v0.3.13

release: v0.3.13 - external benchmark on real repositories

v0.3.12

release: v0.3.12 - durable semantic upgrade and 180/180 semantic gate

v0.3.11

release: v0.3.11 - quick indexing, index_now, and semantic repair

v0.3.10

feat: v0.3.10 release hardening, coverage, docker, and docs

v0.3.9

v0.3.9 - retrieval accuracy, parallel indexing, and hardening

v0.3.8 - Runtime and Distribution Hardening

v0.3.8 - runtime and distribution hardening

v0.3.7

v0.3.7

v0.3.6

v0.3.6 - Distribution and Contract Alignment

v0.3.4

v0.3.4 - offline semantic gate completeness

v0.3.3 - Agentic Retrieval Polish

v0.3.3 - Agentic Retrieval Polish

Token-safe agent output

• MCP tools are now metadata-first: include_body/max_chars opt-in body snippets, server-side limit cap (50), default metadata-only output. Stops agent context blowup on impact_of_change/trace_call_chain/search_code.

Runtime policy + honest eval

• Promoted retrieval policy is wired into runtime: MCP engines load PolicyStore::active() (new_with_active_policy); learned policies now actually affect searches.
• Eval hybrid ranking aligned with production (lexical fallback, spatial score, min_score filter); per-repo graph cache speeds up learning runs.

Reliability

• search_code chunk dedup; trace_call_chain parent-pointer BFS (no path cloning); OpenAI embedder retry-with-warning; fuzzy stable-id drift resolution.
• Trajectory logger records real latency, tool name, request id; MCP sets trajectory context per tool call (Phase 3 feedback ready).
• MCP index writes serialized per project; JSON-RPC errors preserve request id and use -32602/-32603.
• SKILL.md manual MCP config fixed for strict allowlist; CodeQL now analyzes Rust; CI protoc install hardened; self-improve enforces fixture determinism.

Verification

• cargo fmt/clippy/test clean; 4/4 GitHub Actions green (CI, eval, self-improve, CodeQL-Rust).