v0.2.0

Added

• Dependency docs are fetched when not installed locally. Every registry
  publishes the same artifact cgraph reads off disk, so it is downloaded and run
  through the same parsers: npm .tgz (falling back to @types/<pkg>, since a
  large share of npm ships no types of its own), NuGet .nupkg, Maven
  -sources.jar, PyPI wheels. No scraping and no third-party service — the
  result is version-exact because it is the published artifact.
• Maven dependency docs, local and remote, now that a zip reader exists.
• cgraph docs --refresh --guides fetches llms.txt
  where a project publishes one. Package archives carry API reference because
  that is what is in the source; they cannot carry setup guides, which were
  never in the source. This is the only prose source, and coverage is partial.
• tar.gz and zip readers built on node:zlib, with no new dependencies.
• C# language pack. Types, members, properties, namespaces, XML doc
  summaries and access modifiers, with using resolved to the file declaring
  that namespace. LINQ and BCL calls are classified as runtime built-ins rather
  than swamping the unresolved list.
• Project technology discovery. Manifests (*.csproj, package.json,
  pom.xml, pyproject.toml, go.mod, Cargo.toml, …) are read before any
  parsing, so only the packs a project actually needs are loaded and only their
  grammars are fetched. Frameworks — Angular, React, ASP.NET, Entity Framework,
  Spring, Django, FastAPI — are detected from every manifest in the tree, not
  just the root. A file in an unpredicted language still loads its pack on
  sight: discovery decides what loads eagerly, never what gets ignored.
• .NET dependency documentation. NuGet ships compiled assemblies, so there
  is no source to parse — but packages carry an XML documentation file beside
  the DLL with every public member and its summary. docs now reads it,
  preferring the newest target framework present.
• Local embedding models. provider: "local" targets any OpenAI-compatible
  endpoint (Ollama, LM Studio, llama.cpp, vLLM) with baseUrl, model and
  optional dimensions. Nothing leaves the machine and nothing costs money. The
  returned vector width is verified against dimensions on the first batch,
  because a silent mismatch corrupts every similarity score in a way that is
  close to untraceable.
• The index refreshes itself. The MCP server checks for changes before
  answering, so an agent never reads a graph that disagrees with the working
  tree — after an editor save, a git checkout, a rebase, or another agent's
  edit. No watcher, no hook, nothing to remember. Throttled (3s) so a burst of
  tool calls costs one scan, and disabled with
  autoRefresh.enabled: false.
• cgraph hooks install — pre-warms the index after checkout, merge and
  rebase, the operations that change hundreds of files at once. Appends between
  markers and never replaces an existing hook; backgrounded so it cannot delay
  a git command. post-commit is excluded by default because committing does
  not change the working tree.
• cgraph init --agent for Claude Code, GitHub Copilot, opencode, Cursor and
  Windsurf, writing both the MCP registration and an instruction block telling
  the agent to use these tools instead of grep.

Changed

• No token-savings multiplier is published any more. The previous figures
  compared cgraph against an invented grep-then-read baseline: an agent opens
  every file grep matched, capped at ten, and reads each whole. No agent was
  ever run. Change the cap, assume line windows instead of whole files, or
  assume two probes instead of one, and the ratio moves by an order of
  magnitude — the assumption determined the number, not the measurement.
  Responses now report their own size, and where a concrete comparable exists
  (the file an outline describes) both numbers are shown as facts with no
  arithmetic between them. bench/ remains as a regression guard on response
  sizes, comparable only to itself.
• Freshness checks no longer read the repository. A file whose size and
  mtime match the index is skipped without being opened; only files that fail
  that check are read and hashed. On llama.cpp a no-op pass drops from ~520ms
  to ~170ms, which is what makes automatic refresh affordable. The content hash
  still decides, so a file touched but not edited is read once and correctly not
  re-parsed.

Breaking

• SavingsLedger is renamed UsageLedger in the programmatic API, and records
  tokens_returned / tokens_source instead of tokens_baseline /
  tokens_saved. The old name asserted a saving the tool cannot observe. Only
  affects code importing it from cgraph; the CLI and MCP surfaces are
  unchanged.
• cgraph stats no longer prints a reduction factor, and cgraph doctor no
  longer prints "tokens saved". Anything parsing that output needs updating.

---

Install

npm install -g cgraph

Published to npm as cgraph@0.2.0 with build provenance.