f49927888e
feat(02-02): tool registry, executor, and 4 built-in tools
...
- ToolDefinition Pydantic model with JSON Schema parameters + handler
- BUILTIN_TOOLS: web_search, kb_search, http_request, calendar_lookup
- http_request requires_confirmation=True (outbound side effects)
- get_tools_for_agent filters by agent.tool_assignments
- to_litellm_format converts to OpenAI function-calling schema
- execute_tool: jsonschema validation before handler call
- execute_tool: confirmation gate for requires_confirmation=True
- execute_tool: audit logging on every invocation (success + failure)
- web_search: Brave Search API with BRAVE_API_KEY env var
- kb_search: pgvector cosine similarity with HNSW index
- http_request: 30s timeout, 1MB cap, GET/POST/PUT/DELETE only
- calendar_lookup: Google Calendar events.list read-only
- jsonschema dependency added to orchestrator pyproject.toml
- [Rule 1 - Bug] Added missing execute_tool import in test
2026-03-23 14:54:14 -06:00
a025cadc44
feat(02-04): wire escalation into orchestrator pipeline
...
- Add escalation pre-check in _process_message: assistant mode for escalated threads
- Add escalation post-check after LLM response: calls escalate_to_human on rule match
- Load Slack bot token unconditionally (needed for escalation DM, not just placeholders)
- Add keyword-based conversation metadata detector (billing keywords, attempt counter)
- Add no-op audit logger stub (replaced by real AuditLogger from Plan 02 when available)
- Add escalation_assignee and natural_language_escalation fields to Agent model
- Add Alembic migration 003 for new Agent columns
2026-03-23 14:53:45 -06:00
4047b552a7
feat(02-04): implement escalation handler (rule evaluator, transcript, DM delivery)
...
- check_escalation_rules: condition parser for 'keyword AND count > N' and NL phrases
- build_transcript: formats messages as Slack mrkdwn, truncates at 3000 chars
- escalate_to_human: opens DM, posts transcript, sets Redis key, logs audit event
2026-03-23 14:50:56 -06:00
30b9f60668
feat(02-02): audit model, KB model, migration, and audit logger
...
- AuditEvent ORM model with tenant_id, action_type, latency_ms, metadata
- KnowledgeBaseDocument and KBChunk ORM models for vector KB
- Migration 003: audit_events (immutable via REVOKE), kb_documents, kb_chunks
with HNSW index and RLS on all tables
- AuditLogger with log_llm_call, log_tool_call, log_escalation methods
- audit_events immutability enforced at DB level (UPDATE/DELETE rejected)
- [Rule 1 - Bug] Fixed CAST(:metadata AS jsonb) for asyncpg compatibility
2026-03-23 14:50:51 -06:00
45b957377f
feat(02-01): wire two-layer memory into orchestrator pipeline
...
- builder.py: add build_messages_with_memory() — injects pgvector context as
system message + sliding window history before current user turn
- runner.py: accept optional messages parameter; fallback to simple build for
backward compat (existing tests unaffected)
- tasks.py: memory pipeline in _process_message — load short-term + retrieve
long-term before LLM call; append turns to Redis + dispatch embed_and_store
fire-and-forget after response
- tasks.py: add embed_and_store Celery task (sync def + asyncio.run()) for
async pgvector backfill — never blocks the LLM response pipeline
- memory/embedder.py: lazy singleton SentenceTransformer (all-MiniLM-L6-v2)
with embed_text() / embed_texts() helpers
- All 202 tests pass (196 existing + 6 new memory integration tests)
2026-03-23 14:45:21 -06:00
28a5ee996e
feat(02-01): add two-layer memory system — Redis sliding window + pgvector long-term
...
- ConversationEmbedding ORM model with Vector(384) column (pgvector)
- memory_short_key, escalation_status_key, pending_tool_confirm_key in redis_keys.py
- orchestrator/memory/short_term.py: RPUSH/LTRIM sliding window (get_recent_messages, append_message)
- orchestrator/memory/long_term.py: pgvector HNSW cosine search (retrieve_relevant, store_embedding)
- Migration 002: conversation_embeddings table, HNSW index, RLS with FORCE, SELECT/INSERT only
- 10 unit tests (fakeredis), 6 integration tests (pgvector) — all passing
- Auto-fix [Rule 3]: postgres image updated to pgvector/pgvector:pg16 (extension required)
2026-03-23 14:41:57 -06:00
6f30705e1a
feat(01-03): Channel Gateway (Slack adapter) and Message Router
...
- gateway/normalize.py: normalize_slack_event -> KonstructMessage (strips bot mention)
- gateway/channels/slack.py: register_slack_handlers for app_mention + DM events
- rate limit check -> ephemeral rejection on exceeded
- idempotency dedup (Slack retry protection)
- placeholder 'Thinking...' message posted in-thread before Celery dispatch
- auto-follow engaged threads with 30-minute TTL
- HTTP 200 returned immediately; all LLM work dispatched to Celery
- gateway/main.py: FastAPI on port 8001, /slack/events + /health
- router/tenant.py: resolve_tenant workspace_id -> tenant_id (RLS-bypass query)
- router/ratelimit.py: check_rate_limit Redis token bucket, RateLimitExceeded exception
- router/idempotency.py: is_duplicate + mark_processed (SET NX, 24h TTL)
- router/context.py: load_agent_for_tenant with RLS ContextVar setup
- orchestrator/tasks.py: handle_message now extracts placeholder_ts/channel_id,
calls _update_slack_placeholder via chat.update after LLM response
- docker-compose.yml: gateway service on port 8001
- pyproject.toml: added redis, konstruct-router, konstruct-orchestrator deps
2026-03-23 10:27:59 -06:00
8257c554d7
feat(01-02): Celery orchestrator — handle_message task, system prompt builder, LLM pool runner
...
- Create orchestrator/main.py: Celery app with Redis broker/backend, task_acks_late=True, 10-min timeout
- Create orchestrator/tasks.py: SYNC def handle_message (critical pattern: asyncio.run for async work)
- Deserializes KonstructMessage, sets RLS context, loads agent from DB, calls run_agent
- Retries up to 3x on deserialization failure
- Create orchestrator/agents/builder.py: build_system_prompt assembles system_prompt + identity + persona + AI transparency clause
- Create orchestrator/agents/runner.py: run_agent posts to llm-pool /complete via httpx, returns polite fallback on error
- Add Celery[redis] dependency to orchestrator pyproject.toml
- Create tests/integration/test_llm_fallback.py: 7 tests for fallback routing and 503 on total failure (LLM-01)
- Create tests/integration/test_llm_providers.py: 12 tests verifying all three providers configured correctly (LLM-02)
- All 19 integration tests pass
2026-03-23 10:06:44 -06:00
5714acf741
feat(01-foundation-01): monorepo scaffolding, Docker Compose, and shared data models
...
- pyproject.toml: uv workspace with 5 member packages (shared, gateway, router, orchestrator, llm-pool)
- docker-compose.yml: PostgreSQL 16 + Redis 7 + Ollama services on konstruct-net
- .env.example: all required env vars documented, konstruct_app role (not superuser)
- scripts/init-db.sh: creates konstruct_app role at DB init time
- packages/shared/shared/config.py: Pydantic Settings loading all env vars
- packages/shared/shared/models/message.py: KonstructMessage, ChannelType, SenderInfo, MessageContent
- packages/shared/shared/models/tenant.py: Tenant, Agent, ChannelConnection SQLAlchemy 2.0 models
- packages/shared/shared/models/auth.py: PortalUser model for admin portal auth
- packages/shared/shared/db.py: async SQLAlchemy engine, session factory, get_session dependency
- packages/shared/shared/rls.py: current_tenant_id ContextVar and configure_rls_hook with parameterized SET LOCAL
- packages/shared/shared/redis_keys.py: tenant-namespaced key constructors (rate_limit, idempotency, session, engaged_thread)
2026-03-23 09:49:28 -06:00
