v1.0.0 — Customer Service AI Agent (first public release)

First public release of the Customer Service AI Agent multi-agent system.

Internal runtime version is tracked separately as 6.3 (core/config.py). This v1.0.0 marks the first public, documentation-complete release.

PRODUCTION NOT_VERIFIED. This release is not a claim of production validation. Capability state is declared in docs/production-readiness.md: Implemented / Verified (Local|CI) / Not Verified.

Agent architecture

• Four-layer state machine: cache → routing → collaboration → post-response; 9 agent roles, 5 collaboration modes (Sequential / Parallel / Consultation / Hierarchical / ReAct). Single graph definition, deterministic nodes.
• Distributed Agent Runtime (hybrid real-time / async): real-time POST /api/chat and /api/chat/stream stay inline (FastAPI → LangGraph → SSE); long tasks go POST /api/runs → Celery worker → GET /api/runs/{run_id}.
  • PostgreSQL LangGraph checkpoint shared across workers, crash-resumable (resume from checkpoint next, not from scratch).
  • Redis session storage + Redis per-thread lock (owner token + TTL + Lua atomic compare-and-delete).
  • AgentRun is the single source of truth (task_ignore_result=True; Celery result backend is not).
  • at-least-once dispatch + three-layer idempotency (run / thread / tool tool_side_effects ledger), retry backoff, agent_dead_letters + replay_dead_run.py, Redis Stream run events.
• Evidence: Implemented; CI-verified on real PostgreSQL + Redis + multi-process Celery (make runtime-e2e, make runtime-chaos, make runtime-verify).

RAG pipeline

• Qdrant vector store + BM25 hybrid retrieval → RRF fusion → cross-encoder rerank; query rewriting; retrieval contract; deterministic point IDs + migration tooling.
• Formal 649-query metrics: NOT_VERIFIED. The shipped gold set carries no human relevance labels (DATASET_DEFECT). No Hit@K / MRR / NDCG / Recall number is claimed in this release.

MCP tools

• Function-calling tool registry (OpenAI format) + tool-result context engineering (budget / compression / offload / scope-safe reuse).
• MCP adapter is present on main, default-off, read-only-first, with 11 fail-closed policy checks.
• Not verified: real third-party MCP servers (network / auth / multi-replica / write tools).

HITL (human-in-the-loop)

• Governance boundary for high-risk tool side effects: risk grading (only HIGH requires approval; fail-closed on grading failure), interception before execution, durable human_approvals table, reviewer ≠ requester enforced in the service layer, RBAC (admin / supervisor), TTL expiry → EXPIRED (treated as reject, never default-allow), sanitized audit trail, and approval-derived idempotency key for exactly-once effects.
• HITL_ENABLED defaults to false. Not verified: real Kingdee ERP refund/order-write operations (governance mechanism verified via deterministic staging tools only).

Production readiness

• New docs/production-readiness.md + entry docs: architecture, agent-design, evaluation, deployment, limitations.
• Verified (CI): canonical ruff lint, unit/integration/e2e on Python 3.10/3.11/3.12, distributed runtime on real PG + Redis, SIGKILL crash recovery, security scanning (strict mypy + bandit + secret guard).
• Verified (Local): make test (offline), doc-consistency guard, OTel collector smoke, DLQ replay.
• Not verified: production cluster / multi-replica long-run stability / real user traffic, 649-query RAG metrics, real ERP writes, production QPS/P95/P99/token cost, persistent queryable trace backend, Nginx multi-replica scale-out, production worker autoscaling/backpressure, push approval notifications, real human-handoff workbench/ticketing.

Docs

• Every Mermaid diagram now uses one node declaration per line so GitHub renders all of them; a new structural guard (rule AJ) in scripts/audit_doc_consistency.py (+ tests) prevents regression.