you need to inspect remembered assistant instructions before editing or answering.`,
+    },
+    {
+        name: "add_user_info",
+        scope: "user",
+        kind: "write",
+        action: "add",
+        description: "Append a durable fact about the user to user.md.",
+        prompt: `Use for new user facts, preferences, identity details, and profile information. Keep the result at or below ${USER_MEMORY_MAX_CHARS} characters.`,
+    },
+    {
+        name: "add_system_info",
+        scope: "system",
+        kind: "write",
+        action: "add",
+        description: "Append a durable instruction about the assistant to system.md.",
+        prompt: `Use for new assistant identity, style, or behavior instructions. Keep the result at or below ${USER_MEMORY_MAX_CHARS} characters.`,
+    },
+    {
+        name: "remove_user_info",
+        scope: "user",
+        kind: "write",
+        action: "remove",
+        description: "Remove a specific user fact or fragment from user.md.",
+        prompt: `Use when the user asks to forget something about themselves. Keep the result at or below ${USER_MEMORY_MAX_CHARS} characters.`,
+    },
+    {
+        name: "remove_system_info",
+        scope: "system",
+        kind: "write",
+        action: "remove",
+        description: "Remove a specific assistant instruction or fragment from system.md.",
+        prompt: `Use when the user asks to forget something about the assistant. Keep the result at or below ${USER_MEMORY_MAX_CHARS} characters.`,
+    },
+    {
+        name: "replace_user_info",
+        scope: "user",
+        kind: "write",
+        action: "replace",
+        description: "Replace user.md completely with a new compact version.",
+        prompt: `Use when the user wants to overwrite all remembered user info, such as "forget everything about me and remember only this". Keep the result at or below ${USER_MEMORY_MAX_CHARS} characters.`,
+    },
+    {
+        name: "replace_system_info",
+        scope: "system",
+        kind: "write",
+        action: "replace",
+        description: "Replace system.md completely with a new compact version.",
+        prompt: `Use when the user wants to overwrite all remembered assistant info or instructions. Keep the result at or below ${USER_MEMORY_MAX_CHARS} characters.`,
+    },
+    {
+        name: "delete_user_info",
+        scope: "user",
+        kind: "delete",
+        description: "Delete the user memory file user.md.",
+        prompt: `Use when the user asks to delete all remembered user info and remove the memory file entirely.`,
+    },
+    {
+        name: "delete_system_info",
+        scope: "system",
+        kind: "delete",
+        description: "Delete the assistant memory file system.md.",
+        prompt: `Use when the user asks to delete all remembered assistant info and remove the memory file entirely.`,
+    },
+];
+
+export const memoryToolPrompt = [
+    "Use the memory tools to manage persistent per-user memory.",
+    "- `read_*` shows the current file content before editing.",
+    "- `user.md` stores durable facts about the user.",
+    "- `system.md` stores durable facts/instructions about the assistant itself.",
+    "- `add_*` appends a new fact or instruction.",
+    "- `remove_*` removes a specific fact or fragment.",
+    "- `replace_*` rewrites the whole file when the user wants to overwrite memory.",
+    "- `delete_*` removes the file entirely.",
+    `- Keep each file at or below ${USER_MEMORY_MAX_CHARS} characters.`,
+].join("\n");
+
+function createMemoryTool(spec: MemoryToolSpec): AiTool {
+    return {
+        type: "function",
+        function: {
+            name: spec.name,
+            description: spec.description,
+            parameters: {
+                type: "object",
+                properties: spec.kind === "read" || spec.kind === "delete" ? {} : {
+                    content: {
+                        type: "string",
+                        description: spec.action === "remove"
+                            ? "Exact text or fragment to remove from memory."
+                            : "Text to append or replace in memory.",
+                    },
+                },
+                required: spec.kind === "read" || spec.kind === "delete" ? [] : ["content"],
+            },
+        },
+    } satisfies AiTool;
+}
+
+export const memoryTools = MEMORY_TOOL_SPECS.map(createMemoryTool);
+
+function normalizeUserId(userId: number): number | null {
+    return Number.isSafeInteger(userId) && userId > 0 ? userId : null;
+}
+
+function normalizeMemoryText(value: string): string {
+    return value.replaceAll("\r\n", "\n");
+}
+
+function getMemoryUserDir(userId: number): string {
+    return path.join(memoryDir(), String(userId));
+}
+
+export function getMemoryFilePath(userId: number, scope: MemoryScope): string {
+    return path.join(getMemoryUserDir(userId), `${scope}.md`);
+}
+
+async function ensureMemoryDir(userId: number): Promise<string> {
+    const dir = getMemoryUserDir(userId);
+    await mkdir(dir, {recursive: true});
+    return dir;
+}
+
+async function readMemoryFile(userId: number, scope: MemoryScope): Promise<string> {
+    const filePath = getMemoryFilePath(userId, scope);
+    try {
+        return normalizeMemoryText(await readFile(filePath, "utf-8"));
+    } catch (error) {
+        if (error instanceof Error && "code" in error && (error as NodeJS.ErrnoException).code === "ENOENT") {
+            return "";
+        }
+
+        throw error;
+    }
+}
+
+async function writeMemoryFile(userId: number, scope: MemoryScope, content: string): Promise<string> {
+    const normalized = normalizeMemoryText(content);
+    const filePath = getMemoryFilePath(userId, scope);
+    await ensureMemoryDir(userId);
+
+    const tempPath = `${filePath}.tmp-${process.pid}-${Date.now()}`;
+    await writeFile(tempPath, normalized, "utf-8");
+    await rename(tempPath, filePath);
+    return filePath;
+}
+
+function trimToLimit(content: string, limit = USER_MEMORY_MAX_CHARS): string {
+    if (content.length <= limit) return content;
+    return content.slice(0, limit).trimEnd();
+}
+
+function stripCodeFences(content: string): string {
+    const trimmed = content.trim();
+    const fenced = trimmed.match(/^```(?:markdown|md)?\s*([\s\S]*?)\s*```$/i);
+    if (fenced?.[1]) return fenced[1].trim();
+    return trimmed;
+}
+
+function sameTarget(left: AiRuntimeTarget | undefined, right: AiRuntimeTarget | undefined): boolean {
+    if (!left || !right) return false;
+    return left.provider === right.provider
+        && left.model === right.model
+        && (left.baseUrl ?? "") === (right.baseUrl ?? "")
+        && (left.apiKey ?? "") === (right.apiKey ?? "");
+}
+
+async function compressWithTarget(params: {
+    target: AiRuntimeTarget;
+    scope: MemoryScope;
+    currentText: string;
+    limit: number;
+}): Promise<CompressionRunResult> {
+    const {target, scope, currentText, limit} = params;
+
+    const systemPrompt = [
+        "You compress persistent memory for a chat bot.",
+        "Return only the rewritten Markdown text.",
+        "Preserve important facts, preferences, identities, instructions, and durable context.",
+        "Remove noise, duplication, stale details, and low-value filler.",
+        `Keep the result at or below ${limit} characters.`,
+        "Do not add explanations, bullet labels, or code fences.",
+    ].join("\n");
+
+    const userPrompt = [
+        `Memory scope: ${scope}`,
+        `Character limit: ${limit}`,
+        "Current memory:",
+        currentText.trim() || "(empty)",
+        "",
+        "Rewrite it as compact Markdown only.",
+    ].join("\n");
+
+    logger.info("compress.start", {provider: target.provider, model: target.model, scope, chars: currentText.length});
+
+    switch (target.provider) {
+        case AiProvider.OPENAI: {
+            const client = createOpenAiClient(target);
+            const response = await client.chat.completions.create({
+                model: target.model,
+                temperature: 0,
+                messages: [
+                    {role: "system", content: systemPrompt},
+                    {role: "user", content: userPrompt},
+                ],
+            });
+            const text = response.choices[0]?.message?.content ?? "";
+            return {content: stripCodeFences(text)};
+        }
+        case AiProvider.MISTRAL: {
+            const client = createMistralClient(target);
+            const response = await client.chat.complete({
+                model: target.model,
+                temperature: 0,
+                messages: [
+                    {role: "system", content: systemPrompt},
+                    {role: "user", content: userPrompt},
+                ],
+            } as Parameters<typeof client.chat.complete>[0]);
+            const text = extractMistralText(response.choices?.[0]?.message?.content);
+            return {content: stripCodeFences(text)};
+        }
+        case AiProvider.OLLAMA: {
+            const client = createOllamaClient(target);
+            const response = await client.chat({
+                model: target.model,
+                stream: false,
+                options: {temperature: 0},
+                messages: [
+                    {role: "system", content: systemPrompt},
+                    {role: "user", content: userPrompt},
+                ],
+            });
+            const text = typeof response.message?.content === "string" ? response.message.content : "";
+            return {content: stripCodeFences(text)};
+        }
+    }
+}
+
+export async function compressMemoryWithFallback(params: {
+    provider?: AiProvider;
+    currentTarget?: AiRuntimeTarget;
+    scope: MemoryScope;
+    currentText: string;
+    limit?: number;
+}, runner: MemoryCompressionRunner = async (input) => (await compressWithTarget(input)).content): Promise<{content: string; compressed: boolean; usedTarget?: AiRuntimeTarget}> {
+    const limit = params.limit ?? USER_MEMORY_MAX_CHARS;
+    const trimmed = normalizeMemoryText(params.currentText);
+    if (trimmed.length <= limit) {
+        return {content: trimmed, compressed: false};
+    }
+
+    const explicitTarget = params.provider ? resolveOptionalAiRuntimeTarget(params.provider, "memoryCompress") : undefined;
+    const targets = [explicitTarget, params.currentTarget].filter((target, index, list): target is AiRuntimeTarget => !!target && list.findIndex(item => sameTarget(item, target)) === index);
+
+    for (const target of targets) {
+        try {
+            const content = trimToLimit(await runner({target, scope: params.scope, currentText: trimmed, limit}), limit);
+            if (content.length <= limit) {
+                return {content, compressed: true, usedTarget: target};
+            }
+        } catch (error) {
+            logger.warn("compress.failed", {
+                provider: params.provider,
+                scope: params.scope,
+                target: target.model,
+                error: error instanceof Error ? error.message : String(error),
+            });
+        }
+    }
+
+    return {content: trimToLimit(trimmed, limit), compressed: true};
+}
+
+async function compressMemoryIfNeeded(params: {
+    userId: number;
+    scope: MemoryScope;
+    content: string;
+    context?: MemoryRuntimeContext;
+    limit?: number;
+}): Promise<{content: string; compressed: boolean}> {
+    const {scope, context, limit = USER_MEMORY_MAX_CHARS} = params;
+    const result = await compressMemoryWithFallback({
+        provider: context?.provider,
+        currentTarget: context?.runtimeTarget,
+        scope,
+        currentText: params.content,
+        limit,
+    });
+
+    if (!result.compressed) {
+        return result;
+    }
+
+    if (result.content.length > limit) {
+        return {content: trimToLimit(result.content, limit), compressed: true};
+    }
+
+    return {content: result.content, compressed: true};
+}
+
+async function finalizeMemoryWrite(params: {
+    userId: number;
+    scope: MemoryScope;
+    content: string;
+    context?: MemoryRuntimeContext;
+}): Promise<{filePath: string; content: string; compressed: boolean}> {
+    const {userId, scope, context} = params;
+    const compressed = await compressMemoryIfNeeded({userId, scope, content: params.content, context});
+    const filePath = await writeMemoryFile(userId, scope, compressed.content);
+    return {filePath, content: compressed.content, compressed: compressed.compressed};
+}
+
+function findMemoryToolSpec(toolName: string): MemoryToolSpec | undefined {
+    return MEMORY_TOOL_SPECS.find(spec => spec.name === toolName);
+}
+
+function isMemoryWriteTool(spec: MemoryToolSpec): spec is MemoryToolSpec & {kind: "write"; action: MemoryAction} {
+    return spec.kind === "write";
+}
+
+export async function buildUserMemoryPrompt(userId: number | undefined | null): Promise<string | undefined> {
+    const normalizedUserId = typeof userId === "number" ? normalizeUserId(userId) : null;
+    if (!normalizedUserId) return undefined;
+
+    const [userMemoryResult, systemMemoryResult] = await Promise.all([
+        readUserMemory(normalizedUserId, "user"),
+        readUserMemory(normalizedUserId, "system"),
+    ]);
+
+    const userMemory = userMemoryResult.success ? userMemoryResult.content : "";
+    const systemMemory = systemMemoryResult.success ? systemMemoryResult.content : "";
+
+    const blocks: string[] = [];
+    if (systemMemory.trim()) {
+        blocks.push([
+            "## Assistant memory (system.md)",
+            "This is information about the assistant and its behavior.",
+            systemMemory.trim(),
+        ].join("\n"));
+    }
+    if (userMemory.trim()) {
+        blocks.push([
+            "## User memory (user.md)",
+            "This is information about the user.",
+            userMemory.trim(),
+        ].join("\n"));
+    }
+
+    return blocks.length ? blocks.join("\n\n") : undefined;
+}
+
+export async function readUserMemory(userId: number, scope: MemoryScope): Promise<MemoryOperationResult> {
+    const normalizedUserId = normalizeUserId(userId);
+    if (!normalizedUserId) {
+        return {success: false, scope, error: "Invalid userId"};
+    }
+
+    try {
+        const content = await readMemoryFile(normalizedUserId, scope);
+        return {
+            success: true,
+            scope,
+            filePath: getMemoryFilePath(normalizedUserId, scope),
+            content,
+            chars: content.length,
+            compressed: false,
+        };
+    } catch (error) {
+        return {success: false, scope, error: error instanceof Error ? error.message : String(error)};
+    }
+}
+
+export async function updateUserMemory(args: {
+    userId: number;
+    scope: MemoryScope;
+    action: MemoryAction;
+    content?: string;
+    context?: MemoryRuntimeContext;
+}): Promise<MemoryOperationResult> {
+    const normalizedUserId = normalizeUserId(args.userId);
+    if (!normalizedUserId) {
+        return {success: false, scope: args.scope, error: "Invalid userId"};
+    }
+
+    try {
+        const current = await readMemoryFile(normalizedUserId, args.scope);
+        let next = current;
+
+        switch (args.action) {
+            case "add": {
+                const content = normalizeMemoryText(asNonEmptyString(args.content) ?? "");
+                if (!content.trim()) {
+                    return {success: false, scope: args.scope, error: "No content provided"};
+                }
+                next = [current.trimEnd(), content.trim()].filter(Boolean).join(current.trim() ? "\n\n" : "");
+                break;
+            }
+            case "replace": {
+                const content = normalizeMemoryText(asNonEmptyString(args.content) ?? "");
+                next = content;
+                break;
+            }
+            case "remove": {
+                const needle = normalizeMemoryText(asNonEmptyString(args.content) ?? "");
+                if (!needle.trim()) {
+                    return {success: false, scope: args.scope, error: "No text to remove provided"};
+                }
+                if (!current.includes(needle)) {
+                    return {success: false, scope: args.scope, error: "Text not found in memory"};
+                }
+                next = current.split(needle).join("").trim();
+                break;
+            }
+        }
+
+        const finalized = await finalizeMemoryWrite({userId: normalizedUserId, scope: args.scope, content: next, context: args.context});
+        logger.debug("write.done", {
+            userId: normalizedUserId,
+            scope: args.scope,
+            chars: finalized.content.length,
+            compressed: finalized.compressed,
+            filePath: finalized.filePath,
+        });
+
+        return {
+            success: true,
+            scope: args.scope,
+            filePath: finalized.filePath,
+            content: finalized.content,
+            chars: finalized.content.length,
+            compressed: finalized.compressed,
+        };
+    } catch (error) {
+        return {success: false, scope: args.scope, error: error instanceof Error ? error.message : String(error)};
+    }
+}
+
+export async function executeMemoryTool(toolName: MemoryToolName, args: {userId: number; content?: string}, context?: MemoryRuntimeContext): Promise<MemoryOperationResult> {
+    const spec = findMemoryToolSpec(toolName);
+    if (!spec) {
+        return {success: false, scope: "user", error: `Unknown memory tool: ${toolName}`};
+    }
+
+    if (spec.kind === "read") {
+        return readUserMemory(args.userId, spec.scope);
+    }
+
+    if (spec.kind === "delete") {
+        return deleteUserMemory(args.userId, spec.scope);
+    }
+
+    if (!isMemoryWriteTool(spec)) {
+        return {success: false, scope: spec.scope, error: `Unsupported memory tool: ${toolName}`};
+    }
+
+    return updateUserMemory({
+        userId: args.userId,
+        scope: spec.scope,
+        action: spec.action,
+        content: args.content,
+        context,
+    });
+}
+
+export async function deleteUserMemory(userId: number, scope: MemoryScope): Promise<MemoryOperationResult> {
+    const normalizedUserId = normalizeUserId(userId);
+    if (!normalizedUserId) {
+        return {success: false, scope, error: "Invalid userId"};
+    }
+
+    const filePath = getMemoryFilePath(normalizedUserId, scope);
+    try {
+        await rm(filePath, {force: true});
+        return {success: true, scope, filePath, content: "", chars: 0, compressed: false};
+    } catch (error) {
+        return {success: false, scope, error: error instanceof Error ? error.message : String(error)};
+    }
+}

src/ai/tools/utils.ts

@@ -1,5 +1,5 @@
 import {Ollama} from "ollama";
-import {toolsLogger} from "./tool-logger";
+import {toolsLogger} from "./tool-logger.js";
 import {AiJsonObject, AiJsonValue} from "../tool-types";
 import type {BoundaryValue} from "../../common/boundary-types";
 

src/ai/tools/weather.ts

@@ -1,11 +1,11 @@
 import axios from "axios";
-import {toolsLogger} from "./tool-logger";
+import {toolsLogger} from "./tool-logger.js";
 
 const logger = toolsLogger.child("weather");
-import {Environment} from "../../common/environment";
-import {logError} from "../../util/utils";
-import {AiJsonObject, AiTool} from "../tool-types";
-import {asNonEmptyString} from "./utils";
+import {Environment} from "../../common/environment.js";
+import {logError} from "../../util/utils.js";
+import {AiJsonObject, AiTool} from "../tool-types.js";
+import {asNonEmptyString} from "./utils.js";
 
 export const getWeatherTool = {
     type: "function",

src/ai/tools/web-search.ts

@@ -1,11 +1,11 @@
 import axios from "axios";
-import {toolsLogger} from "./tool-logger";
+import {toolsLogger} from "./tool-logger.js";
 
 const logger = toolsLogger.child("brave-search");
-import {Environment} from "../../common/environment";
-import {logError} from "../../util/utils";
-import {AiJsonObject, AiJsonValue, AiTool} from "../tool-types";
-import {asBoolean, asNonEmptyString} from "./utils";
+import {Environment} from "../../common/environment.js";
+import {logError} from "../../util/utils.js";
+import {AiJsonObject, AiJsonValue, AiTool} from "../tool-types.js";
+import {asBoolean, asNonEmptyString} from "./utils.js";
 
 type BraveSearchProfile = {
     name?: string;

src/ai/unified-ai-request-pipeline.ts

@@ -2,7 +2,10 @@ import {AiProvider} from "../model/ai-provider";
 import {AI_VOICE_MODE_TRANSCRIPT, DEFAULT_AI_RESPONSE_LANGUAGE} from "../common/user-ai-settings";
 import {Environment} from "../common/environment";
 import {UserRequestPipeline, type UserRequestPipelineState, type UserRequestPipelineStage} from "./user-request-pipeline";
-import type {AiDownloadedFile} from "./telegram-attachments";
+import {PipelineFallbackNotifier} from "./user-request-pipeline/fallback-notifier";
+import {buildToolRankFallbackTargetDetails} from "./user-request-pipeline/fallback-target-details";
+import {mergeReplyChainDownloads, shouldPreferCurrentDownloads} from "./reply-chain-downloads";
+import {attachmentsToDownloadedFiles, type AiDownloadedFile} from "./telegram-attachments";
 import type {TelegramStreamMessage} from "./telegram-stream-message";
 import type {ChatMessage} from "./chat-messages-types";
 import type {OpenAIChatMessage} from "./openai-chat-message";
@@ -12,6 +15,7 @@ import {prepareDocumentRag} from "./document-rag-pipeline";
 import {persistRagArtifactAttachment} from "./rag-artifact-store";
 import {persistTranscriptArtifactAttachment} from "./transcript-artifact-store";
 import type {ToolRuntimeContext} from "./tools/runtime";
+import {recordPipelineFallback, recordRagRun} from "../common/ai-observability.js";
 import {
     appendTranscriptToChatMessages,
     collectTextMessages,
@@ -21,6 +25,7 @@ import {
     stripAudioFromRunnerMessages,
     toolRuntimeContextFromDownloads,
     transcribeAudioIfNeeded,
+    collectStoredReplyChainAttachments,
     UnifiedRunOptions,
 } from "./unified-ai-runner.shared";
 import {aiLog} from "../logging/ai-logger";
@@ -60,7 +65,7 @@ function runtimeTargetFor(options: UnifiedRunOptions, config: RuntimeConfigSnaps
 
 function createAiRequestPipelineState(options: UnifiedRunOptions): UserRequestPipelineState {
     return {
-        requestId: `ai:${options.msg.chat.id}:${options.msg.message_id}:${Date.now()}`,
+        requestId: options.requestId ?? `ai:${options.msg.chat.id}:${options.msg.message_id}:${Date.now()}`,
         chatId: options.msg.chat.id,
         messageId: options.msg.message_id,
         replyToMessageId: options.msg.reply_to_message?.message_id,
@@ -90,6 +95,12 @@ export async function prepareUnifiedAiRequestPipeline(params: {
     controller: AbortController;
 }): Promise<PreparedUnifiedAiRequest> {
     const {options, config, downloads, streamMessage, controller} = params;
+    const replyChainDownloads = shouldPreferCurrentDownloads(options.text, downloads)
+        ? downloads
+        : mergeReplyChainDownloads(
+            downloads,
+            attachmentsToDownloadedFiles(await collectStoredReplyChainAttachments(options.msg)),
+        );
     const prepared: MutablePreparedContext = {
         chatMessages: [],
         imageCount: 0,
@@ -109,7 +120,7 @@ export async function prepareUnifiedAiRequestPipeline(params: {
                     details: {
                         phase: "ai_request_prepare",
                         provider: options.provider,
-                        downloads: downloads.map(download => ({
+                        downloads: replyChainDownloads.map(download => ({
                             kind: download.kind,
                             fileName: download.fileName,
                             mimeType: download.mimeType,
@@ -126,15 +137,15 @@ export async function prepareUnifiedAiRequestPipeline(params: {
                     options.msg,
                     options.text,
                     options.provider,
-                    downloads,
+                    replyChainDownloads,
                     config,
                     runtimeTargetFor(options, config),
                     options.responseLanguage ?? DEFAULT_AI_RESPONSE_LANGUAGE,
                 );
                 prepared.chatMessages = collected.chatMessages as typeof prepared.chatMessages;
                 prepared.imageCount = collected.imageCount;
-                prepared.firstRoundStatus = initialStatus(downloads, prepared.imageCount);
-                prepared.toolContext = toolRuntimeContextFromDownloads(downloads);
+                prepared.firstRoundStatus = initialStatus(replyChainDownloads, prepared.imageCount);
+                prepared.toolContext = toolRuntimeContextFromDownloads(replyChainDownloads);
 
                 return {
                     stage: "collect_conversation_context",
@@ -169,11 +180,11 @@ export async function prepareUnifiedAiRequestPipeline(params: {
                 prepared.transcript = await transcribeAudioIfNeeded(
                     options.provider,
                     options.msg.from?.id,
-                    downloads,
+                    replyChainDownloads,
                     streamMessage,
                     controller.signal,
                 ).catch(error => {
-                    if (downloads.some(isTranscribableAudioDownload)) throw error;
+                    if (replyChainDownloads.some(isTranscribableAudioDownload)) throw error;
                     return "";
                 });
 
@@ -188,7 +199,7 @@ export async function prepareUnifiedAiRequestPipeline(params: {
                 const transcriptArtifact = await persistTranscriptArtifactAttachment({
                     provider: options.provider,
                     transcript,
-                    downloads,
+                    downloads: replyChainDownloads,
                     chatId: options.msg.chat.id,
                     messageId: options.msg.message_id,
                 });
@@ -233,7 +244,7 @@ export async function prepareUnifiedAiRequestPipeline(params: {
 
                 prepared.preparedDocumentRag = await prepareDocumentRag(
                     options.provider,
-                    downloads,
+                    replyChainDownloads,
                     prepared.chatMessages,
                     streamMessage,
                     config,
@@ -244,7 +255,7 @@ export async function prepareUnifiedAiRequestPipeline(params: {
                 const ragArtifact = await persistRagArtifactAttachment({
                     provider: options.provider,
                     prepared: prepared.preparedDocumentRag,
-                    downloads,
+                    downloads: replyChainDownloads,
                     chatId: options.msg.chat.id,
                     messageId: options.msg.message_id,
                     details: prepared.preparedDocumentRag?.provider === AiProvider.OPENAI
@@ -264,6 +275,10 @@ export async function prepareUnifiedAiRequestPipeline(params: {
                     await streamMessage.storeInternalAttachment(ragArtifact);
                 }
 
+                if (prepared.preparedDocumentRag) {
+                    recordRagRun();
+                }
+
                 return {
                     stage: "document_rag",
                     status: prepared.preparedDocumentRag ? "succeeded" : "skipped",
@@ -290,6 +305,7 @@ export async function prepareUnifiedAiRequestPipeline(params: {
     ];
 
     const state = createAiRequestPipelineState(options);
+    const fallbackNotifier = new PipelineFallbackNotifier(options.msg, options.responseLanguage);
     const pipeline = new UserRequestPipeline({
         stages,
         stageNames: [
@@ -301,6 +317,44 @@ export async function prepareUnifiedAiRequestPipeline(params: {
             "document_rag",
             "audit_finish",
         ],
+        onFallback: async decision => {
+            recordPipelineFallback(decision.action);
+            if (decision.action === "use_alternate_target") {
+                aiLog("warn", "request.fallback.use_alternate_target", {
+                    provider: options.provider,
+                    stage: decision.stage,
+                    reason: decision.reason,
+                    requestId: state.requestId,
+                    ...buildToolRankFallbackTargetDetails(options.provider, config),
+                });
+            }
+
+            if (decision.action === "fail_request") {
+                aiLog("error", "request.fallback.fail_request", {
+                    provider: options.provider,
+                    stage: decision.stage,
+                    reason: decision.reason,
+                    requestId: state.requestId,
+                });
+            }
+
+            const notification = await fallbackNotifier.notify(state.requestId, decision);
+            state.audit.push({
+                stage: decision.stage,
+                status: "fallback",
+                startedAt: nowIso(),
+                finishedAt: nowIso(),
+                details: {
+                    fallbackAction: decision.action,
+                    fallbackNotification: notification.text,
+                    fallbackNotified: notification.notified,
+                    reason: decision.reason,
+                    ...(decision.action === "use_alternate_target"
+                        ? buildToolRankFallbackTargetDetails(options.provider, config)
+                        : {}),
+                },
+            });
+        },
     });
     await pipeline.run(state, controller.signal);
     await streamMessage.storePipelineAudit(state.audit);

src/ai/unified-ai-response-pipeline.ts

@@ -2,6 +2,7 @@ import {AiProvider} from "../model/ai-provider";
 import {Environment} from "../common/environment";
 import {ifTrue, logError} from "../util/utils";
 import {UserRequestPipeline, type UserRequestPipelineState, type UserRequestPipelineStage} from "./user-request-pipeline";
+import {getProviderAdapter} from "./provider-adapters";
 import type {AiDownloadedFile} from "./telegram-attachments";
 import type {TelegramStreamMessage} from "./telegram-stream-message";
 import type {PreparedUnifiedAiRequest} from "./unified-ai-request-pipeline";
@@ -9,15 +10,23 @@ import type {OpenAIChatMessage} from "./openai-chat-message";
 import type {MistralChatMessage} from "./mistral-chat-message";
 import type {ChatMessage} from "./chat-messages-types";
 import {
+    allToolSchemaNames,
     providerName,
     RuntimeConfigSnapshot,
     snapshotModel,
     TELEGRAM_LIMIT,
     UnifiedRunOptions,
 } from "./unified-ai-runner.shared";
+import {runToolRankStage} from "./tool-rank-stage";
 import {runOpenAi} from "./unified-ai-runner.openai";
+import {runOpenAiCompatible} from "./unified-ai-runner.openai-compatible";
 import {runOllama} from "./unified-ai-runner.ollama";
 import {runMistral} from "./unified-ai-runner.mistral";
+import {summarizeModelOutput} from "./response-model-output";
+import {summarizeToolLoop} from "./tool-loop-summary";
+import {persistToolLoopSummaryArtifactAttachment} from "./tool-loop-artifact-store";
+import {PipelineFallbackNotifier} from "./user-request-pipeline/fallback-notifier";
+import {buildToolRankFallbackTargetDetails} from "./user-request-pipeline/fallback-target-details";
 import {
     resolveTextToSpeechProviderForUser,
     sendSynthesizedSpeech,
@@ -26,6 +35,7 @@ import {
 } from "./text-to-speech";
 import {persistFinalTextArtifactAttachment} from "./final-response-artifact-store";
 import {aiLog} from "../logging/ai-logger";
+import {recordPipelineFallback, recordTtsRun} from "../common/ai-observability.js";
 
 function nowIso(): string {
     return new Date().toISOString();
@@ -33,7 +43,7 @@ function nowIso(): string {
 
 function createResponsePipelineState(options: UnifiedRunOptions): UserRequestPipelineState {
     return {
-        requestId: `ai-response:${options.msg.chat.id}:${options.msg.message_id}:${Date.now()}`,
+        requestId: options.requestId ?? `ai-response:${options.msg.chat.id}:${options.msg.message_id}:${Date.now()}`,
         chatId: options.msg.chat.id,
         messageId: options.msg.message_id,
         replyToMessageId: options.msg.reply_to_message?.message_id,
@@ -71,6 +81,21 @@ async function runProviderModelCall(params: {
 
     switch (options.provider) {
         case AiProvider.OPENAI:
+            if (config.openAiBackend === "compatible") {
+                await runOpenAiCompatible(
+                    options.msg,
+                    prepared.chatMessages as OpenAIChatMessage[],
+                    streamMessage,
+                    signal,
+                    options.stream ?? true,
+                    options.msg,
+                    config,
+                    prepared.toolContext,
+                    downloads,
+                );
+                return;
+            }
+
             await runOpenAi(
                 options.msg,
                 prepared.chatMessages as OpenAIChatMessage[],
@@ -159,6 +184,10 @@ export async function runUnifiedAiResponsePipeline(params: {
 }): Promise<void> {
     const {options, config, downloads, prepared, streamMessage, controller} = params;
     const state = createResponsePipelineState(options);
+    const fallbackNotifier = new PipelineFallbackNotifier(options.msg, options.responseLanguage);
+    const adapter = getProviderAdapter(options.provider);
+    let selectedToolNames: string[] = [];
+    let filteredTools: unknown[] = [];
 
     const stages: UserRequestPipelineStage[] = [
         {
@@ -177,6 +206,62 @@ export async function runUnifiedAiResponsePipeline(params: {
                 };
             },
         },
+        {
+            name: "tool_rank",
+            async run() {
+                const availableTools = adapter.rankTools(config, {
+                    forCreator: options.msg.from?.id === Environment.CREATOR_ID,
+                    vectorStoreIds: prepared.preparedDocumentRag?.provider === AiProvider.OPENAI
+                        ? prepared.preparedDocumentRag.vectorStoreIds
+                        : [],
+                });
+
+                const rankResult = await runToolRankStage({
+                    provider: options.provider,
+                    model: snapshotModel(options.provider, config),
+                    round: state.toolRankDecisions.length,
+                    config,
+                    availableTools,
+                    messages: prepared.chatMessages,
+                    streamMessage,
+                    signal: controller.signal,
+                });
+
+                selectedToolNames = rankResult.selectedToolNames;
+                filteredTools = rankResult.filteredTools;
+                state.toolRankDecisions.push({
+                    provider: options.provider,
+                    round: state.toolRankDecisions.length,
+                    availableTools: allToolSchemaNames(availableTools),
+                    selectedTools: selectedToolNames,
+                    usedRanker: rankResult.usedRanker,
+                });
+
+                return {
+                    stage: "tool_rank",
+                    status: "succeeded",
+                    details: {
+                        selectedTools: selectedToolNames,
+                        usedRanker: rankResult.usedRanker,
+                        availableTools: allToolSchemaNames(availableTools),
+                        toolRankDecision: state.toolRankDecisions.at(-1),
+                    },
+                };
+            },
+        },
+        {
+            name: "filter_tools",
+            async run() {
+                return {
+                    stage: "filter_tools",
+                    status: "succeeded",
+                    details: {
+                        selectedTools: selectedToolNames,
+                        filteredToolCount: filteredTools.length,
+                    },
+                };
+            },
+        },
         {
             name: "model_call",
             async run() {
@@ -192,6 +277,13 @@ export async function runUnifiedAiResponsePipeline(params: {
                 return {
                     stage: "model_call",
                     status: "succeeded",
+                    details: {
+                        modelOutput: summarizeModelOutput({
+                            text: streamMessage.getText(),
+                            toolExecutions: streamMessage.getToolExecutions(),
+                            outputAttachments: streamMessage.getOutputAttachments(),
+                        }),
+                    },
                 };
             },
         },
@@ -199,33 +291,31 @@ export async function runUnifiedAiResponsePipeline(params: {
             name: "tool_loop",
             async run() {
                 const executions = streamMessage.getToolExecutions();
+                const outputAttachments = streamMessage.getOutputAttachments();
+                const summary = summarizeToolLoop({
+                    text: streamMessage.getText(),
+                    executions,
+                    outputAttachments,
+                });
+                const persisted = await persistToolLoopSummaryArtifactAttachment({
+                    chatId: options.msg.chat.id,
+                    messageId: options.msg.message_id,
+                    text: streamMessage.getText(),
+                    executions,
+                    outputAttachments,
+                });
+
+                if (persisted) {
+                    await streamMessage.storeInternalAttachment(persisted);
+                }
+
                 return {
                     stage: "tool_loop",
-                    status: executions.length ? "succeeded" : "skipped",
-                    fallbackAction: executions.length ? undefined : "continue_without_stage",
+                    ...summary,
                     details: {
-                        count: executions.length,
-                        tools: executions.map(execution => ({
-                            toolName: execution.toolName,
-                            callId: execution.callId,
-                            resultChars: execution.resultChars,
-                        })),
+                        ...summary.details,
+                        persistedSummaryArtifact: !!persisted,
                     },
-                    artifacts: executions.length ? [{
-                        kind: "tool_result",
-                        stage: "tool_loop",
-                        createdAt: new Date().toISOString(),
-                        toolName: "summary",
-                        callId: "tool_loop_summary",
-                        resultText: JSON.stringify({
-                            count: executions.length,
-                            tools: executions.map(execution => ({
-                                toolName: execution.toolName,
-                                callId: execution.callId,
-                                resultChars: execution.resultChars,
-                            })),
-                        }),
-                    }] : undefined,
                 };
             },
         },
@@ -284,6 +374,7 @@ export async function runUnifiedAiResponsePipeline(params: {
             name: "text_to_speech",
             async run() {
                 const status = await synthesizeResponseIfRequested({options, config, streamMessage});
+                recordTtsRun(status);
                 return {
                     stage: "text_to_speech",
                     status,
@@ -312,6 +403,8 @@ export async function runUnifiedAiResponsePipeline(params: {
         stages,
         stageNames: [
             "audit_start",
+            "tool_rank",
+            "filter_tools",
             "model_call",
             "tool_loop",
             "output_size_gate",
@@ -320,6 +413,44 @@ export async function runUnifiedAiResponsePipeline(params: {
             "persist_output_artifacts",
             "audit_finish",
         ],
+        onFallback: async decision => {
+            recordPipelineFallback(decision.action);
+            if (decision.action === "use_alternate_target") {
+                aiLog("warn", "response.fallback.use_alternate_target", {
+                    provider: options.provider,
+                    stage: decision.stage,
+                    reason: decision.reason,
+                    requestId: state.requestId,
+                    ...buildToolRankFallbackTargetDetails(options.provider, config),
+                });
+            }
+
+            if (decision.action === "fail_request") {
+                aiLog("error", "response.fallback.fail_request", {
+                    provider: options.provider,
+                    stage: decision.stage,
+                    reason: decision.reason,
+                    requestId: state.requestId,
+                });
+            }
+
+            const notification = await fallbackNotifier.notify(state.requestId, decision);
+            state.audit.push({
+                stage: decision.stage,
+                status: "fallback",
+                startedAt: new Date().toISOString(),
+                finishedAt: new Date().toISOString(),
+                details: {
+                    fallbackAction: decision.action,
+                    fallbackNotification: notification.text,
+                    fallbackNotified: notification.notified,
+                    reason: decision.reason,
+                    ...(decision.action === "use_alternate_target"
+                        ? buildToolRankFallbackTargetDetails(options.provider, config)
+                        : {}),
+                },
+            });
+        },
     });
 
     try {

src/ai/unified-ai-runner.mistral.ts

@@ -1,30 +1,29 @@
 import {Environment} from "../common/environment";
-import {getMistralTools} from "./tool-mappers";
 import {TelegramStreamMessage} from "./telegram-stream-message";
 import {ToolRuntimeContext} from "./tools/runtime";
 import {MistralChatMessage} from "./mistral-chat-message";
 import {createMistralClient} from "./ai-runtime-target";
 import {aiLog, aiLogDuration, aiLogProviderTarget, aiLogToolCall} from "../logging/ai-logger";
 import {AiProvider} from "../model/ai-provider";
-import {ToolRanker} from "./unified-ai-runner.tool-ranker";
+import {getProviderAdapter} from "./provider-adapters";
+import {runToolRankStage} from "./tool-rank-stage";
+import {ensureToolsSelected} from "./tool-mappers.js";
+import {MEMORY_TOOL_NAMES} from "./tools/user-memory.js";
 
 import {
-    contentFromMistralDelta,
-    executeToolBatch,
     MAX_TOOL_ROUNDS,
-    MistralDeltaLike,
     MistralDocumentReference,
-    mistralToolCalls,
-    normalizeMistralToolCalls,
     roundStatus,
     RuntimeConfigSnapshot,
     StreamingToolCallAccumulator,
     ToolCallData,
     ToolExecutionMemory
 } from "./unified-ai-runner.shared";
+import {executeToolBatchWithAdapter} from "./tool-batch-runner";
+import {decideToolLoopContinuation} from "./tool-loop-control";
+import {runToolLoopRounds} from "./tool-loop-runner";
+import {runSingleModelRequest} from "./model-call-stage";
 import {Message} from "typescript-telegram-bot-api";
-import {filterRankedTools, latestUserTextFromMessages} from "./tool-ranker-pipeline";
-import {storeToolRankAudit} from "./tool-rank-audit";
 
 export async function runMistral(
     msg: Message,
@@ -39,8 +38,9 @@ export async function runMistral(
 ): Promise<void> {
     const runnerStartedAt = Date.now();
     const mistralAi = createMistralClient(config.mistralChatTarget);
-    const toolRanker = new ToolRanker(config);
-    const availableTools = getMistralTools(msg.from?.id === Environment.CREATOR_ID);
+    const adapter = getProviderAdapter(AiProvider.MISTRAL);
+    const availableTools = adapter.rankTools(config, {forCreator: msg.from?.id === Environment.CREATOR_ID});
+    const requestMessages = adapter.mapMessages([...messages]) as unknown as MistralChatMessage[];
     aiLog("info", "mistral.run.start", {
         stream,
         target: aiLogProviderTarget(config.mistralChatTarget),
@@ -50,49 +50,25 @@ export async function runMistral(
     });
 
     const toolMemory: ToolExecutionMemory = new Map();
-
-    for (let round = 0; round < MAX_TOOL_ROUNDS; round++) {
+    try {
+        await runToolLoopRounds({
+            maxRounds: MAX_TOOL_ROUNDS,
+            onRound: async (round) => {
             const roundStartedAt = Date.now();
             aiLog("debug", "mistral.round.start", {round, messages: messages.length, stream});
             if (signal.aborted) throw new Error("Aborted");
 
-        streamMessage.setStatus(Environment.getSelectingToolsText());
-        await streamMessage.flush();
-        const toolRankStartedAt = Date.now();
-        const toolRankStartedAtIso = new Date().toISOString();
-        const rankerSelection = await toolRanker.selectTools({
+            const rankResult = await runToolRankStage({
                 provider: AiProvider.MISTRAL,
-                userQuery: latestUserTextFromMessages(messages),
+                model: config.mistralChatTarget.model,
+                round,
+                config,
                 availableTools,
-                round,
+                messages,
+                streamMessage,
                 signal,
-            })
-            .catch(async error => {
-                streamMessage.clearStatus();
-                await streamMessage.flush();
-                await storeToolRankAudit({
-                    streamMessage,
-                    provider: AiProvider.MISTRAL,
-                    model: config.mistralChatTarget.model,
-                    round,
-                    startedAt: toolRankStartedAt,
-                    startedAtIso: toolRankStartedAtIso,
-                    error,
             });
-                throw error;
-            });
-        streamMessage.clearStatus();
-        await streamMessage.flush();
-        await storeToolRankAudit({
-            streamMessage,
-            provider: AiProvider.MISTRAL,
-            model: config.mistralChatTarget.model,
-            round,
-            startedAt: toolRankStartedAt,
-            startedAtIso: toolRankStartedAtIso,
-            selectedTools: rankerSelection.toolNames,
-        });
-        const filteredTools = filterRankedTools(availableTools, rankerSelection.toolNames);
+            const filteredTools = ensureToolsSelected(availableTools, rankResult.filteredTools, MEMORY_TOOL_NAMES);
             const requestTools = filteredTools.length ? filteredTools : undefined;
 
             streamMessage.setStatus(roundStatus(round, firstRoundStatus) ?? "");
@@ -101,22 +77,24 @@ export async function runMistral(
             if (!stream) {
                 const request = {
                     model: config.mistralChatTarget.model,
-                messages,
+                    messages: requestMessages,
                     tools: requestTools,
                     documents: documents
                 } as Parameters<typeof mistralAi.chat.complete>[0];
-            const response = await mistralAi.chat.complete(request, {signal});
+                const response = await runSingleModelRequest({
+                    execute: () => adapter.callModel(request, () => mistralAi.chat.complete(request, {signal})),
+                });
                 const message = response.choices?.[0]?.message;
                 const text = typeof message?.content === "string" ? message.content : JSON.stringify(message?.content ?? "");
                 streamMessage.append(text);
-            const calls = normalizeMistralToolCalls(mistralToolCalls(message));
+                const calls = adapter.extractToolCalls(message);
                 aiLog(calls.length ? "info" : "success", calls.length ? "mistral.tool_calls" : "mistral.run.done", {
                     round,
                     duration: calls.length ? aiLogDuration(roundStartedAt) : aiLogDuration(runnerStartedAt),
                     textChars: text.length,
                     calls: calls.map(aiLogToolCall),
                 });
-            if (!calls.length) return;
+                if (!calls.length) return {shouldContinue: false};
                 messages.push({
                     role: "assistant",
                     content: text,
@@ -125,25 +103,50 @@ export async function runMistral(
                         function: {name: call.name, arguments: call.argumentsText},
                     })),
                 });
-            const toolResults = await executeToolBatch(msg.from?.id, calls, streamMessage, toolContext, toolMemory);
-            for (const [index, call] of calls.entries()) {
-                messages.push({
-                    role: "tool",
-                    name: call.name,
-                    toolCallId: call.id,
-                    content: toolResults[index] ?? "",
+                requestMessages.push({
+                    role: "assistant",
+                    content: text,
+                    toolCalls: calls.map(call => ({
+                        id: call.id,
+                        function: {name: call.name, arguments: call.argumentsText},
+                    })),
+                });
+                await executeToolBatchWithAdapter({
+                    userId: msg.from?.id,
+                    toolCalls: calls,
+                    streamMessage,
+                    toolContext: {
+                        ...toolContext,
+                        provider: AiProvider.MISTRAL,
+                        runtimeTarget: config.mistralChatTarget,
+                    },
+                    toolMemory,
+                    adapter,
+                    appendTargets: [messages, requestMessages],
+                });
+                const continuation = decideToolLoopContinuation({
+                    round,
+                    maxRounds: MAX_TOOL_ROUNDS,
+                    toolCalls: calls,
+                });
+                if (!continuation.continue && continuation.reason === "max_rounds_reached") {
+                    aiLog("warn", "mistral.tool_loop.max_rounds_reached", {
+                        round,
+                        maxRounds: MAX_TOOL_ROUNDS,
                     });
                 }
-            continue;
+                return {shouldContinue: true};
             }
 
             const request = {
                 model: config.mistralChatTarget.model,
-            messages,
+                messages: requestMessages,
                 tools: requestTools,
                 documents: documents
             } as Parameters<typeof mistralAi.chat.stream>[0];
-        const streamResponse = await mistralAi.chat.stream(request, {signal});
+            const streamResponse = await runSingleModelRequest({
+                execute: () => adapter.callModel(request, () => mistralAi.chat.stream(request, {signal})),
+            });
             aiLog("debug", "mistral.stream.open", {round});
             let calls: ToolCallData[] = [];
             const roundTextStart = streamMessage.getText().length;
@@ -154,11 +157,10 @@ export async function runMistral(
 
                 const choice = event.data?.choices?.[0];
                 const delta = choice?.delta;
-            const mistralDelta = delta as MistralDeltaLike;
+                const mistralDelta = delta;
+                streamMessage.append(adapter.extractTextDelta(mistralDelta));
 
-            streamMessage.append(contentFromMistralDelta(mistralDelta));
-
-            const rawDeltaCalls = mistralToolCalls(mistralDelta);
+                const rawDeltaCalls = adapter.extractStreamingToolCalls(mistralDelta);
                 if (rawDeltaCalls.length) {
                     calls = toolCallAccumulator.add(rawDeltaCalls);
                     streamMessage.setStatus(Environment.getUseToolText(calls));
@@ -171,21 +173,46 @@ export async function runMistral(
                 textChars: streamMessage.getText().slice(roundTextStart).length,
                 calls: calls.map(aiLogToolCall),
             });
-        if (!calls.length) return;
+            if (!calls.length) return {shouldContinue: false};
             const roundText = streamMessage.getText().slice(roundTextStart);
             messages.push({
                 role: "assistant",
                 content: roundText,
                 toolCalls: calls.map(c => ({id: c.id, function: {name: c.name, arguments: c.argumentsText}}))
             });
-        const toolResults = await executeToolBatch(msg.from?.id, calls, streamMessage, toolContext, toolMemory);
-        for (const [index, call] of calls.entries()) {
-            messages.push({
-                role: "tool",
-                name: call.name,
-                toolCallId: call.id,
-                content: toolResults[index] ?? "",
+            requestMessages.push({
+                role: "assistant",
+                content: roundText,
+                toolCalls: calls.map(c => ({id: c.id, function: {name: c.name, arguments: c.argumentsText}}))
+            });
+            await executeToolBatchWithAdapter({
+                userId: msg.from?.id,
+                toolCalls: calls,
+                streamMessage,
+                toolContext: {
+                    ...toolContext,
+                    provider: AiProvider.MISTRAL,
+                    runtimeTarget: config.mistralChatTarget,
+                },
+                toolMemory,
+                adapter,
+                appendTargets: [messages, requestMessages],
+            });
+            const continuation = decideToolLoopContinuation({
+                round,
+                maxRounds: MAX_TOOL_ROUNDS,
+                toolCalls: calls,
+            });
+                if (!continuation.continue && continuation.reason === "max_rounds_reached") {
+                    aiLog("warn", "mistral.tool_loop.max_rounds_reached", {
+                        round,
+                        maxRounds: MAX_TOOL_ROUNDS,
                     });
                 }
+                return {shouldContinue: true};
+        },
+        });
+    } finally {
+        await adapter.finalize().catch(() => undefined);
     }
 }

src/ai/unified-ai-runner.ollama.ts

@@ -5,7 +5,6 @@ import {Environment} from "../common/environment";
 import type {BoundaryValue} from "../common/boundary-types";
 import {bot, notesDir} from "../index";
 import {clamp, logError} from "../util/utils";
-import {getOllamaTools} from "./tool-mappers";
 import {TelegramStreamMessage} from "./telegram-stream-message";
 import {ChatMessage} from "./chat-messages-types";
 import {ChatRequest, Tool} from "ollama";
@@ -14,20 +13,20 @@ import {enqueueTelegramApiCall} from "../util/telegram-api-queue";
 import {loadOllamaModel, unloadAllOllamaModels} from "./tools/utils";
 import {createOllamaClient} from "./ai-runtime-target";
 import {aiLog, aiLogDuration, aiLogMessageIdentity, aiLogProviderTarget, aiLogToolCall} from "../logging/ai-logger";
+import {getProviderAdapter} from "./provider-adapters";
+import {runToolRankStage} from "./tool-rank-stage";
+import {ensureToolsSelected} from "./tool-mappers.js";
+import {MEMORY_TOOL_NAMES} from "./tools/user-memory.js";
 
 import {
     allToolSchemaNames,
-    appendOllamaToolResults,
     dedupeToolCalls,
     DEFAULT_OLLAMA_CONTEXT_SIZE,
-    executeToolBatch,
     isOllamaModelActive,
     isRecord,
     MAX_OLLAMA_CONTEXT_SIZE,
     MAX_TOOL_ROUNDS,
     MIN_OLLAMA_CONTEXT_SIZE,
-    normalizeOllamaToolCalls,
-    OllamaToolCallLike,
     roundStatus,
     RuntimeConfigSnapshot,
     safeJsonParseObject,
@@ -35,14 +34,15 @@ import {
     ToolCallData,
     ToolExecutionMemory
 } from "./unified-ai-runner.shared";
-import {ToolRanker} from "./unified-ai-runner.tool-ranker";
+import {executeToolBatchWithAdapter} from "./tool-batch-runner";
+import {decideToolLoopContinuation} from "./tool-loop-control";
+import {runToolLoopRounds} from "./tool-loop-runner";
+import {runSingleModelRequest} from "./model-call-stage";
 import {getToolPrompts} from "./tools/registry";
-import {filterRankedTools, latestUserTextFromMessages} from "./tool-ranker-pipeline";
 import {GetNoteFileResult, GetNoteFileResultSchema} from "./tools/notes";
 import {getModelCapabilities} from "./provider-model-runtime";
 import {AiProvider} from "../model/ai-provider";
 import {Message} from "typescript-telegram-bot-api";
-import {storeToolRankAudit} from "./tool-rank-audit";
 
 export async function runOllama(
     msg: Message,
@@ -157,9 +157,12 @@ export async function runOllama(
     }
 
     const toolMemory: ToolExecutionMemory = new Map();
+    const adapter = getProviderAdapter(AiProvider.OLLAMA);
 
     try {
-        for (let round = 0; round < MAX_TOOL_ROUNDS; round++) {
+        await runToolLoopRounds({
+            maxRounds: MAX_TOOL_ROUNDS,
+            onRound: async (round) => {
             const roundStartedAt = Date.now();
             aiLog("debug", "ollama.round.start", {
                 round,
@@ -183,7 +186,7 @@ export async function runOllama(
 
             let activeToolNames: string[] = [];
             if ((await getModelCapabilities(AiProvider.OLLAMA, model, "tools"))?.tools?.supported) {
-                const availableOllamaTools: Tool[] = getOllamaTools(msg.from?.id === Environment.CREATOR_ID) as Tool[];
+                const availableOllamaTools: Tool[] = adapter.rankTools(config, {forCreator: msg.from?.id === Environment.CREATOR_ID}) as Tool[];
 
                 aiLog("debug", "ollama.tools.available", {
                     round,
@@ -191,44 +194,18 @@ export async function runOllama(
                     rankerEnabled: !!config.ollamaToolRankerTarget,
                 });
 
-                streamMessage.setStatus(Environment.getSelectingToolsText());
-                await streamMessage.flush();
-                const toolRankStartedAt = Date.now();
-                const toolRankStartedAtIso = new Date().toISOString();
-                const rankerSelection = await new ToolRanker(config).selectTools({
+                const rankResult = await runToolRankStage({
                     provider: AiProvider.OLLAMA,
-                        userQuery: latestUserTextFromMessages(messages),
+                    model,
+                    round,
+                    config,
                     availableTools: availableOllamaTools,
-                        round,
+                    messages,
+                    streamMessage,
                     signal,
-                    })
-                    .catch(async error => {
-                        streamMessage.clearStatus();
-                        await streamMessage.flush();
-                        await storeToolRankAudit({
-                            streamMessage,
-                            provider: AiProvider.OLLAMA,
-                            model,
-                            round,
-                            startedAt: toolRankStartedAt,
-                            startedAtIso: toolRankStartedAtIso,
-                            error,
-                        });
-                        throw error;
-                    });
-                streamMessage.clearStatus();
-                await streamMessage.flush();
-                await storeToolRankAudit({
-                    streamMessage,
-                    provider: AiProvider.OLLAMA,
-                    model,
-                    round,
-                    startedAt: toolRankStartedAt,
-                    startedAtIso: toolRankStartedAtIso,
-                    selectedTools: rankerSelection.toolNames,
                 });
 
-                const filteredTools = [...new Set(filterRankedTools(availableOllamaTools, rankerSelection.toolNames))];
+                const filteredTools = [...new Set(ensureToolsSelected(availableOllamaTools, rankResult.filteredTools as Tool[], MEMORY_TOOL_NAMES) as Tool[])];
                 activeToolNames = filteredTools.map(t => t.function.name ?? "");
                 if (filteredTools.length > 0) {
                     request.tools = [...filteredTools];
@@ -256,24 +233,23 @@ export async function runOllama(
                     round,
                     tools: activeToolNames,
                     count: activeToolNames.length,
-                    usedRanker: rankerSelection.usedRanker,
+                    usedRanker: rankResult.usedRanker,
                 });
             }
 
             if (!stream) {
-                const response = await ollama.chat({
+                const response = await runSingleModelRequest({
+                    execute: () => adapter.callModel(request, () => ollama.chat({
                         ...request,
                         stream: false
+                    })),
                 });
 
                 const message = response.message;
                 const rawContent = message?.content ?? "";
 
                 const nativeCalls = dedupeToolCalls(
-                    normalizeOllamaToolCalls(
-                        message?.tool_calls as readonly OllamaToolCallLike[] | undefined,
-                        round,
-                    ),
+                    adapter.extractToolCalls(message),
                 );
 
                 const responseText = rawContent;
@@ -298,10 +274,10 @@ export async function runOllama(
 
                 if (!nativeCalls.length) {
                     aiLog("success", "ollama.run.done", {round, duration: aiLogDuration(runnerStartedAt)});
-                    break;
+                    return {shouldContinue: false};
                 }
 
-                const calls = nativeCalls;
+                const calls = adapter.extractToolCalls(message).length ? adapter.extractToolCalls(message) : nativeCalls;
 
                 aiLog("info", "ollama.tool_calls", {
                     round,
@@ -319,22 +295,44 @@ export async function runOllama(
                     })),
                 });
 
-                appendOllamaToolResults(
-                    messages,
-                    calls,
-                    await executeToolBatch(msg.from?.id, calls, streamMessage, toolContext, toolMemory),
-                );
+                await executeToolBatchWithAdapter({
+                    userId: msg.from?.id,
+                    toolCalls: calls,
+                    streamMessage,
+                    toolContext: {
+                        ...toolContext,
+                        provider: AiProvider.OLLAMA,
+                        runtimeTarget: target,
+                    },
+                    toolMemory,
+                    adapter,
+                    appendTargets: [messages],
+                });
 
-                continue;
+                const continuation = decideToolLoopContinuation({
+                    round,
+                    maxRounds: MAX_TOOL_ROUNDS,
+                    toolCalls: calls,
+                });
+                if (!continuation.continue && continuation.reason === "max_rounds_reached") {
+                    aiLog("warn", "ollama.tool_loop.max_rounds_reached", {
+                        round,
+                        maxRounds: MAX_TOOL_ROUNDS,
+                    });
+                }
+
+                return {shouldContinue: true};
             }
 
             aiLog("debug", "ollama.stream.messages", {
                 round,
                 messageCount: request.messages?.length ?? 0,
             });
-            const response = await ollama.chat({
+            const response = await runSingleModelRequest({
+                execute: () => adapter.callModel(request, () => ollama.chat({
                     ...request,
                     stream: true
+                })),
             });
 
             aiLog("debug", "ollama.stream.open", {round});
@@ -354,10 +352,7 @@ export async function runOllama(
 
                     const localToolCalls: ToolCallData[] = [];
 
-                    localToolCalls.push(...normalizeOllamaToolCalls(
-                        chunk.message.tool_calls as readonly OllamaToolCallLike[] | undefined,
-                        round,
-                    ));
+                    localToolCalls.push(...adapter.extractStreamingToolCalls(chunk.message));
 
                     const newStatus = roundStatus(round, firstRoundStatus, chunk.message.content, localToolCalls, !!chunk.message.thinking);
                     const previousStatus = streamMessage.getStatus();
@@ -377,13 +372,10 @@ export async function runOllama(
                     }
 
                     if (!(chunk.message?.thinking && streamMessage.getStatus() !== Environment.reasoningText)) {
-                        streamMessage.append(chunk.message?.content ?? "");
+                        streamMessage.append(adapter.extractTextDelta(chunk));
                     }
 
-                    calls.push(...normalizeOllamaToolCalls(
-                        chunk.message?.tool_calls as readonly OllamaToolCallLike[] | undefined,
-                        round,
-                    ));
+                    calls.push(...adapter.extractStreamingToolCalls(chunk.message));
 
                     if (chunk.done) {
                         aiLog("debug", "ollama.stream.done", {
@@ -416,7 +408,7 @@ export async function runOllama(
                     duration: aiLogDuration(runnerStartedAt),
                 });
 
-                break;
+                return {shouldContinue: false};
             }
 
             calls.splice(0, calls.length, ...dedupeToolCalls(calls));
@@ -439,7 +431,31 @@ export async function runOllama(
                 })),
             });
 
-            const toolResults = await executeToolBatch(msg.from?.id, calls, streamMessage, toolContext, toolMemory);
+            const toolResults = await executeToolBatchWithAdapter({
+                userId: msg.from?.id,
+                toolCalls: calls,
+                streamMessage,
+                toolContext: {
+                    ...toolContext,
+                    provider: AiProvider.OLLAMA,
+                    runtimeTarget: target,
+                },
+                toolMemory,
+                adapter,
+                appendTargets: [messages],
+            });
+
+            const continuation = decideToolLoopContinuation({
+                round,
+                maxRounds: MAX_TOOL_ROUNDS,
+                toolCalls: calls,
+            });
+            if (!continuation.continue && continuation.reason === "max_rounds_reached") {
+                aiLog("warn", "ollama.tool_loop.max_rounds_reached", {
+                    round,
+                    maxRounds: MAX_TOOL_ROUNDS,
+                });
+            }
 
             let successGetNoteFileResult: GetNoteFileResult | undefined = undefined;
 
@@ -471,9 +487,11 @@ export async function runOllama(
                 }).catch(logError);
             }
 
-            appendOllamaToolResults(messages, calls, toolResults);
-        }
+            return {shouldContinue: true};
+        },
+        });
     } finally {
         if (interval) clearInterval(interval);
+        await adapter.finalize().catch(() => undefined);
     }
 }

src/ai/unified-ai-runner.openai-compatible.ts

@@ -0,0 +1,419 @@
+import {Message} from "typescript-telegram-bot-api";
+import type {
+    ChatCompletionCreateParamsNonStreaming,
+    ChatCompletionCreateParamsStreaming,
+    ChatCompletionTool,
+} from "openai/resources/chat/completions";
+import {Environment} from "../common/environment.js";
+import {TelegramStreamMessage} from "./telegram-stream-message";
+import {ToolRuntimeContext} from "./tools/runtime";
+import {OpenAIChatMessage, OpenAICompatibleChatMessage} from "./openai-chat-message";
+import {createOpenAiClient} from "./ai-runtime-target";
+import {aiLog, aiLogDuration, aiLogMessageIdentity, aiLogProviderTarget, aiLogToolCall} from "../logging/ai-logger";
+import type {BoundaryValue} from "../common/boundary-types.js";
+import {
+    AsyncIterableStream,
+    buildSystemInstruction,
+    MAX_TOOL_ROUNDS,
+    OpenAiChatCompletionResponseLike,
+    OpenAiChatCompletionStreamChunkLike,
+    RuntimeConfigSnapshot,
+    safeJsonParseObject,
+    ToolCallData,
+    ToolExecutionMemory,
+} from "./unified-ai-runner.shared";
+import {mergeToolCallChunks, normalizeStreamingTextDelta} from "./provider-adapter-contract.js";
+import {buildUserMemoryPrompt} from "./tools/user-memory.js";
+import {executeToolBatchWithAdapter} from "./tool-batch-runner";
+import {decideToolLoopContinuation} from "./tool-loop-control";
+import {runToolLoopRounds} from "./tool-loop-runner";
+import {runSingleModelRequest} from "./model-call-stage";
+import {ensureToolsSelected, getOpenAICompatibleTools} from "./tool-mappers.js";
+import {MEMORY_TOOL_NAMES} from "./tools/user-memory.js";
+import {logError} from "../util/utils";
+import {DEFAULT_AI_RESPONSE_LANGUAGE} from "../common/user-ai-settings";
+import {AiDownloadedFile} from "./telegram-attachments";
+import {AiProvider} from "../model/ai-provider";
+import {getProviderAdapter} from "./provider-adapters";
+import {runToolRankStage} from "./tool-rank-stage";
+import type {AiProviderAdapter} from "./provider-adapters.js";
+import {tryToUploadFiles} from "./openai-upload-files.js";
+import {buildAssistantToolMessage, openAiResponseMessagesToChatCompletions} from "./openai-chat-completions.js";
+
+function describeOpenAiCompatibleError(error: unknown): Record<string, unknown> {
+    const err = error as {
+        message?: unknown;
+        status?: unknown;
+        code?: unknown;
+        type?: unknown;
+        error?: unknown;
+    } | undefined;
+
+    return {
+        errorSummary: typeof err?.message === "string" ? err.message : String(error),
+        httpStatus: err?.status,
+        errorCode: err?.code,
+        errorType: err?.type,
+    };
+}
+
+async function executeChatCompletionWithOptionalToolFallback<T>(params: {
+    openAi: ReturnType<typeof createOpenAiClient>;
+    request: ChatCompletionCreateParamsNonStreaming | ChatCompletionCreateParamsStreaming;
+    signal: AbortSignal;
+    stream: boolean;
+}): Promise<T> {
+    try {
+        return await params.openAi.chat.completions.create(params.request as never, {signal: params.signal}) as T;
+    } catch (error) {
+        const requestWithTools = params.request as {tools?: unknown[]};
+        if (!requestWithTools.tools || !Array.isArray(requestWithTools.tools) || requestWithTools.tools.length === 0) {
+            aiLog("error", "openai_compatible.request.failed", {
+                stream: params.stream,
+                hasTools: false,
+                error: describeOpenAiCompatibleError(error),
+            });
+            throw error;
+        }
+
+        aiLog("warn", "openai_compatible.tools.retry_without_tools", {
+            stream: params.stream,
+            error: describeOpenAiCompatibleError(error),
+        });
+
+        const retryRequest = {...params.request} as ChatCompletionCreateParamsNonStreaming | ChatCompletionCreateParamsStreaming & {tools?: unknown[]};
+        delete retryRequest.tools;
+
+        try {
+            return await params.openAi.chat.completions.create(retryRequest as never, {signal: params.signal}) as T;
+        } catch (retryError) {
+            aiLog("error", "openai_compatible.request.retry_without_tools.failed", {
+                stream: params.stream,
+                hasTools: true,
+                error: describeOpenAiCompatibleError(retryError),
+            });
+            throw retryError;
+        }
+    }
+}
+
+function makeChatCompletionAdapter(): AiProviderAdapter {
+    const baseAdapter = getProviderAdapter(AiProvider.OPENAI);
+
+    return {
+        ...baseAdapter,
+        callModel: baseAdapter.callModel.bind(baseAdapter),
+        mapMessages(messages: readonly unknown[]): unknown[] {
+            return openAiResponseMessagesToChatCompletions(messages as OpenAIChatMessage[]);
+        },
+        rankTools(config: RuntimeConfigSnapshot, options?: {forCreator?: boolean; vectorStoreIds?: string[]}): readonly BoundaryValue[] {
+            void config;
+            void options?.vectorStoreIds;
+            return getOpenAICompatibleTools(options?.forCreator) as BoundaryValue[];
+        },
+        extractTextDelta(input: unknown): string {
+            const chunk = input as OpenAiChatCompletionStreamChunkLike | undefined;
+            return chunk?.choices?.[0]?.delta?.content ?? "";
+        },
+        extractToolCalls(input: unknown): ToolCallData[] {
+            const response = input as OpenAiChatCompletionResponseLike | undefined;
+            const toolCalls = response?.choices?.[0]?.message?.tool_calls ?? [];
+
+            return toolCalls
+                .map((call, index) => ({
+                    id: typeof call?.id === "string" && call.id.trim().length > 0 ? call.id : `openai_chat_${index}`,
+                    name: typeof call?.function?.name === "string" ? call.function.name : typeof call?.name === "string" ? call.name : "",
+                    argumentsText: typeof call?.function?.arguments === "string"
+                        ? call.function.arguments
+                        : JSON.stringify(call?.function?.arguments ?? call?.arguments ?? {}),
+                }))
+                .filter(call => call.name.length > 0);
+        },
+        extractStreamingToolCalls(input: unknown): ToolCallData[] {
+            const chunk = input as OpenAiChatCompletionStreamChunkLike | undefined;
+            const toolCalls = chunk?.choices?.[0]?.delta?.tool_calls ?? [];
+
+            return toolCalls
+                .map((call, index) => ({
+                    id: typeof call?.id === "string" && call.id.trim().length > 0
+                        ? call.id
+                        : `openai_chat_${typeof call?.index === "number" ? call.index : index}`,
+                    name: typeof call?.function?.name === "string" ? call.function.name : typeof call?.name === "string" ? call.name : "",
+                    argumentsText: typeof call?.function?.arguments === "string"
+                        ? call.function.arguments
+                        : call?.function?.arguments
+                            ? JSON.stringify(call.function.arguments)
+                            : typeof call?.arguments === "string"
+                                ? call.arguments
+                                : "",
+                }))
+                .filter(call => call.id.length > 0);
+        },
+        appendToolResults(messages: unknown[], calls: ToolCallData[], results: string[]): void {
+            for (const [index, call] of calls.entries()) {
+                messages.push({
+                    role: "tool",
+                    tool_call_id: call.id,
+                    content: results[index] ?? "",
+                });
+            }
+        },
+        finalize: baseAdapter.finalize.bind(baseAdapter),
+    };
+}
+
+export async function runOpenAiCompatible(
+    msg: Message,
+    messages: OpenAIChatMessage[],
+    streamMessage: TelegramStreamMessage,
+    signal: AbortSignal,
+    stream: boolean,
+    sourceMessage: Message,
+    config: RuntimeConfigSnapshot,
+    toolContext: ToolRuntimeContext,
+    downloads: AiDownloadedFile[] = [],
+): Promise<void> {
+    void downloads;
+    const runnerStartedAt = Date.now();
+    const openAi = createOpenAiClient(config.openAiChatTarget);
+    const adapter = makeChatCompletionAdapter();
+    const systemPrompt = buildSystemInstruction(
+        config,
+        DEFAULT_AI_RESPONSE_LANGUAGE,
+        false,
+        config.openAiChatTarget.systemPromptAdditions,
+        await buildUserMemoryPrompt(msg.from?.id),
+    );
+    let conversationMessages = [...openAiResponseMessagesToChatCompletions(messages)];
+
+    if (systemPrompt.trim().length) {
+        conversationMessages.unshift({role: "system", content: systemPrompt});
+    }
+
+    const availableTools = getOpenAICompatibleTools(msg.from?.id === Environment.CREATOR_ID) as ChatCompletionTool[];
+
+    aiLog("info", "openai_compatible.run.start", {
+        stream,
+        target: aiLogProviderTarget(config.openAiChatTarget),
+        inputMessages: messages.length,
+        sourceMessage: aiLogMessageIdentity(sourceMessage),
+        hasToolInputFiles: !!toolContext.pythonInputFiles?.length,
+        backend: config.openAiBackend,
+    });
+
+    const toolMemory: ToolExecutionMemory = new Map();
+
+    try {
+        await runToolLoopRounds({
+            maxRounds: MAX_TOOL_ROUNDS,
+            onRound: async (round) => {
+                const roundStartedAt = Date.now();
+                aiLog("debug", "openai_compatible.round.start", {round, inputMessages: conversationMessages.length, stream});
+
+                const rankResult = await runToolRankStage({
+                    provider: AiProvider.OPENAI,
+                    model: config.openAiChatTarget.model,
+                    round,
+                    config,
+                    availableTools: availableTools as readonly BoundaryValue[],
+                    messages,
+                    streamMessage,
+                    signal,
+                });
+
+                const requestTools = ensureToolsSelected(
+                    availableTools,
+                    rankResult.filteredTools as ChatCompletionTool[],
+                    MEMORY_TOOL_NAMES,
+                );
+
+                if (!stream) {
+                    const request: ChatCompletionCreateParamsNonStreaming = {
+                        model: config.openAiChatTarget.model,
+                        messages: conversationMessages,
+                        tools: requestTools.length ? requestTools : undefined,
+                    };
+
+                    const response = await runSingleModelRequest({
+                        execute: () => adapter.callModel(request, () => executeChatCompletionWithOptionalToolFallback<OpenAiChatCompletionResponseLike>({
+                            openAi,
+                            request,
+                            signal,
+                            stream: false,
+                        })),
+                    }) as OpenAiChatCompletionResponseLike;
+
+                    const message = response.choices?.[0]?.message;
+                    const responseText = typeof message?.content === "string" ? message.content : "";
+                    streamMessage.append(responseText);
+                    aiLog("debug", "openai_compatible.response.received", {
+                        round,
+                        duration: aiLogDuration(roundStartedAt),
+                        textChars: responseText.length,
+                        hasToolCalls: !!message?.tool_calls?.length,
+                    });
+
+                    const calls = adapter.extractToolCalls(response);
+                    aiLog(calls.length ? "info" : "success", calls.length ? "openai_compatible.tool_calls" : "openai_compatible.run.done", {
+                        round,
+                        duration: calls.length ? aiLogDuration(roundStartedAt) : aiLogDuration(runnerStartedAt),
+                        calls: calls.map(call => ({
+                            id: call.id,
+                            name: call.name,
+                            arguments: safeJsonParseObject(call.argumentsText)
+                        })),
+                    });
+                    if (!calls.length) return {shouldContinue: false};
+
+                    const toolCalls = calls.map(call => ({
+                        id: call.id,
+                        name: call.name,
+                        argumentsText: call.argumentsText,
+                    }));
+                    const toolMessages: OpenAICompatibleChatMessage[] = [];
+                    const toolResults = await executeToolBatchWithAdapter({
+                        userId: msg.from?.id,
+                        toolCalls,
+                        streamMessage,
+                        toolContext: {
+                            ...toolContext,
+                            provider: AiProvider.OPENAI,
+                            runtimeTarget: config.openAiChatTarget,
+                        },
+                        toolMemory,
+                        adapter,
+                        appendTargets: [toolMessages],
+                    });
+
+                    const uploadFilesResult = await tryToUploadFiles(msg, toolResults);
+                    if (uploadFilesResult.found && !uploadFilesResult.uploaded && uploadFilesResult.toolIndex >= 0) {
+                        const toolMessage = toolMessages[uploadFilesResult.toolIndex];
+                        if (toolMessage && toolMessage.role === "tool") {
+                            toolMessage.content = "Error: " + uploadFilesResult.error;
+                        }
+                    }
+
+                    const continuation = decideToolLoopContinuation({
+                        round,
+                        maxRounds: MAX_TOOL_ROUNDS,
+                        toolCalls: calls,
+                    });
+                    if (!continuation.continue && continuation.reason === "max_rounds_reached") {
+                        aiLog("warn", "openai_compatible.tool_loop.max_rounds_reached", {
+                            round,
+                            maxRounds: MAX_TOOL_ROUNDS,
+                        });
+                    }
+
+                    conversationMessages = [...conversationMessages, buildAssistantToolMessage(calls, responseText), ...toolMessages];
+                    return {shouldContinue: true};
+                }
+
+                const request: ChatCompletionCreateParamsStreaming = {
+                    model: config.openAiChatTarget.model,
+                    messages: conversationMessages,
+                    stream: true,
+                    tools: requestTools.length ? requestTools : undefined,
+                };
+
+                const response = await runSingleModelRequest({
+                    execute: () => adapter.callModel(request, () => executeChatCompletionWithOptionalToolFallback<AsyncIterableStream<OpenAiChatCompletionStreamChunkLike>>({
+                        openAi,
+                        request,
+                        signal,
+                        stream: true,
+                    })),
+                }) as AsyncIterableStream<OpenAiChatCompletionStreamChunkLike>;
+
+                aiLog("debug", "openai_compatible.stream.open", {round});
+
+                let responseText = "";
+                let toolCallState: ToolCallData[] = [];
+                for await (const chunk of response) {
+                    if (signal.aborted) throw new Error("Aborted");
+
+                    const deltaText = adapter.extractTextDelta(chunk);
+                    if (deltaText) {
+                        const appendedText = normalizeStreamingTextDelta(responseText, deltaText);
+                        responseText += appendedText;
+                        streamMessage.append(appendedText);
+                    }
+
+                    const streamedCalls = adapter.extractStreamingToolCalls(chunk);
+                    if (streamedCalls.length) {
+                        toolCallState = mergeToolCallChunks(toolCallState, streamedCalls);
+                        const activeCalls = toolCallState.filter(call => call.name.length > 0);
+                        aiLog("info", "openai_compatible.stream.tool_call.added", {
+                            round,
+                            toolCalls: activeCalls.map(aiLogToolCall),
+                        });
+                        streamMessage.setStatus(Environment.getUseToolText(activeCalls));
+                        await streamMessage.flush();
+                    }
+                }
+
+                const calls = toolCallState.filter(call => call.name.length > 0);
+                aiLog(calls.length ? "info" : "success", calls.length ? "openai_compatible.tool_calls" : "openai_compatible.stream.done", {
+                    round,
+                    duration: aiLogDuration(roundStartedAt),
+                    textChars: responseText.length,
+                    calls: calls.map(call => ({
+                        id: call.id,
+                        name: call.name,
+                        arguments: safeJsonParseObject(call.argumentsText)
+                    })),
+                });
+                if (!calls.length) return {shouldContinue: false};
+
+                streamMessage.clearStatus();
+                await streamMessage.flush();
+
+                const toolMessages: OpenAICompatibleChatMessage[] = [];
+                const toolResults = await executeToolBatchWithAdapter({
+                    userId: msg.from?.id,
+                    toolCalls: calls,
+                    streamMessage,
+                    toolContext: {
+                        ...toolContext,
+                        provider: AiProvider.OPENAI,
+                        runtimeTarget: config.openAiChatTarget,
+                    },
+                    toolMemory,
+                    adapter,
+                    appendTargets: [toolMessages],
+                });
+
+                const uploadFilesResult = await tryToUploadFiles(msg, toolResults);
+                if (uploadFilesResult.found && !uploadFilesResult.uploaded && uploadFilesResult.toolIndex >= 0) {
+                    const toolMessage = toolMessages[uploadFilesResult.toolIndex];
+                    if (toolMessage && toolMessage.role === "tool") {
+                        toolMessage.content = "Error: " + uploadFilesResult.error;
+                    }
+                }
+
+                const continuation = decideToolLoopContinuation({
+                    round,
+                    maxRounds: MAX_TOOL_ROUNDS,
+                    toolCalls: calls,
+                });
+                if (!continuation.continue && continuation.reason === "max_rounds_reached") {
+                    aiLog("warn", "openai_compatible.tool_loop.max_rounds_reached", {
+                        round,
+                        maxRounds: MAX_TOOL_ROUNDS,
+                    });
+                }
+
+                conversationMessages = [...conversationMessages, buildAssistantToolMessage(calls, responseText), ...toolMessages];
+                return {shouldContinue: true};
+            },
+        });
+    } catch (error) {
+        aiLog("error", "openai_compatible.run.failed", {
+            duration: aiLogDuration(runnerStartedAt),
+            error: describeOpenAiCompatibleError(error),
+        });
+        throw error;
+    } finally {
+        await adapter.finalize().catch(logError);
+    }
+}

src/ai/unified-ai-runner.openai.ts

@@ -12,16 +12,14 @@ import type {
 } from "openai/resources/responses/responses";
 import {createOpenAiClient} from "./ai-runtime-target";
 import {aiLog, aiLogDuration, aiLogMessageIdentity, aiLogProviderTarget, aiLogToolCall} from "../logging/ai-logger";
+import {buildUserMemoryPrompt} from "./tools/user-memory.js";
 
 import {
     AsyncIterableStream,
     buildSystemInstruction,
     collectOpenAiResponseCodeInterpreterCalls,
-    collectOpenAiResponseFunctionCalls,
     collectOpenAiResponseImages,
     collectOpenAiResponseText,
-    executeToolBatch,
-    getOpenAIResponsesToolsWithImage,
     MAX_TOOL_ROUNDS,
     OPENAI_IMAGE_PARTIALS,
     openAiResponseItemCallId,
@@ -32,20 +30,21 @@ import {
     showOpenAiGeneratedImage,
     ToolCallData,
     ToolExecutionMemory,
-    errorMessage,
     allToolSchemaNames
 } from "./unified-ai-runner.shared";
-import {bot} from "../index";
-import fs from "node:fs";
-import path from "node:path";
+import {executeToolBatchWithAdapter} from "./tool-batch-runner";
+import {decideToolLoopContinuation} from "./tool-loop-control";
+import {runToolLoopRounds} from "./tool-loop-runner";
+import {runSingleModelRequest} from "./model-call-stage";
+import {ensureToolsSelected} from "./tool-mappers.js";
+import {MEMORY_TOOL_NAMES} from "./tools/user-memory.js";
 import {logError} from "../util/utils";
-import {SendFileAttachmentResult, SendFileAttachmentResultSchema} from "./tools/files";
 import {DEFAULT_AI_RESPONSE_LANGUAGE} from "../common/user-ai-settings";
 import {AiDownloadedFile} from "./telegram-attachments";
-import {ToolRanker} from "./unified-ai-runner.tool-ranker";
 import {AiProvider} from "../model/ai-provider";
-import {filterRankedTools, latestUserTextFromMessages} from "./tool-ranker-pipeline";
-import {storeToolRankAudit} from "./tool-rank-audit";
+import {getProviderAdapter} from "./provider-adapters";
+import {runToolRankStage} from "./tool-rank-stage";
+import {tryToUploadFiles} from "./openai-upload-files.js";
 
 export async function runOpenAi(
     msg: Message,
@@ -60,22 +59,22 @@ export async function runOpenAi(
     documentRag?: OpenAiDocumentRagContext,
 ): Promise<void> {
     const runnerStartedAt = Date.now();
-    let responseInput: Array<ResponseInputItem | OpenAiResponseOutputItem> = [...messages] as Array<ResponseInputItem | OpenAiResponseOutputItem>;
     const openAi = createOpenAiClient(config.openAiChatTarget);
     const ownsDocumentRag = !documentRag;
     const preparedDocumentRag = documentRag ?? await prepareOpenAiDocumentRag(openAi, downloads.filter(download => download.kind === "document"));
-    const toolRanker = new ToolRanker(config);
-    const availableTools = getOpenAIResponsesToolsWithImage(
-        config,
-        msg.from?.id === Environment.CREATOR_ID,
-        preparedDocumentRag?.vectorStoreIds ?? [],
-    );
+    const adapter = getProviderAdapter(AiProvider.OPENAI);
+    let responseInput: Array<ResponseInputItem | OpenAiResponseOutputItem> = adapter.mapMessages(messages) as unknown as Array<ResponseInputItem | OpenAiResponseOutputItem>;
+    const availableTools = adapter.rankTools(config, {
+        forCreator: msg.from?.id === Environment.CREATOR_ID,
+        vectorStoreIds: preparedDocumentRag?.vectorStoreIds ?? [],
+    });
 
     const systemPrompt = buildSystemInstruction(
         config,
         DEFAULT_AI_RESPONSE_LANGUAGE,
         false,
         config.openAiChatTarget.systemPromptAdditions,
+        await buildUserMemoryPrompt(msg.from?.id),
     );
 
     aiLog("info", "openai.run.start", {
@@ -90,46 +89,22 @@ export async function runOpenAi(
     const toolMemory: ToolExecutionMemory = new Map();
 
     try {
-        for (let round = 0; round < MAX_TOOL_ROUNDS; round++) {
+        await runToolLoopRounds({
+            maxRounds: MAX_TOOL_ROUNDS,
+            onRound: async (round) => {
             const roundStartedAt = Date.now();
             aiLog("debug", "openai.round.start", {round, inputItems: responseInput.length, stream});
-        streamMessage.setStatus(Environment.getSelectingToolsText());
-        await streamMessage.flush();
-        const toolRankStartedAt = Date.now();
-        const toolRankStartedAtIso = new Date().toISOString();
-        const rankerSelection = await toolRanker.selectTools({
+            const rankResult = await runToolRankStage({
                 provider: AiProvider.OPENAI,
-                userQuery: latestUserTextFromMessages(messages),
+                model: config.openAiChatTarget.model,
+                round,
+                config,
                 availableTools,
-                round,
+                messages,
+                streamMessage,
                 signal,
-            })
-            .catch(async error => {
-                streamMessage.clearStatus();
-                await streamMessage.flush();
-                await storeToolRankAudit({
-                    streamMessage,
-                    provider: AiProvider.OPENAI,
-                    model: config.openAiChatTarget.model,
-                    round,
-                    startedAt: toolRankStartedAt,
-                    startedAtIso: toolRankStartedAtIso,
-                    error,
             });
-                throw error;
-            });
-        streamMessage.clearStatus();
-        await streamMessage.flush();
-        await storeToolRankAudit({
-            streamMessage,
-            provider: AiProvider.OPENAI,
-            model: config.openAiChatTarget.model,
-            round,
-            startedAt: toolRankStartedAt,
-            startedAtIso: toolRankStartedAtIso,
-            selectedTools: rankerSelection.toolNames,
-        });
-        const filteredTools = filterRankedTools(availableTools, rankerSelection.toolNames);
+            const filteredTools = rankResult.filteredTools;
             const requestTools = preparedDocumentRag?.vectorStoreIds.length
                 ? (() => {
                     const tools = [...filteredTools];
@@ -140,9 +115,13 @@ export async function runOpenAi(
                             tools.unshift(fileSearchTool);
                         }
                     }
-                return tools.length ? tools : undefined;
+                    const withMemory = ensureToolsSelected(availableTools, tools, MEMORY_TOOL_NAMES);
+                    return withMemory.length ? withMemory : undefined;
                 })()
-            : (filteredTools.length ? filteredTools : undefined);
+                : (() => {
+                    const withMemory = ensureToolsSelected(availableTools, filteredTools, MEMORY_TOOL_NAMES);
+                    return withMemory.length ? withMemory : undefined;
+                })();
 
             if (!stream) {
                 const request: ResponseCreateParamsNonStreaming = {
@@ -151,7 +130,9 @@ export async function runOpenAi(
                     tools: requestTools as ResponseCreateParamsNonStreaming["tools"],
                     instructions: systemPrompt,
                 };
-            const response = await openAi.responses.create(request, {signal}) as OpenAiResponseLike;
+                const response = await runSingleModelRequest({
+                    execute: () => adapter.callModel(request, () => openAi.responses.create(request, {signal})),
+                }) as OpenAiResponseLike;
 
                 const responseText = collectOpenAiResponseText(response);
                 streamMessage.append(responseText);
@@ -188,29 +169,37 @@ export async function runOpenAi(
                     });
                 }
 
-            const calls = collectOpenAiResponseFunctionCalls(response);
+                const calls = adapter.extractToolCalls(response);
                 aiLog(calls.length ? "info" : "success", calls.length ? "openai.tool_calls" : "openai.run.done", {
                     round,
                     duration: calls.length ? aiLogDuration(roundStartedAt) : aiLogDuration(runnerStartedAt),
                     calls: calls.map(call => ({
-                    id: call.callId,
+                        id: call.id,
                         name: call.name,
                         arguments: safeJsonParseObject(call.argumentsText)
                     })),
                 });
-            if (!calls.length) return;
+                if (!calls.length) return {shouldContinue: false};
 
                 const toolCalls = calls.map(call => ({
-                id: call.callId,
+                    id: call.id,
                     name: call.name,
                     argumentsText: call.argumentsText,
                 }));
-            const toolResults = await executeToolBatch(msg.from?.id, toolCalls, streamMessage, toolContext, toolMemory);
-            const toolOutputs = calls.map((call, index) => ({
-                type: "function_call_output" as const,
-                call_id: call.callId,
-                output: toolResults[index] ?? "",
-            }));
+                const toolOutputs: Array<{type: "function_call_output"; call_id: string; output: string}> = [];
+                const toolResults = await executeToolBatchWithAdapter({
+                    userId: msg.from?.id,
+                    toolCalls,
+                    streamMessage,
+                    toolContext: {
+                        ...toolContext,
+                        provider: AiProvider.OPENAI,
+                        runtimeTarget: config.openAiChatTarget,
+                    },
+                    toolMemory,
+                    adapter,
+                    appendTargets: [toolOutputs],
+                });
 
                 const uploadFilesResult = await tryToUploadFiles(msg, toolResults);
                 if (uploadFilesResult.found) {
@@ -230,8 +219,20 @@ export async function runOpenAi(
                     }
                 }
 
+                const continuation = decideToolLoopContinuation({
+                    round,
+                    maxRounds: MAX_TOOL_ROUNDS,
+                    toolCalls: calls,
+                });
+                if (!continuation.continue && continuation.reason === "max_rounds_reached") {
+                    aiLog("warn", "openai.tool_loop.max_rounds_reached", {
+                        round,
+                        maxRounds: MAX_TOOL_ROUNDS,
+                    });
+                }
+
                 responseInput = [...responseInput, ...(response.output ?? []), ...toolOutputs];
-            continue;
+                return {shouldContinue: true};
             }
 
             let completedResponse: OpenAiResponseLike | null = null;
@@ -243,7 +244,9 @@ export async function runOpenAi(
                 parallel_tool_calls: true,
                 instructions: systemPrompt
             };
-        const response = await openAi.responses.create(request, {signal}) as AsyncIterableStream<ResponseStreamEvent>;
+            const response = await runSingleModelRequest({
+                execute: () => adapter.callModel(request, () => openAi.responses.create(request, {signal})),
+            }) as AsyncIterableStream<ResponseStreamEvent>;
 
             aiLog("debug", "openai.stream.open", {round});
 
@@ -253,7 +256,7 @@ export async function runOpenAi(
 
                 switch (event.type) {
                     case "response.output_text.delta":
-                    streamMessage.append(event.delta ?? "");
+                        streamMessage.append(adapter.extractTextDelta(event));
                         break;
                     case "response.image_generation_call.in_progress":
                         streamMessage.setStatus(Environment.startingImageGenText);
@@ -301,14 +304,11 @@ export async function runOpenAi(
                     case "response.code_interpreter_call_code.done":
                         break;
                     case "response.output_item.added":
-                    if (event.item.type === "function_call" && event.item.name) {
-                        const item = event.item as OpenAiResponseOutputItem & { id?: string };
-                        localToolCalls.push({
-                            id: openAiResponseItemCallId(item),
-                            name: item.name ?? "",
-                            argumentsText: item.arguments ?? "{}",
-                        });
-
+                        {
+                            const streamedCalls = adapter.extractStreamingToolCalls(event);
+                            if (streamedCalls.length) {
+                                localToolCalls.push(...streamedCalls);
+                            }
                             aiLog("info", "openai.stream.tool_call.added", {
                                 round,
                                 toolCalls: localToolCalls.map(aiLogToolCall)
@@ -383,29 +383,37 @@ export async function runOpenAi(
                 });
             }
 
-        const calls = collectOpenAiResponseFunctionCalls(completedResponse);
+            const calls = adapter.extractToolCalls(completedResponse);
             aiLog(calls.length ? "info" : "success", calls.length ? "openai.tool_calls" : "openai.run.done", {
                 round,
                 duration: calls.length ? aiLogDuration(roundStartedAt) : aiLogDuration(runnerStartedAt),
                 calls: calls.map(call => ({
-                id: call.callId,
+                    id: call.id,
                     name: call.name,
                     arguments: safeJsonParseObject(call.argumentsText)
                 })),
             });
-        if (!calls.length) return;
+            if (!calls.length) return {shouldContinue: false};
 
             const toolCalls = calls.map(call => ({
-            id: call.callId,
+                id: call.id,
                 name: call.name,
                 argumentsText: call.argumentsText,
             }));
-        const toolResults = await executeToolBatch(msg.from?.id, toolCalls, streamMessage, toolContext, toolMemory);
-        const toolOutputs = calls.map((call, index) => ({
-            type: "function_call_output",
-            call_id: call.callId,
-            output: toolResults[index] ?? "",
-        }));
+            const toolOutputs: Array<{type: "function_call_output"; call_id: string; output: string}> = [];
+            const toolResults = await executeToolBatchWithAdapter({
+                userId: msg.from?.id,
+                toolCalls,
+                streamMessage,
+                toolContext: {
+                    ...toolContext,
+                    provider: AiProvider.OPENAI,
+                    runtimeTarget: config.openAiChatTarget,
+                },
+                toolMemory,
+                adapter,
+                appendTargets: [toolOutputs],
+            });
 
             const uploadFilesResult = await tryToUploadFiles(msg, toolResults);
             if (uploadFilesResult.found) {
@@ -425,12 +433,27 @@ export async function runOpenAi(
                 }
             }
 
-        responseInput = [...responseInput, ...(completedResponse.output ?? []), ...toolOutputs];
+            const continuation = decideToolLoopContinuation({
+                round,
+                maxRounds: MAX_TOOL_ROUNDS,
+                toolCalls: calls,
+            });
+            if (!continuation.continue && continuation.reason === "max_rounds_reached") {
+                aiLog("warn", "openai.tool_loop.max_rounds_reached", {
+                    round,
+                    maxRounds: MAX_TOOL_ROUNDS,
+                });
             }
+
+            responseInput = [...responseInput, ...(completedResponse.output ?? []), ...toolOutputs];
+            return {shouldContinue: true};
+        },
+        });
     } finally {
         if (ownsDocumentRag) {
             await preparedDocumentRag?.cleanup().catch(logError);
         }
+        await adapter.finalize().catch(logError);
     }
 }
 
@@ -493,72 +516,6 @@ async function cleanupOpenAiDocumentRag(openAi: OpenAI, vectorStoreId: string, f
     }
 }
 
-async function tryToUploadFiles(
-    msg: Message,
-    toolResults: string[]
-): Promise<
-    | { found: false }
-    | { found: true, uploaded: true }
-    | { found: boolean, uploaded: false, error: string, toolIndex: number }
-> {
-    let sendFileAttachment: {
-        result: SendFileAttachmentResult & { success: true },
-        toolIndex: number
-    } | null = null;
-
-    let found = false;
-
-    try {
-        for (const [index, toolResult] of toolResults.entries()) {
-            const raw = JSON.parse(toolResult);
-            const res = SendFileAttachmentResultSchema.safeParse(raw);
-
-            if (res.success) {
-                found = true;
-
-                if (res.data.success) {
-                    sendFileAttachment = {result: res.data, toolIndex: index};
-                }
-            }
-        }
-
-        if (!found) {
-            return {found: false};
-        }
-
-        const attachmentRoot = Environment.FILE_TOOLS_ROOT_DIR;
-        const attachmentPath = attachmentRoot
-            ? path.join(
-                attachmentRoot,
-                String(msg.from?.id),
-                sendFileAttachment?.result?.attachment?.relativePath ?? "",
-            )
-            : "";
-
-        if (!fs.existsSync(attachmentPath)) {
-            throw new Error(`Attachment file does not exist: ${attachmentPath}`);
-        }
-
-        await bot.sendDocument({
-            chat_id: msg.chat.id,
-            reply_parameters: {
-                message_id: msg.message_id,
-            },
-            document: fs.createReadStream(attachmentPath),
-        });
-
-        return {found: true, uploaded: true};
-    } catch (e) {
-        logError(e instanceof Error ? e : String(e));
-        return {
-            found: found,
-            uploaded: false,
-            error: errorMessage(e instanceof Error ? e : String(e)),
-            toolIndex: sendFileAttachment?.toolIndex ?? -1
-        };
-    }
-}
-
 // function openAiResponseContentToText(content: string | readonly { text?: string; refusal?: string }[]): string {
 //     if (typeof content === "string") return content;
 //     if (!Array.isArray(content)) return "";

src/ai/unified-ai-runner.shared.ts

@@ -2,40 +2,39 @@ import {Message} from "typescript-telegram-bot-api";
 import * as fs from "node:fs";
 import path from "node:path";
 import type {BoundaryValue} from "../common/boundary-types";
-import {AiProvider} from "../model/ai-provider";
-import {ToolRankerFallbackPolicy} from "../common/policies";
-import {Environment} from "../common/environment";
-import {photoGenDir} from "../index";
-import {delay, logError, replyToMessage} from "../util/utils";
-import {MessageStore} from "../common/message-store";
-import type {OpenAiResponseTool} from "./tool-mappers";
-import {AiProviderName, getOpenAICodeInterpreterTool, getOpenAIResponsesTools} from "./tool-mappers";
-import {TelegramArtifactFile, TelegramStreamMessage} from "./telegram-stream-message";
-import {AiDownloadedFile} from "./telegram-attachments";
-import {getRuntimeCapabilities} from "./provider-model-runtime";
-import {StoredAttachment} from "../model/stored-attachment";
-import {AiChatMessage, ChatMessage} from "./chat-messages-types";
+import {AiProvider} from "../model/ai-provider.js";
+import {ToolRankerFallbackPolicy} from "../common/policies.js";
+import {Environment, type OpenAiBackend} from "../common/environment.js";
+import {delay, logError, replyToMessage} from "../util/utils.js";
+import {MessageStore} from "../common/message-store.js";
+import type {OpenAiResponseTool} from "./tool-mappers.js";
+import {AiProviderName, getOpenAICodeInterpreterTool, getOpenAIResponsesTools} from "./tool-mappers.js";
+import {TelegramArtifactFile, TelegramStreamMessage} from "./telegram-stream-message.js";
+import {AiDownloadedFile} from "./telegram-attachments.js";
+import {getRuntimeCapabilities} from "./provider-model-runtime.js";
+import {StoredAttachment} from "../model/stored-attachment.js";
+import {AiChatMessage, ChatMessage} from "./chat-messages-types.js";
 import {ListResponse, Ollama} from "ollama";
-import {executeToolCall, ToolRuntimeContext} from "./tools/runtime";
-import {MessageImagePart, MessagePart} from "../common/message-part";
-import {KeyedAsyncLock} from "../util/async-lock";
-import {type AiRequestQueueTarget} from "./provider-request-queue";
-import {PYTHON_INTERPRETER_TOOL_NAME, pythonInterpreterToolPrompt} from "./tools/python-interpretator";
-import {getResponseLanguageInstruction, UserAiResponseLanguage, UserAiVoiceMode} from "../common/user-ai-settings";
+import {executeToolCall, ToolRuntimeContext} from "./tools/runtime.js";
+import {MessageImagePart, MessagePart} from "../common/message-part.js";
+import {KeyedAsyncLock} from "../util/async-lock.js";
+import {type AiRequestQueueTarget} from "./provider-request-queue.js";
+import {PYTHON_INTERPRETER_TOOL_NAME, pythonInterpreterToolPrompt} from "./tools/python-interpretator.js";
+import {getResponseLanguageInstruction, UserAiResponseLanguage, UserAiVoiceMode} from "../common/user-ai-settings.js";
 import {
     isTranscribableAudioDownload,
     resolveSpeechToTextProviderForUser,
     transcribeSpeechDownloads
-} from "./speech-to-text";
+} from "./speech-to-text.js";
 import type {ChatCompletionMessageParam} from "openai/resources/chat/completions";
-import {MistralChatMessage} from "./mistral-chat-message";
-import {prepareTelegramMarkdownV2} from "../util/markdown-v2-renderer";
-import {AiRuntimeTarget, createMistralClient, resolveAiRuntimeTarget} from "./ai-runtime-target";
-import {aiLog, aiLogDuration, aiLogProviderTarget, aiLogToolCall} from "../logging/ai-logger";
-import {buildConversationSnapshot, serializeConversationSnapshot} from "./conversation-pipeline";
+import {MistralChatMessage} from "./mistral-chat-message.js";
+import {prepareTelegramMarkdownV2} from "../util/markdown-v2-renderer.js";
+import {AiRuntimeTarget, createMistralClient, resolveAiRuntimeTarget} from "./ai-runtime-target.js";
+import {aiLog, aiLogDuration, aiLogProviderTarget, aiLogToolCall} from "../logging/ai-logger.js";
+import {buildConversationSnapshot, serializeConversationSnapshot} from "./conversation-pipeline.js";
 import type {ResponseInputMessageContentList} from "openai/resources/responses/responses";
-import {persistToolResultArtifactAttachment} from "./tool-result-artifact-store";
-import {filterUserVisibleStoredAttachments} from "../common/stored-attachment-utils";
+import {persistToolResultArtifactAttachment} from "./tool-result-artifact-store.js";
+import {filterUserInputStoredAttachments} from "../common/attachment-visibility.js";
 
 export type {Message} from "typescript-telegram-bot-api";
 export type {AiRuntimeTarget} from "./ai-runtime-target";
@@ -72,9 +71,14 @@ export const MAX_OLLAMA_CONTEXT_SIZE = 262144;
 export const DEFAULT_OLLAMA_CONTEXT_SIZE = 32768;
 export const toolResourceLocks = new KeyedAsyncLock();
 
+function photoGenDir(): string {
+    return path.join(Environment.DATA_PATH, "cache", "photo", "gen");
+}
+
 export type UnifiedRunOptions = {
     provider: AiProvider;
     msg: Message;
+    requestId?: string;
     isGuestMsg?: boolean;
     text: string;
     stream?: boolean;
@@ -270,6 +274,7 @@ export type RuntimeConfigSnapshot = {
     openAiChatTarget: AiRuntimeTarget;
     openAiImageTarget: AiRuntimeTarget;
     openAiToolRankerTarget?: AiRuntimeTarget;
+    openAiBackend: OpenAiBackend;
 };
 
 export function snapshotRuntimeConfig(): RuntimeConfigSnapshot {
@@ -303,9 +308,14 @@ export function snapshotRuntimeConfig(): RuntimeConfigSnapshot {
         openAiChatTarget: resolveAiRuntimeTarget(AiProvider.OPENAI, "chat"),
         openAiImageTarget: resolveAiRuntimeTarget(AiProvider.OPENAI, "outputImages"),
         openAiToolRankerTarget: resolveAiRuntimeTarget(AiProvider.OPENAI, "toolRank"),
+        openAiBackend: Environment.OPENAI_BACKEND,
     };
 }
 
+export function isOpenAiCompatibleBackend(config: RuntimeConfigSnapshot): boolean {
+    return config.openAiBackend === "compatible";
+}
+
 export function getMessageImageParts(part: MessagePart): MessageImagePart[] {
     if (part.imageParts?.length) return part.imageParts;
     return (part.images ?? []).map(data => ({data, mimeType: "image/jpeg"}));
@@ -378,11 +388,13 @@ export function buildSystemInstruction(
     responseLanguage: UserAiResponseLanguage,
     includePythonToolPrompt: boolean,
     additions?: string | null,
+    memoryInstruction?: string | null,
 ): string {
     return [
         config.useSystemPrompt ? getResponseLanguageInstruction(responseLanguage) : null,
         config.systemPrompt && config.useSystemPrompt ? config.systemPrompt : null,
         additions?.trim() ? additions.trim() : null,
+        memoryInstruction?.trim() ? memoryInstruction.trim() : null,
         includePythonToolPrompt ? pythonInterpreterToolPrompt : null,
     ].filter(Boolean).join("\n\n");
 }
@@ -512,13 +524,13 @@ export function addMessageAttachmentKinds(msg: Message | undefined, kinds: Set<A
     if (msg.video) kinds.add("video");
 }
 
-export async function collectStoredReplyChainAttachments(msg: Message, limit: number = 1): Promise<StoredAttachment[]> {
+export async function collectStoredReplyChainAttachments(msg: Message, limit: number = 40): Promise<StoredAttachment[]> {
     const attachments: StoredAttachment[] = [];
     const seen = new Set<string>();
     let current = await MessageStore.get(msg.chat.id, msg.message_id);
 
     for (let i = 0; current && i < limit; i++) {
-        for (const attachment of filterUserVisibleStoredAttachments(current?.attachments ?? [])) {
+        for (const attachment of filterUserInputStoredAttachments(current?.attachments ?? [])) {
             const key = [
                 attachment.kind,
                 attachment.fileUniqueId || attachment.fileId,
@@ -1113,19 +1125,31 @@ export async function executeTool(
     }
 }
 
-export function toolResourceKeys(toolCall: ToolCallData): string[] {
+export function toolResourceKeys(toolCall: ToolCallData, userId?: number | undefined | null): string[] {
     const args = safeJsonParseObject(toolCall.argumentsText);
     const pathValue = typeof args.path === "string" ? args.path : undefined;
     const sourcePath = typeof args.sourcePath === "string" ? args.sourcePath : undefined;
     const targetPath = typeof args.targetPath === "string" ? args.targetPath : undefined;
+    const memoryScope = toolCall.name.endsWith("_user_info") ? "user"
+        : toolCall.name.endsWith("_system_info") ? "system"
+            : undefined;
 
     switch (toolCall.name) {
+        case "read_user_info":
+        case "read_system_info":
         case "get_datetime":
         case "web_search":
         case "get_weather":
         case "read_file":
         case "list_directory":
             return [];
+        case "add_user_info":
+        case "add_system_info":
+        case "remove_user_info":
+        case "remove_system_info":
+        case "replace_user_info":
+        case "replace_system_info":
+            return userId && memoryScope ? [`memory:${userId}:${memoryScope}`] : [];
         case "create_file":
         case "create_directory":
         case "update_file":
@@ -1158,7 +1182,7 @@ export async function executeScheduledTool(
     message: TelegramStreamMessage,
     context: ToolRuntimeContext,
 ): Promise<string> {
-    const keys = toolResourceKeys(toolCall);
+    const keys = toolResourceKeys(toolCall, userId);
     if (!keys.length) return executeTool(userId, toolCall, message, context);
     return runWithToolLocks(keys, () => executeTool(userId, toolCall, message, context));
 }
@@ -1523,7 +1547,7 @@ export function writeOpenAiGeneratedImage(sourceMessage: Message, b64: string, l
 } {
     const buffer = Buffer.from(b64, "base64");
     const fileName = `${sourceMessage.chat.id}_${sourceMessage.message_id}_${Date.now()}_${label}.png`;
-    const cachePath = path.join(photoGenDir, fileName);
+    const cachePath = path.join(photoGenDir(), fileName);
     fs.writeFileSync(cachePath, buffer);
     return {buffer, cachePath, fileName};
 }

src/ai/unified-ai-runner.tool-ranker.ts

@@ -1,20 +1,21 @@
-import {ChatCompletionMessageParam} from "openai/resources/chat/completions";
+import type {ChatCompletionCreateParamsNonStreaming, ChatCompletionMessageParam} from "openai/resources/chat/completions";
 import {ChatRequest} from "ollama";
-import {BoundaryValue} from "../common/boundary-types";
-import {ToolRankerFallbackPolicy} from "../common/policies";
-import {AiProvider} from "../model/ai-provider";
-import {createMistralClient, createOllamaClient, createOpenAiClient, sameRuntimeEndpoint} from "./ai-runtime-target";
-import {aiLog, aiLogDuration, aiLogProviderTarget} from "../logging/ai-logger";
-import {providerChatTarget, RuntimeConfigSnapshot} from "./unified-ai-runner.shared";
+import {BoundaryValue} from "../common/boundary-types.js";
+import {ToolRankerFallbackPolicy} from "../common/policies.js";
+import {AiProvider} from "../model/ai-provider.js";
+import {createMistralClient, createOllamaClient, createOpenAiClient, sameRuntimeEndpoint} from "./ai-runtime-target.js";
+import {aiLog, aiLogDuration, aiLogProviderTarget} from "../logging/ai-logger.js";
+import {providerChatTarget, RuntimeConfigSnapshot} from "./unified-ai-runner.shared.js";
 import {
     buildRankerContext,
     buildRankerTarget,
     buildToolRankerPrompt,
     filterRankedTools,
     ToolRankerSelection,
-} from "./tool-ranker-pipeline";
-import {allToolSchemaNames} from "./unified-ai-runner.shared";
-import {sanitizeToolRankerResult} from "./tool-ranker-metadata";
+} from "./tool-ranker-pipeline.js";
+import {allToolSchemaNames} from "./unified-ai-runner.shared.js";
+import {sanitizeToolRankerResult} from "./tool-ranker-metadata.js";
+import {resolveToolRankerFallbackSelection} from "./tool-ranker-fallback.js";
 
 export class ToolRanker {
     constructor(private readonly config: RuntimeConfigSnapshot) {
@@ -27,8 +28,15 @@ export class ToolRanker {
         round: number;
         signal: AbortSignal;
         messages?: readonly { role?: string; content?: string | readonly { text?: string }[] }[];
+        runRanker?: (
+            provider: AiProvider,
+            target: NonNullable<ReturnType<typeof buildRankerTarget>>,
+            prompt: string,
+            userQuery: string,
+        ) => Promise<string>;
     }): Promise<ToolRankerSelection> {
         const {availableTools, provider, round, signal, userQuery} = args;
+        const runRanker = args.runRanker ?? this.runRanker.bind(this);
         const availableNames = allToolSchemaNames(availableTools);
         const fallbackPolicy = this.config.toolRankerFallbackPolicy;
         const configuredTarget = buildRankerTarget(this.config, provider);
@@ -41,11 +49,10 @@ export class ToolRanker {
         const target = configuredTarget ?? (fallbackPolicy === ToolRankerFallbackPolicy.MAIN_MODEL ? mainModelTarget : undefined);
 
         if (!target) {
-            if (fallbackPolicy === ToolRankerFallbackPolicy.NO_TOOLS) {
-                return {toolNames: [], usedRanker: false};
-            }
-
-            return {toolNames: availableNames, usedRanker: false};
+            return resolveToolRankerFallbackSelection({
+                fallbackPolicy,
+                availableToolNames: availableNames,
+            });
         }
 
         const startedAt = Date.now();
@@ -63,7 +70,7 @@ export class ToolRanker {
 
         try {
             if (signal.aborted) throw new Error("Aborted");
-            const raw = await this.runRanker(provider, target, ranker.prompt, userQuery);
+            const raw = await runRanker(provider, target, ranker.prompt, userQuery);
             if (signal.aborted) throw new Error("Aborted");
             const selectedNames = sanitizeToolRankerResult({
                 raw,
@@ -100,13 +107,13 @@ export class ToolRanker {
                         target: aiLogProviderTarget(target),
                         fallbackTarget: aiLogProviderTarget(mainModelTarget),
                         duration: aiLogDuration(startedAt),
-                        error: failureMessage,
+                        errorSummary: failureMessage,
                     });
 
                     const fallbackRanker = buildToolRankerPrompt(
                         buildRankerContext(this.config, provider, mainModelTarget, round, userQuery, availableTools),
                     );
-                    const raw = await this.runRanker(provider, mainModelTarget, fallbackRanker.prompt, userQuery);
+                    const raw = await runRanker(provider, mainModelTarget, fallbackRanker.prompt, userQuery);
                     const selectedNames = sanitizeToolRankerResult({
                         raw,
                         availableToolNames: availableNames,
@@ -135,7 +142,7 @@ export class ToolRanker {
                         target: aiLogProviderTarget(target),
                         fallbackTarget: aiLogProviderTarget(mainModelTarget),
                         duration: aiLogDuration(startedAt),
-                        error: fallbackErrorMessage,
+                        errorSummary: fallbackErrorMessage,
                     });
 
                     failureMessage = fallbackErrorMessage;
@@ -148,17 +155,13 @@ export class ToolRanker {
                 target: aiLogProviderTarget(target),
                 fallbackPolicy,
                 duration: aiLogDuration(startedAt),
-                error: failureMessage,
+                errorSummary: failureMessage,
             });
 
-            if (fallbackPolicy === ToolRankerFallbackPolicy.NO_TOOLS) {
-                return {toolNames: [], usedRanker: false};
-            }
-
-            return {
-                toolNames: availableNames,
-                usedRanker: false,
-            };
+            return resolveToolRankerFallbackSelection({
+                fallbackPolicy,
+                availableToolNames: availableNames,
+            });
         }
     }
 
@@ -224,12 +227,19 @@ export class ToolRanker {
                     {role: "user", content: userQuery},
                 ] satisfies ChatCompletionMessageParam[];
 
-                // gpt-5 family ranker targets reject temperature=0; use the model default instead.
-                const response = await openAi.chat.completions.create({
+                // OpenAI-compatible servers often reject `response_format`, so keep JSON mode
+                // only for official OpenAI endpoints.
+                const request: ChatCompletionCreateParamsNonStreaming = {
                     model: target.model,
                     messages,
-                    response_format: {type: "json_object"},
-                });
+                };
+
+                if (!target.baseUrl) {
+                    // gpt-5 family ranker targets reject temperature=0; use the model default instead.
+                    request.response_format = {type: "json_object"};
+                }
+
+                const response = await openAi.chat.completions.create(request);
 
                 return response.choices[0]?.message?.content?.trim() ?? "";
             }

src/ai/unified-ai-runner.ts

@@ -35,6 +35,7 @@ import {persistErrorArtifactAttachment} from "./final-response-artifact-store";
 import {runUnifiedAiResponsePipeline} from "./unified-ai-response-pipeline";
 import {AiRequestStore} from "../common/ai-request-store";
 import type {StoredAiRequestStatus} from "../model/stored-ai-request";
+import {recordAiRequestFinish, recordAiRequestStart} from "../common/ai-observability.js";
 
 export type {ToolCallData} from "./unified-ai-runner.shared";
 export {snapshotModel, providerTargets, ollamaModelNames} from "./unified-ai-runner.shared";
@@ -49,6 +50,7 @@ async function executeUnifiedAiRequest(
     const requestStartedAt = Date.now();
     let preparedRequest: Awaited<ReturnType<typeof prepareUnifiedAiRequestPipeline>> | undefined;
     aiLog("info", "request.execute.start", {
+        requestId: options.requestId,
         provider: providerName(options.provider),
         stream: options.stream ?? true,
         think: options.think,
@@ -74,6 +76,7 @@ async function executeUnifiedAiRequest(
     if (preparedRequest.finishAfterTranscript) return;
 
     aiLog("debug", "request.messages.collected", {
+        requestId: options.requestId,
         provider: providerName(options.provider),
         chatMessages: preparedRequest.chatMessages.length,
         imageCount: preparedRequest.imageCount,
@@ -91,6 +94,7 @@ async function executeUnifiedAiRequest(
             controller,
         });
         aiLog("success", "request.execute.done", {
+            requestId: options.requestId,
             provider: providerName(options.provider),
             duration: aiLogDuration(requestStartedAt),
             responseChars: streamMessage.getText().length,
@@ -99,6 +103,7 @@ async function executeUnifiedAiRequest(
         return;
     } catch (e) {
         aiLog("error", "request.execute.failed", {
+            requestId: options.requestId,
             provider: providerName(options.provider),
             duration: aiLogDuration(requestStartedAt),
             error: e instanceof Error ? e : String(e),
@@ -117,6 +122,7 @@ export async function runUnifiedAi(options: UnifiedRunOptions): Promise<void> {
     const requestedAttachmentKinds = await collectRequestedAttachmentKinds(options.msg);
 
     aiLog("info", "run.start", {
+        requestId: options.requestId ?? `pending:${options.msg.chat.id}:${options.msg.message_id}`,
         provider: providerName(options.provider),
         model: snapshotModel(options.provider, config),
         message: aiLogMessageIdentity(options.msg),
@@ -133,6 +139,7 @@ export async function runUnifiedAi(options: UnifiedRunOptions): Promise<void> {
 
     if (await rejectUnsupportedAttachments(options.provider, snapshotModel(options.provider, config), options.msg, config, requestedAttachmentKinds)) {
         aiLog("warn", "run.rejected.unsupported_attachment", {
+            requestId: options.requestId ?? `pending:${options.msg.chat.id}:${options.msg.message_id}`,
             provider: providerName(options.provider),
             requestedAttachmentKinds: [...requestedAttachmentKinds],
         });
@@ -150,6 +157,7 @@ export async function runUnifiedAi(options: UnifiedRunOptions): Promise<void> {
             text: Environment.getAttachmentMissingFromCacheText(cached.missing[0].fileName),
         }).catch(logError);
         aiLog("warn", "run.rejected.missing_attachment_cache", {
+            requestId: options.requestId ?? `pending:${options.msg.chat.id}:${options.msg.message_id}`,
             missing: cached.missing.map(a => ({kind: a.kind, fileName: a.fileName, cachePath: a.cachePath})),
         });
         return;
@@ -166,6 +174,8 @@ export async function runUnifiedAi(options: UnifiedRunOptions): Promise<void> {
         provider: providerName(options.provider),
         controller
     });
+    options.requestId ??= cancel.id;
+    const requestId = options.requestId;
     const streamMessage = new TelegramStreamMessage(
         options.msg,
         cancel.id,
@@ -180,10 +190,11 @@ export async function runUnifiedAi(options: UnifiedRunOptions): Promise<void> {
     );
     cancel.onCancel = () => streamMessage.cancel(cancel.provider);
     const queueTarget = resolveAiRequestQueueTarget(options, config, requestedAttachmentKinds);
-    aiLog("debug", "run.queue.target", {target: aiLogProviderTarget(queueTarget), cancelId: cancel.id});
+    aiLog("debug", "run.queue.target", {requestId, target: aiLogProviderTarget(queueTarget), cancelId: cancel.id});
     const aiRequestStartedAt = new Date().toISOString();
+    recordAiRequestStart();
     await AiRequestStore.put({
-        requestId: cancel.id,
+        requestId,
         chatId: options.msg.chat.id,
         messageId: options.msg.message_id,
         fromId: options.msg.from?.id ?? 0,
@@ -197,7 +208,7 @@ export async function runUnifiedAi(options: UnifiedRunOptions): Promise<void> {
         const queueMessage = await streamMessage.start(Environment.waitThinkText);
         responseMessageId = queueMessage.message_id;
         await AiRequestStore.put({
-            requestId: cancel.id,
+            requestId,
             chatId: options.msg.chat.id,
             messageId: options.msg.message_id,
             responseMessageId,
@@ -207,8 +218,9 @@ export async function runUnifiedAi(options: UnifiedRunOptions): Promise<void> {
             status: "running",
             startedAt: aiRequestStartedAt,
         }).catch(logError);
-        setAiCancelMessageId(cancel.id, queueMessage.message_id);
+        setAiCancelMessageId(requestId, queueMessage.message_id);
         aiLog("info", "run.queue.enter", {
+            requestId,
             cancelId: cancel.id,
             queueMessageId: queueMessage.message_id,
             target: aiLogProviderTarget(queueTarget),
@@ -217,15 +229,16 @@ export async function runUnifiedAi(options: UnifiedRunOptions): Promise<void> {
         await aiProviderRequestQueue.enqueue(queueTarget, {
             signal: controller.signal,
             onPositionChange: async requestsBefore => {
-                aiLog("debug", "run.queue.position", {cancelId: cancel.id, requestsBefore});
+                aiLog("debug", "run.queue.position", {requestId, cancelId: cancel.id, requestsBefore});
                 streamMessage.setStatus(Environment.getAiQueueText(options.provider, requestsBefore));
                 await streamMessage.flush();
             },
             run: async (): Promise<null> => {
                 const queueWaitFinishedAt = Date.now();
-                aiLog("info", "run.queue.dequeued", {cancelId: cancel.id});
+                aiLog("info", "run.queue.dequeued", {requestId, cancelId: cancel.id});
                 const downloads = attachmentsToDownloadedFiles(cached.attachments);
                 aiLog("debug", "run.downloads.ready", {
+                    requestId,
                     count: downloads.length,
                     downloads: downloads.map(d => ({
                         kind: d.kind,
@@ -239,12 +252,13 @@ export async function runUnifiedAi(options: UnifiedRunOptions): Promise<void> {
                     await executeUnifiedAiRequest(options, config, downloads, controller, streamMessage);
                     aiRequestStatus = "succeeded";
                     aiLog("success", "run.queue.task.done", {
+                        requestId,
                         cancelId: cancel.id,
                         duration: aiLogDuration(queueWaitFinishedAt),
                     });
                 } finally {
                     cleanupDownloads(downloads);
-                    aiLog("debug", "run.downloads.cleaned", {cancelId: cancel.id, count: downloads.length});
+                    aiLog("debug", "run.downloads.cleaned", {requestId, cancelId: cancel.id, count: downloads.length});
                 }
                 return null;
             },
@@ -253,13 +267,13 @@ export async function runUnifiedAi(options: UnifiedRunOptions): Promise<void> {
         if (controller.signal.aborted || isAbortError(e instanceof Error ? e : String(e))) {
             aiRequestStatus = "aborted";
             aiRequestError = e instanceof Error ? e.message : String(e);
-            aiLog("warn", "run.aborted", {cancelId: cancel.id, duration: aiLogDuration(startedAt), error: e instanceof Error ? e : String(e)});
+            aiLog("warn", "run.aborted", {requestId, cancelId: cancel.id, duration: aiLogDuration(startedAt), error: e instanceof Error ? e : String(e)});
             streamMessage.replaceText(streamMessage.getText());
             await streamMessage.finish();
         } else {
             aiRequestStatus = "failed";
             aiRequestError = e instanceof Error ? e.message : String(e);
-            aiLog("error", "run.failed", {cancelId: cancel.id, duration: aiLogDuration(startedAt), error: e instanceof Error ? e : String(e)});
+            aiLog("error", "run.failed", {requestId, cancelId: cancel.id, duration: aiLogDuration(startedAt), error: e instanceof Error ? e : String(e)});
             const errorMessage = e instanceof Error ? e.message : String(e);
             await streamMessage.fail(e instanceof Error ? e : String(e));
             try {
@@ -279,7 +293,7 @@ export async function runUnifiedAi(options: UnifiedRunOptions): Promise<void> {
     } finally {
         clearTimeout(timeout);
         await AiRequestStore.put({
-            requestId: cancel.id,
+            requestId,
             chatId: options.msg.chat.id,
             messageId: options.msg.message_id,
             responseMessageId,
@@ -291,8 +305,10 @@ export async function runUnifiedAi(options: UnifiedRunOptions): Promise<void> {
             finishedAt: new Date().toISOString(),
             error: aiRequestError,
         }).catch(logError);
-        finishAiRequest(cancel.id);
+        recordAiRequestFinish(aiRequestStatus);
+        finishAiRequest(requestId);
         aiLog("success", "run.finished", {
+            requestId,
             cancelId: cancel.id,
             provider: providerName(options.provider),
             duration: aiLogDuration(startedAt),

src/ai/user-request-pipeline/fallback-failure.ts

@@ -0,0 +1,12 @@
+import type {PipelineFallbackDecision} from "./fallback-executor.js";
+
+export class PipelineRequestFailure extends Error {
+    constructor(public readonly decision: PipelineFallbackDecision, message: string) {
+        super(message);
+        this.name = "PipelineRequestFailure";
+    }
+}
+
+export function raisePipelineRequestFailure(decision: PipelineFallbackDecision, stageName: string): never {
+    throw new PipelineRequestFailure(decision, `Pipeline send failed at stage ${stageName} with fallback action ${decision.action}`);
+}

src/ai/user-request-pipeline/fallback-notifier-registry.ts

@@ -0,0 +1,16 @@
+import type {PipelineFallbackDecision} from "./fallback-executor.js";
+
+export function fallbackNotificationKey(requestId: string, decision: PipelineFallbackDecision): string {
+    return `${requestId}:${decision.stage}:${decision.action}`;
+}
+
+export class PipelineFallbackNotificationRegistry {
+    private readonly notifiedKeys = new Set<string>();
+
+    claim(requestId: string, decision: PipelineFallbackDecision): boolean {
+        const key = fallbackNotificationKey(requestId, decision);
+        if (this.notifiedKeys.has(key)) return false;
+        this.notifiedKeys.add(key);
+        return true;
+    }
+}

src/ai/user-request-pipeline/fallback-notifier-text.ts

@@ -0,0 +1,26 @@
+import {Localization} from "../../common/localization.js";
+import type {PipelineFallbackAction, PipelineStageName} from "./types.js";
+
+export function resolvePipelineFallbackText(
+    stage: PipelineStageName,
+    action: PipelineFallbackAction,
+    locale?: string,
+): string | undefined {
+    if (action === "continue_without_stage") return undefined;
+    if (action === "fail_request") return Localization.text("pipelineFallback.failRequest", {}, "⚠️ I could not finish this request.", locale);
+
+    switch (stage) {
+        case "speech_to_text":
+            return Localization.text("pipelineFallback.speechToText", {}, "⚠️ Speech transcription failed, so I will continue without the audio transcript.", locale);
+        case "document_rag":
+            return Localization.text("pipelineFallback.documentRag", {}, "⚠️ Document retrieval failed, so I will answer without RAG.", locale);
+        case "tool_loop":
+            return Localization.text("pipelineFallback.toolLoop", {}, "⚠️ Tool execution failed, so I will continue without that tool.", locale);
+        case "text_to_speech":
+            return Localization.text("pipelineFallback.textToSpeech", {}, "⚠️ Text-to-speech failed, so I will continue without audio output.", locale);
+        default:
+            return action === "notify_user"
+                ? Localization.text("pipelineFallback.notifyUser", {}, "⚠️ I hit a problem and need to continue with a fallback.", locale)
+                : Localization.text("pipelineFallback.generic", {}, "⚠️ I had to skip part of the request, but I can continue.", locale);
+    }
+}

src/ai/user-request-pipeline/fallback-notifier.ts

@@ -0,0 +1,43 @@
+import type {Message} from "typescript-telegram-bot-api";
+import {Localization} from "../../common/localization.js";
+import {replyToMessage, logError} from "../../util/utils.js";
+import type {PipelineFallbackDecision} from "./fallback-executor.js";
+import {PipelineFallbackNotificationRegistry} from "./fallback-notifier-registry.js";
+import {resolvePipelineFallbackText} from "./fallback-notifier-text.js";
+
+export class PipelineFallbackNotifier {
+    private readonly registry = new PipelineFallbackNotificationRegistry();
+
+    constructor(
+        private readonly sourceMessage: Message,
+        private readonly responseLanguage?: string,
+        private readonly sendFallbackMessage: (text: string) => Promise<void> = async text => {
+            await replyToMessage({
+                message: this.sourceMessage,
+                text,
+            });
+        },
+    ) {}
+
+    async notify(requestId: string, decision: PipelineFallbackDecision): Promise<{notified: boolean; text?: string}> {
+        if (!this.registry.claim(requestId, decision)) {
+            return {notified: false};
+        }
+
+        const locale = this.responseLanguage === "default"
+            ? Localization.currentLocale()
+            : Localization.normalizeLocale(this.responseLanguage) ?? Localization.currentLocale();
+        const text = resolvePipelineFallbackText(decision.stage, decision.action, locale);
+        if (!text) {
+            return {notified: false};
+        }
+
+        try {
+            await this.sendFallbackMessage(text);
+            return {notified: true, text};
+        } catch (error) {
+            logError(error instanceof Error ? error : String(error));
+            return {notified: false, text};
+        }
+    }
+}

src/ai/user-request-pipeline/fallback-target-details.ts

@@ -0,0 +1,15 @@
+import {AiProvider} from "../../model/ai-provider.js";
+import type {RuntimeConfigSnapshot} from "../unified-ai-runner.shared.js";
+import {aiLogProviderTarget} from "../../logging/ai-logger.js";
+import {buildRankerTarget} from "../tool-ranker-pipeline.js";
+import {providerChatTarget} from "../unified-ai-runner.shared.js";
+
+export function buildToolRankFallbackTargetDetails(provider: AiProvider, config: RuntimeConfigSnapshot) {
+    const sourceTarget = buildRankerTarget(config, provider);
+    const alternateTarget = providerChatTarget(provider, config);
+
+    return {
+        sourceTarget: aiLogProviderTarget(sourceTarget),
+        alternateTarget: aiLogProviderTarget(alternateTarget),
+    };
+}

src/ai/user-request-pipeline/pipeline.ts

@@ -1,5 +1,6 @@
 import {DEFAULT_PIPELINE_FALLBACK_POLICIES, USER_REQUEST_PIPELINE_STAGES} from "./blueprint.js";
 import {decidePipelineFallback, type PipelineFallbackDecision} from "./fallback-executor.js";
+import {raisePipelineRequestFailure} from "./fallback-failure.js";
 import type {
     PipelineAuditEvent,
     PipelineFallbackPolicy,
@@ -66,7 +67,7 @@ export class UserRequestPipeline {
                     },
                 }));
                 if (decision.shouldFailRequest) {
-                    throw new Error(`Required pipeline stage is not registered: ${stageName}`);
+                    raisePipelineRequestFailure(decision, stageName);
                 }
                 continue;
             }
@@ -112,7 +113,7 @@ export class UserRequestPipeline {
                     error: error instanceof Error ? error.message : String(error),
                 }));
                 if (decision.shouldFailRequest) {
-                    throw error;
+                    raisePipelineRequestFailure(decision, stageName);
                 }
             }
         }

src/commands/ai-audit.ts

@@ -0,0 +1,33 @@
+import {Message} from "typescript-telegram-bot-api";
+import {Command} from "../base/command.js";
+import {Requirements} from "../base/requirements.js";
+import {Requirement} from "../base/requirement.js";
+import {Environment} from "../common/environment.js";
+import {buildAiAuditReport, replyWithTrimmedText, resolveAuditTarget} from "./ai-observability.js";
+import {logError, sendErrorPlaceholder} from "../util/utils.js";
+
+export class AIAudit extends Command {
+    command = ["aiaudit", "audit"];
+    argsMode = "optional" as const;
+
+    requirements = Requirements.Build(Requirement.BOT_ADMIN);
+
+    title = Environment.commandTitles.aiAudit;
+    description = Environment.commandDescriptions.aiAudit;
+
+    async execute(msg: Message, match?: RegExpExecArray | null): Promise<void> {
+        try {
+            const target = resolveAuditTarget(msg, match?.[3] ?? null);
+            if (!target) {
+                await replyWithTrimmedText(msg, "Usage: reply to a message or pass messageId, or chatId messageId.");
+                return;
+            }
+
+            const text = await buildAiAuditReport(target);
+            await replyWithTrimmedText(msg, text);
+        } catch (error) {
+            logError(error instanceof Error ? error : String(error));
+            await sendErrorPlaceholder(msg).catch(logError);
+        }
+    }
+}

src/commands/ai-metrics.ts

@@ -0,0 +1,27 @@
+import {Message} from "typescript-telegram-bot-api";
+import {Command} from "../base/command.js";
+import {Requirements} from "../base/requirements.js";
+import {Requirement} from "../base/requirement.js";
+import {Environment} from "../common/environment.js";
+import {buildAiMetricsReport, replyWithTrimmedText} from "./ai-observability.js";
+import {logError, sendErrorPlaceholder} from "../util/utils.js";
+
+export class AIMetrics extends Command {
+    command = ["aimetrics", "metrics"];
+    argsMode = "none" as const;
+
+    requirements = Requirements.Build(Requirement.BOT_ADMIN);
+
+    title = Environment.commandTitles.aiMetrics;
+    description = Environment.commandDescriptions.aiMetrics;
+
+    async execute(msg: Message): Promise<void> {
+        try {
+            const text = await buildAiMetricsReport();
+            await replyWithTrimmedText(msg, text);
+        } catch (error) {
+            logError(error instanceof Error ? error : String(error));
+            await sendErrorPlaceholder(msg).catch(logError);
+        }
+    }
+}

src/commands/ai-observability.ts

@@ -0,0 +1,155 @@
+import {Message} from "typescript-telegram-bot-api";
+import {DatabaseManager} from "../db/database-manager.js";
+import type {AttachmentDbRow} from "../db/db-types.js";
+import {replyToMessage} from "../util/utils.js";
+import {snapshotAiObservability} from "../common/ai-observability.js";
+
+export type AuditTarget = {
+    chatId: number;
+    messageId: number;
+};
+
+export function resolveAuditTarget(msg: Message, argsText?: string | null): AuditTarget | null {
+    if (msg.reply_to_message) {
+        return {
+            chatId: msg.chat.id,
+            messageId: msg.reply_to_message.message_id,
+        };
+    }
+
+    const args = argsText?.trim().split(/\s+/).filter(Boolean) ?? [];
+    if (!args.length) return null;
+
+    if (args.length === 1) {
+        const messageId = Number(args[0]);
+        if (!Number.isFinite(messageId)) return null;
+        return {
+            chatId: msg.chat.id,
+            messageId,
+        };
+    }
+
+    const chatId = Number(args[0]);
+    const messageId = Number(args[1]);
+    if (!Number.isFinite(chatId) || !Number.isFinite(messageId)) return null;
+
+    return {chatId, messageId};
+}
+
+function formatSize(bytes: number | null | undefined): string {
+    if (!Number.isFinite(bytes ?? NaN)) return "n/a";
+    const value = Number(bytes);
+    if (value >= 1024 * 1024) return `${(value / (1024 * 1024)).toFixed(1)} MB`;
+    if (value >= 1024) return `${(value / 1024).toFixed(1)} KB`;
+    return `${value} B`;
+}
+
+function clip(value: string | null | undefined, max = 120): string {
+    const text = (value ?? "").trim();
+    if (!text) return "n/a";
+    return text.length <= max ? text : `${text.slice(0, max)}…`;
+}
+
+function formatAttachmentLine(index: number, attachment: AttachmentDbRow): string {
+    return [
+        `${index + 1}.`,
+        attachment.direction,
+        attachment.kind,
+        attachment.fileName,
+        `size=${formatSize(attachment.sizeBytes)}`,
+        attachment.artifactKind ? `artifact=${attachment.artifactKind}` : null,
+    ].filter(Boolean).join(" ");
+}
+
+export async function buildAiAuditReport(target: AuditTarget): Promise<string> {
+    const [request, audits, artifacts, attachments] = await Promise.all([
+        DatabaseManager.getAiRequestByMessage(target.chatId, target.messageId),
+        DatabaseManager.getRequestAuditsByMessage(target.chatId, target.messageId),
+        DatabaseManager.getArtifactsByMessage(target.chatId, target.messageId),
+        DatabaseManager.getAttachmentsByMessage(target.chatId, target.messageId),
+    ]);
+
+    const lines: string[] = [
+        "AI observability audit",
+        `chatId: ${target.chatId}`,
+        `messageId: ${target.messageId}`,
+        "",
+        "AI request:",
+    ];
+
+    if (request) {
+        lines.push(
+            `  requestId: ${request.requestId}`,
+            `  provider: ${request.provider}`,
+            `  model: ${request.model}`,
+            `  status: ${request.status}`,
+            `  startedAt: ${request.startedAt}`,
+            `  finishedAt: ${request.finishedAt ?? "n/a"}`,
+            `  error: ${clip(request.error, 240)}`,
+        );
+    } else {
+        lines.push("  not found");
+    }
+
+    lines.push("", `Pipeline audits: ${audits.length}`);
+    audits.slice(0, 12).forEach((audit, index) => {
+        lines.push(
+            `  ${index + 1}. ${audit.stage} ${audit.status}` +
+            `${audit.durationMs !== null ? ` ${audit.durationMs}ms` : ""}` +
+            `${audit.provider ? ` provider=${audit.provider}` : ""}` +
+            `${audit.model ? ` model=${audit.model}` : ""}` +
+            `${audit.error ? ` error=${clip(audit.error, 120)}` : ""}`,
+        );
+    });
+    if (audits.length > 12) {
+        lines.push(`  … and ${audits.length - 12} more`);
+    }
+
+    lines.push("", `Artifacts: ${artifacts.length}`);
+    artifacts.slice(0, 12).forEach((artifact, index) => {
+        lines.push(
+            `  ${index + 1}. ${artifact.kind} stage=${artifact.stage}` +
+            `${artifact.attachmentId ? ` attachmentId=${artifact.attachmentId}` : ""}` +
+            `${artifact.createdAt ? ` createdAt=${artifact.createdAt}` : ""}`,
+        );
+    });
+    if (artifacts.length > 12) {
+        lines.push(`  … and ${artifacts.length - 12} more`);
+    }
+
+    lines.push("", `Attachments: ${attachments.length}`);
+    attachments.slice(0, 12).forEach((attachment, index) => {
+        lines.push(`  ${formatAttachmentLine(index, attachment)}`);
+    });
+    if (attachments.length > 12) {
+        lines.push(`  … and ${attachments.length - 12} more`);
+    }
+
+    return lines.join("\n");
+}
+
+export async function buildAiMetricsReport(): Promise<string> {
+    const snapshot = snapshotAiObservability();
+    const [aiRequests, attachments, artifacts, requestAudits] = await Promise.all([
+        DatabaseManager.getAllAiRequests(),
+        DatabaseManager.getAllAttachments(),
+        DatabaseManager.getAllArtifacts(),
+        DatabaseManager.getAllRequestAudits(),
+    ]);
+
+    return [
+        "AI observability metrics",
+        `requests: total=${snapshot.requests.total} succeeded=${snapshot.requests.succeeded} failed=${snapshot.requests.failed} aborted=${snapshot.requests.aborted}`,
+        `fallbacks: total=${snapshot.fallbacks.total} ignore=${snapshot.fallbacks.ignore} notify_user=${snapshot.fallbacks.notifyUser} continue_without_stage=${snapshot.fallbacks.continueWithoutStage} use_alternate_target=${snapshot.fallbacks.useAlternateTarget} fail_request=${snapshot.fallbacks.failRequest}`,
+        `tool calls: ${snapshot.toolCalls}`,
+        `RAG runs: ${snapshot.ragRuns}`,
+        `TTS runs: total=${snapshot.ttsRuns.total} succeeded=${snapshot.ttsRuns.succeeded} failed=${snapshot.ttsRuns.failed} skipped=${snapshot.ttsRuns.skipped}`,
+        `db rows: ai_requests=${aiRequests.length} attachments=${attachments.length} artifacts=${artifacts.length} request_audit=${requestAudits.length}`,
+    ].join("\n");
+}
+
+export async function replyWithTrimmedText(msg: Message, text: string): Promise<void> {
+    const maxLength = 3800;
+    const nextText = text.length <= maxLength ? text : `${text.slice(0, maxLength)}\n… (trimmed)`;
+    await replyToMessage({message: msg, text: nextText});
+}

src/commands/ai-requests.ts

@@ -0,0 +1,51 @@
+import {Message} from "typescript-telegram-bot-api";
+import {Command} from "../base/command.js";
+import {Requirements} from "../base/requirements.js";
+import {Requirement} from "../base/requirement.js";
+import {Environment} from "../common/environment.js";
+import {DatabaseManager} from "../db/database-manager.js";
+import {logError, sendErrorPlaceholder} from "../util/utils.js";
+import {replyWithTrimmedText} from "./ai-observability.js";
+
+function formatRequestLine(index: number, request: Awaited<ReturnType<typeof DatabaseManager.getAllAiRequests>>[number]): string {
+    return [
+        `${index + 1}.`,
+        `requestId=${request.requestId}`,
+        `chatId=${request.chatId}`,
+        `messageId=${request.messageId}`,
+        request.responseMessageId ? `responseMessageId=${request.responseMessageId}` : null,
+        `provider=${request.provider}`,
+        `model=${request.model}`,
+        `status=${request.status}`,
+        `startedAt=${request.startedAt}`,
+        request.finishedAt ? `finishedAt=${request.finishedAt}` : null,
+        request.error ? `error=${request.error}` : null,
+    ].filter(Boolean).join(" ");
+}
+
+export class AIRequests extends Command {
+    command = ["airequests"];
+    argsMode = "none" as const;
+
+    requirements = Requirements.Build(Requirement.BOT_ADMIN);
+
+    title = Environment.commandTitles.aiRequests;
+    description = Environment.commandDescriptions.aiRequests;
+
+    async execute(msg: Message): Promise<void> {
+        try {
+            const requests = (await DatabaseManager.getAllAiRequests()).slice(-10).reverse();
+            const lines = [
+                "Recent AI requests",
+                `count: ${requests.length}`,
+                "",
+                ...requests.map((request, index) => formatRequestLine(index, request)),
+            ];
+
+            await replyWithTrimmedText(msg, lines.join("\n"));
+        } catch (error) {
+            logError(error instanceof Error ? error : String(error));
+            await sendErrorPlaceholder(msg).catch(logError);
+        }
+    }
+}

src/common/ai-observability.ts

@@ -0,0 +1,123 @@
+import type {PipelineFallbackAction} from "../ai/user-request-pipeline";
+import type {StoredAiRequestStatus} from "../model/stored-ai-request.js";
+
+type CounterSnapshot = {
+    total: number;
+    succeeded: number;
+    failed: number;
+    aborted: number;
+};
+
+export type AiObservabilitySnapshot = {
+    requests: CounterSnapshot;
+    fallbacks: {
+        total: number;
+        ignore: number;
+        notifyUser: number;
+        continueWithoutStage: number;
+        useAlternateTarget: number;
+        failRequest: number;
+    };
+    toolCalls: number;
+    ragRuns: number;
+    ttsRuns: {
+        total: number;
+        succeeded: number;
+        failed: number;
+        skipped: number;
+    };
+};
+
+const requestCounters = {
+    total: 0,
+    succeeded: 0,
+    failed: 0,
+    aborted: 0,
+};
+
+const fallbackCounters = {
+    total: 0,
+    ignore: 0,
+    notifyUser: 0,
+    continueWithoutStage: 0,
+    useAlternateTarget: 0,
+    failRequest: 0,
+};
+
+const ttsCounters = {
+    total: 0,
+    succeeded: 0,
+    failed: 0,
+    skipped: 0,
+};
+
+let toolCalls = 0;
+let ragRuns = 0;
+
+function incrementFallback(action: PipelineFallbackAction): void {
+    fallbackCounters.total += 1;
+    switch (action) {
+        case "ignore":
+            fallbackCounters.ignore += 1;
+            break;
+        case "notify_user":
+            fallbackCounters.notifyUser += 1;
+            break;
+        case "continue_without_stage":
+            fallbackCounters.continueWithoutStage += 1;
+            break;
+        case "use_alternate_target":
+            fallbackCounters.useAlternateTarget += 1;
+            break;
+        case "fail_request":
+            fallbackCounters.failRequest += 1;
+            break;
+    }
+}
+
+export function recordAiRequestStart(): void {
+    requestCounters.total += 1;
+}
+
+export function recordAiRequestFinish(status: StoredAiRequestStatus): void {
+    switch (status) {
+        case "succeeded":
+            requestCounters.succeeded += 1;
+            break;
+        case "failed":
+            requestCounters.failed += 1;
+            break;
+        case "aborted":
+            requestCounters.aborted += 1;
+            break;
+        case "running":
+            break;
+    }
+}
+
+export function recordPipelineFallback(action: PipelineFallbackAction): void {
+    incrementFallback(action);
+}
+
+export function recordToolCall(): void {
+    toolCalls += 1;
+}
+
+export function recordRagRun(): void {
+    ragRuns += 1;
+}
+
+export function recordTtsRun(status: "succeeded" | "failed" | "skipped"): void {
+    ttsCounters.total += 1;
+    ttsCounters[status] += 1;
+}
+
+export function snapshotAiObservability(): AiObservabilitySnapshot {
+    return {
+        requests: {...requestCounters},
+        fallbacks: {...fallbackCounters},
+        toolCalls,
+        ragRuns,
+        ttsRuns: {...ttsCounters},
+    };
+}

src/common/attachment-visibility.ts

@@ -0,0 +1,9 @@
+import type {StoredAttachment} from "../model/stored-attachment";
+
+export function filterUserVisibleStoredAttachments(attachments: StoredAttachment[]): StoredAttachment[] {
+    return attachments.filter(attachment => attachment.scope !== "internal_artifact");
+}
+
+export function filterUserInputStoredAttachments(attachments: StoredAttachment[]): StoredAttachment[] {
+    return attachments.filter(attachment => attachment.scope === "user_input" || attachment.scope === undefined);
+}

src/common/environment.ts

@@ -3,18 +3,28 @@ import os from "node:os";
 import path from "node:path";
 import {parse as parseDotEnv} from "dotenv";
 import {z} from "zod";
-import {appLogger} from "../logging/logger";
+import {appLogger} from "../logging/logger.js";
 import type {BoundaryValue, ErrorLike} from "./boundary-types";
 
-import {saveData} from "../db/database";
-import {Answers} from "../model/answers";
-import {ifTrue} from "../util/utils";
-import {AiProvider} from "../model/ai-provider";
-import {ImageHandleFallbackPolicy, ImageHandlePolicy, RateLimitFallbackPolicy} from "./policies";
-import {ToolRankerFallbackPolicy} from "./policies";
-import type {ToolCallData} from "../ai/unified-ai-runner";
-import {PYTHON_INTERPRETER_TOOL_NAME} from "../ai/tools/python-interpretator";
-import {Localization, type LocalizationParams} from "./localization";
+import {Answers} from "../model/answers.js";
+import {AiProvider} from "../model/ai-provider.js";
+import {ImageHandleFallbackPolicy, ImageHandlePolicy, RateLimitFallbackPolicy} from "./policies.js";
+import {ToolRankerFallbackPolicy} from "./policies.js";
+import type {ToolCallData} from "../ai/unified-ai-runner.js";
+import {PYTHON_INTERPRETER_TOOL_NAME} from "../ai/tools/python-interpretator.js";
+import {Localization, type LocalizationParams} from "./localization.js";
+
+export const OpenAiBackendModes = {
+    OFFICIAL: "official",
+    COMPATIBLE: "compatible",
+} as const;
+
+export type OpenAiBackend = typeof OpenAiBackendModes[keyof typeof OpenAiBackendModes];
+
+function parseBooleanLike(value: string): boolean {
+    const normalized = value.trim().toLowerCase();
+    return ["true", "t", "y", "1"].includes(normalized);
+}
 
 type EnvRecord = Record<string, string>;
 type StringEnumLike = Record<string, string>;
@@ -53,7 +63,7 @@ function booleanWithDefaultSchema(defaultValue: boolean) {
                 return defaultValue;
             }
 
-            return ifTrue(normalized);
+            return parseBooleanLike(normalized);
         }, z.boolean())
         .default(defaultValue)
         .catch(defaultValue);
@@ -62,7 +72,7 @@ function booleanWithDefaultSchema(defaultValue: boolean) {
 const optionalBooleanSchema = z
     .preprocess(value => {
         const normalized = normalizeString(value as BoundaryValue);
-        return normalized === undefined ? undefined : ifTrue(normalized);
+        return normalized === undefined ? undefined : parseBooleanLike(normalized);
     }, z.boolean().optional())
     .optional()
     .catch(undefined);
@@ -211,6 +221,10 @@ const RuntimeEnvSchema = z.object({
     SEND_TIME_TOOK: optionalBooleanSchema,
 
     ENABLE_PYTHON_INTERPRETER: optionalBooleanSchema,
+    DISABLE_LOCAL_TOOLS: optionalBooleanSchema,
+    LOCAL_TOOL_ALLOWLIST: optionalStringSchema,
+    LOCAL_TOOL_DENYLIST: optionalStringSchema,
+    MCP_SERVERS: optionalStringSchema,
 
     OLLAMA_API_KEY: optionalStringSchema,
     OLLAMA_ADDRESS: optionalStringSchema,
@@ -238,6 +252,10 @@ const RuntimeEnvSchema = z.object({
 
     OPENAI_BASE_URL: optionalStringSchema,
     OPENAI_API_KEY: optionalStringSchema,
+    OPENAI_BACKEND: enumWithDefaultSchema(
+        OpenAiBackendModes,
+        OpenAiBackendModes.OFFICIAL,
+    ),
     OPENAI_MODEL: stringWithDefaultSchema("gpt-4.1-nano"),
     OPENAI_IMAGE_MODEL: stringWithDefaultSchema("gpt-image-1-mini"),
     OPENAI_TRANSCRIPTION_MODEL: stringWithDefaultSchema("gpt-4o-mini-transcribe"),
@@ -305,6 +323,10 @@ export class Environment {
     static SEND_TIME_TOOK: boolean = false;
 
     static ENABLE_PYTHON_INTERPRETER: boolean = false;
+    static DISABLE_LOCAL_TOOLS: boolean = false;
+    static LOCAL_TOOL_ALLOWLIST?: string;
+    static LOCAL_TOOL_DENYLIST?: string;
+    static MCP_SERVERS?: string;
 
     static OLLAMA_API_KEY?: string;
     static OLLAMA_ADDRESS?: string;
@@ -332,6 +354,7 @@ export class Environment {
 
     static OPENAI_BASE_URL?: string;
     static OPENAI_API_KEY?: string;
+    static OPENAI_BACKEND: OpenAiBackend = OpenAiBackendModes.OFFICIAL;
     static OPENAI_MODEL: string = "";
     static OPENAI_IMAGE_MODEL: string = "";
     static OPENAI_TRANSCRIPTION_MODEL: string = "";
@@ -820,6 +843,34 @@ export class Environment {
         return this.text("noTextToSynthesizeText", "No text to synthesize.");
     }
 
+    static get pipelineFallbackGenericText() {
+        return this.text("pipelineFallbackGenericText", "⚠️ I had to skip part of the request, but I can continue.");
+    }
+
+    static get pipelineFallbackNotifyText() {
+        return this.text("pipelineFallbackNotifyText", "⚠️ I hit a problem and need to continue with a fallback.");
+    }
+
+    static get pipelineFallbackFailText() {
+        return this.text("pipelineFallbackFailText", "⚠️ I could not finish this request.");
+    }
+
+    static get pipelineFallbackRagText() {
+        return this.text("pipelineFallbackRagText", "⚠️ Document retrieval failed, so I will answer without RAG.");
+    }
+
+    static get pipelineFallbackSpeechToTextText() {
+        return this.text("pipelineFallbackSpeechToTextText", "⚠️ Speech transcription failed, so I will continue without the audio transcript.");
+    }
+
+    static get pipelineFallbackTextToSpeechText() {
+        return this.text("pipelineFallbackTextToSpeechText", "⚠️ Text-to-speech failed, so I will continue without audio output.");
+    }
+
+    static get pipelineFallbackToolText() {
+        return this.text("pipelineFallbackToolText", "⚠️ Tool execution failed, so I will continue without that tool.");
+    }
+
     static get mistralTtsNoAudioDataText() {
         return this.text("mistralTtsNoAudioDataText", "Mistral TTS did not return audioData.");
     }
@@ -960,6 +1011,9 @@ export class Environment {
         choice: "/choice a, b, ..., c",
         coin: "/coin",
         debug: "/debug",
+        aiRequests: "/aiRequests",
+        aiAudit: "/aiAudit [reply|messageId|chatId messageId]",
+        aiMetrics: "/aiMetrics",
         dice: "/dice",
         distort: "/distort [amp] [wavelength]",
         help: "/help",
@@ -1010,6 +1064,9 @@ export class Environment {
             choice: this.text("commandDescriptions.choice", "Choose a random value"),
             coin: this.text("commandDescriptions.coin", "Heads or tails"),
             debug: this.text("commandDescriptions.debug", "Returns msg (or reply) as json"),
+            aiRequests: this.text("commandDescriptions.aiRequests", "Show recent AI requests"),
+            aiAudit: this.text("commandDescriptions.aiAudit", "Inspect AI request audit and artifacts"),
+            aiMetrics: this.text("commandDescriptions.aiMetrics", "Show AI observability counters"),
             dice: this.text("commandDescriptions.dice", "Sends random or specific dice"),
             distort: this.text("commandDescriptions.distort", "Distortion of picture"),
             help: this.text("commandDescriptions.help", "Show list of commands"),
@@ -1805,6 +1862,10 @@ export class Environment {
         Environment.SEND_TIME_TOOK = env.SEND_TIME_TOOK ?? false;
 
         Environment.ENABLE_PYTHON_INTERPRETER = env.ENABLE_PYTHON_INTERPRETER ?? false;
+        Environment.DISABLE_LOCAL_TOOLS = env.DISABLE_LOCAL_TOOLS ?? false;
+        Environment.LOCAL_TOOL_ALLOWLIST = env.LOCAL_TOOL_ALLOWLIST;
+        Environment.LOCAL_TOOL_DENYLIST = env.LOCAL_TOOL_DENYLIST;
+        Environment.MCP_SERVERS = env.MCP_SERVERS;
 
         Environment.OLLAMA_API_KEY = env.OLLAMA_API_KEY;
         Environment.OLLAMA_ADDRESS = env.OLLAMA_ADDRESS;
@@ -1832,6 +1893,7 @@ export class Environment {
 
         Environment.OPENAI_BASE_URL = env.OPENAI_BASE_URL;
         Environment.OPENAI_API_KEY = env.OPENAI_API_KEY;
+        Environment.OPENAI_BACKEND = env.OPENAI_BACKEND;
         Environment.OPENAI_MODEL = env.OPENAI_MODEL;
         Environment.OPENAI_IMAGE_MODEL = env.OPENAI_IMAGE_MODEL;
         Environment.OPENAI_TRANSCRIPTION_MODEL = env.OPENAI_TRANSCRIPTION_MODEL;
@@ -1939,6 +2001,7 @@ export class Environment {
 
         if (!has) {
             this.ADMIN_IDS.add(id);
+            const {saveData} = await import("../db/database.js");
             await saveData();
         }
 
@@ -1950,6 +2013,7 @@ export class Environment {
 
         if (has) {
             this.ADMIN_IDS.delete(id);
+            const {saveData} = await import("../db/database.js");
             await saveData();
         }
 
@@ -1966,6 +2030,7 @@ export class Environment {
         }
 
         this.MUTED_IDS.add(id);
+        const {saveData} = await import("../db/database.js");
         await saveData();
         return true;
     }
@@ -1976,6 +2041,7 @@ export class Environment {
         }
 
         this.MUTED_IDS.delete(id);
+        const {saveData} = await import("../db/database.js");
         await saveData();
         return true;
     }
@@ -2028,6 +2094,10 @@ export class Environment {
         this.OPENAI_API_KEY = newAIApiKey;
     }
 
+    static setOpenAIBackend(newBackend: OpenAiBackend): void {
+        this.OPENAI_BACKEND = newBackend;
+    }
+
     static setOpenAIModel(newModel: string): void {
         this.OPENAI_MODEL = newModel;
     }

src/common/localization.ts

@@ -1,7 +1,7 @@
 import {AsyncLocalStorage} from "node:async_hooks";
 import fs from "node:fs";
 import path from "node:path";
-import {appLogger} from "../logging/logger";
+import {appLogger} from "../logging/logger.js";
 
 const logger = appLogger.child("localization");
 

src/common/message-part.ts

@@ -20,6 +20,9 @@ export type MessagePart = {
     audios?: string[];
     audioParts?: MessageAudioPart[];
     documents?: string[];
+    documentNames?: string[];
     videos?: string[];
     videoNotes?: string[];
+    videoNames?: string[];
+    videoNoteNames?: string[];
 }

src/common/stored-attachment-utils.ts

@@ -1,6 +1,7 @@
 import path from "node:path";
 import {Environment} from "./environment";
 import {StoredAttachment} from "../model/stored-attachment";
+export {filterUserVisibleStoredAttachments} from "./attachment-visibility";
 
 export function photoCachePathForUniqueId(uniqueId: string): string {
     return path.join(Environment.DATA_PATH, "cache", "photo", `${uniqueId}.jpg`);
@@ -44,7 +45,3 @@ export function uniqueStoredAttachments(attachments: StoredAttachment[]): Stored
 
     return result;
 }
-
-export function filterUserVisibleStoredAttachments(attachments: StoredAttachment[]): StoredAttachment[] {
-    return attachments.filter(attachment => attachment.scope !== "internal_artifact");
-}

src/db/database-manager.ts

@@ -2055,7 +2055,24 @@ export class DatabaseManager {
     }
 
     private static async migrateLegacyNormalizedTables(): Promise<void> {
-        const messages = await DatabaseManager.getAllMessages();
+        // Do not call getAllMessages() here: it awaits DatabaseManager.ready, which
+        // is the promise currently waiting on ensureSchema(). That creates a
+        // self-deadlock during startup migrations.
+        const messages = await DatabaseManager.query<MessageDbRow>(`
+            SELECT
+                "id",
+                "chatId",
+                "replyToMessageId",
+                "fromId",
+                "text",
+                "quoteText",
+                "date",
+                "deletedByBotAt",
+                "attachments",
+                "pipelineAudit"
+            FROM "messages"
+            ORDER BY "chatId", "id"
+        `);
         const attachments = messages.flatMap(message => DatabaseManager.attachmentRowsFromMessageRow(message));
         const artifacts = messages.flatMap(message => DatabaseManager.artifactRowsFromMessageRow(message));
         const requestAudits = messages.flatMap(message => DatabaseManager.requestAuditRowsFromMessageRow(message));

src/db/database.ts

@@ -1,9 +1,9 @@
 import * as fs from "fs";
-import {Environment} from "../common/environment";
-import {logError} from "../util/utils";
-import {Answers} from "../model/answers";
+import {Environment} from "../common/environment.js";
+import {logError} from "../util/utils.js";
+import {Answers} from "../model/answers.js";
 import path from "node:path";
-import {KeyedAsyncLock} from "../util/async-lock";
+import {KeyedAsyncLock} from "../util/async-lock.js";
 
 type DataJsonFile = {
     admins: number[]

src/index.ts

@@ -1,9 +1,9 @@
 import "dotenv/config";
-import {appLogger} from "./logging/logger";
-import {Environment} from "./common/environment";
+import {appLogger} from "./logging/logger.js";
+import {Environment} from "./common/environment.js";
 import {BotCommand, TelegramBot, User} from "typescript-telegram-bot-api";
-import {Command} from "./base/command";
-import type {LogDetails} from "./logging/logger";
+import {Command} from "./base/command.js";
+import type {LogDetails} from "./logging/logger.js";
 import {
     initSystemSpecs,
     logError,
@@ -13,68 +13,73 @@ import {
     processInlineQuery,
     processMyChatMember,
     processNewMessage
-} from "./util/utils";
-import {Ae} from "./commands/ae";
-import {Help} from "./commands/help";
-import {Ignore} from "./commands/ignore";
-import {Unignore} from "./commands/unignore";
-import {Ping} from "./commands/ping";
-import {RandomString} from "./commands/random-string";
-import {SystemInfo} from "./commands/system-info";
-import {Test} from "./commands/test";
-import {readData, retrieveAnswers} from "./db/database";
-import {Uptime} from "./commands/uptime";
-import {WhatBetter} from "./commands/what-better";
-import {When} from "./commands/when";
-import {RandomInt} from "./commands/random-int";
-import {Ban} from "./commands/ban";
-import {Quote} from "./commands/quote";
-import {OllamaSearch} from "./commands/ollama-search";
-import {Id} from "./commands/id";
-import {AdminsAdd} from "./commands/admins-add";
-import {AdminsRemove} from "./commands/admins-remove";
-import {Shutdown} from "./commands/shutdown";
-import {Leave} from "./commands/leave";
-import {OllamaChat} from "./commands/ollama-chat";
-import {Start} from "./commands/start";
-import {Choice} from "./commands/choice";
-import {Coin} from "./commands/coin";
-import {Qr} from "./commands/qr";
-import {Distort} from "./commands/distort";
-import {Dice} from "./commands/dice";
-import {Unban} from "./commands/unban";
-import {Title} from "./commands/title";
-import {MessageDao} from "./db/message-dao";
-import {DatabaseManager} from "./db/database-manager";
-import {UserDao} from "./db/user-dao";
-import {UserStore} from "./common/user-store";
-import {CallbackCommand} from "./base/callback-command";
-import {AiCancel} from "./callback_commands/ai-cancel";
-import {AiRegenerate} from "./callback_commands/ai-regenerate";
-import {MistralChat} from "./commands/mistral-chat";
-import {Transliteration} from "./commands/transliteration";
-import {OllamaListModels} from "./commands/ollama-list-models";
-import {OllamaGetModel} from "./commands/ollama-get-model";
-import {OllamaSetModel} from "./commands/ollama-set-model";
-import {MistralGetModel} from "./commands/mistral-get-model";
-import {MistralSetModel} from "./commands/mistral-set-model";
-import {MistralListModels} from "./commands/mistral-list-models";
-import {Debug} from "./commands/debug";
+} from "./util/utils.js";
+import {Ae} from "./commands/ae.js";
+import {Help} from "./commands/help.js";
+import {Ignore} from "./commands/ignore.js";
+import {Unignore} from "./commands/unignore.js";
+import {Ping} from "./commands/ping.js";
+import {RandomString} from "./commands/random-string.js";
+import {SystemInfo} from "./commands/system-info.js";
+import {Test} from "./commands/test.js";
+import {readData, retrieveAnswers} from "./db/database.js";
+import {Uptime} from "./commands/uptime.js";
+import {WhatBetter} from "./commands/what-better.js";
+import {When} from "./commands/when.js";
+import {RandomInt} from "./commands/random-int.js";
+import {Ban} from "./commands/ban.js";
+import {Quote} from "./commands/quote.js";
+import {OllamaSearch} from "./commands/ollama-search.js";
+import {Id} from "./commands/id.js";
+import {AdminsAdd} from "./commands/admins-add.js";
+import {AdminsRemove} from "./commands/admins-remove.js";
+import {Shutdown} from "./commands/shutdown.js";
+import {Leave} from "./commands/leave.js";
+import {OllamaChat} from "./commands/ollama-chat.js";
+import {Start} from "./commands/start.js";
+import {Choice} from "./commands/choice.js";
+import {Coin} from "./commands/coin.js";
+import {Qr} from "./commands/qr.js";
+import {Distort} from "./commands/distort.js";
+import {Dice} from "./commands/dice.js";
+import {Unban} from "./commands/unban.js";
+import {Title} from "./commands/title.js";
+import {MessageDao} from "./db/message-dao.js";
+import {DatabaseManager} from "./db/database-manager.js";
+import {UserDao} from "./db/user-dao.js";
+import {UserStore} from "./common/user-store.js";
+import {CallbackCommand} from "./base/callback-command.js";
+import {AiCancel} from "./callback_commands/ai-cancel.js";
+import {AiRegenerate} from "./callback_commands/ai-regenerate.js";
+import {MistralChat} from "./commands/mistral-chat.js";
+import {Transliteration} from "./commands/transliteration.js";
+import {OllamaListModels} from "./commands/ollama-list-models.js";
+import {OllamaGetModel} from "./commands/ollama-get-model.js";
+import {OllamaSetModel} from "./commands/ollama-set-model.js";
+import {MistralGetModel} from "./commands/mistral-get-model.js";
+import {MistralSetModel} from "./commands/mistral-set-model.js";
+import {MistralListModels} from "./commands/mistral-list-models.js";
+import {Debug} from "./commands/debug.js";
 import fs from "node:fs";
 import path from "node:path";
-import {OpenAIChat} from "./commands/openai-chat";
-import {OpenAIListModels} from "./commands/openai-list-models";
-import {OpenAIGetModel} from "./commands/openai-get-model";
-import {OpenAISetModel} from "./commands/openai-set-model";
-import {Info} from "./commands/info";
-import {AdminsList} from "./commands/admins-list";
-import {ExportDb} from "./commands/export-db";
-import {ImportDb} from "./commands/import-db";
-import {Settings} from "./commands/settings";
-import {UserSettingsCallback} from "./callback_commands/user-settings";
-import {TextToSpeech} from "./commands/text-to-speech";
-import {SpeechToText} from "./commands/speech-to-text";
-import {cleanupInternalArtifactCache} from "./ai/internal-artifact-store";
+import {OpenAIChat} from "./commands/openai-chat.js";
+import {OpenAIListModels} from "./commands/openai-list-models.js";
+import {OpenAIGetModel} from "./commands/openai-get-model.js";
+import {OpenAISetModel} from "./commands/openai-set-model.js";
+import {Info} from "./commands/info.js";
+import {AdminsList} from "./commands/admins-list.js";
+import {ExportDb} from "./commands/export-db.js";
+import {ImportDb} from "./commands/import-db.js";
+import {Settings} from "./commands/settings.js";
+import {UserSettingsCallback} from "./callback_commands/user-settings.js";
+import {TextToSpeech} from "./commands/text-to-speech.js";
+import {SpeechToText} from "./commands/speech-to-text.js";
+import {cleanupInternalArtifactCache} from "./ai/internal-artifact-store.js";
+import {AIAudit} from "./commands/ai-audit.js";
+import {AIMetrics} from "./commands/ai-metrics.js";
+import {AIRequests} from "./commands/ai-requests.js";
+import {cleanupStaleRagProviderState} from "./ai/rag-retention.js";
+import {initializeMcpTools, shutdownMcpTools} from "./ai/mcp/mcp-registry.js";
 
 process.setUncaughtExceptionCaptureCallback(logError);
 
@@ -119,6 +124,9 @@ export const commands: Command[] = [
     new Settings(),
     new TextToSpeech(),
     new SpeechToText(),
+    new AIRequests(),
+    new AIAudit(),
+    new AIMetrics(),
 
     new AdminsAdd(),
     new AdminsRemove(),
@@ -186,6 +194,7 @@ export const filesDir = path.join(Environment.DATA_PATH, "files");
 export const NOTES_HEADER = "## Notes\n";
 export const notesDir = path.join(Environment.DATA_PATH, "notes");
 export const notesRootFile = path.join(notesDir, "index.md");
+export const memoryDir = path.join(Environment.DATA_PATH, "memory");
 
 const logger = appLogger.child("main");
 
@@ -227,6 +236,11 @@ export async function shutdown(signal: NodeJS.Signals | "manual") {
         await bot.stopPolling();
     } catch (error) {
         logError(error instanceof Error ? error : String(error));
+    } finally {
+        try {
+            await shutdownMcpTools();
+        } catch (error) {
+            logError(error instanceof Error ? error : String(error));
         } finally {
             try {
                 await DatabaseManager.close();
@@ -235,6 +249,7 @@ export async function shutdown(signal: NodeJS.Signals | "manual") {
             }
             process.exit(0);
         }
+    }
 }
 
 async function main() {
@@ -248,7 +263,7 @@ async function main() {
     });
 
     await measureStartupStep("environment.load", () => Environment.load());
-    const dirsToCheck = [cacheDir, photoDir, photoGenDir, documentDir, audioDir, videoDir, videoNotesDir, videoTempDir, notesDir, filesDir];
+    const dirsToCheck = [cacheDir, photoDir, photoGenDir, documentDir, audioDir, videoDir, videoNotesDir, videoTempDir, notesDir, memoryDir, filesDir];
     await measureStartupStep("prepare_directories", () => {
         const created: string[] = [];
         for (const dir of dirsToCheck) {
@@ -272,6 +287,23 @@ async function main() {
     }, () => ({notesRootFilePath}));
 
     await measureStartupStep("cleanup_internal_artifacts", () => cleanupInternalArtifactCache(), () => ({retentionDays: 14}));
+    await measureStartupStep("cleanup_stale_rag_provider_state", () => cleanupStaleRagProviderState(), () => ({retentionDays: 14}));
+    await measureStartupStep("mcp.initialize", () => initializeMcpTools());
+    await measureStartupStep("observability.snapshot", async () => {
+        const [aiRequests, attachments, artifacts, requestAudits] = await Promise.all([
+            DatabaseManager.getAllAiRequests(),
+            DatabaseManager.getAllAttachments(),
+            DatabaseManager.getAllArtifacts(),
+            DatabaseManager.getAllRequestAudits(),
+        ]);
+
+        return {
+            aiRequests: aiRequests.length,
+            attachments: attachments.length,
+            artifacts: artifacts.length,
+            requestAudits: requestAudits.length,
+        };
+    }, () => ({tables: ["ai_requests", "attachments", "artifacts", "request_audit"]}));
 
     const cmds = await measureStartupStep("build_commands", () => commands.filter(cmd => {
         return cmd.title && cmd.title.startsWith("/") && cmd.title.split(" ").length === 1 && cmd.description;

src/logging/ai-logger.ts

@@ -1,5 +1,5 @@
 import {Message} from "typescript-telegram-bot-api";
-import {createLogger, formatDuration, LogDetails, LogLevel} from "./logger";
+import {createLogger, formatDuration, LogDetails, LogLevel} from "./logger.js";
 
 export type AiRunnerLogLevel = LogLevel;
 export type AiRunnerLogDetails = LogDetails;

src/model/ai-capability-info.ts

@@ -1,4 +1,4 @@
-import {AiProvider} from "./ai-provider";
+import {AiProvider} from "./ai-provider.js";
 
 export type AiEndpointInfo = {
     provider?: AiProvider;

src/model/ai-model-capabilities.ts

@@ -1,4 +1,4 @@
-import {AiCapabilityInfo} from "./ai-capability-info";
+import {AiCapabilityInfo} from "./ai-capability-info.js";
 
 export class AiModelCapabilities {
     chat: AiCapabilityInfo | undefined;

src/util/utils.ts

@@ -1,7 +1,7 @@
 import * as si from "systeminformation";
-import {appLogger} from "../logging/logger";
-import {Command} from "../base/command";
-import {CallbackCommand} from "../base/callback-command";
+import {appLogger} from "../logging/logger.js";
+import {Command} from "../base/command.js";
+import {CallbackCommand} from "../base/callback-command.js";
 import {
     CallbackQuery,
     ChatMember,
@@ -15,39 +15,40 @@ import {
     TelegramBot,
     User
 } from "typescript-telegram-bot-api";
-import {Environment} from "../common/environment";
-import {TelegramError} from "typescript-telegram-bot-api/dist/errors";
-import {bot, botUser, callbackCommands, commands, messageDao, photoDir} from "../index";
+import {Environment} from "../common/environment.js";
+import {TelegramError} from "typescript-telegram-bot-api/dist/errors.js";
+import {bot, botUser, callbackCommands, commands, messageDao, photoDir} from "../index.js";
 import os from "os";
 import axios from "axios";
-import {MessageAudioPart, MessageImagePart, MessagePart} from "../common/message-part";
-import {StoredMessage} from "../model/stored-message";
+import {MessageAudioPart, MessageImagePart, MessagePart} from "../common/message-part.js";
+import {StoredMessage} from "../model/stored-message.js";
 import sharp from "sharp";
-import {UserStore} from "../common/user-store";
+import {UserStore} from "../common/user-store.js";
 import fs from "node:fs";
 import path from "node:path";
-import {MessageStore} from "../common/message-store";
-import {SystemInfo} from "../commands/system-info";
-import {PrefixResponse} from "../commands/prefix-response";
-import {ChatCommand} from "../base/chat-command";
-import {AiProvider} from "../model/ai-provider";
-import {SendOptions} from "../model/send-options";
-import {EditOptions} from "../model/edit-options";
-import {StoredUser} from "../model/stored-user";
-import {StoredAttachment} from "../model/stored-attachment";
-import {AiDownloadedFile} from "../ai/telegram-attachments";
-import {runUnifiedAi} from "../ai/unified-ai-runner";
-import {enqueueTelegramApiCall} from "./telegram-api-queue";
-import {AsyncSemaphore, KeyedAsyncLock} from "./async-lock";
-import {resolveEffectiveAiProviderForUser, resolveInterfaceLocaleForUser} from "../common/user-ai-settings";
-import {Localization} from "../common/localization";
-import {createOllamaClient, resolveAiRuntimeTarget} from "../ai/ai-runtime-target";
-import {RandomUtils} from "./random-utils";
-import {HtmlUtils} from "./html-utils";
-import {ShellCommandResult, ShellCommandRunner} from "./shell-command-runner";
-import type {BoundaryValue, ErrorLike} from "../common/boundary-types";
-import {createStoredImageAttachment, photoCachePathForUniqueId, uniqueStoredAttachments} from "../common/stored-attachment-utils";
-import {runTelegramMessageAttachmentPipeline} from "../ai/user-request-pipeline";
+import {MessageStore} from "../common/message-store.js";
+import {filterUserInputStoredAttachments} from "../common/attachment-visibility.js";
+import {SystemInfo} from "../commands/system-info.js";
+import {PrefixResponse} from "../commands/prefix-response.js";
+import {ChatCommand} from "../base/chat-command.js";
+import {AiProvider} from "../model/ai-provider.js";
+import {SendOptions} from "../model/send-options.js";
+import {EditOptions} from "../model/edit-options.js";
+import {StoredUser} from "../model/stored-user.js";
+import {StoredAttachment} from "../model/stored-attachment.js";
+import {AiDownloadedFile} from "../ai/telegram-attachments.js";
+import {runUnifiedAi} from "../ai/unified-ai-runner.js";
+import {enqueueTelegramApiCall} from "./telegram-api-queue.js";
+import {AsyncSemaphore, KeyedAsyncLock} from "./async-lock.js";
+import {resolveEffectiveAiProviderForUser, resolveInterfaceLocaleForUser} from "../common/user-ai-settings.js";
+import {Localization} from "../common/localization.js";
+import {createOllamaClient, resolveAiRuntimeTarget} from "../ai/ai-runtime-target.js";
+import {RandomUtils} from "./random-utils.js";
+import {HtmlUtils} from "./html-utils.js";
+import {ShellCommandResult, ShellCommandRunner} from "./shell-command-runner.js";
+import type {BoundaryValue, ErrorLike} from "../common/boundary-types.js";
+import {createStoredImageAttachment, photoCachePathForUniqueId, uniqueStoredAttachments} from "../common/stored-attachment-utils.js";
+import {runTelegramMessageAttachmentPipeline} from "../ai/user-request-pipeline/index.js";
 
 const imageProcessingSemaphore = new AsyncSemaphore(2);
 const fileWriteLocks = new KeyedAsyncLock();
@@ -1487,12 +1488,13 @@ export async function collectReplyChainText(options: ReplyChainOptions): Promise
             const cleanText = cutPrefix ? cutPrefixes(rawText) : rawText;
             const imageNames = await loadImagesIfExists(msg);
             const messageDownloads = includeDownloads ? downloads : [];
-            const storedImageAttachments = isStoredMessage(msg)
-                ? (msg.attachments ?? []).filter(attachment => attachment.kind === "image" && fs.existsSync(attachment.cachePath))
+            const storedAttachments = isStoredMessage(msg)
+                ? filterUserInputStoredAttachments(msg.attachments ?? []).filter(attachment => fs.existsSync(attachment.cachePath))
                 : [];
+            const storedImageAttachments = storedAttachments.filter(attachment => attachment.kind === "image");
 
             if (!cleanText && !quoteText && textRequired) return;
-            if (!cleanText && !quoteText && !imageNames?.length && !storedImageAttachments.length && !messageDownloads.length) return;
+            if (!cleanText && !quoteText && !imageNames?.length && !storedAttachments.length && !messageDownloads.length) return;
 
             const fromId = isStoredMessage(msg) ? msg.fromId : msg.from?.id;
             const user = await UserStore.get(isStoredMessage(msg) ? msg.fromId : msg.from?.id ?? -1);
@@ -1527,11 +1529,19 @@ export async function collectReplyChainText(options: ReplyChainOptions): Promise
             });
             const imageParts = [...photoImageParts, ...cachedImageParts];
 
+            const storedDocumentAttachments = storedAttachments.filter(attachment => attachment.kind === "document");
+            const storedVideoAttachments = storedAttachments.filter(attachment => attachment.kind === "video");
+            const storedVideoNoteAttachments = storedAttachments.filter(attachment => attachment.kind === "video-note");
+            const storedAudioAttachments = storedAttachments.filter(attachment => attachment.kind === "audio");
+
             const audios: string[] = [];
             const audioParts: MessageAudioPart[] = [];
             const documents: string[] = [];
+            const documentNames: string[] = [];
             const videos: string[] = [];
+            const videoNames: string[] = [];
             const videoNotes: string[] = [];
+            const videoNoteNames: string[] = [];
 
             if (messageDownloads.length) {
                 messageDownloads
@@ -1544,21 +1554,51 @@ export async function collectReplyChainText(options: ReplyChainOptions): Promise
 
                 messageDownloads
                     .filter(d => d.kind === "document")
-                    .forEach(d => documents.push(d.buffer.toString("base64")));
+                    .forEach(d => {
+                        documents.push(d.buffer.toString("base64"));
+                        documentNames.push(d.fileName);
+                    });
 
                 messageDownloads
                     .filter(d => d.kind === "video")
-                    .forEach(v => videos.push(v.buffer.toString("base64")));
+                    .forEach(v => {
+                        videos.push(v.buffer.toString("base64"));
+                        videoNames.push(v.fileName);
+                    });
 
                 messageDownloads
                     .filter(d => d.kind === "video-note")
                     .forEach(v => {
                         const data = v.buffer.toString("base64");
                         videoNotes.push(data);
+                        videoNoteNames.push(v.fileName);
                         audioParts.push({data, mimeType: mimeTypeFromAudioDownload(v)});
                     });
             }
 
+            storedAudioAttachments.forEach(attachment => {
+                const data = Buffer.from(fs.readFileSync(attachment.cachePath)).toString("base64");
+                audios.push(data);
+                audioParts.push({data, mimeType: attachment.mimeType || "audio/ogg"});
+            });
+
+            storedDocumentAttachments.forEach(attachment => {
+                documents.push(Buffer.from(fs.readFileSync(attachment.cachePath)).toString("base64"));
+                documentNames.push(attachment.fileName);
+            });
+
+            storedVideoAttachments.forEach(attachment => {
+                videos.push(Buffer.from(fs.readFileSync(attachment.cachePath)).toString("base64"));
+                videoNames.push(attachment.fileName);
+            });
+
+            storedVideoNoteAttachments.forEach(attachment => {
+                const data = Buffer.from(fs.readFileSync(attachment.cachePath)).toString("base64");
+                videoNotes.push(data);
+                videoNoteNames.push(attachment.fileName);
+                audioParts.push({data, mimeType: attachment.mimeType || "video/mp4"});
+            });
+
             const content = [
                 quoteText ? `[citation]:\n${quoteText}\n\n[message]:\n` : "",
                 cleanText ?? ""
@@ -1576,8 +1616,11 @@ export async function collectReplyChainText(options: ReplyChainOptions): Promise
                 audios: audios.length ? audios : undefined,
                 audioParts: audioParts.length ? audioParts : undefined,
                 documents: documents.length ? documents : undefined,
+                documentNames: documentNames.length ? documentNames : undefined,
                 videos: videos.length ? videos : undefined,
+                videoNames: videoNames.length ? videoNames : undefined,
                 videoNotes: videoNotes.length ? videoNotes : undefined,
+                videoNoteNames: videoNoteNames.length ? videoNoteNames : undefined,
             });
         }
     };

test/ai-observability.test.mjs

@@ -0,0 +1,24 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const observability = await import("../dist/common/ai-observability.js");
+
+test("ai observability snapshot counts recorded events", () => {
+    const before = observability.snapshotAiObservability();
+
+    observability.recordAiRequestStart();
+    observability.recordAiRequestFinish("succeeded");
+    observability.recordPipelineFallback("notify_user");
+    observability.recordToolCall();
+    observability.recordRagRun();
+    observability.recordTtsRun("skipped");
+
+    const after = observability.snapshotAiObservability();
+
+    assert.equal(after.requests.total, before.requests.total + 1);
+    assert.equal(after.requests.succeeded, before.requests.succeeded + 1);
+    assert.equal(after.fallbacks.notifyUser, before.fallbacks.notifyUser + 1);
+    assert.equal(after.toolCalls, before.toolCalls + 1);
+    assert.equal(after.ragRuns, before.ragRuns + 1);
+    assert.equal(after.ttsRuns.skipped, before.ttsRuns.skipped + 1);
+});

test/model-call-stage.test.mjs

@@ -0,0 +1,17 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const {runSingleModelRequest} = await import("../dist/ai/model-call-stage.js");
+
+test("single model request wrapper executes exactly once", async () => {
+    let calls = 0;
+    const result = await runSingleModelRequest({
+        async execute() {
+            calls += 1;
+            return "ok";
+        },
+    });
+
+    assert.equal(result, "ok");
+    assert.equal(calls, 1);
+});

test/openai-backend.test.mjs

@@ -0,0 +1,14 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const {Environment} = await import("../dist/common/environment.js");
+
+test("openai backend defaults to official", () => {
+    assert.equal(Environment.OPENAI_BACKEND, "official");
+});
+
+test("openai backend setter updates runtime config", () => {
+    Environment.setOpenAIBackend("compatible");
+    assert.equal(Environment.OPENAI_BACKEND, "compatible");
+    Environment.setOpenAIBackend("official");
+});

test/openai-compatible-integration.test.mjs

@@ -0,0 +1,43 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import {OpenAI} from "openai";
+
+const {extractOpenAiChatToolCalls} = await import("../dist/ai/provider-adapter-contract.js");
+
+const baseURL = process.env.OPENAI_COMPATIBLE_TEST_BASE_URL;
+const model = process.env.OPENAI_COMPATIBLE_TEST_MODEL;
+const apiKey = process.env.OPENAI_COMPATIBLE_TEST_API_KEY ?? process.env.OPENAI_API_KEY ?? "test";
+
+test("openai-compatible chat.completions tool loop works on a real server", {skip: !baseURL || !model}, async () => {
+    const client = new OpenAI({baseURL, apiKey});
+
+    const response = await client.chat.completions.create({
+        model,
+        temperature: 0,
+        messages: [
+            {role: "system", content: "You must call the ping tool exactly once. Do not answer in plain text."},
+            {role: "user", content: "ping"},
+        ],
+        tools: [{
+            type: "function",
+            function: {
+                name: "ping",
+                description: "Return a ping token.",
+                parameters: {
+                    type: "object",
+                    properties: {},
+                    additionalProperties: false,
+                },
+            },
+        }],
+        tool_choice: {
+            type: "function",
+            function: {name: "ping"},
+        },
+    });
+
+    const calls = extractOpenAiChatToolCalls(response);
+
+    assert.equal(calls.length, 1);
+    assert.equal(calls[0].name, "ping");
+});

test/pipeline-fallback-notifier.test.mjs

@@ -0,0 +1,33 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const {PipelineFallbackNotificationRegistry} = await import("../dist/ai/user-request-pipeline/fallback-notifier-registry.js");
+const {resolvePipelineFallbackText} = await import("../dist/ai/user-request-pipeline/fallback-notifier-text.js");
+
+test("pipeline fallback text maps notify_user to a user-facing message", () => {
+    assert.match(resolvePipelineFallbackText("document_rag", "notify_user"), /RAG/i);
+    assert.match(resolvePipelineFallbackText("speech_to_text", "notify_user"), /transcription/i);
+    assert.match(resolvePipelineFallbackText("tool_loop", "notify_user"), /tool/i);
+});
+
+test("pipeline fallback text is localized when locale is provided", () => {
+    assert.match(resolvePipelineFallbackText("document_rag", "notify_user", "ru"), /RAG|документ/i);
+    assert.match(resolvePipelineFallbackText("text_to_speech", "notify_user", "ua"), /аудіо|мовлення/i);
+});
+
+test("pipeline fallback text stays silent for continue_without_stage", () => {
+    assert.equal(resolvePipelineFallbackText("document_rag", "continue_without_stage"), undefined);
+    assert.equal(resolvePipelineFallbackText("tool_loop", "continue_without_stage"), undefined);
+});
+
+test("pipeline fallback notification registry deduplicates one request-stage-action", () => {
+    const registry = new PipelineFallbackNotificationRegistry();
+    const decision = {
+        stage: "tool_loop",
+        action: "notify_user",
+    };
+
+    assert.equal(registry.claim("request-1", decision), true);
+    assert.equal(registry.claim("request-1", decision), false);
+    assert.equal(registry.claim("request-2", decision), true);
+});

test/pipeline-integration.test.mjs

@@ -0,0 +1,398 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const {UserRequestPipeline} = await import("../dist/ai/user-request-pipeline/pipeline.js");
+const {PIPELINE_ATTACHMENT_LIMIT_BYTES} = await import("../dist/ai/user-request-pipeline/types.js");
+
+class FakeTelegramStreamMessage {
+    constructor() {
+        this.status = "";
+        this.text = "";
+        this.toolExecutions = [];
+        this.outputAttachments = [];
+        this.internalAttachments = [];
+        this.pipelineAudits = [];
+        this.finished = false;
+        this.failed = false;
+    }
+
+    setStatus(status) {
+        this.status = status;
+    }
+
+    clearStatus() {
+        this.status = "";
+    }
+
+    append(delta) {
+        this.text += delta;
+    }
+
+    replaceText(text) {
+        this.text = text;
+    }
+
+    getText() {
+        return this.text;
+    }
+
+    recordToolExecution(record) {
+        this.toolExecutions.push(record);
+    }
+
+    getToolExecutions() {
+        return [...this.toolExecutions];
+    }
+
+    recordOutputAttachment(record) {
+        this.outputAttachments.push(record);
+    }
+
+    getOutputAttachments() {
+        return [...this.outputAttachments];
+    }
+
+    async storeInternalAttachment(attachment) {
+        this.internalAttachments.push(attachment);
+    }
+
+    async storePipelineAudit(events) {
+        this.pipelineAudits.push(...events);
+    }
+
+    async finish() {
+        this.finished = true;
+    }
+
+    async fail() {
+        this.failed = true;
+    }
+}
+
+class FakeProviderAdapter {
+    constructor() {
+        this.calls = [];
+    }
+
+    async callModel(request, execute) {
+        this.calls.push(request);
+        return await execute();
+    }
+
+    appendToolResults(messages, calls, results) {
+        for (const [index, call] of calls.entries()) {
+            messages.push({
+                role: "tool",
+                name: call.name,
+                content: results[index] ?? "",
+            });
+        }
+    }
+}
+
+class FakeMemoryStore {
+    constructor() {
+        this.rows = [];
+    }
+
+    persist(state) {
+        this.rows.push({
+            requestId: state.requestId,
+            audit: [...state.audit],
+            artifacts: [...state.artifacts],
+            outputAttachments: [...state.outputAttachments],
+        });
+    }
+}
+
+function createBaseState() {
+    return {
+        requestId: "integration-request-1",
+        chatId: 10,
+        messageId: 20,
+        fromId: 30,
+        receivedAt: new Date().toISOString(),
+        text: "process my attachments",
+        settings: {
+            provider: "OLLAMA",
+            responseLanguage: "en",
+            voiceMode: "execute",
+            imageOutputMode: "photo",
+        },
+        inputAttachments: [],
+        outputAttachments: [],
+        artifacts: [],
+        toolRankDecisions: [],
+        audit: [],
+    };
+}
+
+function artifact(kind, stage, extra = {}) {
+    return {
+        kind,
+        stage,
+        createdAt: "2026-05-18T00:00:00.000Z",
+        ...extra,
+    };
+}
+
+function outputAttachment(fileName, kind = "file") {
+    return {
+        direction: "output",
+        kind,
+        fileId: `${fileName}-file-id`,
+        fileName,
+        sizeBytes: 1024,
+        cachePath: `/tmp/${fileName}`,
+    };
+}
+
+test("integration pipeline rejects oversized attachment before later stages", async () => {
+    const stream = new FakeTelegramStreamMessage();
+    const state = createBaseState();
+    state.inputAttachments.push({
+        direction: "input",
+        kind: "document",
+        fileId: "doc-oversized",
+        fileName: "big.pdf",
+        sizeBytes: PIPELINE_ATTACHMENT_LIMIT_BYTES + 1,
+        cachePath: "/tmp/big.pdf",
+    });
+
+    const pipeline = new UserRequestPipeline({
+        stages: [{
+            name: "input_size_gate",
+            async run() {
+                stream.setStatus("Checking size");
+                const tooLarge = state.inputAttachments.some(attachment => attachment.sizeBytes > PIPELINE_ATTACHMENT_LIMIT_BYTES);
+                stream.clearStatus();
+
+                return {
+                    stage: "input_size_gate",
+                    status: tooLarge ? "fallback" : "succeeded",
+                    fallbackAction: tooLarge ? "notify_user" : undefined,
+                };
+            },
+        }],
+        stageNames: ["input_size_gate"],
+    });
+
+    await pipeline.run(state, new AbortController().signal);
+
+    assert.equal(state.audit.at(-1)?.status, "fallback");
+    assert.equal(state.audit.at(-1)?.details?.fallbackAction, "notify_user");
+    assert.equal(stream.status, "");
+});
+
+test("integration pipeline carries artifacts through fake document, voice, tool and tts stages", async () => {
+    const stream = new FakeTelegramStreamMessage();
+    const adapter = new FakeProviderAdapter();
+    const store = new FakeMemoryStore();
+    const state = createBaseState();
+    state.inputAttachments.push(
+        {
+            direction: "input",
+            kind: "document",
+            fileId: "doc-1",
+            fileName: "contract.pdf",
+            sizeBytes: 1024,
+            cachePath: "/tmp/contract.pdf",
+        },
+        {
+            direction: "input",
+            kind: "audio",
+            fileId: "audio-1",
+            fileName: "voice.ogg",
+            sizeBytes: 2048,
+            cachePath: "/tmp/voice.ogg",
+        },
+    );
+
+    const pipeline = new UserRequestPipeline({
+        stages: [
+            {
+                name: "input_size_gate",
+                async run() {
+                    return {
+                        stage: "input_size_gate",
+                        status: "succeeded",
+                    };
+                },
+            },
+            {
+                name: "document_rag",
+                async run() {
+                    stream.setStatus("RAG");
+                    stream.clearStatus();
+                    return {
+                        stage: "document_rag",
+                        status: "succeeded",
+                        artifacts: [artifact("rag", "document_rag", {
+                            provider: "OLLAMA",
+                            sourceAttachmentIds: ["doc-1"],
+                            extractedText: "contract text",
+                        })],
+                    };
+                },
+            },
+            {
+                name: "speech_to_text",
+                async run() {
+                    return {
+                        stage: "speech_to_text",
+                        status: "succeeded",
+                        artifacts: [artifact("transcript", "speech_to_text", {
+                            text: "transcribed voice",
+                            sourceAttachmentIds: ["audio-1"],
+                            model: "fake-stt",
+                        })],
+                    };
+                },
+            },
+            {
+                name: "model_call",
+                async run() {
+                    const reply = await adapter.callModel({provider: "OLLAMA", model: "fake-model"}, async () => {
+                        stream.append("final answer");
+                        return "final answer";
+                    });
+
+                    return {
+                        stage: "model_call",
+                        status: "succeeded",
+                        artifacts: [artifact("final_text", "model_call", {
+                            text: reply,
+                        })],
+                    };
+                },
+            },
+            {
+                name: "tool_loop",
+                async run() {
+                    const calls = [{id: "tool-call-1", name: "read_file", argumentsText: "{\"path\":\"docs/a.md\"}"}];
+                    const results = ["tool result"];
+                    adapter.appendToolResults([], calls, results);
+                    stream.recordToolExecution({
+                        toolName: "read_file",
+                        callId: "tool-call-1",
+                        argumentsText: "{\"path\":\"docs/a.md\"}",
+                        resultChars: results[0].length,
+                        startedAt: "2026-05-18T00:00:00.000Z",
+                        finishedAt: "2026-05-18T00:00:01.000Z",
+                    });
+
+                    return {
+                        stage: "tool_loop",
+                        status: "succeeded",
+                        artifacts: [artifact("tool_result", "tool_loop", {
+                            toolName: "read_file",
+                            callId: "tool-call-1",
+                            resultText: results[0],
+                        })],
+                    };
+                },
+            },
+            {
+                name: "persist_output_artifacts",
+                async run() {
+                    const generatedFile = outputAttachment("report.txt", "file");
+                    stream.recordOutputAttachment({
+                        artifactKind: "generated_file",
+                        fileName: generatedFile.fileName,
+                        mimeType: "text/plain",
+                        sizeBytes: generatedFile.sizeBytes,
+                        messageId: 321,
+                    });
+
+                    return {
+                        stage: "persist_output_artifacts",
+                        status: "succeeded",
+                        artifacts: [artifact("generated_file", "persist_output_artifacts", {
+                            attachmentId: generatedFile.fileId,
+                        })],
+                        attachments: [generatedFile],
+                    };
+                },
+            },
+            {
+                name: "text_to_speech",
+                async run() {
+                    stream.recordOutputAttachment({
+                        artifactKind: "tts_audio",
+                        fileName: "answer.ogg",
+                        mimeType: "audio/ogg",
+                        sizeBytes: 4096,
+                        messageId: 322,
+                    });
+
+                    return {
+                        stage: "text_to_speech",
+                        status: "succeeded",
+                        artifacts: [artifact("tts_audio", "text_to_speech", {
+                            attachmentId: "tts-audio-id",
+                        })],
+                        attachments: [outputAttachment("answer.ogg", "audio")],
+                    };
+                },
+            },
+            {
+                name: "audit_finish",
+                async run() {
+                    store.persist(state);
+                    return {
+                        stage: "audit_finish",
+                        status: "succeeded",
+                    };
+                },
+            },
+        ],
+        stageNames: [
+            "input_size_gate",
+            "document_rag",
+            "speech_to_text",
+            "model_call",
+            "tool_loop",
+            "persist_output_artifacts",
+            "text_to_speech",
+            "audit_finish",
+        ],
+    });
+
+    await pipeline.run(state, new AbortController().signal);
+
+    assert.equal(adapter.calls.length, 1);
+    assert.equal(stream.getText(), "final answer");
+    assert.equal(stream.getToolExecutions().length, 1);
+    assert.equal(stream.getOutputAttachments().length, 2);
+    assert.equal(state.artifacts.some(entry => entry.kind === "rag"), true);
+    assert.equal(state.artifacts.some(entry => entry.kind === "transcript"), true);
+    assert.equal(state.artifacts.some(entry => entry.kind === "final_text"), true);
+    assert.equal(state.artifacts.some(entry => entry.kind === "tool_result"), true);
+    assert.equal(state.artifacts.some(entry => entry.kind === "generated_file"), true);
+    assert.equal(state.artifacts.some(entry => entry.kind === "tts_audio"), true);
+    assert.equal(store.rows.length, 1);
+    assert.equal(store.rows[0].artifacts.length >= 6, true);
+});
+
+test("integration pipeline stops on fail_request fallback", async () => {
+    const stream = new FakeTelegramStreamMessage();
+    const state = createBaseState();
+    const pipeline = new UserRequestPipeline({
+        stages: [{
+            name: "input_size_gate",
+            async run() {
+                stream.setStatus("Boom");
+                throw new Error("boom");
+            },
+        }],
+        stageNames: ["input_size_gate", "document_rag"],
+        fallbackPolicies: [{
+            stage: "input_size_gate",
+            onUnavailable: "fail_request",
+            onFailed: "fail_request",
+        }],
+    });
+
+    await assert.rejects(() => pipeline.run(state, new AbortController().signal), /PipelineRequestFailure/);
+    assert.equal(state.audit.some(entry => entry.stage === "document_rag"), false);
+});

test/provider-adapter-contract.test.mjs

@@ -0,0 +1,144 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const {
+    extractOpenAiToolCalls,
+    extractOpenAiStreamingToolCalls,
+    extractOpenAiTextDelta,
+    extractOpenAiChatToolCalls,
+    extractOpenAiChatStreamingToolCalls,
+    extractOpenAiChatTextDelta,
+    mergeToolCallChunks,
+    normalizeStreamingTextDelta,
+    extractMistralToolCalls,
+    extractMistralTextDelta,
+    extractOllamaToolCalls,
+    extractOllamaTextDelta,
+} = await import("../dist/ai/provider-adapter-contract.js");
+
+test("openai contract extracts text delta and function calls", () => {
+    assert.equal(extractOpenAiTextDelta({type: "response.output_text.delta", delta: "hello"}), "hello");
+
+    const calls = extractOpenAiToolCalls({
+        output: [{
+            type: "function_call",
+            call_id: "call-1",
+            name: "read_file",
+            arguments: "{\"path\":\"src/index.ts\"}",
+        }],
+    });
+
+    assert.equal(calls.length, 1);
+    assert.equal(calls[0].id, "call-1");
+    assert.equal(calls[0].name, "read_file");
+
+    const streamed = extractOpenAiStreamingToolCalls({
+        type: "response.output_item.added",
+        item: {
+            type: "function_call",
+            id: "call-2",
+            name: "search_files",
+            arguments: "{\"query\":\"sendMessage\"}",
+        },
+    });
+
+    assert.equal(streamed.length, 1);
+    assert.equal(streamed[0].id, "call-2");
+    assert.equal(streamed[0].name, "search_files");
+});
+
+test("openai chat contract extracts text delta and tool calls", () => {
+    assert.equal(extractOpenAiChatTextDelta({choices: [{delta: {content: "hello chat"}}]}), "hello chat");
+    assert.equal(normalizeStreamingTextDelta("hel", "hello"), "lo");
+    assert.equal(normalizeStreamingTextDelta("hel", "lo"), "lo");
+
+    const calls = extractOpenAiChatToolCalls({
+        choices: [{
+            message: {
+                tool_calls: [{
+                    id: "chat-1",
+                    function: {
+                        name: "read_user_info",
+                        arguments: "{\"userId\":123}",
+                    },
+                }],
+            },
+        }],
+    });
+
+    assert.equal(calls.length, 1);
+    assert.equal(calls[0].id, "chat-1");
+    assert.equal(calls[0].name, "read_user_info");
+
+    const streamed = extractOpenAiChatStreamingToolCalls({
+        choices: [{
+            delta: {
+                tool_calls: [{
+                    index: 0,
+                    id: "chat-2",
+                    function: {
+                        name: "write_note",
+                        arguments: "{\"text\":",
+                    },
+                }],
+            },
+        }],
+    });
+
+    assert.equal(streamed.length, 1);
+    assert.equal(streamed[0].id, "chat-2");
+    assert.equal(streamed[0].name, "write_note");
+    assert.equal(streamed[0].argumentsText, "{\"text\":");
+
+    const merged = mergeToolCallChunks([
+        {id: "chat-2", name: "", argumentsText: "{\"text\":"},
+    ], [{
+        id: "chat-2",
+        name: "write_note",
+        argumentsText: "\"hello\"}",
+    }]);
+
+    assert.equal(merged.length, 1);
+    assert.equal(merged[0].name, "write_note");
+    assert.equal(merged[0].argumentsText, "{\"text\":\"hello\"}");
+});
+
+test("mistral contract extracts content and tool calls", () => {
+    assert.equal(extractMistralTextDelta({
+        content: [{text: "hello"}, {text: " world"}],
+    }), "hello world");
+
+    const calls = extractMistralToolCalls({
+        toolCalls: [{
+            id: "m-1",
+            function: {
+                name: "get_weather",
+                arguments: {location: "Moscow"},
+            },
+        }],
+    });
+
+    assert.equal(calls.length, 1);
+    assert.equal(calls[0].id, "m-1");
+    assert.equal(calls[0].name, "get_weather");
+});
+
+test("ollama contract extracts content and tool calls", () => {
+    assert.equal(extractOllamaTextDelta({
+        message: {content: "hello from ollama"},
+    }), "hello from ollama");
+
+    const calls = extractOllamaToolCalls({
+        tool_calls: [{
+            id: "o-1",
+            function: {
+                name: "web_search",
+                arguments: {query: "openai docs"},
+            },
+        }],
+    });
+
+    assert.equal(calls.length, 1);
+    assert.equal(calls[0].id, "o-1");
+    assert.equal(calls[0].name, "web_search");
+});

test/rag-artifact.test.mjs

@@ -1,32 +1,13 @@
-import test, {after} from "node:test";
+import test from "node:test";
 import assert from "node:assert/strict";
-import fs from "node:fs";
-import os from "node:os";
-import path from "node:path";
 
-const tempRoot = fs.mkdtempSync(path.join(os.tmpdir(), "tg-chat-bot-rag-"));
-process.env.BOT_TOKEN = process.env.BOT_TOKEN ?? "test-token";
-process.env.CREATOR_ID = process.env.CREATOR_ID ?? "1";
-process.env.DATA_PATH = tempRoot;
-process.env.DB_PATH = `file:${path.join(tempRoot, "test.sqlite")}`;
-process.env.TEST_ENVIRONMENT = "true";
-
-const {Environment} = await import("../dist/common/environment.js");
-Environment.load();
-
-const {DatabaseManager} = await import("../dist/db/database-manager.js");
-DatabaseManager.init();
-await DatabaseManager.ready;
-
-const {ArtifactStore} = await import("../dist/common/artifact-store.js");
-const {filterUserVisibleStoredAttachments} = await import("../dist/common/stored-attachment-utils.js");
+const {
+    buildRagArtifactPayload,
+} = await import("../dist/ai/rag-artifact-payload.js");
+const {
+    filterUserVisibleStoredAttachments,
+} = await import("../dist/common/attachment-visibility.js");
 const {AiProvider} = await import("../dist/model/ai-provider.js");
-const {persistRagArtifactAttachment} = await import("../dist/ai/rag-artifact-store.js");
-
-after(async () => {
-    await DatabaseManager.close().catch(() => undefined);
-    fs.rmSync(tempRoot, {recursive: true, force: true});
-});
 
 test("internal artifacts are not treated as user-visible attachments", () => {
     const visible = filterUserVisibleStoredAttachments([
@@ -50,65 +31,26 @@ test("internal artifacts are not treated as user-visible attachments", () => {
     assert.equal(visible[0].fileId, "visible");
 });
 
-test("RAG artifacts persist structured ollama metadata", async () => {
-    const chatId = 42;
-    const messageId = 7;
-
-    const attachment = await persistRagArtifactAttachment({
+test("RAG artifact payload keeps ollama retrieval metadata", () => {
+    const payload = buildRagArtifactPayload({
         provider: AiProvider.OLLAMA,
-        prepared: {
-            provider: AiProvider.OLLAMA,
-            prepared: true,
-            cleanup: async () => undefined,
-            artifact: {
-                query: "What is in the file?",
-                extractedDocuments: [
-                    {documentIndex: 0, fileName: "report.txt", textChars: 120},
-                ],
-                selectedChunks: [
-                    {
-                        sourceId: "doc1-1",
-                        documentIndex: 0,
-                        documentName: "report.txt",
-                        chunkIndex: 0,
-                        chunkCount: 1,
-                        textChars: 120,
-                        score: 0.91,
-                    },
-                ],
-                skippedDocuments: [
-                    {documentIndex: 1, fileName: "ignored.bin", reason: "unsupported format"},
-                ],
-                providerState: {
-                    embeddingModel: "nomic-embed-text:latest",
-                    topK: 8,
-                    chunkSize: 1400,
-                    chunkOverlap: 220,
-                    maxContextChars: 14000,
-                    minScore: 0.12,
-                    maxArchiveFiles: 200,
-                    maxArchiveBytes: 50 * 1024 * 1024,
-                    maxArchiveDepth: 2,
-                },
-            },
-        },
-        downloads: [{
-            kind: "document",
+        createdAt: "2026-01-01T00:00:00.000Z",
+        sources: [{
             fileId: "file-1",
             fileName: "report.txt",
-            buffer: Buffer.from("hello world"),
-            path: path.join(tempRoot, "report.txt"),
+            mimeType: "text/plain",
+            sizeBytes: 12,
+            sha256: "abc123",
+            uploadedFileId: "uploaded-1",
         }],
-        chatId,
-        messageId,
-        details: {
+        providerState: {
+            provider: AiProvider.OLLAMA,
+            prepared: true,
             embeddingModel: "nomic-embed-text:latest",
             topK: 8,
             chunkSize: 1400,
             chunkOverlap: 220,
             maxContextChars: 14000,
-            artifact: {
-                query: "What is in the file?",
             extractedDocuments: [
                 {documentIndex: 0, fileName: "report.txt", textChars: 120},
             ],
@@ -126,29 +68,20 @@ test("RAG artifacts persist structured ollama metadata", async () => {
             skippedDocuments: [
                 {documentIndex: 1, fileName: "ignored.bin", reason: "unsupported format"},
             ],
-                providerState: {
-                    embeddingModel: "nomic-embed-text:latest",
-                    topK: 8,
-                    chunkSize: 1400,
-                    chunkOverlap: 220,
-                    maxContextChars: 14000,
             minScore: 0.12,
             maxArchiveFiles: 200,
             maxArchiveBytes: 50 * 1024 * 1024,
             maxArchiveDepth: 2,
-                },
-            },
+            query: "What is in the file?",
         },
     });
 
-    assert.equal(attachment?.artifactKind, "rag");
-    assert.equal(fs.existsSync(attachment.cachePath), true);
-
-    const stored = await ArtifactStore.getByMessage(chatId, messageId);
-    assert.equal(stored.length, 1);
-    assert.equal(stored[0].kind, "rag");
-    assert.equal(stored[0].payload.providerState.query, "What is in the file?");
-    assert.equal(stored[0].payload.providerState.selectedChunks[0].score, 0.91);
-    assert.equal(stored[0].payload.providerState.skippedDocuments[0].reason, "unsupported format");
-    assert.equal(stored[0].payload.providerState.ollama.embeddingModel, "nomic-embed-text:latest");
+    assert.equal(payload.artifactKind, "rag");
+    assert.equal(payload.provider, AiProvider.OLLAMA);
+    assert.equal(payload.sources[0].uploadedFileId, "uploaded-1");
+    assert.equal(payload.providerState.provider, AiProvider.OLLAMA);
+    assert.equal(payload.providerState.query, "What is in the file?");
+    assert.equal(payload.providerState.selectedChunks[0].score, 0.91);
+    assert.equal(payload.providerState.skippedDocuments[0].reason, "unsupported format");
+    assert.equal(payload.providerState.embeddingModel, "nomic-embed-text:latest");
 });

test/rag-retention-planner.test.mjs

@@ -0,0 +1,53 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const {buildStaleRagCleanupPlan} = await import("../dist/ai/rag-retention-planner.js");
+
+test("stale rag cleanup plan selects only older rag artifacts", () => {
+    const plan = buildStaleRagCleanupPlan([
+        {
+            id: "recent-openai",
+            createdAt: "2026-05-18T00:00:00.000Z",
+            payload: JSON.stringify({
+                artifactKind: "rag",
+                providerState: {
+                    provider: "OPENAI",
+                    vectorStoreIds: ["vs_1"],
+                    uploadedFileIds: ["file_1"],
+                },
+            }),
+        },
+        {
+            id: "stale-openai",
+            createdAt: "2026-04-01T00:00:00.000Z",
+            payload: JSON.stringify({
+                artifactKind: "rag",
+                providerState: {
+                    provider: "OPENAI",
+                    vectorStoreIds: ["vs_2"],
+                    uploadedFileIds: ["file_2"],
+                },
+            }),
+        },
+        {
+            id: "stale-ollama",
+            createdAt: "2026-04-01T00:00:00.000Z",
+            payload: JSON.stringify({
+                artifactKind: "rag",
+                providerState: {
+                    provider: "OLLAMA",
+                    prepared: true,
+                },
+            }),
+        },
+    ], 14, new Date("2026-05-18T00:00:00.000Z"));
+
+    assert.equal(plan.targets.length, 1);
+    assert.deepEqual(plan.targets[0], {
+        artifactId: "stale-openai",
+        createdAt: "2026-04-01T00:00:00.000Z",
+        provider: "OPENAI",
+        vectorStoreIds: ["vs_2"],
+        uploadedFileIds: ["file_2"],
+    });
+});