ci: ignore RUSTSEC-2025-0134 (rustls-pemfile unmaintained)#20
Merged
Conversation
Introduce KIMETSU_CONFIG_VERSION (the project.toml format version), separate from KIMETSU_SCHEMA_VERSION (the brain.db schema). The DB schema can now advance via migrations without forcing project.toml rewrites or breaking existing projects on open. Prep for the schema-migration runner. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add kimetsu-brain/src/migrate.rs: a forward-only migration runner that advances brain.db from its stored schema_info version to the binary's target, each migration applied with its schema_info bump inside one transaction (crash-safe, resumable). Empty migration set for now; the v1->v2 fold and the initialize wiring land next. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Split schema::initialize into create_baseline + the migration runner. The historical add_column_if_missing patches, citations/conflicts tables, and FTS reshapes now live in a single idempotent v1->v2 migration; existing brain.db files (all stamped v1) and fresh ones converge to the v2 shape and are version-stamped. Replaces the old run kimetsu brain rebuild hard-error with automatic forward migration on open. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
run_with snapshots the live DB to a brain.db.bak-<from>-<to>-<ts> sidecar (SQLite online backup, WAL-consistent) before applying any version- advancing migration, and prunes to the newest 3 backups on success. No backup for in-memory DBs or no-op opens. Recoverable if a migration fails. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ration) schema::validate now hard-errors only when the DB is NEWER than the binary; an OLDER DB returns a typed SchemaNeedsMigration. open_user_brain_readonly treats that as Ok(None) (a stale user brain no longer breaks unrelated read-only ops; the next read-write open migrates it); load_project_readonly surfaces it so the next read-write load_project migrates. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
apply_event now dispatches through upcast_event, the localized point where a future EVENT_SCHEMA_VERSION bump normalizes older payloads to the current shape (identity today). Lock the projector's missing-field durability with per-kind tests: every event kind replays from an empty payload without panic, so old/sparse trace events project losslessly. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ackup Skip the pre-migration backup when the brain has no memories to protect (no more useless brain.db.bak on fresh installs); log the migration via tracing; add the brain.db schema version to `brain status`; reframe `brain rebuild` help (schema upgrades are automatic on open). End-to-end test: an existing populated project brain AND user brain upgrade v1->v2 with a backup sidecar and preserved data. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Golden tests for every install merge writer (Claude/Codex hooks, MCP config, CLAUDE.md): pre-existing user entries survive, Kimetsu's are added alongside, and re-install is byte-idempotent. Full install-path tests per target x scope preserve seeded user content; an upgrade-idempotency test proves a second install can't corrupt managed files. Add windows-latest to the PR test matrix. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add kimetsu-brain/src/analytics.rs: read-only proof-of-value metrics (proposal acceptance, corpus health, harvest yield, usefulness trend, citation rate, token economy) over a recent-runs window. Extend the context.injected event with used_tokens + capsule_count so token economy is measurable. Retrieval hit-rate / skip-rate land with context.served (C7). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Surface the analytics module as a CLI command (human table + --json, --last-n-runs/--since/--top) and a matching MCP tool with an interpretation string. Retrieval hit-rate shows n/a until context.served lands (C7). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Log every retrieval (hit or miss) as a context.served event: from the UserPromptSubmit context-hook (best-effort, hashed query, opt-out via KIMETSU_BRAIN_LOG_RETRIEVAL=0) and from the autonomous pipeline. Analytics now computes retrieval hit-rate, avg top score, and skip-rate over a time window (hook events have no run, so they're counted by timestamp). Removes the n/a-until-C7 placeholders. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add the sqlite-vec dependency under the embeddings feature and register it once before every brain DB open. Proven on the MSVC host: a vec0 virtual table creates and a KNN MATCH query runs. No-op on the lean build. Foundation for ANN candidate generation (D1b-D1c). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a derived vec0 index (memory_vec) reconciled on demand from memories, and union a KNN semantic-recall pass into memory_candidates on embeddings builds — finding memories whose meaning matches the query even when no words overlap (replacing the recency-bounded fallback). Rebuilds on embedder change; lean (FTS-only) retrieval is unchanged. Derived structure, not a numbered migration, so lean/embeddings DB schema stays identical. D1b: ensure_vec_index() creates memory_vec (vec0, TEXT PRIMARY KEY mapping via approach 1: vec0(memory_id TEXT PRIMARY KEY, embedding float[D])) plus memory_vec_meta tracking the active (model_id, dim). Drops and recreates on embedder change; incremental INSERT/DELETE reconciliation on each call. D1c: memory_ann_candidates() runs a KNN query (MATCH + ORDER BY distance + LIMIT 80) on the vec0 index, fetches full memory rows, builds Candidates via memory_row_to_candidate. memory_candidates() on embeddings builds now unions FTS + ANN candidates, deduping by memory_id (higher raw_relevance wins). D1d: four new tests — ANN semantic match FTS misses (oracle embedder), vec index rebuild on model-id change, dedup guarantee, lean-unchanged path with NoopEmbedder. 151/151 tests pass on embeddings build; all workspace tests green; lean build clean. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Embedding-MMR collapses true paraphrase near-duplicates (cosine redundancy, Jaccard fallback); an absolute semantic-relevance floor drops genuinely off-topic candidates so irrelevant queries hit the zero-capsule path. Render caps honor the configured budget/max_capsules instead of a hard-coded 12, and defaults drop now that selection is more precise. Lean (FTS-only) retrieval unchanged. Token economy proven lower for the same queries with the relevant capsule preserved. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
No not_implemented subcommands ship in 1.0: `config edit` opens $EDITOR on project.toml and re-validates on save; `run abort <id>` cleanly finalizes a dangling run (run.aborted event + lock release), erroring on unknown/terminal runs. `kimetsu doctor --selftest` proves the brain works end-to-end (record → retrieve) hermetically, and the README gains a 5-minute quickstart. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fix all workspace clippy warnings on the lean and embeddings builds (derivable impls, doc indentation, idiomatic iterator/string fixes, a real MSRV-1.85 portability fix, targeted allows on retrieval arg-count). Flip the CI clippy job from advisory to a hard gate denying warnings on both flavors. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
In-memory per-run bookkeeping of what the brain has surfaced/injected: is_injected/mark_injected with once-counted token accounting (F1 cross-stage dedup) and is_surfaced/mark_surfaced (E1/E2 proactive dedup). Pure leaf data structure; consumers wire it into the pipeline next. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A brain capsule rendered into an earlier stage's prompt is back-referenced (not re-rendered in full) in later stages of the same run, and its token cost is counted once. Brain overhead now shrinks as a task spans more stages. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…mpt (E1) Before the first implementation attempt, a tight failure_pattern/convention retrieval surfaces known pitfalls for the task as a "Known pitfalls" prompt block — proactive recall, not just post-failure. ~Zero tokens when nothing matches (high min_score + kind filter); the run recall ledger stops a pitfall being re-surfaced on retries. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Classify each task once (deterministic keyword classifier, no model call) as Debug/Feature/Refactor/Docs/Investigation, threaded on ContextRequest. A task-kind weight layer composes over the stage weights (renormalized) and biases kind preference (Debug→failure_pattern, Refactor→convention, Investigation→fact/preference) so recall fits the task. Feature is the neutral default — existing retrieval behavior is unchanged. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
… (F2) Inject top-confidence capsules in full and the long tail as ~1-line headlines the agent expands on demand via a new expand_capsule(handle) tool (resolves memory:/file: handles to full text, emits capsule.expanded for measurement). Up-front tokens drop sharply while recall stays high — the marginal cost of breadth approaches zero. Composes with F1 (no double-charge). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Capsule sort tiebreaks used capsule.id, a fresh random ULID minted per retrieval, so two retrievals over the same corpus could order score+freshness ties differently — non-reproducible retrieval. Tiebreak on expansion_handle (memory:<id> / file:<path>), the stable identity, matching search_repo_files. Surfaced by the E3 task-kind-neutral test flaking under feature-unified runs. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…io (F3) Replace the flat 6000/stage budget with one that grows sublinearly with a defined task-size signal (task + localized-context tokens), floored so small tasks aren't starved and globally capped per run via the recall ledger so total brain tokens grow far slower than task size. TokenEconomy exposes the overhead ratio (brain tokens / total run tokens); on the small vs 5x-larger fixtures brain tokens grow <2x and the ratio strictly falls — the measured form of "cheaper on larger tasks." Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Bump workspace + inter-crate versions 0.9.0 -> 1.0.0 and document the v1.0.0 release: durable schema migrations, proof-of-value analytics, sqlite-vec semantic retrieval, proactive + cost-shrinking agent recall, install hardening, and polish (implemented subcommands, doctor --selftest, enforced clippy). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…1.1/W1.2) reset_projection no longer wipes events (and now also clears the previously- missed citations/conflicts); add rebuild_in_place that replays the events table into the derived tables without re-inserting (no duplication). Factor project_event out of apply_event. Foundation for dropping per-write run dirs. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
kimetsu brain rebuild now replays brain.db's events table in place, with a --from-traces flag (and an auto-fallback when the table is empty but on-disk traces exist) to recover history from legacy run traces. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
add_memory/propose/ingest/accept/invalidate/reject stop creating a TraceWriter run directory — their events persist to brain.db's durable events table (and the runs-table row, blame, usefulness are all kept). A brain-only .kimetsu/ no longer fills with per-write run dirs. Also fixes an orphan empty-dir on dedup. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Stop eagerly creating the runs/ dir on init; create the .kimetsu/ dir itself (needed for project.toml + brain.db). runs/ is now created lazily only by the agent pipeline. A brain-only install never grows a runs/ tree. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Lock the core invariant of the durable-events-log change: a pipeline run's events land in brain.db and survive rebuild_projection (which now replays the events table, not trace.jsonl). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add [embedder] enabled, [broker] ambient, and [kimetsu] use_user_brain project.toml fields (serde-default true) so every optional feature can be turned off durably, honored at runtime with env override > config > default. Existing/partial project.toml files load unchanged. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Wire three new BrainCommand variants (EmbedDaemon, Warm, Daemon) into main.rs with real handlers behind #[cfg(feature = embeddings)] and no-op lean stubs, completing the warm-embedder daemon CLI surface. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The UserPromptSubmit hook now tries the warm embedder daemon first (semantic retrieval) and falls back to floored-FTS when the daemon is disabled, unreachable, or cold. The `retrieval_path` field (daemon or fts_fallback) is recorded in the `context.served` telemetry event. Adds `ContextCapsule::wire_minimal` for constructing render-only capsules from daemon wire data, and a unit test covering the adapter. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Wire `kimetsu brain warm` to each harness startup event so the embedder daemon pre-warms when the harness starts: - Claude Code: SessionStart hook group (upsert_kimetsu_hook, idempotent) - Pi: kimetsuExec([brain,warm]) prepended in the session_start handler - OpenClaw: kimetsuExec([brain,warm]) called at plugin registration time - Codex: no session-start event; daemon warms lazily on first prompt (comment) Three new tests (claude_hooks_include_sessionstart_warm, pi_extension_ts_session_start_includes_warm, openclaw_plugin_ts_register_includes_warm) all pass; 122 total pass. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- gate brain_warm() on config.embedder.warm_on_start before resolving model - self-connect after Shutdown to unblock accept() without a manual nudge - treat PermissionDenied/AlreadyExists as benign Windows race exits (exit 0) - add 50ms backoff in accept-loop error arm to avoid busy-spin - spawn_daemon directly on retrieval miss instead of re-pinging via ensure_daemon Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…capsules Adds v1.0.0 cross-encoder reranking to the kimetsu-brain retrieval pipeline: - `Reranker` trait (Send+Sync, doc-order scores) + `StubReranker` (token-overlap based, deterministic, no-model-download) near existing `Embedder`/`StubEmbedder` - `FastembedReranker` (behind `#[cfg(feature="embeddings")]`) wrapping `Mutex<TextRerank>` for the 4 curated BGE/JINA models; sigmoid-normalizes raw logits, maps results back to document order via `.index` - `open_reranker_for_model`: ungated signature, gated body mirroring `open_embedder_for_model`; off/none/empty → None, lean → always None - `rerank_capsules` in context.rs: final-stage rerank by summary, overwrites score, stable-sort descending, floor filter, cap truncation, fail-open on error - 9 new tests (5 rerank_capsules + 4 StubReranker), all passing; clippy clean Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- config: add `embedder.reranker` field (default `jina-reranker-v1-turbo-en`, serde-defaulted so older configs load cleanly with the reranker on) - server: `DaemonState` gains `reranker: Option<Box<dyn Reranker>>`; `retrieve` over-fetches a 12-capsule pool when reranking, bumps budget to ≥6000 tokens, then applies `rerank_capsules` with a 0.30 floor - cli: `EmbedDaemonArgs` gains `--reranker`; `brain_embed_daemon` opens the reranker inside the timed block; new `resolve_daemon_reranker` reads config; `brain_warm` and `try_daemon_retrieve` pass the reranker through `ensure_daemon` / `spawn_daemon` - client: `REQUEST_BUDGET` 750 → 300ms (embed ~10ms + ANN + rerank-of-12 ~30–80ms fits; anything slower falls back to FTS by design) - test: `retrieve_with_stub_reranker_reorders_and_caps` — hermetic temp brain, NoopEmbedder + StubReranker, verifies the rust/tokio memory beats the python/django one and exactly 1 capsule returns Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…tually wired min_semantic_score defaulted to 0.0 AND was never populated from config in any production path, so the cosine floor was dormant. Default it to 0.35 and wire it from BrokerSection into the two embeddings-capable retrieval paths (MCP tool + warm daemon), same if-zero-then-config pattern as min_lexical_coverage. Set 0.0 in project.toml to restore keep-everything behaviour. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…tic|rerank Adds `kimetsu brain eval` which measures recall@2, recall@4, and MRR for three retrieval modes (lexical-FTS, semantic ANN+cosine, semantic+rerank) against a committed 42-memory/26-case fixture, giving a quantitative signal for threshold and ranking changes instead of vibes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ness Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three measured fixes from the latency investigation: - Bind the socket BEFORE loading models in the daemon entrypoint, so a redundant spawn exits in milliseconds instead of after a multi-second model load. - Windows: clear HANDLE_FLAG_INHERIT on the hook's std handles before spawning the daemon. The long-lived child otherwise inherits the harness's stdout pipe and holds it open forever, stalling the first prompt of a session until the host's hook timeout. - Default `[embedder] reranker` to "off": full-fidelity reranking costs ~0.5-1s (beyond the 300ms hook budget) and forcing it under budget via snippet truncation + smaller pools regressed eval recall@4 from 0.83 to 0.66 (below FTS). The default hook path is semantic + the 0.35 cosine floor (recall@4 0.90, MRR 0.91, ~230ms warm end-to-end); reranking stays as a quality-over-latency opt-in at quality-optimal settings (full summaries, pool 12). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds HuggingFace Hub download path for non-curated reranker models (jina-tiny, ms-marco TinyBERT-L-2-v2, ms-marco MiniLM-L-4-v2) and a --rerankers flag to `kimetsu brain eval` that measures load latency, per-query rerank latency (mean + max), recall@2/4, MRR, and noise suppression side-by-side for all four cross-encoder candidates. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…fault Benchmark (eval --rerankers, release build): jina-reranker-v1-tiny-en beats turbo on both quality (recall@4 0.896 / MRR 0.938 vs 0.833 / 0.875) and speed (~95ms vs ~137ms/query). On the real brain's long memories a pool-12 rerank still exceeds the 300ms hook budget (measured ~670ms+ totals = timeout fallbacks), and semantic+floor already matches tiny's recall@4 at ~230ms warm — so the default stays off and tiny becomes the documented opt-in recommendation. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The pool-6 experiment closed the loop: with FULL summaries, pool 6 matches pool-12 quality exactly on the eval fixture (recall@2/4 0.882/0.896, MRR 0.938, noise 0) at half the rerank latency (~44ms/query), and on the real brain a warm pool-6 jina-tiny rerank answers inside the 300ms hook budget (~265ms measured; totals 422-473ms end-to-end and descending). The earlier pool-shrink regression was the snippet truncation, not the pool. - default [embedder] reranker: "off" -> "jina-reranker-v1-tiny-en" - daemon RERANK_POOL: 12 -> 6 (quality-equal, half the cost) - eval gains --pool for pool-size sweeps Slower machines degrade gracefully to floored-FTS for the turn; set reranker = "off" to opt out. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…path
Two gaps surfaced by real-world usage:
- Inflected query words matched nothing: FTS prefix (benchmarked*) and the
IDF document-frequency LIKE both miss "benchmark", so the query's one
discriminating token got df=0/zero weight and the relevance floor went
blind ("how is kimetsu benchmarked?" surfaced off-topic memories that
merely shared "kimetsu"). Light query-side stemming (strip ing/ed/es/s,
keep >=4-char stems, haystacks raw) fixes FTS, lexical scoring, and IDF
at once. Eval fixture improves: FTS recall@4 0.715 -> 0.778, MRR 0.812 ->
0.875; semantic MRR 0.910 -> 0.931.
- retrieve_proactive (PreToolUse/PostToolUse recall) never got the lexical
floor, so an off-topic memory sharing a ubiquitous token with the command
line could take the single proactive slot. Wire min_lexical_coverage from
config like the other retrieval paths.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The brain's prescribed workflow (record lessons via kimetsu_brain_record, Stop-hook harvest cue) contradicted the default-deny write gate: agents were nagged into a hard-blocked call every session. Make the gate config-driven for the local stdio server via kimetsu.mcp_write_tools (default true, settable with `kimetsu config set kimetsu.mcp_write_tools false`). Precedence is a pure, unit-tested decision: env var when set always wins (both directions) > remote deny > local config (default allow). The remote server stays env-only default-deny — a cloned repo's project.toml is untrusted input and must never be able to enable writes (covered by a dispatch-level test using allowed_tools as the remote marker). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…/RAM grid Adds `kimetsu brain bench` subcommand (embeddings-gated; lean build prints a notice). Orchestrator mode spawns one child process per embedder×reranker combo for honest per-combo RSS measurement; each child writes a JSON file then exits. After all combos finish the orchestrator reads the JSON files, sorts by MRR, and writes bench/results/summary.md. Per-combo worker: loads embedder + reranker (measuring Windows WorkingSet RSS via K32GetProcessMemoryInfo), seeds a hermetic temp brain with KIMETSU_BRAIN_EMBEDDER set to the combo embedder so stored vectors match query vectors, then runs all 45 dataset cases with min_score=0.0/min_lexical_coverage=0.0 for unfiltered recall. Cargo.toml: adds Win32_System_ProcessStatus + Win32_System_Threading to windows-sys features for K32GetProcessMemoryInfo / GetCurrentProcess. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…nchmark Defaults chosen by kimetsu brain bench on real exported memories: the code-tuned embedder recovers oblique dev-phrased queries bge-small never gets into the candidate pool and injects ~4x less off-topic noise; TinyBERT reranks the pool in ~43ms (any reranker beats none; reranker-vs-reranker deltas are within noise at this corpus size). Docs: new "Retrieval models & benchmarking" section in HOW-KIMETSU-WORKS with the grid table + model-swap instructions (reindex required on embedder change), README results blurb. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…-remote Adds `kimetsu brain bench --remote` which benchmarks the kimetsu-remote HTTP MCP server with the same 100-case dataset.json used by the local bench. Measures per-case recall@2/4 and MRR (sequential), HTTP latency (sequential and 4-way concurrent), server-process RSS (warm + peak), across embedders. Writes per-embedder remote-<emb>.json and remote-summary.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… over-filtering The 0.35 cosine floor was calibrated on bge-family distributions; the new kimetsu-remote benchmark showed it killing relevant jina-v2 results outright on every production path (MRR 0.90 -> 0.77, recall@2 == recall@4). The config default becomes -1.0 = AUTO, resolved per session embedder: 0.35 for bge-family, disabled for others (jina-v2's own precision already keeps noise at ~1.2 capsules). Explicit non-negative config values are respected as-is. Re-run confirms jina-v2 remote recovers to MRR 0.903 / recall@4 0.966. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…lt jina-tiny Add a cross-encoder rerank stage to kimetsu-remote. The server now reranks every kimetsu_brain_context call with a configurable cross-encoder (--reranker, default jina-reranker-v1-tiny-en, off disables). Implementation: - kimetsu-chat: refactor kimetsu_brain_context into pub brain_context_tool() taking Option<&dyn Reranker>; None path (stdio) is byte-identical. Exports REMOTE_RERANK_POOL=6, REMOTE_RERANK_FLOOR=0.30 mirroring the embed daemon constants. - kimetsu-remote: add --reranker ServeArgs field (default jina-tiny), load once at startup into AppState::reranker (Arc<dyn Reranker>), intercept tools/call for kimetsu_brain_context in rpc.rs when the reranker is present (same pattern as the ingest interception). - Tests: brain_context_tool_with_stub_reranker_reorders_and_caps (unit), reranker_in_appstate_intercepts_brain_context (roundtrip). All 141 pass. - Docs: HOW-KIMETSU-WORKS §7a "Retrieval models on the server", README results blurb, CHANGELOG v1.0.0 entry. - Bench: re-ran 100-case remote benchmark WITH reranker; results: jina-v2-base-code MRR 0.906 (+0.002), bge-small MRR 0.909 (+0.008). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…guard Truthfulness audit: README + HOW-KIMETSU-WORKS quoted 18-memory-corpus metrics (0.938/0.966/43ms) as 100-memory results — replaced with the real 100-memory grid (0.914/0.949/132ms for the default combo, full table); CHANGELOG had sed-duplicated entries in every release section (deduped to v1.0.0), a stale "off" reranker default, pre-reranker remote numbers, and the old 750ms budget. Code audit (no criticals): rerank_capsules now fails open on a score-count mismatch from a custom reranker instead of silently dropping the unscored tail; the remote bench wraps its spawned server in a kill-on-drop guard so error paths can't orphan a live server holding the port and temp dir. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
fmt was never run during the retrieval/daemon/bench work (13 files reformatted). The lean --no-default-features --all-targets clippy flagged embeddings-only items compiled ungated: gate the daemon wire protocol (proto, PROTOCOL_MAJOR) and the bench RSS helpers behind the embeddings feature. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
CI clippy runs on Linux, where the Windows-only update preflight (decide_preflight_action, PreflightAction, the Scheduled outcomes, stopped_mcp counter) and the WMI/CSV process parsers are dead code. Annotate them with #[cfg_attr(not(windows), allow(dead_code/unused_mut))] — the established pattern here — instead of cfg-gating the fns, so cross-platform unit tests keep compiling everywhere. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…erver) Informational unmaintained notice, same class as the already-ignored paste advisory: transitive via axum-server (kimetsu-remote TLS), no first-party usage, no parent release without it, zero actual vulnerabilities. Re-evaluate when axum-server migrates to rustls-pki-types. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
rustls-pemfileis unmaintained.pastenotice: transitive via axum-server (kimetsu-remote in-process TLS), no first-party usage, no parent release without it, zero actual vulnerabilities.rustsec/audit-checkignore list with a documented re-evaluation trigger (axum-server migrating to rustls-pki-types).Test Plan
cargo auditlocally: only allowed informational warnings, exit 0🤖 Generated with Claude Code