Comparing changes

…e dist-count gate - user-guide: semantic search section (enable/test/status/disable, costs, fusion principles, measured 33%->0%) and right-to-be-forgotten section (archive-vs-forget rule of thumb, dual anti-resurrection barriers) + quick-reference card entries - README.en: semantic + forget feature sections, tests 125->185, roadmap cleaned (semantic shipped); README.zh roadmap cleaned likewise - pack-e2e: assert installed dist file count <= 60 — regression gate for the orphan-chunk incident (0.3.0 shipped 94 dead files; gate caught at 37)

…gn v1, 3 review rounds, 9 revisions) - dual-track: vector similarity (semantic-enabled; active anchor embedded on-the-fly, never persisted) + topics/tags/files overlap fallback; fused output with explainable reasons per item - stateless by design: no new tables/config; never writes session_links (anti-echo-chamber, documented ruling) - contract invariant upgraded: exact 16-tool list + name blacklist scan (delete/remove/forget/purge fails) — the real intent of 'no AI deletion' no longer tied to a count - scope does NOT gate recommendations (navigation aid) — pinned by test with try/finally cleanup (fixture pollution root-caused: confirmSession overwrites topics, so the spec seeds its own overlap pair) - tests: 7 unit (scoring hierarchy/active candidates/isolation/hard cap/ cosine order/on-the-fly embed/fusion) + 2 contract; 194 green total - real-model acceptance 4/4 (top1 plausibility, scripts/related-real-verify.ts); small-corpus tail noise honestly documented in design §9

- README second tool table (zh) missing related row (String.replace only replaces first occurrence); README.en row was silently never inserted (pattern mismatch, script only validated zh patterns) — both fixed with miss-detection - CLI --limit abc produced NaN -> silently empty results; now validates and exits 2 - user-guide quick-reference card gained srelay related - design-related §9: score is dual-scale (cosine vs overlap points), ordering safe by construction, cross-group comparison meaningless — documented

Self-review round 2 caught: failedSkip filtering happened AFTER the SQL LIMIT, so poison pills at the head of the queue (recent last_event_at) consumed the whole per-cycle budget — newly confirmed sessions starved until process restart. Fix: fetch 2x limit when skip-set is non-empty, filter, then slice. markSemanticFailed() exported as test/ops hook. C6 regression test proven by sabotage (red with old code: expected 1 to be 2). 195 green.

…ssertions (release 0.4.0) - service.ts: macOS launchd LaunchAgent + Linux systemd user unit (with best-effort linger), aligned with Windows registry run key semantics (login autostart, no crash-restart); uninstall/status symmetric - CI verification: macOS runner does a real install->status->uninstall cycle (launchd is the macOS init); Linux content assertions only (runner user-systemd unreliable, documented) - check-artifacts.mjs (npm run check:artifacts, wired into CI test job and prepublishOnly): MCP tool names fully covered in docs, no stale tool-count wording, CHANGELOG top version matches package.json — hardening from the two release-audit incidents - watch.ts slimmed to daemon entry; windows implementation unchanged - tests: 6 service cases (content assertions all platforms, live launchd cycle darwin-only); 200 green + 1 platform-skipped

Navigation Menu

Repository navigation

Comparing changes

Open a pull request

Commits on Sep 6, 2026

This comparison is taking too long to generate.

Uh oh!