v0.2.3

Added

• Automatic query-time freshness: a search detects edited files and re-indexes what changed with no git and no editor hook required. A stat gate (mtime + size per file) means unchanged files are never read, so the poll is cheap; it runs at most once per refresh_ttl (default 2s). On by default; disable with INTENT_CODE_AUTO_REFRESH=0 or tune with INTENT_CODE_REFRESH_TTL.

Changed

• Incremental indexing skips reading files whose mtime and size are unchanged, and a poll that finds no changes does no writes (no graph or repo-map rebuild, no manifest save).

Designed by Harsha Rahul

v0.2.2

Added

• Repo-scoped search for indexes that span several repositories. Each symbol and chunk is tagged with its enclosing git repository (the nearest ancestor holding .git; deepest wins for nested repos and submodules). Every search result carries its repo, and a search can be scoped with search --repo <name> (CLI) or filters={"repo": "<name>"} (MCP). stats lists the repositories in the index with their file counts.

Notes

• Run index --full once to backfill repo tags on an index built before 0.2.2; the tag is location-derived, so unchanged files are not otherwise re-touched.

Designed by Harsha Rahul

v0.2.1

Added

• intent-code init --agent claude|copilot|gemini|all: wire the index into each agent's native config (.mcp.json, .vscode/mcp.json, or .gemini/settings.json) and write the protocol into its instruction file (CLAUDE.md, .github/copilot-instructions.md, or GEMINI.md). Repeatable, defaults to claude. Idempotent: managed block + JSON key-merge never clobber user content.
• intent-code install-hooks: git hooks (post-commit, post-merge, post-checkout, post-rewrite) that flag the index stale so the next query re-indexes incrementally. Editor-agnostic; appends to an existing shell hook and leaves a non-shell hook untouched.

Changed

• The indexer no longer indexes the agent instruction files it generates (CLAUDE.md, GEMINI.md, AGENTS.md, .github/copilot-instructions.md, *.prompt.md, .mcp.json, and the .gemini/ directory), avoiding a re-index feedback loop.

Designed by Harsha Rahul

v0.2.0

Added

• Comprehension tools that return code, not just locations:
  • code_read (CLI read): a symbol's full source span, untruncated, resolved by doc_key, qualname, or name.
  • code_context (CLI context): the full bodies of a symbol and what it calls, in call order, packed to a token budget, so an agent can follow a flow end to end in a single call.
  • code_flow (CLI flow): the ordered call sequence inside a function, each call with its resolved target location.
• --local flag on the note subcommands.
• A clear warning when no embedder is configured and Ollama is unreachable, so it is obvious that search has fallen back to the weaker lexical hashing embedder.

Changed

• Call edges now preserve source order, so flow and context reflect the real execution sequence.
• Re-indexing removes stale symbols in a single batched delete (intent-db's delete_many) instead of rebuilding the matrix once per key.
• Requires intent-vector-db>=0.2.3.

Designed by Harsha Rahul