feat: proactive recall, selectable embedding model, full MCP control (v0.8.0)

[RodCor]

Summary

Makes the brain proactive and gives the agent (and user) full control over it from inside Claude Code / Codex.

Proactive recall (mid-work)

The brain now surfaces a relevant memory while the agent works, not just on prompt — the way a memory "comes to you" rather than you fetching it. New PreToolUse / PostToolUse hooks (matcher Bash):

• after a failed Bash command → surface a matching failure_pattern / command fix;
• before a risky Bash command → warn from a matching failure_pattern;
• on a repeated failing command (loop) → lower the score floor and bypass the throttle.

Discipline keeps it near-zero-cost: lexical-FTS-only (no embedding-model load), high score floor (0.45; loop 0.35), one capsule, per-session dedupe, and a refractory throttle. Silent (zero tokens) when nothing qualifies. Wired into both Claude Code (.claude/settings.json) and Codex (.codex/hooks.json); opt out with kimetsu plugin install --no-proactive (or proactive:false over MCP). Per-session state lives in <repo>/.kimetsu/proactive/<session_id>.json, GC'd after 7 days.

Selectable embedding model

• New [embedder] table in project.toml (precedence: KIMETSU_BRAIN_EMBEDDER env > config > default).
• kimetsu brain model list / model set <id> + MCP kimetsu_brain_model_list / kimetsu_brain_model_set. model set re-embeds the corpus with the new model via a fresh, uncached embedder (reindex_all_with_embedder), independent of the process's cached default. Curated built-ins: bge-small-en-v1.5 (default), bge-m3, jina-v2-base-code.

Full MCP control surface

New tools so an agent can manage the brain without leaving the harness: kimetsu_brain_model_list / _model_set, _reindex, _memory_search (FTS), _conflict_resolve, _prune, _config_show. _memory_list and _memory_proposals gained limit/offset pagination. (21 → 28 tools.)

Fix: test isolation (+ a latent production bug)

discover_repo_root runs git rev-parse --show-toplevel. On a machine whose $HOME is itself a git repo, temp project roots created under AppData//tmp climbed to $HOME, so parallel tests shared one brain.db + project.lock — and the benchmark harness would init a fixture's brain at that ancestor, leaking fixture memories into the enclosing repo. New kimetsu_core::paths::git_init_boundary gives each fixture/test root its own git boundary (also a real correctness fix for bench). Plain cargo test is now hermetic with no env overrides.

Verification

• cargo test (plain, no TMP / KIMETSU_USER_BRAIN overrides): agent 86, brain 106, chat 51, core 8, cli 18 + 7 smoke, e2e 8, 3 ignored — 0 failures.
• cargo check --features embeddings typechecks; cargo fmt clean.
• New unit tests: embedder resolution, proactive_state (dedupe/refractory/loop), capsule_matches_kind, search_memories pagination, retrieve_proactive kind filtering, reindex_all_with_embedder, config round-trip, and a Codex proactive-hook wiring + --no-proactive opt-out test.
• Manual: proactive hooks (failure → inject, dedupe → silent, PreTool warn, non-Bash/success → silent), model list/set, and an MCP JSON-RPC smoke of all 7 new tools.

Bumps workspace + path-dep versions to 0.8.0; CHANGELOG.md and docs/HOW-KIMETSU-WORKS.md updated.

🤖 Generated with Claude Code