Yuriy 764003fc09 Commit X4: Auto inject Online Research summary into plan context

This commit implements fully automatic injection of online research results into the LLM prompt without user clicks.

## Backend

### Environment Variables
- Added `PAPAYU_ONLINE_AUTO_USE_AS_CONTEXT=1` (default: 0) to enable automatic injection of online research results into subsequent `proposeActions` calls.
- Added `is_online_auto_use_as_context()` helper function in `online_research/mod.rs`.

### Command Changes
- **`propose_actions` command**: Added `online_fallback_reason: Option<String>` parameter to track the error code that triggered online fallback.
- **`llm_planner::plan` function**: Added `online_fallback_reason: Option<&str>` parameter for tracing.
- **Trace Enhancements**: Added `online_fallback_reason` field to trace when `online_fallback_executed` is true.

### Module Exports
- Made `extract_error_code_prefix` public in `online_research/fallback.rs` for frontend use.

## Frontend

### Project Settings
- Added `onlineAutoUseAsContext` state (persisted in `localStorage` as `papa_yu_online_auto_use_as_context`).
- Initialized from localStorage or defaults to `false`.
- Auto-saved to localStorage on change.

### Auto-Chain Flow
- When `plan.ok === false` and `plan.online_fallback_suggested` is present:
  - If `onlineAutoUseAsContext === true` and not already attempted for this goal (cycle protection via `lastGoalWithOnlineFallbackRef`):
    - Automatically calls `researchAnswer(query)`.
    - Truncates result to `8000` chars and `10` sources (frontend-side limits).
    - Immediately calls `proposeActions` again with:
      - `online_context_md`
      - `online_context_sources`
      - `online_fallback_executed: true`
      - `online_fallback_reason: error_code`
      - `online_fallback_attempted: true`
    - Displays the new plan/error without requiring "Use as context" button click.
  - If `onlineAutoUseAsContext === false` or already attempted:
    - Falls back to manual mode (shows online research block with "Use as context (once)" button).

### Cycle Protection
- `lastGoalWithOnlineFallbackRef` tracks the last goal that triggered online fallback.
- If the same goal triggers fallback again, auto-chain is skipped to prevent infinite loops.
- Maximum 1 auto-chain per user query.

### UI Enhancements
- **Online Research Block**:
  - When `onlineAutoUseAsContext === true`: displays "Auto-used ✓" badge.
  - Hides "Use as context (once)" button when auto-use is enabled.
  - Adds "Disable auto-use" button (red) to disable auto-use for the current project.
  - When disabled, shows system message: "Auto-use отключён для текущего проекта."

### API Updates
- **`proposeActions` in `tauri.ts`**: Added `onlineFallbackReason?: string | null` parameter.

## Tests

- **`online_context_auto_test.rs`**: Added unit tests for:
  - `test_is_online_auto_use_disabled_by_default`
  - `test_is_online_auto_use_enabled_when_set`
  - `test_extract_error_code_prefix_timeout`
  - `test_extract_error_code_prefix_schema`
  - `test_extract_error_code_prefix_empty_when_no_prefix`

All tests pass.

## Documentation

### README.md
- Added "Auto-use (X4)" subsection under "Online Research":
  - Describes `PAPAYU_ONLINE_AUTO_USE_AS_CONTEXT=1` env var (default: 0).
  - Explains cycle protection: maximum 1 auto-chain per goal.
  - Documents UI behavior: "Auto-used ✓" badge and "Disable auto-use" button.

## Behavior Summary

**Without auto-use (default):**
1. `proposeActions` → error + `online_fallback_suggested`
2. UI calls `researchAnswer`
3. UI displays online research block with "Use as context (once)" button
4. User clicks button → sets `onlineContextPending` → next `proposeActions` includes context

**With auto-use enabled (`PAPAYU_ONLINE_AUTO_USE_AS_CONTEXT=1`):**
1. `proposeActions` → error + `online_fallback_suggested`
2. UI calls `researchAnswer` automatically
3. UI displays online research block with "Auto-used ✓" badge
4. UI immediately calls `proposeActions` again with online context → displays new plan
5. If still fails → no retry (cycle protection)

## Build Status

- ✅ Backend: `cargo build --lib` (2 warnings about unused code for future features)
- ✅ Frontend: `npm run build`
- ✅ Tests: `cargo test online_context_auto_test --lib` (5 passed)

Co-authored-by: Cursor <cursoragent@cursor.com>

2026-01-31 14:39:40 +03:00

4.8 KiB

Raw Blame History

Protocol v1 — контракт papa-yu

Краткий документ (1–2 страницы): что гарантируется, лимиты, логирование, PLAN→APPLY, strict/best-effort.

Версионирование

schema_version: 1
schema_hash: sha256 от llm_response_schema.json (в trace)
При изменении контракта — увеличивать schema_version; v2 — новый документ.

Default protocol: v2; Apply может fallback на v1 при специфичных кодах ошибок (см. PROTOCOL_V2_PLAN.md).

Гарантии

JSON: ответ LLM парсится; при неудаче — 1 repair-ретрай с подсказкой.
Валидация: path (no ../, absolute, ~), конфликты действий, content (no NUL, pseudo-binary).
UPDATE base: в APPLY каждый UPDATE_FILE — только для файлов, прочитанных в Plan.
Protected paths: .env, *.pem, *.key, id_rsa*, **/secrets/** — запрещены.
Apply: snapshot → apply → auto_check; при падении check — rollback.

Лимиты

Область	Переменная	По умолчанию
path_len	—	240
actions	—	200
total_content_bytes	—	5MB
context_files	PAPAYU_CONTEXT_MAX_FILES	8
file_chars	PAPAYU_CONTEXT_MAX_FILE_CHARS	20000
context_total	PAPAYU_CONTEXT_MAX_TOTAL_CHARS	120000

Логирование

Событие	Где
LLM_REQUEST_SENT	stderr (model, schema_version, provider, token_budget, input_chars)
LLM_RESPONSE_OK, LLM_RESPONSE_REPAIR	stderr
VALIDATION_FAILED	stderr (code, reason)
CONTEXT_CACHE_HIT, CONTEXT_CACHE_MISS	stderr (key)
CONTEXT_DIET_APPLIED	stderr (files, dropped, truncated, total_chars)
APPLY_SUCCESS, APPLY_ROLLBACK, PREVIEW_READY	stderr

Trace (PAPAYU_TRACE=1): .papa-yu/traces/<trace_id>.json — config_snapshot, context_stats, cache_stats, validated_json, schema_version, schema_hash.

PLAN → APPLY

Plan: plan: <текст> или «исправь/почини» → LLM возвращает actions: [], summary, context_requests.
Apply: apply: <текст> или «ok»/«применяй» при сохранённом плане → LLM применяет план с тем же context.
NO_CHANGES: при пустом actions в APPLY — summary обязан начинаться с NO_CHANGES:.

Strict / best-effort

strict (PAPAYU_LLM_STRICT_JSON=1): response_format: json_schema в API; при ошибке schema — repair-ретрай.
best-effort: без response_format; извлечение из json ... ; при ошибке — repair-ретрай.
Capability detection: при 4xx/5xx с упоминанием response_format — retry без него.

Кеш контекста

read_file, search, logs, env кешируются в plan-цикле. Ключ Logs: {source, last_n} — разные last_n не пересекаются.

Контекст-диета

При превышении лимитов — урезание: search → logs → файлы низкого приоритета; FILE (read_file) — последними; для priority=0 гарантия минимум 4k chars.

Provider Compatibility

Provider	Endpoint	`response_format: json_schema`	`strict`	OneOf (array/object)	Режим
OpenAI	`/v1/chat/completions`	✅	✅	✅	strict + local validate
OpenAI-compatible (часть)	разные	⚠️	⚠️	⚠️	best-effort + local validate + repair
Ollama	`/api/chat`	❌ (часто)	❌	⚠️	best-effort + local validate + repair

Поведенческие гарантии:

Если response_format не поддержан провайдером → fallback и лог LLM_RESPONSE_FORMAT_FALLBACK.
Локальная schema validation выполняется всегда (если schema compile ok).
Repair-ретрай выполняется один раз при невалидном JSON.
Если после repair невалидно → Err.
Capability detection: при 4xx/5xx с упоминанием response_format/json_schema → retry без него.

Ускоряет диагностику: если Ollama/локальная модель «тупит» — выключи PAPAYU_LLM_STRICT_JSON или оставь пустым.

4.8 KiB Raw Blame History Unescape Escape