What's new in graphchat

Full CLI tool (gph) shipped with 12 commands, API-key + JWT exchange authentication, and a complete REST API surface for programmatic access.

• Feature
  New gph CLI — 12 commands: login, logout, status, repos, index, search, query, path, explain, report, export, watch.
• Feature
  API key management: mint, list, and revoke named scoped keys from the dashboard or via POST /api/auth/keys.
• Feature
  JWT exchange endpoint (POST /api/auth/exchange) — trades an API key for short-lived access + refresh tokens, enabling fully non-interactive CLI authentication.
• Feature
  Auto token refresh — the CLI silently refreshes the access token before expiry using POST /api/auth/refresh without requiring re-login.
• Feature
  gph search --agent flag emits a compact Markdown context block ready to paste directly into an AI chat prompt.
• Feature
  gph watch --on-commit installs git post-commit/post-checkout hooks for automatic re-index on every commit.
• Feature
  gph path <source> <target> finds the shortest graph path between any two named symbols and prints every hop.
• Feature
  gph explain <label> produces an AI-generated explanation of a node including its community, callers, callees, and confidence tiers.
• Improvement
  Graph report (gph report) now includes surprise edge rankings and file coverage percentage per community.
• Improvement
  POST /api/auth/logout now accepts the refresh token in the body and revokes it server-side for clean CLI logout.
• Security
  API key hashing: only the key hash is stored in the database; the plaintext is returned once at creation and never logged.
• Security
  Refresh token deduplication: concurrent auto-refresh races are resolved server-side to prevent duplicate token issuance.

Graph analysis and community detection improvements, Redis community cache, and the interactive graph visualisation.

• Feature
  Leiden community detection via graspologic — partitions graphs into well-connected clusters without pre-trained embeddings.
• Feature
  Redis community cache: bundles are serialised once per graph version and served instantly on repeated queries (zero stale-data risk via version-keyed TTL).
• Feature
  GET /api/graph/community/:id/prompt returns a token-bounded, agent-ready context bundle for a single community.
• Feature
  God-node detection: flags the highest-degree symbols by degree centrality and betweenness — surfaced in the graph UI and reports.
• Feature
  Surprise edge highlighting: inter-community edges are ranked and shown with a distinct colour in the interactive graph.
• Improvement
  Graph sync is now incremental — only files changed since the last sync are re-parsed and re-embedded.
• Improvement
  Cache hit rate is exposed in GET /api/health for observability.
• Fix
  Fixed a race condition in concurrent graph sync jobs that could produce duplicate nodes for the same file.

Multi-modal extraction support and edge confidence tiers added to the graph build pipeline.

• Feature
  Tree-sitter AST parsing for Python, JavaScript/TypeScript, Go, Java, Rust, Ruby, C/C++, Kotlin, and Swift.
• Feature
  LLM-pass extraction for prose files (Markdown, RST, PDF text) — identifies concepts, entities, and inter-document references.
• Feature
  Vision model pass for image assets — diagram descriptions are chunked and vectorised alongside code.
• Feature
  Edge confidence tiers: EXTRACTED (1.0), INFERRED (0.6–0.9), SPECULATIVE (<0.6) — stored as edge attributes, surfaced in search and reports.
• Feature
  Budget trimming: speculative edges are dropped first when results exceed the requested token budget.
• Improvement
  Chunk size and overlap are now configurable per file type via the admin settings panel.
• Fix
  Fixed incorrect line-number offsets in Python function extraction when files used Windows line endings.