v0.7.0-mcp-rc2
Pre-release
Release candidate for real-world agent dogfooding and stabilization.
Summary
- 14 MCP tools (exact count preserved)
- 3140 tests passing, 0 failed, 11 ignored
- MCP-first architecture — host agent owns planning/orchestration; CodeBro owns engineering truth/infrastructure
- Stabilization / dogfooding phase — M6 has NOT started
M1–M5 Complete
M1: Engineering Memory Trust
Trust scores on memory entries computed from confidence, importance, and freshness. engineering_memory responses include per-entry trust metadata. memory_stats reports average trust. Backward-compatible.
M2: Change Invalidation Advisory
apply_change now returns needs_reindex advisory when source files in the fact store are affected. Impacted fact IDs are correlation metadata (not independently verified causal evidence).
M3: Lightweight Evidence Chaining
impact_analyze returns directed relationship edges with provenance metadata. Evidence chain: engineering_facts → impact_analyze → apply_change → invalidation advisory → reindex → fresh FactStore → sandbox_test/sandbox_build → VerificationResult.
M4: MCP Reindex Trigger
reindex tool performs full fact-store regeneration via the existing codebro init pipeline. Full rebuild, not incremental indexing.
M5: Repository Health MCP Tool
repository_health exposes the existing codebro doctor capability as MCP tool #14. Read-only; structured JSON with exit_code, status, checks, and summary. All 6 doctor checks preserved.
What This Is Not
- NOT a stable release
- NOT production-ready
- M6 has NOT started
- No architecture rule checking, semantic search, incremental indexing, or autonomous orchestration added