Add hit_processing.py: HitProcessor with model_opt (LLM-generated) and direct (concatenated chunks) modes, with in-process cache
Add settings.py: KBSettings/KBSettingsUpdate models, KBSettingsStore with async CRUD
Add KB settings endpoints to kb_management.py: GET/PUT /kb-management/kbs/{kb_id}/settings with owner-only modification
Tests: 43 new tests (25 hit_processing + 18 settings), 293 total passing
When disclosure_level=0, system prompt only injects skill name + description
(summary mode). SkillDetailTool is injected into the tool set, allowing the
LLM to load full instructions on-demand via skill_detail(query). This reduces
context window consumption when many skills are registered.
Add PipelineCheckpoint for stage-level crash recovery with Redis-first
+ memory fallback. TeamOrchestrator saves checkpoints after each phase
finalizes and supports resume(plan_id) to continue from the last
completed phase. New POST /api/v1/tasks/{id}/resume endpoint recreates
the team from saved plan and calls resume.
U4: ExpertTeam accepts redis_client, passes to SharedWorkspace. After phase
completion, full result is written to workspace and in-memory phase.result
is replaced with a 500-char summary + _ref_key. Dependency output reading
resolves offloaded content from workspace on demand, with graceful fallback
to summary on read failure.
Tests: 8 scenarios (offload creation, short content, dependency resolution,
workspace failure fallback, non-offloaded passthrough, redis_client wiring,
memory dict fallback, pipeline integration) — all pass.
U3: ContextCompressor now accepts model_context_limit, headroom_threshold,
and min_tokens. should_compress() triggers when token ratio exceeds 0.8 of
model limit OR exceeds min_tokens (8000 fallback). ReActEngine._should_compress
delegates to compressor when available, checks is_available() first.
Tests: 6 scenarios (headroom trigger, min_tokens guard, small model,
unavailable compressor, delegation, fallback) — all pass.
U2: Add asyncio.Semaphore to bound concurrent phase execution and debate
argument generation. Default limit=3, configurable via max_concurrent_phases.
Prevents LLM rate-limit spikes when many phases run in the same layer.
Tests: 5 scenarios (happy path, 5-phase edge case, serial mode, failure
release, debate integration) — all pass.
U1: Sliding window hash detection in ReAct loop. When the same tool is
called with identical arguments >= threshold times (default 2), injects
a correction message first, then raises LoopDetectedError if the LLM
doesn't change strategy. Covers both _execute_loop and execute_stream.
Implement _execute_debate_phase() with Lead-facilitated structured debate:
- Lead opens with divergence point + dependency context
- Experts argue in parallel per round (asyncio.gather)
- Lead summarizes each round, then adjudicates final verdict
- Verdict produces decision (adopt/compromise/shelve/inconclusive) + conclusion
- Conclusion written to SharedWorkspace for downstream phases
Escape hatches:
- debate_config.skip=true short-circuits with template text
- MAX_DEBATE_ROUNDS=4 hard cap on rounds
- User /stop intervention ends debate early (U4-compatible via getattr fallback)
- LLM unavailable falls back to template verdict, no crash
New events: debate_started, expert_argument, debate_round_summary,
debate_resolved (plus existing phase_completed for consistency).
Phase dispatcher (_execute_phase) routes by phase_type:
EXECUTION to _execute_execution_phase, DEBATE to _execute_debate_phase.
36 new tests in test_orchestrator_debate.py covering happy path (2 rounds,
2 experts), max_rounds=1 boundary, empty participants, user stop, skip
escape hatch, LLM unavailable, SharedWorkspace integration, event
broadcasting, intervention channel compatibility, and helper methods.
All 377 expert tests pass.
Also includes planning artifacts (brainstorm requirements + implementation
plan with 6 units U1-U6).
U1: Data model foundation for structured debate collaboration.
- Add PhaseType enum (EXECUTION | DEBATE)
- Add phase_type and debate_config fields to PlanPhase
- Update to_dict/from_dict for serialization with backward compatibility
- Add tests for PhaseType, debate phase creation, serialization, and
mixed EXECUTION+DEBATE topological sort
OutlookSyncProvider implementing AbstractSyncProvider for
bidirectional Outlook Calendar sync. Uses Graph API delta query
for incremental pull, auto-refreshes OAuth tokens on 401, and
converts Outlook recurrence patterns to RRULE. Same conflict
resolution as CalDAV (last-write-wins + WS notification).
- src/agentkit/calendar/sync/outlook_provider.py — OutlookSyncProvider
- tests/unit/calendar/test_sync_outlook.py — 8 tests
ICSProvider parses .ics files (icalendar library) and creates local
CalendarEvents, skipping duplicate UIDs. Export builds an iCalendar
from events in a date range, deduplicating recurring event
occurrences back to a single VEVENT with RRULE. REST endpoints:
POST /import-ics (multipart upload), GET /export-ics (download).
- src/agentkit/calendar/sync/__init__.py — sync subpackage init
- src/agentkit/calendar/sync/ics_provider.py — ICSProvider (import/export)
- src/agentkit/calendar/db.py — added get_event_by_external_id() for dedup
- src/agentkit/server/routes/calendar.py — import-ics and export-ics endpoints
- pyproject.toml — added icalendar>=5.0 dependency
- tests/unit/calendar/test_ics_provider.py — 8 tests
Adds PostProcessingExtractor — a zero-LLM keyword gate (Chinese +
English time words) followed by LLM extraction for ambiguous cases.
Events created from extraction carry source="post_extract" so the UI
can style them distinctly (R33). LLM gateway is optional to keep the
constructor testable without a live provider.
- src/agentkit/calendar/extraction.py — PostProcessingExtractor
- tests/unit/calendar/test_extraction.py — 13 tests with MockLLMGateway
Adds CalendarTool implementing the Tool ABC so the ReAct engine can
create, query, update, and delete events autonomously. Resolves
event_type_name and tag_names (look up or create), sets
source="agent" to distinguish agent-created events from manual ones.
- src/agentkit/tools/calendar_tool.py — CalendarTool(Tool)
- tests/unit/tools/test_calendar_tool.py — 13 tests covering all actions
Deploy to Production / deploy (push) Waiting to runDetails
The skills tab mixed generic execution-engine templates (react/direct/
rewoo/...) with business-domain skills (monitor/geo_optimizer/...) with
no visual or data distinction. Adds a derived `category` field to the
SkillInfo/SkillDetail API models and groups the frontend display.
Backend:
- SkillInfo/SkillDetail: add category (Literal), agent_type, execution_mode,
task_mode fields
- _skill_to_info: derive category from explicit _ENGINE_TEMPLATE_NAMES set
(not name suffix — trend_agent/deai_agent are business skills despite
the _agent suffix)
- Simplify repetitive hasattr pattern with getattr
Frontend:
- ISkillInfo/ISkillDetail: add category + mode fields
- skills store: agentTemplates/businessSkills computed getters
(businessSkills is defensive: anything not explicitly engine template)
- SkillsView: group into 执行引擎 / 业务技能 sections with counts
- SkillCard: type badge (引擎/技能), category-based icon, mode display,
dark-mode-aware accent color
Tests:
- test_category_derived_from_name_suffix: verifies field exposure
- test_category_no_orphans: invariant — every skill has a valid category
- test_trend_agent_classified_as_business_skill: regression guard for
the _agent suffix misclassification bug
Code review (ce-code-review): 2 P1 + 5 P2 findings applied.
All config file writes now use the write-temp + fsync + os.replace
pattern (KTD-4) so a crash mid-write leaves the original file intact.
- Add src/agentkit/server/utils/atomic_write.py with write_text_atomic
- settings.py: _write_yaml_config and _write_env_var use atomic write
- skill_service.py: import_skill uses atomic write
- skill_service.py: update_skill_config uses atomic write + fcntl.flock
around the read-modify-write cycle to serialize concurrent updates
- Add 11 unit tests covering happy path, crash safety, concurrency, errors
The whoami route accepted rotated/old refresh tokens for cold-start
because it only checked session revocation status, not the token hash.
Now when token_type == "refresh", the route computes hash_token(token)
and compares it with the session's stored refresh_token_hash using
hmac.compare_digest (constant-time). Mismatch returns 401.
- Add SessionService.get_stored_refresh_hash(session_id) helper
- Add hash verification in whoami route (R9)
- Add TestWhoamiTokenHash with 5 integration tests
SkillService: enable/disable (persisted in skill_states table, schema
v4), import from YAML (with path traversal + name validation), reload
from file, update config. GET /skills now filters disabled skills.
KbService: list/upload/delete documents with department_id binding.
Added department_id field to KnowledgeSource + UploadedDocument.
Department visibility: (bound to user depts) ∪ (global = None).
10 new admin endpoints: skill enable/disable/import/reload/update,
KB documents CRUD, source sync/rebuild. All guarded by _require_admin.
Implemented reload stub in skill_management.py (was no-op).
54 new tests (26 unit + 28 integration). Fixed 4 pre-existing lint
errors. 357 admin tests pass, no regressions.
U1: Bump _SCHEMA_VERSION to 3, add 5 department tables (departments,
user_departments, department_skill_bindings, department_kb_bindings,
department_quotas) + 5 ORM models + helpers.
U2: DepartmentService (12 async methods: CRUD + bind/unbind skill/KB +
count_users). Mount admin_router in app.py. 36 unit + 28 integration tests.
U4: DepartmentContext FastAPI dependency (per-route, admin bypasses
filtering). filter_skills_by_department / filter_kb_sources_by_department
helpers. Applied to GET /skills and GET /kb-management/* routes.
15 integration tests for department isolation.
Also includes brainstorm + plan docs. 108 new tests, all pass.
Add create_user method to LocalAuthProvider (bcrypt hash + INSERT,
raises ValueError on duplicate username/email).
Add UserService with 9 async methods: create/list/get/update/delete
(soft)/reset_password/assign_department/remove_department/list_user_departments. reset_password revokes all sessions via SessionService.
delete_user is soft (is_active=0, row preserved).
Add 9 user endpoints to routes/admin.py: POST/GET/PATCH/DELETE users,
reset-password, assign/remove department, list departments. All
guarded by _require_admin.
Tests: 40 unit + 37 integration = 77 new tests. Full admin suite
170 tests pass, no regressions.
Adds the central business-logic layer for ``auth_sessions`` so routes,
the auth middleware, and the admin endpoints can call a single service
instead of touching the table directly.
Server
- session_service.SessionService: CRUD + lifecycle for auth_sessions.
- create() enforces the per-user cap (default 10): the oldest
active session is evicted with reason=session_cap_eviction.
- rotate() swaps a refresh token, adds the old hash to the
denylist, and raises SessionReuseDetected (revoking all sessions
for the user) when the old token is replayed.
- revoke() / revoke_by_refresh_token() / revoke_all_for_user()
with explicit reasons: user_terminated, admin_revoked,
password_changed, reuse_detected, session_cap_eviction.
- touch() bumps last_active_at (called on /auth/whoami).
- session_cache.SessionValidationCache: bounded LRU+TTL wrapper
(default 30s/1k entries) around SessionService.is_session_valid.
The middleware hits this on every request carrying a V2 sid claim;
one SQLite round-trip per 30s per session instead of per request.
- get_session_service() / get_validation_cache() module-level
singletons overridable in tests via set_session_service() /
set_validation_cache().
Tests
- tests/unit/auth/test_session_service.py: 15 cases covering
create/rotate/revoke/list/cap-eviction/reuse-detection/expired
sessions.
Refs: U3 in docs/plans/2026-06-20-002-feat-centralized-auth-token-persistence-plan.md
Adds V2 JWT claim schema that closes the kicked-out window and enables
refresh-token rotation with reuse detection.
Server
- jwt_utils.create_token_pair now takes ``session_id`` and ``remember_me``
kwargs. When ``session_id`` is provided, both tokens carry a ``sid``
claim and the access token also carries a ``jti`` claim; the refresh
token's jti is intentionally absent (rotation uses the token hash).
- New ``REFRESH_TOKEN_TTL_REMEMBER_ME = 30d`` (default 7d) selected by
the ``remember_me`` flag.
- ``verify_token`` now supports an optional ``expected_type`` filter
(e.g. ``"access"`` / ``"refresh"``); when omitted, both types pass
(used by /auth/whoami's cold-start path).
- New ``auth.denylist`` module: ``InMemoryRecentlyRevoked`` (default for
the Tauri sidecar / dev mode) and ``RedisRecentlyRevoked`` (multi-
process server). Bounded LRU with auto-expiry via ``time.monotonic()``.
Backwards-compat
- Tokens issued before U2 (no ``sid``) are still accepted by
``verify_token``; validation falls through to the legacy
``user_sessions`` table via the U10 shim (next commit).
Tests
- tests/unit/auth/test_jwt_utils.py: 12 cases — V1/V2 claim presence,
default + remember-me TTL, expected_type filter, expiry, wrong secret.
- tests/unit/auth/test_denylist.py: 6 cases — add/contains, TTL expiry,
LRU eviction, re-add refresh, clear, hash stability.
Refs: U2 in docs/plans/2026-06-20-002-feat-centralized-auth-token-persistence-plan.md
- Settings API: reverse-resolve env vars to preserve ${VAR} refs in yaml,
write new API keys to .env instead of agentkit.yaml, extract env_key
from existing ${VAR} reference when updating providers
- Onboarding: merge-update instead of overwrite when config exists,
use config_arg to determine output path, .env merge instead of overwrite
- Unified templates: bailian-coding provider name, full model_aliases,
docker-compose with postgres, expanded .env.example
- Optional ruamel.yaml for comment/format preservation in Settings API
- clients.yaml: add _deep_resolve for ${VAR} env var references
- All CLI commands use load_config_with_dotenv() consistently
- Tests: mock find_config_path and CWD auto-discovery to avoid env leaks
Key improvements:
- Low-complexity queries (<0.3) now try IntentRouter keyword match
before falling back to DIRECT_CHAT, fixing 0% F1 on keyword_match
- SemanticRouter similarity_low lowered from 0.6 to 0.4
- Short text (<20 chars) uses effective_low = max(0.25, low - 0.15)
- Short text with no semantic match forces LLM classify fallback
- Added colloquial keywords to 7 skill YAMLs
- Fixed code_reviewer.yaml output_schema placement
- Fixed SemanticRouter build in e2e tests
- Fixed base_url detection for bailian-coding API keys
Results: keyword_match F1 0->60.87%, colloquial F1 0->100%, mixed_lang F1 0->100%
- Add HeuristicClassifier to replace LLM quick_classify with zero-cost
local heuristic (keyword/length/code-pattern scoring), gated by
router.classifier config (default: heuristic)
- Add parallel tool execution in ReActEngine via asyncio.gather for
multiple independent tool_calls, gated by parallel_tools param
- Add AsyncWriteQueue for non-blocking session persistence with WAL
buffer, gated by async_writes param on SessionManager
- Add httpx.Limits connection pool config to all LLM providers
- Add router config section to ServerConfig and agentkit.yaml
- All optimizations have config switches for safe rollback
Critical:
- C1: Add verifier_timeout_seconds for independent Verifier timeout
- C2: Verifier parse failure raises RuntimeError instead of dead-loop
Major:
- M1: Inject previous_output into Worker retry context
- M2: Add Pydantic ge/le constraint on ReviewFeedback.score
- M3: Use Literal type for feedback_mode enum validation
- M4: Use Literal types for ReviewIssue severity and category
- M5: Merge error messages when escalation agent also fails
Tests: 8 new test cases added (19 total), all passing
chiguyong