Claude Synergy

A local, queryable mirror of every Anthropic + adjacent AI dev tool changelog — plus a curated Synergy layer describing cross-product workflows — so the LLM agent inside the harness knows what the harness can do.

$ hk query redact
2026-05-11  anthropic-cli@1.7.1            [changed]  redact api-key headers in debug logs
2026-05-11  anthropic-sdk-java@2.31.0      [changed]  redact api-key headers in debug logs
2026-05-11  anthropic-sdk-go@1.42.0        [changed]  redact api-key headers in debug logs
2026-05-07  anthropic-sdk-typescript@0.95.1 [changed] redact api-key headers in debug logs

4 results

A single FTS query surfaces a coordinated cross-SDK security fix that no individual changelog flagged as a CVE. That's the killer demo: patterns emerge when every changelog is side-by-side.

Repo: github.com/mcp-tool-shop-org/claude-synergy

The problem

Claude Code CLI ships ~daily. Claude API ships almost as often. SDKs ship per-CLI-release. Claude Design, Cowork, Chat, and Mobile feed through the unified Help Center. The MCP ecosystem ships ~200-300 new servers per week. Plus there are 7 major AI dev tool surfaces (Cursor, Aider, Continue, Copilot, Cody, Windsurf) all shipping their own changelogs at their own cadences.

The LLM agent inside any one of these has a frozen training cutoff. The gap widens every day. Features ship that the agent doesn't know exist. Bugs get fixed that the agent still works around. Env vars and flags get added that the agent never suggests. Cross-product workflows that compose multiple surfaces remain undiscovered.

This repo closes the gap. The Synergy section makes it a product instead of a mirror.

What's here

claude-synergy/
├── products/                # 44 product directories (1,186 release files)
│   ├── claude-code/             # Anthropic CLI — 117 releases
│   ├── claude-agent-sdk-{python,typescript}/  # Agent SDKs
│   ├── anthropic-sdk-{python,typescript,go,java,csharp,ruby,php}/  # 7 language SDKs
│   ├── claude-api/              # Platform release notes
│   ├── anthropic-apps/          # Design / Cowork / Chat / Mobile (Help Center feed)
│   ├── claude-code-action/      # GitHub Action
│   ├── anthropic-cli/           # `ant` CLI
│   ├── mcp-{python,typescript,go,java,csharp,kotlin,ruby,swift,rust,php}-sdk/
│   ├── mcp-{spec,inspector,registry,mcpb,conformance}/
│   ├── cursor/                  # RSS feed
│   ├── aider/                   # raw HISTORY.md
│   ├── continue-{dev,cli}/      # GH releases
│   ├── cody-enterprise/         # filtered Sourcegraph RSS
│   ├── github-copilot/          # HTML scrape (github.blog)
│   ├── vscode-copilot-chat/     # HTML scrape (code.visualstudio.com)
│   ├── windsurf/                # Playwright fetcher (CSR-only changelog)
│   ├── skills/                  # Anthropic Skills catalog
│   └── plugins-{official,community,knowledge-work}/  # Plugin marketplaces
├── synergies/               # 12 curated cross-product workflows
├── src/                     # TypeScript implementation
├── test/                    # 508 tests (unit, integration, regression, smoke)
├── data/claude-synergy.db   # SQLite database (created by `hk init`)
├── schema.sql               # Tier 2a tables (products, releases, changes, entities, FTS5, …)
├── schema-vec.sql           # Tier 2b tables (chunks, chunks_vec, chunks_fts)
├── SOURCES.md               # 5-tier source landscape with fetch strategies
└── URGENT_FINDINGS.md       # 23 actionable items surfaced from the corpus

Live numbers (as of v1.2.1): 44 products / 1,171 release files / 6,573 changes / 1,260 entities / 12 synergies / 519 tests / 13 MCP tools / 17 CLI commands. (Corpus refreshed via sync_now on 2026-05-24.)

Status — all tiers shipped

Roadmap beyond v1.3: tracked in URGENT_FINDINGS.md and issues.

Security & data model

This tool runs locally. Data touched: a derived SQLite database and markdown release files — all re-creatable. Network: outbound HTTPS only when you run hk fetch/hk sync (GitHub API, RSS feeds, changelog pages, MCP registries) or hk embed with a remote provider (Voyage, Cohere). Secrets: reads GITHUB_TOKEN, VOYAGE_API_KEY, COHERE_API_KEY, ANTHROPIC_API_KEY from env vars — never logged, never stored to disk. No telemetry. See SECURITY.md for reporting policy.

Install

git clone https://github.com/mcp-tool-shop-org/claude-synergy
cd claude-synergy
pnpm install
pnpm build       # produces dist/cli.js + dist/mcp-server.js
npm link         # makes `hk` and `claude-synergy-mcp` available globally

For dev without building, use npx tsx src/cli.ts ... directly. pnpm 10 quirk: pnpm dev swallows CLI flags after --; use npx tsx for development.

CLI surface — 17 commands

# DB lifecycle
hk init                              # create DB with schema
hk ingest                            # parse products/*/releases/*.md → DB (idempotent)
hk embed                             # generate chunks + embeddings (sqlite-vec)
hk fetch [--product X]               # incremental pull from sources
hk sync                              # combined fetch → ingest → embed (cron-friendly)
hk seed-markers                      # one-time setup after initial corpus

# Search
hk query "managed agents"            # FTS5 keyword search (+ --until <date>)
hk hybrid "credential exfiltration"  # FTS5 + vec hybrid via RRF (+ --rerank, --until)

# Windowed change browsing
hk diff [product] --since 7d         # what changed in a window, grouped by product+version
hk breaking --since 30d              # filter-browse of breaking changes (no search term)

# Entity lookups
hk env-var CLAUDE_CODE_WORKFLOWS     # when introduced + history
hk command code-review               # slash command + rename history
hk model claude-opus-4-7             # model launch + mentions across products
hk cve CVE-2025-66414                # CVE references in corpus

# Browsing
hk latest [--product X] [--limit N]  # recent releases (+ --since <date>)
hk products                          # list all 44 with counts
hk top env_var                       # most-mentioned by entity type
                                     # (env_var, slash_command, cli_option,
                                     #  model_id, beta_header, cve, ghsa,
                                     #  hook_event, setting_key)

New in v1.1: hk diff and hk breaking answer "what changed lately?" without needing a search term. Date bounds are uniform: every browsing command takes --since and --until in YYYY-MM-DD (or full ISO 8601), or a relative form (7d, 2w, 3m, 1y).

Example workflows

Find when a Claude Code env var was introduced:

$ hk env-var CLAUDE_CODE_WORKFLOWS
env var CLAUDE_CODE_WORKFLOWS — 1 mention:

2026-05-21  claude-code@2.1.147  [added]
  Added the `Workflow` tool for deterministic multi-agent orchestration.
  It is off by default — set `CLAUDE_CODE_WORKFLOWS=1` to enable

Track a cross-SDK breaking change:

$ hk query TodoWrite --limit 5
2026-05-15  claude-agent-sdk-python@0.2.82       [breaking]   Headless and SDK sessions now use Task tools...
2026-05-14  claude-agent-sdk-typescript@0.3.142  [breaking]   Headless and SDK sessions now use Task tools...
2026-05-08  claude-agent-sdk-typescript@0.2.136  [deprecated] Deprecated TodoWrite tool...

Plan a model migration:

$ hk model claude-opus-4-20250514
model id claude-opus-4-20250514 — 2 mentions:

2026-04-14  anthropic-sdk-python@0.94.0  [deprecated]
  Deprecation of the Claude Sonnet 4 model and the Claude Opus 4 model,
  with retirement on the Claude API scheduled for June 15, 2026...

Semantic search across the whole corpus:

$ hk hybrid "credential exfiltration" --limit 3
2026-03-25  claude-code@2.1.83  [added]          vec#5 rrf=0.0154
  Added `CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1` to strip Anthropic and
  cloud provider credentials from subprocess environments...

The query never says "env_scrub" — vec surfaces it by semantic similarity. Pure FTS5 misses it entirely.

What changed in claude-code this week:

$ hk diff claude-code --since 7d
claude-code@2.1.147  2026-05-21  (3 changes)
  [added]     Added the `Workflow` tool for deterministic multi-agent orchestration.
  [changed]   Slash commands now lazy-load until first invocation.
  [fixed]     Race condition in MCP server discovery on Windows.

claude-code@2.1.146  2026-05-19  (1 change)
  [fixed]     Restored `--debug` flag accidentally removed in 2.1.144.

Browse breaking changes across the corpus:

$ hk breaking --since 30d --limit 5
2026-05-15  claude-agent-sdk-python@0.2.82       Headless and SDK sessions now use Task tools by default.
2026-05-14  claude-agent-sdk-typescript@0.3.142  Headless and SDK sessions now use Task tools by default.
2026-05-08  anthropic-sdk-go@1.42.0              Removed deprecated `client.Beta()` namespace.
2026-04-29  cursor@0.49.0                        MCP server config moved from `cursor.json` to `.cursor/mcp.json`.
2026-04-22  windsurf@1.10.0                      Removed `cascade.run` JSON-RPC method.

No search term needed — hk breaking is the answer to "did anything load-bearing change recently?"

MCP server — give your agents access to this corpus

claude-synergy-mcp exposes 13 tools over stdio. Wire into Claude Code (or any MCP host) via ~/.claude/.mcp.json or your project's .mcp.json:

{
  "mcpServers": {
    "claude-synergy": {
      "command": "claude-synergy-mcp",
      "env": {
        "CLAUDE_SYNERGY_DB": "/path/to/claude-synergy/data/claude-synergy.db"
      }
    }
  }
}

For GitHub Copilot's .vscode/mcp.json, use the servers wrapper instead of mcpServers (see synergy 12).

Tools exposed:

The v1.1 tools mirror hk diff / hk breaking and the version-comparison workflow that previously required scripting. The v1.2 sync tools close the seam where a session could query the corpus but not refresh it — sync_status reports staleness, sync_now runs the pipeline. See handbook → MCP server for full input schemas.

Sources — 5 tiers, 7 fetcher strategies

Full landscape in SOURCES.md.

• Tier 1 (GitHub Releases) — gh api repos/<owner>/<repo>/releases for 23 products including Anthropic SDKs (7 languages), Agent SDKs (2), ant CLI, claude-code (now auto-synced via gh-releases as of v1.1 — previously manually seeded), claude-code-action, claude-code-security-review, and 15 MCP ecosystem SDKs

• Tier 2 (raw markdown) — Aider-AI/aider/HISTORY.md. The generic keep-a-changelog parser (v1.1+) is also available for any product whose source is a CHANGELOG.md in Keep-a-Changelog format — configure via one entry in products.yaml.

• Tier 3 (HTML / RSS) — platform.claude.com/docs/release-notes, support.claude.com/articles/12138966, cursor.com/changelog/rss.xml, sourcegraph.com/changelog/featured.rss (filtered), github.blog/changelog/label/copilot/, code.visualstudio.com/updates/v1_NNN

• Tier 4 (catalog) — anthropics/skills, claude-plugins-{official,community}, knowledge-work-plugins

• Tier 5 (advisory) — @ClaudeCodeLog X account; marckrenn changelog mirror

Fetch strategies: gh-releases | rss | raw-changelog | keep-a-changelog | html-scrape | catalog | playwright. New product = one entry in products.yaml.

Synergies — what gets unlocked

12 curated cross-product workflows. Each describes a composition pattern, the trigger that makes it the right answer, and the changelog evidence that enables it. Examples:

• 08 — Universal SKILL.md format (Code + Cursor + Codex): one skill author, three agents read it

• 09 — MCP across seven surfaces (Code + Cursor + Continue + Copilot + Windsurf + Cody + API): one binary, every agent

• 10 — Anthropic BYOK across surfaces: one API key powers Claude in 7 editors with unified billing

• 11 — Claude Code orchestrates Aider: cost-shift heavy edits to a cheap model while Claude plans

• 12 — MCP config format gotcha: Copilot uses servers; everyone else uses mcpServers

Full index in synergies/INDEX.md.

Datasets and fine-tuned models

The corpus has a downstream layer: curated datasets derived from change bullets, and fine-tuned models trained on those datasets. The dataset is the artifact in this repo; the model is downstream.

dataset/changelog-actions/v1 — 301 entries

Each entry pairs one change bullet from the corpus with a hand-written action item: {kind, severity, subject, action_text, deadline, tags}. 8-enum kind taxonomy (breaking | deprecation | security | feature | fix | performance | docs | unknown), stratified 80/20 train/holdout, A3c-reviewed by human + cross-family qwen3:8b judge (92.4% agreement). See dataset/README.md for the distribution table, SCHEMA.md for the field-by-field contract, and STYLE.md for the action_text writing rules.

cs-actions:v1 — fine-tuned synthesizer

First model trained on the dataset. Qwen2.5-7B-Instruct + LoRA (rank-256 / all-linear, 100 steps QLoRA on 242 training entries), q8_0 GGUF, deployed via two-stage Ollama Modelfile (cs-actions-base + cs-actions:v1).

Release-gate eval (59-entry stratified holdout, three runs, all passed, zero parse errors):

Release pass criterion qwen3-vs-cs-actions ≥ qwen3-vs-GT (0.780 ≥ 0.780) — PASS ✓. Full per-entry verdicts in eval-report.v1.json, also attached to the GitHub Release as a single-file download.

Documented v1 limitation — anti-unknown bias. unknown is the only class where cs-actions:v1 underperforms qwen3:8b (F1 0.444 vs 0.545). Precision 1.000 / recall 0.286 — when input is genuinely ambiguous, the model commits to a specific kind rather than abstain. Inherited qwen-family prior, not fully overridden by the LoRA. Downstream consumers should treat a lower-than-expected kind: "unknown" rate as a signal to route batches to human review. The v2 plan addresses this with unknown-class augmentation plus hint-randomization.

Not distributed through this repo — the q8_0 GGUF is ~5 GB and the merged checkpoint is ~15 GB. The dataset and the build pipeline are what's published; the model rebuilds locally per TRAINING.md (4-stage manual chain, ~88 min warm on RTX 5080 Laptop 16 GB VRAM). The three-run release-gate contract lives in EVAL.md.

Full usage walkthrough — including local Ollama invocation, the format: "json" per-request caller requirement, and the rebuild pipeline — is on the handbook → cs-actions:v1 page.

Testing

Vitest suite covers unit / integration / regression / smoke tiers. test-spec-3.md is the current authority as of v0.7.0; test-spec.md (v1) and test-spec-2.md (v2) remain in the repo as historical record of the design lineage.

pnpm test               # unit + integration + regression (~36s, 519 tests)
pnpm test:watch         # interactive
pnpm test:coverage      # generate coverage/index.html (thresholds: 78/75/85/78)
pnpm test:smoke         # opt-in full-corpus smoke (RUN_SMOKE=1)

Layout:

No network in tests by default — provider HTTP is mocked via vi.spyOn(global, 'fetch'). Real SQLite in temp files (not :memory:) because sqlite-vec extension load semantics differ across versions and on-disk is the canonical path. Playwright is loaded via dynamic import and mocked via vi.doMock('playwright', ...) so tests pass without a real browser install.

CI: .github/workflows/test.yml runs pnpm test:coverage on push and PR.

Troubleshooting

"Database locked" or WAL errors

Another hk process (or a stale MCP server) is holding the SQLite database open. Close other hk processes, then retry. If the issue persists, check for a -wal or -shm file alongside data/claude-synergy.db — these are normal WAL-mode files and will be cleaned up when the last connection closes. Do not delete them while another process has the DB open.

"sqlite-vec extension not found" / sqlite-vec load failed

The sqlite-vec native extension failed to load. Common causes:

1. Node version too old — claude-synergy requires Node 22+. Check with node -v.

2. Native module needs rebuild — run npm rebuild better-sqlite3 (or pnpm rebuild better-sqlite3).

3. Platform mismatch — on Windows/ARM, better-sqlite3 needs a C++ build toolchain. Install the windows-build-tools or Visual Studio Build Tools with "Desktop development with C++".

Note: sqlite-vec is optional. FTS5 keyword search (hk query) works without it. Only hk embed and hk hybrid require the vector extension.

"Sync failed for product X" / fetch errors

hk fetch and hk sync call external APIs. Common causes:

• GitHub rate limit — the gh-releases strategy shells out to gh api, which uses your GITHUB_TOKEN. Unauthenticated requests hit 60 req/hr; authenticate with gh auth login or set GITHUB_TOKEN in your environment.

• Network / proxy — RSS and HTML fetchers use fetch(). Check connectivity and any corporate proxy settings (HTTPS_PROXY).

• Unknown product — hk fetch --product foo only works for products listed in products.yaml. Run hk products to see all available names.

Sync is idempotent — safe to re-run after a partial failure. Already-fetched releases are skipped.

"Embedding provider not responding"

hk embed calls an external embedding service:

• Ollama (default, 768-dim) — ensure Ollama is running (ollama serve) and the embedding model is pulled (ollama pull nomic-embed-text).

• Voyage (1024-dim) — set VOYAGE_API_KEY in your environment. Check your API key at dash.voyageai.com.

• OpenAI (1536-dim default, configurable) — set OPENAI_API_KEY. Default model is text-embedding-3-small; override with OPENAI_EMBED_MODEL (e.g. text-embedding-3-large for 3072-dim). Use via hk hybrid --embed openai or hk embed --embed openai.

Embedding dimension mismatch on provider switch

Each provider produces vectors of a fixed dimension (Ollama 768, Voyage 1024, OpenAI 1536 by default — OpenAI supports configurable dim within the model's native size). The DB stores the active dimension in schema_meta.embedding_dim. Switching providers across different dimensions while chunks exist raises EMBEDDING_DIM_MISMATCH (AppError) rather than silently corrupting the vector table. To switch:

rm data/claude-synergy.db data/claude-synergy.db-wal data/claude-synergy.db-shm
hk init
hk ingest
hk embed --embed openai     # new provider, new dim, fresh chunks_vec

For OpenAI Matryoshka truncation (smaller-than-native dim), set OPENAI_EMBED_MODEL and pass the desired dim through hk embed's provider construction — see the handbook embedding section for details.

Schema version mismatch / corrupted database

If the DB was created with an older schema version and migration fails, or if data looks wrong after a crash:

rm data/claude-synergy.db data/claude-synergy.db-wal data/claude-synergy.db-shm
hk init
hk ingest
hk embed --context structured --embedding ollama   # optional, for vector search

This is safe — the DB is a derived cache. All source data lives in products/*/releases/*.md files.

Related files

• CONTRIBUTING.md — how to add products, run tests, submit PRs

• URGENT_FINDINGS.md — 23 actionable items (security CVEs, model retirements, breaking changes, config gotchas)

• SOURCES.md — 5-tier source landscape with fetch strategies

• synergies/INDEX.md — 12 curated cross-product workflows

• schema.sql + schema-vec.sql — SQLite + sqlite-vec schemas

• test-spec-3.md (current) + test-spec-2.md, test-spec.md (historical) — test suite specs

License

MIT. Built by MCP Tool Shop.