@cachly-dev/mcp-server 0.10.65 → 0.10.67

package/CHANGELOG.md

@@ -1,653 +0,0 @@
-# Changelog – cachly SDK (mcp)
-
-**Language:** MCP (Model Context Protocol)  
-**Package:** `@cachly-dev/mcp-server` on **npm**
-
-> Full cross-SDK release notes: [../CHANGELOG.md](../CHANGELOG.md)
-
----
-
-## [0.10.65] – 2026-05-29
-
-### Flat-file head-to-head in the benchmark (the proof gets honest)
-
-- **Cachly-Bench now compares three rankers**, not two: a `flatfile` ranker
-  (naive term-overlap, no IDF, no length norm, **no quality signal**, recency
-  tiebreak) joins the BM25 `baseline` and `cachly`. The flat-file ranker is an
-  honest — in fact *charitable* — stand-in for "an LLM reading its own memory files."
-- **Result:** cachly wins the metrics an agent depends on — **Precision@1 +10.0 %,
-  MRR +4.4 % vs. flat-file** (and +22.2 % / +10.9 % vs. BM25). The report and BENCH.md
-  are honest about *why* flat-file's Precision@3 / Recall@3 look higher (it ranks the
-  whole corpus; we did not rig that away).
-- **CI-defended:** a new regression test asserts cachly ≥ flat-file on Precision@1
-  and MRR, and that the flat-file simulation isn't trivially weak (MRR > 0.5).
-
-## [0.10.64] – 2026-05-29
-
-### The three decisive metrics, now measurable
-
-- **`brain_metrics(instance_id)`** — new tool (101st). Reports the three metrics
-  that decide whether the Brain delivers value:
-  1. **Time-to-first-recall** — `born_at` (first learn) → `first_recall_at` (first
-     proven recall), both `SET NX` so only the first event wins.
-  2. **Recall-lift** — published Cachly-Bench headline (+22.2 % P@1), CI-defended.
-  3. **Team-knowledge-reuse** — % of proven recalls that used a *teammate's* lesson.
-- **Cross-author reuse tracking.** `smart_recall` gains an optional `author`. When
-  you recall a lesson written by someone else, cachly increments
-  `cross_author_recalls`, records the distinct reuse pair, and surfaces a
-  "👥 Team knowledge reuse" banner inline — the value only a shared brain delivers.
-
-## [0.10.63] – 2026-05-29
-
-### Graceful-degradation stability
-
-- New utilities: `withTimeout`, `scanKeys` (capped + timed-out scan), `normalizeGitPath`.
-- `skill_gaps` / `brain_coverage` / `team_expertise_map` / `brain_file_map` now scan
-  via `scanKeys` — a huge keyspace can no longer hang the agent turn (3 s cap + key cap).
-- `brain_from_git` normalizes git rename paths (`src/{old => new}/f.ts`) so file nodes
-  don't fragment on renames. Removed a dead person-node scan in `skill_gaps`.
-- +13 utils tests.
-
-## [0.10.62] – 2026-05-29
-
-### Hardened Phase 3 tools
-
-- `brain_who_knows`: reject empty/undefined `topic` (was crashing in `ckgSlug`); clamp `limit` to [1,50].
-- `brain_file_map`: filter empty path strings, handle non-array input gracefully.
-- +5 stability tests.
-
-## [0.10.61] – 2026-05-29
-
-### Phase 3C: 100 MCP tools milestone — zero-setup knowledge graph
-
-- **`brain_from_git` enhanced.** Now auto-builds Person nodes + File nodes from git
-  history. Every commit that brain_from_git ingests now upserts the author as a
-  PersonNode and each changed file as a FileNode, with `authored` and `touched` CKG
-  edges. Run once and your entire org-wide knowledge graph is populated retroactively
-  — no `author` fields needed in individual `learn_from_attempts` calls.
-  The git log format is updated to `--name-only` so file attribution happens without
-  extra git subprocess calls per commit.
-- **`skill_gaps()`** — new tool. Scans all lessons and surfaces knowledge blind spots:
-  - 🔴 **critical**: domains with ≥1 failure and 0 success lessons
-  - 🟡 **warn**: fewer failures than threshold but still unresolved  
-  - 🔵 **info**: domains with ≥3 lessons but no attribution (brain_who_knows can't help)
-  Private lessons are excluded from gap analysis.
-- **`brain_coverage()`** — new tool. Knowledge-coverage health score 0–100 based on
-  4 equally-weighted factors: lesson volume, success ratio, attribution completeness,
-  and team engagement. Also computes file coverage ratio vs `git ls-files` when run
-  inside a git repo. Use `skill_gaps` to find what to improve.
-- **+6 tests** → 400 total.
-- **100 MCP tools** (was 98). 🎯
-
-## [0.10.60] – 2026-05-29
-
-### Phase 3B: File knowledge map + team expertise overview + visibility scopes
-
-- **`brain_file_map(file_paths=[...])`** — before touching a file, see who has
-  worked on it (from `learn_from_attempts` author+file_paths history) and which
-  lessons reference it. Experts shown with medal rankings 🥇🥈🥉 + touch count +
-  recency. Related lessons shown with outcome + severity + author badges.
-- **`team_expertise_map()`** — full team skills matrix as a markdown table:
-  contributor · lesson count · top domains · last active. Zero setup: auto-built
-  from `author` fields. Useful for onboarding and knowledge-gap detection.
-- **`visibility: "private" | "team" | "public"`** on `learn_from_attempts`:
-  `private` lessons are never surfaced in `smart_recall` or team views — only
-  accessible via exact `recall_best_solution(topic=...)`. Default remains `"team"`.
-- **Resilient `smart_recall`**: `apiFetch` for instance info now has a null guard
-  so semantic search degrades gracefully when API is unreachable.
-- **+8 tests** covering visibility storage, private filtering in smart_recall,
-  brain_file_map expert attribution, and team_expertise_map → 394 total.
-- **98 MCP tools** (was 96).
-
-## [0.10.59] – 2026-05-29
-
-### Phase 3A: Org-wide knowledge graph — "Who Knows What?"
-
-- **`brain_who_knows(topic="...")`** — new tool (96th). Queries the org-wide
-  knowledge graph to find your team's top experts on any topic. Returns a ranked
-  list with lesson count, confidence %, primary domains, and last-active recency.
-  Medal rankings: 🥇 🥈 🥉 for top 3.
-- **Person nodes auto-built.** Every `learn_from_attempts(author="name", ...)` call
-  now upserts a `PersonNode` in the CKG and creates an `authored` edge from the
-  person to the lesson concept. No extra setup needed.
-- **File nodes auto-built.** `file_paths` in `learn_from_attempts` upsert `FileNode`
-  entries + `touched` edges from the author → file. Enables future "who worked on X
-  file?" queries.
-- **Author attribution in `smart_recall`.** Results now show `👤 author-handle`
-  inline whenever the matched lesson carries an author field.
-- **`+7` tests** covering `ckgUpsertPersonNode`, `ckgUpsertFileNode`, and the full
-  `brain_who_knows` flow including ranking order correctness.
-- **96 MCP tools** (was 95).
-
-## [0.10.58] – 2026-05-29
-
-### Zero lint warnings
-
-- Systematic unused-import cleanup across all handler files and index.ts
-- ESLint exits clean: 0 errors, 0 warnings
-- `eslint.config.js`: added `varsIgnorePattern: '^_'`
-
-## [0.10.57] – 2026-05-29
-
-### Critical: fixed broken npm entry point + version hygiene
-
-- **Fixed the package entry point.** When the shared `telegram-notify` package was
-  added to the TypeScript build, tsc began emitting to `dist/src/index.js`, but
-  `bin`/`main`/`start` still pointed at `dist/index.js`. Versions **0.10.50–0.10.52
-  shipped a non-existent entry point** (`npx @cachly-dev/mcp-server` failed). All
-  paths now correctly point to `dist/src/index.js`, verified by running the built
-  binary.
-- **Version re-sync.** `package.json`, `server.json` (MCP registry manifest, was
-  stuck at 0.10.23), `package-lock.json` (was 0.10.49), and `MCP_VERSION` are now
-  all `0.10.57`. Bumped ahead of the previously published 0.10.55/0.10.56 so
-  `latest` points forward again instead of regressing to 0.10.52.
-- **Accurate tool count (95).** README badge, `server.json`, `package.json`, and CLI
-  banners claimed 80/89 tools; corrected to the actual 95.
-
-## [0.10.52] – 2026-05-29
-
-- `smart_recall`: CKG traversal as a 3rd retrieval signal — surfaces lessons that
-  *fixed* causal-graph-similar problems even when vocabulary differs (🕸️ badge).
-- Cachly-Bench corpus expanded with a governance adversarial pair; headline lift
-  rose to **Precision@1 +22.2%, MRR +10.9%**.
-
-## [0.10.51] – 2026-05-29
-
-- `team_confirm`: human review (🛡️ senior / ✔️ peer) raises a lesson's recall
-  ranking — confirmed knowledge outranks unreviewed auto-learned entries.
-
-## [0.10.50] – 2026-05-29
-
-- `smart_recall`: unified keyword + semantic results into one hybrid-ranked list.
-- "Brain saved you here" banner surfaces estimated time saved per proven recall.
-
-## [0.10.49] – 2026-05-29
-
-- Persist JWT after device flow (telemetry showed zero authenticated calls after
-  restarts). Quality-aware reranking + Cachly-Bench (the moat proof) introduced.
-
-> Note: 0.10.39–0.10.48 and 0.10.53–0.10.56 were interim/parallel publishes; see
-> git history for details. 0.10.57 supersedes all of them.
-
----
-
-## [0.10.38] – 2026-05-25
-
-### `cachly status` — Brain health at a glance
-
-- New CLI command `npx @cachly-dev/mcp-server@latest status`: shows lessons, recalls, quality score, Brain level, team contributors (with count), and your personal invite link — all in one terminal card.
-
----
-
-## [0.10.37] – 2026-05-25
-
-### Viral wow moments — invite link + first-recall tweet button
-
-- **`cachly invite` fixed** — previously called a non-existent `/api/v1/team/invite` endpoint (always 404). Now calls `GET /api/v1/referral/me` and shows your personal referral link with pre-written Slack DM + tweet text. One command = shareable invite.
-- **First-recall email** — the "Your AI just remembered" email now includes a one-click `𝕏 Share this moment` tweet button and a `🤝 Share with your team` section with your personal referral link (shows only if referral code exists). The highest-value wow moment now has a viral exit.
-
----
-
-## [0.10.36] – 2026-05-25
-
-### ESLint flat-config (ESLint 10 / v9 API)
-
-- Added `eslint` v9 and `typescript-eslint` v8 as devDependencies.
-- Created `eslint.config.js` using ESLint flat-config API — replaces the legacy `--ext .ts` syntax that broke with ESLint 10.
-- Updated lint script: `eslint src --ext .ts` → `eslint src`.
-- Rules: `no-explicit-any` warn, `no-unused-vars` warns but ignores `_`-prefixed params.
-
----
-
-## [0.10.35] – 2026-05-25
-
-### Docs catch-up
-
-- Backfilled CHANGELOG entries for 0.10.25–0.10.34 (they were missing).
-- README now documents the `index` and `learn-git` CLI commands plus the PR-merge auto-learn GitHub Action tip.
-
----
-
-## [0.10.34] – 2026-05-25
-
-### `learn-git` CLI — auto-learn from commits (PR-merge ready)
-
-- New command `npx @cachly-dev/mcp-server learn-git [./repo] [--max-commits N]` runs `brain_from_git` over recent commits and fires the `brain_from_git` telemetry event. Building block for PR-merge auto-learning in CI.
-
----
-
-## [0.10.33] – 2026-05-25
-
-### Close the setup funnel
-
-- The setup wizard now fires `setup_completed` at the end of a successful run (with `instance_id` + JWT attribution), so the onboarding funnel `setup_started → setup_completed` is finally traceable.
-
----
-
-## [0.10.32] – 2026-05-25
-
-### Viral moments: shareable digest card + team first-briefing
-
-- `cachly digest` now prints a tweet-ready card with a pre-filled X/Twitter URL (lessons, recalls, tokens saved, top lesson, brain level).
-- **Team-virality**: when `session_start` runs for a user who has lessons from teammates but has never seen them, a one-time "Your team's AI brain has been briefing you" section surfaces exactly which teammate solved what.
-
----
-
-## [0.10.31] – 2026-05-25
-
-### Unblock npm publish — tests green
-
-- `vitest.config.ts` excludes `**/__tests__/e2e/**` from the unit-test suite (E2E tests require live credentials and were failing CI, blocking every publish since 0.10.27).
-- CLI with no args now writes the setup banner to **stdout** and exits 0 (was stderr + exit 1, invisible to callers).
-- `process.exit(0)` guarded with an `_isMain` check so importing modules never trigger early exit.
-
----
-
-## [0.10.30] – 2026-05-24
-
-### CLAUDE.md delivers a visible briefing
-
-- CLAUDE.md template now calls `session_start` (with `workspace_path`) as the first action of every conversation and `session_end` as the last, so the brain briefing is visible to the user instead of running silently.
-
----
-
-## [0.10.29] – 2026-05-24
-
-### Fix: recall counting + git bootstrap actually fires
-
-- `smart_recall` (the primary recall tool in CLAUDE.md) now also fires `recall_best_solution` when it returns lessons, so `BrainRecallCount` reflects real usage.
-- Auto-session-start now passes `process.cwd()` as `workspace_path` so the git auto-bootstrap introduced in 0.10.28 actually triggers.
-
----
-
-## [0.10.28] – 2026-05-24
-
-### Auto-bootstrap the brain from git on first session
-
-- On a first session with a `workspace_path`, `brain_from_git` runs automatically and its summary is appended inline to the `session_start` briefing — the brain learns from your repo history with zero setup.
-
----
-
-## [0.10.27] – 2026-05-24
-
-### Fix: session_start counts as a recall
-
-- `session_start` now fires `recall_best_solution` when the brain has lessons (not the first-session welcome), so the dashboard activation nudge clears and the first-recall email fires for active users.
-
----
-
-## [0.10.26] – 2026-05-14
-
-### Fix: brain_from_git + brain_predict telemetry
-
-- `brain_from_git` and `brain_predict` telemetry events were dead code paths and never fired; both now correctly send per-tool telemetry.
-
----
-
-## [0.10.25] – 2026-05-14
-
-### Per-tool telemetry events for all brain tools
-
-- Every brain tool (`session_start`, `session_end`, `learn_from_attempts`, `smart_recall`, `recall_best_solution`) now emits a dedicated telemetry event for accurate funnel + report breakdowns.
-
----
-
-## [0.10.24] – 2026-05-14
-
-### Funnel visibility + server.json version sync
-
-- `first_call_success` now includes `instance_id` for full per-tenant attribution in the `mcp_events` table.
-- `server.json` version bumped to 0.10.24 — Smithery and MCP directory indexes now show the current version.
-
----
-
-## [0.10.23] – 2026-05-14
-
-### Full activation funnel + UTM tracking
-
-- **`sendFunnelEvent` carries JWT**: all funnel events (`setup_started`, `device_flow_completed`, `first_call_success`) now include the JWT so the backend can resolve them to a tenant. Activation funnel is now fully attributable per user.
-- **`cachly share` UTM link**: the generated tweet now includes `https://cachly.dev?ref=share&utm_source=x&utm_medium=social&utm_campaign=cli-share` so shares can be traced back to new signups.
-
----
-
-## [0.10.22] – 2026-05-14
-
-### Activation telemetry, health Redis ping, join backup
-
-- **`first_call_success` telemetry**: fires once per process on the first successful brain/context tool call. Combined with `setup_started → device_flow_completed → first_call_success`, the full activation funnel is now visible.
-- **`cachly health` Redis PING**: section 3 now opens an actual Redis connection and sends `PING` — verifies end-to-end connectivity, not just instance metadata.
-- **`cachly join` corrupt JSON**: backs up corrupted config files to `.bak` before overwriting, matching the `mergeMcpConfig` behavior.
-
----
-
-## [0.10.21] – 2026-05-14
-
-### Stability hardening — corrupt configs, silent auth failures, upgrade validation
-
-- **`mergeMcpConfig` corrupt JSON**: instead of silently overwriting, backs up the corrupted file to `.bak` and prints a warning so the user knows their previous config was preserved.
-- **`getConnection` auth failure**: if the `/connection` endpoint returns 401/403, throws an actionable "run `cachly setup` to refresh credentials" error instead of silently attempting a password-less Redis connect that fails with cryptic `NOAUTH`.
-- **`upgrade` version validation**: validates that the npm registry response contains a valid semver string before comparing — prevents false "update available" from a malformed registry response.
-
----
-
-## [0.10.20] – 2026-05-14
-
-### Edge case hardening — silent failures eliminated
-
-- **`getConnection` status messages**: `failed`, `suspended`, and `pending_payment` now each get a specific actionable message instead of the generic "not reachable" fallback. Users know exactly what happened and where to go to fix it.
-- **`setup` write errors**: config file write failures now print a clear warning with the OS error and a recovery hint instead of silently succeeding.
-- **`join` write errors**: permission and unexpected filesystem errors are surfaced per-file instead of silently swallowed. Expected "editor not installed" cases are still quiet.
-- **Auto-provision telemetry**: if free-tier auto-provision returns a non-2xx response, a `auto_provision_failed` telemetry event fires so the team can detect new-user activation failures in real time.
-
----
-
-## [0.10.19] – 2026-05-14
-
-### Fix: accurate provisioning time estimate
-
-- `getConnection` no longer says "30 seconds" — the message now correctly says "1–3 minutes" for first-time provisioning. This prevents users from thinking their instance failed just because the first tool call waited 25 s and still saw `provisioning`.
-- Dashboard instance detail page provisioning banner updated to match ("1–3 minutes" instead of "under 60 seconds").
-
----
-
-## [0.10.17] – 2026-05-14
-
-### `cachly upgrade` — check for newer versions
-
-- **`cachly upgrade`** — fetches the latest version from the npm registry and compares with the running version. If outdated, shows current vs. latest and the one-liner to update (`npx @cachly-dev/mcp-server@latest setup`). No auth required.
-- Added `upgrade` to no-auth bypass list and splash screen.
-
-### Dashboard — Brain Badge widget
-
-- Instance detail page now shows a live badge preview + copy-ready Markdown and HTML snippets directly in the dashboard. No CLI needed.
-- Badge renders as `![cachly Brain](https://api.cachly.dev/api/v1/badge/<id>)` — one click to copy.
-
----
-
-## [0.10.16] – 2026-05-14
-
-### `cachly join` — accept a Brain invite in one command
-
-- **`cachly join <token>`** — the missing half of the invite loop. Accepts a shared Brain invite link:
-  1. Fetches invite info (public, no auth) — shows Brain name, tier, expiry
-  2. Prompts for confirmation
-  3. If no JWT: starts Device Flow sign-in (browser, 10 s)
-  4. Writes `CACHLY_BRAIN_INSTANCE_ID` into every detected editor MCP config (`.mcp.json`, `.cursor/mcp.json`, `.windsurf/mcp.json`, `.vscode/mcp.json`, `~/.claude/mcp.json`)
-  5. Suggests `cachly badge` as next step
-- Added `join` to no-auth bypass list
-- Splash screen updated to list `join` command
-
-### MCP Directory metadata
-
-- **`smithery.yaml`** — added `name` + `description` block for Smithery listing quality
-- **`glama.json`** — added keywords: `cline`, `zed`, `brain-from-git`, `failure-prediction`, `memory-crystals`, `readme-badge`, `incremental-learning`
-- **`server.json`** — version bumped to `0.10.15`
-
----
-
-## [0.10.15] – 2026-05-14
-
-### Brain Badge
-
-- **`cachly badge`** — new CLI command that outputs the Markdown + HTML snippet for embedding a live lesson-count badge in any README or website. Badge SVG served by `GET /api/v1/badge/:instanceId` (public, no auth, rate-limited 30 req/min, result cached 1 h in instance Valkey). Badge shows lesson count only — no topic names or content.
-- **Setup CTA** — both setup flows now print `npx ... badge` as the next suggested action after successful configuration.
-- **README** — Brain Badge section added with copy-paste Markdown snippet and explanation. Badge shield added to header.
-
----
-
-## [0.10.8] – 2026-05-14
-
-### Viral CLI — 6 new commands
-
-- **`cachly demo`** — zero-signup Brain preview. Reads local `.git/` history, classifies commits by type (fix/feat/security/deploy/refactor/perf), and renders an ASCII table showing commit count, date range, contributors, per-category bar charts, and sample security + bug-fix lessons. No account or token required — works before any sign-in.
-
-- **`cachly share`** — shareable ASCII stats card. Fetches live Brain stats (lessons, recalls, tokens/cost saved, Brain level) and generates a tweet-ready text block with `#AIMemory #ClaudeCode #Cursor` hashtags. Requires `CACHLY_JWT`.
-
-- **`cachly digest`** — weekly Brain summary. Displays a formatted summary table: lessons stored, recalls performed, tokens saved, estimated cost saved, Brain level, top team contributors, and top-recalled lessons. Includes a crontab snippet for automated Monday morning delivery. Requires `CACHLY_JWT`.
-
-- **`cachly invite [email]`** — team referral. Accepts an email address as argument or via interactive prompt, then POSTs to `/api/v1/team/invite`. Handles 409 (already a member) gracefully. Requires `CACHLY_JWT`.
-
-- **`cachly health`** — connection check. Validates JWT (decode + expiry), tests API reachability, checks Brain instance status, scans all editor MCP config files (Claude Code, Cursor, Windsurf, VS Code, Zed, Continue), and verifies the git post-commit hook. Exits 1 if any checks fail — CI-friendly.
-
-- **No-args splash screen** — running `npx @cachly-dev/mcp-server@latest` in a TTY now shows a violet-framed help card listing all commands instead of silently starting the MCP server.
-
-### brain_from_git — Incremental mode + progress feedback
-
-- **Incremental by default** — after the first full run, subsequent calls only process commits since the last ingested SHA (stored in Redis with 90-day TTL). Repeated calls on active repos are near-instant.
-- **Progress output** — reports progress to stderr every ~10% of commits (`⏳ Processing N/total…`) and prints a completion summary (`✅ brain_from_git complete: N/total commits ingested`).
-- New `incremental` boolean parameter (default `true`). Pass `false` to force a full reprocess.
-
-### README
-
-- Complete rewrite: `cachly demo` featured at the very top with embedded ASCII preview output, all 6 CLI commands documented, Zed manual-setup section added, competitor table updated, FAQ expanded with incremental `brain_from_git` Q&A.
-
----
-
-## [0.10.0] – 2026-05-08
-
-### Zero-Friction Onboarding
-
-- **Global Claude Code config** — `setup` now writes `~/.claude/mcp.json` (global) in addition to project-level `.mcp.json`. cachly is available in every Claude Code project after running setup once — no per-project re-configuration needed. Existing MCP servers in the global config are preserved (merge, not overwrite).
-
-- **Device Flow authentication** — No credentials required in MCP config. On the first tool call without a `CACHLY_JWT`, the server starts an OAuth Device Flow: the AI shows a one-click sign-in URL in chat, polls for completion, exchanges the token for a long-lived API key, and auto-provisions a Brain instance. Fully transparent to the user.
-
-- **Fully automatic sessions** — `session_start` and `session_end` are called automatically (on first tool call and on SIGTERM). Users never need to call them manually.
-
-- **Auto project indexing** — On session start, the project is indexed in the background if the last index is older than 24 hours. Zero extra calls.
-
-### CLAUDE.md Rules — Mandatory Brain Behavior
-
-The CLAUDE.md block written by `setup` now instructs the AI with binding rules (not suggestions):
-
-- `smart_recall` before every task
-- `causal_trace` before grepping/reading files when debugging
-- `brain_predict` before deploys, migrations, and dependency upgrades
-- `remember_context` before editing any file (WIP registry)
-- `learn_from_attempts` after every fix, deploy, or discovery
-
-### SEO & Search AI Optimization
-
-- **`llms.txt`** added — AI crawler standard file with complete tool catalog (89 tools + descriptions), competitor comparison, architecture overview. Indexed by Perplexity, ChatGPT Search, Claude web search.
-- **README** — FAQ section, competitor table (cachly vs mem0 vs MemGPT vs CLAUDE.md), `autopilot` tool documented prominently.
-- **GitHub topics** set: `mcp`, `mcp-server`, `ai-memory`, `persistent-memory`, `claude-code`, `cursor`, `github-copilot`, `causal-trace`, `developer-tools`, `typescript`.
-- **package.json** — description and keywords updated: `causal-trace`, `root-cause-analysis`, `failure-prediction`, `long-term-memory`, `llm-memory`, `claude-code-memory`.
-
-### Auth Module
-
-- `jwtExpiryMs`, `checkJwt`, `handleApiError` extracted from `index.ts` into `src/auth.ts`
-- 23 unit tests added for auth module (Vitest)
-- Publish workflow now runs tests before building
-
-### Bug Fixes
-
-- `safeJsonParse` applied to all `JSON.parse` calls in `advanced.ts` and `syndicate.ts`
-- Removed dead warm-up code
-
----
-
-## [0.9.3] – 2026-05-06
-
-### 🧠 10x Vision Phase 2 — Layers 3, 5 & 6
-
-#### Layer 3 — Multi-Agent Deliberation Chamber (MADC)
-
-New tool: **`madc_deliberate`** — When conflicting lessons exist for a topic, run deliberation between 6 specialist expert agents (InfraAgent, AuthAgent, DeployAgent, DatabaseAgent, DebugAgent, APIAgent). Each agent votes based on its CKG domain coverage. Unanimous → loser superseded. Split → contested flag. Resolutions stored as permanent CKG nodes.
-
-Auto-trigger: `learn_from_attempts` now detects contradictions, writes a `contradicts` CKG edge, and suggests `madc_deliberate` automatically.
-
-#### Layer 5 — Continuous Learning Stream (CLS)
-
-Two new tools for learning **without `session_end`**:
-
-- **`cls_ingest`** — Ingest learning signals from 3 sources: `git_commit` (commit message + files → CKG edges + lesson), `ci_outcome` (green after red → confirmed fix at 75% confidence), `ide_diagnostic` (compiler error + fix pair → instant lesson + CKG `fixes` edge).
-
-- **`cls_install_hooks`** — Outputs a ready-to-paste git post-commit hook + GitHub Actions step. Once installed: every commit and CI run feeds the brain automatically. Zero `session_end` required.
-
-#### Layer 6 — Federation Protocol (FedBrain)
-
-Four new tools for AI-to-AI knowledge transfer with cryptographic provenance:
-
-- **`fedbrain_contribute`** — Contribute a lesson to the global commons with a HMAC-signed knowledge certificate (domain fingerprint + confidence + outcome chain hash).
-
-- **`fedbrain_search`** — Context-weighted search: your domain fingerprint (from contributed lessons + `context_hints`) weights results from brains with matching tech stacks higher. Shows Gold Standard badges (10+ confirms).
-
-- **`fedbrain_confirm`** — Confirm that a syndicated lesson worked. Propagates back to global commons (increments `confirm_count`). Updates local CKG confidence. At 10 independent confirmations → 🏆 Gold Standard.
-
-- **`fedbrain_status`** — Dashboard: contributed lessons, confirm history, Gold Standard count, pending propagations.
-
-#### Summary
-
-| Layer | Tool(s) | What it does |
-|-------|---------|--------------|
-| 3 (MADC) | `madc_deliberate` | Expert agent voting on contradicting beliefs |
-| 5 (CLS) | `cls_ingest`, `cls_install_hooks` | Continuous learning without session_end |
-| 6 (FedBrain) | `fedbrain_contribute`, `fedbrain_search`, `fedbrain_confirm`, `fedbrain_status` | Federated knowledge with crypto certificates |
-
-Total tools: **63** (was 54 in v0.9.2)
-
----
-
-## [0.9.2] – 2026-05-06
-
-### 🧠 10x Vision Phase 1 — Layers 1, 2, 4 & 7
-
-- **Layer 1 (CKG):** Redis-backed Causal Knowledge Graph. Typed edges: `fixes`, `requires`, `co-occurs`, `causes`, `contradicts`.
-- **Layer 2 (BUE):** `learn_from_attempts` now writes CKG edges with Bayesian confidence `(s+1)/(t+2)`.
-- **Layer 4 (PPE):** `brain_predict` — CKG traversal + text fallback for failure prediction.
-- **Layer 7 (MCM):** Domain coverage map in `session_start` with confidence bars.
-- New: `brain_search` (BM25+ over all brain data), `ckg_inspect` (BFS with confidence bars).
-- `causal_trace` upgraded: CKG graph-first, text similarity fallback.
-
----
-
-
-
-### 🌐 Knowledge Syndication — The Global AI Brain
-
-The first collective intelligence layer for AI memory.
-Every instance contributes. Every instance learns. Privacy-preserving by design.
-
-#### New Tools
-
-- **`syndicate`** — Contribute a verified lesson to the **global Knowledge Commons**. Your identity is a one-way HMAC hash — completely anonymous. The lesson is immediately searchable by every other AI brain on the planet. Call this after any `learn_from_attempts` that is worth sharing: critical bugs, deployment gotchas, architecture discoveries. This is how individual knowledge becomes collective intelligence.
-
-- **`syndicate_search`** — Search the **global Knowledge Commons** for solutions discovered by the entire community. Results ranked by `confirm_count` (trust score) then recency. Use this *before* debugging any unknown issue — someone in the global brain likely solved it already. Returns: topic, what worked, what failed, trust bar `████████░░ ×47`.
-
-#### Privacy Design
-- Contributors identified only by HMAC-SHA256 of user_id — irreversible, not linkable to any identity
-- Absolute file paths stripped from all content before storage
-- Community flagging: lessons with 3+ flags are hidden globally
-- Trust scoring: `confirm_count` rises as independent instances verify a lesson works
-
----
-
-## [0.6.1] – 2026-05-04
-
-### 🧠 Cognitive Cache — v0.6 Major Feature
-
-The first AI Memory system with **reasoning over its own knowledge**. No other cache has ever done this.
-
-#### New Tools
-
-- **`memory_consolidate`** — Weekly garbage collector for your AI Brain. Detects contradictions (same topic, different outcomes), merges duplicate lesson clusters, flags stale entries (0 recalls in N days). Like `git gc` for knowledge. Returns a full consolidation report with before/after counts.
-
-- **`brain_diff`** — `git log` for your AI Brain. Shows exactly what changed since a given time window (`"7d"`, `"30d"`, ISO-8601). New lessons, updated lessons, recalled lessons. Perfect for weekly reviews: *"What did my AI learn this week?"*
-
-- **`causal_trace`** — Root Cause Analysis through memory. Given a problem description, scores all lessons by relevance, reconstructs the failure chain (root → intermediate → symptom), and surfaces the exact solution that worked before. *"auth breaks after restart"* → root: `k8s:namespace-terminating` → via: `keycloak:jwks-race` → fix: `PollUntilContextTimeout 3min`. No other system can do this.
-
-- **`knowledge_decay`** — Confidence scoring per lesson. Age × recall-frequency × outcome = decay score (0–100%). Visual bars: `████░░░░░░ 40%`. Lessons recalled recently score high; 90-day-old untouched lessons score low. Run before a big refactor to know which knowledge to trust.
-
-- **`autopilot`** — Generates a `CLAUDE.md` / `copilot-instructions.md` that turns any AI (Claude, Cursor, Copilot, Windsurf, Gemini) into a self-managing Brain operator. No manual `session_start`, `learn_from_attempts`, or `session_end` calls ever again. One command. Every AI. Forever.
-
----
-
-## [0.5.80] – 2026-05-01
-
-
-
-### Added
-
-- **RTL language support (Arabic, Hebrew)** — word-level tokenization with Unicode ranges U+0590–U+05FF (Hebrew) and U+0600–U+06FF (Arabic); Arabic light stemmer strips definite article `ال` and single-char prefix particles (`و`,`ب`,`ل`,`ف`,`ك`)
-- **Arabic and Hebrew stopwords** — ~60 high-frequency function words per language added to the STOPWORDS set
-- **Romanization matching** — katakana segments now additionally emit Hepburn romaji tokens at index time (e.g. `デプロイ` → `depuroi`), so users can query Japanese docs using romaji
-- **Full katakana→romaji converter** — handles digraphs (シャ→sha, チェ→che, ファ→fa), voiced consonants, geminate consonants (ッ), long vowel marks (ー), and loanword combinations
-- **Cross-language retrieval** — 130+ tech term synonyms spanning EN↔JA↔ZH↔KO↔AR↔HE; searching `deploy` now finds docs containing `デプロイ`, `部署`, `배포` and vice versa; applies bidirectionally at tokenize time (zero runtime overhead)
-- **73 unit tests** all passing — new test suites for `katakanaToRomaji`, `arabicLightStem`, `expandCrossLingual`, RTL tokenization, and cross-lingual expansion
-
----
-
-## [0.5.36] – 2026-04-22
-
-### Added
-
-- **Roadmap tools** — `roadmap_add`, `roadmap_update`, `roadmap_list`, `roadmap_next` for persistent project planning inside the Brain
-- `session_start` now shows open roadmap items automatically
-
----
-
-## [0.5.35] – 2026-04-20
-
-### Added
-
-- **CJK language support** — Chinese (Simplified + Traditional), Japanese, Korean
-- Character bigram extraction for CJK Unicode ranges
-- ~140 CJK stopwords (Chinese particles, Japanese hiragana particles, Korean postpositions)
-
----
-
-
-
-### Added
-
-- **`setup` command** — interactive zero-arg CLI wizard (`npx @cachly-dev/mcp-server setup`):
-  - Reads `CACHLY_JWT` from env or prompts interactively via readline
-  - Fetches instances from API; presents list when multiple exist
-  - Auto-detects installed editors (Cursor, Windsurf, VS Code, Continue.dev) by checking for their config directories
-  - Writes the correct MCP config file for each detected editor in one step
-  - Always writes/updates `CLAUDE.md` (idempotent via `<!-- cachly-brain-start/end -->` markers)
-- **`init` command** now idempotent — re-running `npx @cachly-dev/mcp-server init` updates the brain block in `CLAUDE.md` instead of appending a duplicate
-- Shared helpers: `buildMcpConfig()`, `buildClaudeMdBlock()`, `writeClaudeMd()` — used by both `setup` and `init`
-
-### Fixed
-
-- `init` no longer duplicates the brain block in `CLAUDE.md` when run multiple times
-- Correct package name `@cachly-dev/mcp-server` used consistently (was `@cachly-dev/mcp` in generated configs)
-
----
-
-## [0.1.1] – 2026-04-07
-
-### Fixed
-
-- Broken `index_project` tool schema – properties were accidentally placed outside the `TOOLS` array
-- Unused `openai` variable removed from `cache_warmup` handler
-- `readdir` type mismatch (`Dirent<string>` vs. `NonSharedBuffer`) fixed
-- Version bumped to `0.3.0` in server metadata
-
----
-
-## [0.1.0] – 2026-04-07
-
-Initial release.
-
-### Added
-
-- MCP tool: `cache_set` – store a value with optional TTL
-- MCP tool: `cache_get` – retrieve a cached value
-- MCP tool: `cache_delete` – remove a key
-- MCP tool: `semantic_search` – vector-similarity lookup for LLM response caching
-- MCP tool: `cache_clear` – flush namespace or entire cache
-- Compatible with Claude Desktop, Cursor, Windsurf, and any MCP-capable host
-- API-key-based authentication
-- EU data residency (German servers, DSGVO compliant)
-
-### Known limitations
-
-- ~~Streaming tools not yet supported~~ (tracked for a future release)
-
----
-
-## [Unreleased]
-
-See [../CHANGELOG.md](../CHANGELOG.md) for upcoming features.
-

package/Dockerfile

@@ -1,16 +0,0 @@
-FROM node:22-alpine AS builder
-WORKDIR /app
-COPY package*.json tsconfig*.json ./
-RUN npm ci --ignore-scripts
-COPY src/ ./src/
-RUN npm run build
-
-FROM node:22-alpine
-WORKDIR /app
-COPY package*.json ./
-RUN npm ci --omit=dev --ignore-scripts
-COPY --from=builder /app/dist/ ./dist/
-ENV NODE_ENV=production
-ENV PORT=3000
-EXPOSE 3000
-ENTRYPOINT ["node", "dist/index.js"]