v0.10.111

Added

• Org-wide ROI in cache_org_stats — calls the new GET /api/v1/orgs/:id/savings endpoint and appends org-aggregated savings: instance count, total hits, hits last 24h, total + monthly USD savings, top-5 per-instance breakdown. Falls back to Redis-only output when the endpoint is unavailable or the caller is not an org member.
• GET /api/v1/orgs/:id/savings (Go API) — sums SavingsEstimate across all instances with org_id = :id. Caller must be an accepted org member (403 otherwise). Instances whose stats fail are skipped with a warn log.
• src/ci-integration/brain-from-ci-gitlab.yml — GitLab CI integration: two .post jobs (on_success/on_failure, since GitLab does not expose final pipeline status to a running job), allow_failure: true, same push-ci-outcome.mjs helper.
• .github/workflows/brain-ci-feedback.yml — dogfooding: cachly's own repo now pushes every CI outcome to the Brain (requires CACHLY_JWT + CACHLY_BRAIN_INSTANCE_ID repo secrets; skips cleanly until set).

Fixed

• GitHub Action template — removed job-level if referencing secrets.* (the secrets context is not available in job-level if expressions; the workflow would fail to parse). PREV_STATUS now derived from run_attempt > 1 instead of previous_attempt_url.

---

See the CHANGELOG for full details.

v0.10.107

See CHANGELOG.md for details.
See the CHANGELOG for full details.

v0.10.104

Added

• brain_briefing — flips the Brain from pull-only to push. Call it on file_open (with the file path), pr_open (with the PR title/body), or deploy (with a short description) and the Brain proactively surfaces up to 5 ranked warnings — confidence, severity, what to watch, and a known fix — before anything breaks. Returns an overall risk_level (low/medium/high) derived from the strongest match. Optional threshold (default 0.6) tunes signal-to-noise.
• New API endpoint POST /api/v1/instances/:id/briefing — tokenises the event context (camelCase-aware, same tokeniser as PR scan), matches it against the instance's lessons above the confidence threshold, and ranks results.

---

See the CHANGELOG for full details.

v0.10.85

See CHANGELOG.md for details.
See the CHANGELOG for full details.

v0.10.84

See CHANGELOG.md for details.
See the CHANGELOG for full details.

v0.10.83

See CHANGELOG.md for details.
See the CHANGELOG for full details.

v0.10.82

See CHANGELOG.md for details.
See the CHANGELOG for full details.

v0.10.81

Onboarding-Bench (time-to-first-recall, measured) + external dashboard plan

Measuring the v0.10.80 starter-corpus impact. Recall-lift was already proven
(npm run bench); this release measures the other decisive metric —
time-to-first-recall — as a cold-start hit rate through the real search engine.
New: npm run bench:onboarding (src/bench/onboarding-bench.ts)

• 16 realistic first-session queries, phrased the way a frustrated developer types
  them (NOT the lesson's topic slug), each mapped to a starter topic.
• Two scenarios over the real keyword search engine: cold (empty Brain) vs
  seeded (starter corpus, which auto-seeds on first session when git history is empty).
• Result: first-query hit@1 0% → 87.5%, hit@3 0% → 100%, MRR 0% → 93.8%, answered 0% → 100%.
• Interpretation: an empty Brain returns nothing on the first query (user must do work
  • learn before any recall is possible — time-to-first-recall spans a full work cycle);
    the seeded Brain hits immediately (collapses to seconds). Documented in BENCH.md.
    5 new CI-guard tests (src/__tests__/onboarding-bench.test.ts): cold must answer
    0%; seeded must keep hit@3 ≥ 90%, hit@1 ≥ 70%, MRR ≥ 80%; seeding must lift every
    metric; every cold-start query maps to a real starter topic; the bench exercises the
    whole corpus (no flattering subset).
    New: PLAN-DASHBOARD.md — complete build spec for the external Team-Knowledge-Reuse
    dashboard (cachly-insights, a separate repo): grounded in the actual funnel events and
    cachly:stats:* keys the MCP server already emits. Covers data sources, architecture,
    TimescaleDB schema, the five dashboard views (activation funnel, TTFR with seeded-vs-cold
    cohort split, per-org reuse, brain health, marketplace), GDPR constraints, repo bootstrap
    steps, milestones, and the one cross-repo prerequisite (resolved telemetry stream in the
    cachly API).
    No new tools (still 113). Build: clean tsc. Tests: 567 passing (14 suites). Lint: 0 warnings.
    See the CHANGELOG for full details.

v0.10.80

Starter corpus — first recall in seconds, attacking time-to-first-recall

The onboarding gap: a brand-new Brain (fresh repo, shallow clone, no fix-commits in git history) returned nothing on the user's very first smart_recall — so time-to-first-recall never started ticking and the "aha" moment was delayed by a whole session.
New: curated starter corpus (src/starter-corpus.ts) — 16 universal, high-value, stack-agnostic engineering lessons with proven fixes: Docker layer cache, git force-push safety, flaky-test timing, ESM/CJS interop, JWT clock skew, Postgres migration locks, Redis eviction policy, K8s OOM limits, CORS preflight, dotenv precedence, unhandled rejections, SQL N+1, HTTP retry idempotency, cache stampede, secrets-in-logs, TLS cert expiry. No secrets, no PII, no project-specific paths.
New tool: brain_seed_starter(instance_id, topic_filter?, force?, dry_run?)

• Seeds the corpus so the very first query hits. Tagged source:"starter", never overrides your own lessons, idempotent (won't double-seed without force), stamps born_at so the metric starts counting.
• topic_filter matches by topic OR tag; dry_run previews; force re-seeds + overwrites.
  Auto-seeding on first session: when session_start auto-bootstraps from git history and that yields zero lessons (new/empty repo), the starter corpus is seeded automatically — the user's first real query returns a useful answer instead of an empty brain. The empty-brain welcome also surfaces the brain_seed_starter command explicitly.
  113 MCP tools (up from 112).
  18 new tests (src/__tests__/starter-seed.test.ts): corpus integrity (unique topics, required fields, no-PII scan), seeding to best+history keys, source:"starter" tagging, born_at stamping, idempotency, force re-seed, user-lesson protection, topic_filter by topic + tag, no-match path, dry_run, and — critically — the real keyword search engine surfacing the right seeded lesson for natural queries ("docker build slow", "pod OOMKilled", "jwt token rejected").
  Build: clean tsc. Tests: 562 passing (all 13 suites). Lint: 0 warnings.
  See the CHANGELOG for full details.

v0.10.79

Brain Marketplace — full shareable Brain lifecycle + stress tests

Three more MCP tools completing the share/import lifecycle:

• brain_share_list(instance_id) — lists all Brain snapshots you've shared, with share IDs, lesson counts, visibility, and creation date. Checks the cachly API first; falls back to the local provenance log.
• brain_unshare(instance_id, share_id) — revokes and permanently deletes a public share. Link goes dead immediately. Gracefully removes from local log when API is unreachable.
• brain_discover(query?, topic?, limit?) — searches the cachly Brain marketplace for publicly shared Brain snapshots. Returns ranked results with lesson counts, topics, import counts, and a one-line brain_import command. Shows a "marketplace coming soon" message with alternative paths when the API isn't live yet.
  publish CLI command:
• npx @cachly-dev/mcp-server@latest publish [--public] [--title "My Patterns"] — creates a public Brain snapshot and prints a formatted card with the share URL + import command. --public makes it discoverable in the marketplace; default is unlisted (link-only).
  112 MCP tools (up from 109).
  24 new tests covering:
• brain_share_list: empty state, API results, local-log fallback, missing args
• brain_unshare: success, 404, ECONNREFUSED local-log-only removal, missing args
• brain_discover: results display, topic filter, empty state, marketplace-not-live fallback
• Full lifecycle integration: share → list → unshare in a single test
• team_expertise_map stress test: 50 contributors, top_n cap, sort order, empty/single contributor, 0-lesson contributor, default cap
• brain_who_knows regression: multi-contributor ranking, inbound edge traversal
• brain_coverage regression: high score with attribution, low score with only failures
  Build: clean tsc. Tests: 544 passing (all 12 suites). Lint: 0 warnings.
  See the CHANGELOG for full details.