repowitness

# RepoWitness RepoWitness is a local-first Rust CLI and stdio MCP server that indexes Rust, Go, TypeScript, TSX, and Python source into bounded, evidence-backed retrieval and engineering-memory workflows. The product loop is deliberately small: ```text source change -> atomic index generation -> memory revalidation -> evidence-backed context ``` Every result identifies its generation, source evidence, coverage, and limits. ## Install From a checkout: ```text make install ~/.local/bin/repowitness --help ``` `make install` builds the locked release binary and atomically replaces `~/.local/bin/repowitness`. Set `INSTALL_PATH` to choose another destination. ## Basic CLI Index into a database outside the repository: ```text repowitness index \ --repository-id rwi1:h:<64-lowercase-hex> \ --database /path/outside/repository/repowitness.sqlite3 \ /path/to/repository ``` Useful read commands are `search`, `symbol-search`, `symbol-get`, `locate-relevant-paths`, `architecture-map`, `architecture-overview`, `repository-topology`, `graph`, `context-build`, `diagnostics`, `verify`, and `memory-recall`. Use `repowitness <command> --help` for exact bounds. `onboard --root <repository>` is the explicit private-state shortcut. It uses a fast source-only index by default; add `--full` when graph evidence is needed. `watch` keeps one repository current in the foreground; it never starts a daemon. Memory writes are explicit and local: ```text repowitness memory-manage --help ``` ## MCP RepoWitness can expose one MCP connection for one repository or a private local catalog containing all repositories onboarded by RepoWitness: ```text repowitness mcp-serve \ --repository-id rwi1:h:<64-lowercase-hex> \ --database /path/outside/repository/repowitness.sqlite3 \ --root /path/to/repository ``` For one MCP connection across repositories: ```text repowitness mcp-serve --catalog ``` Run `repowitness onboard --root <repository>` once per repository. Catalog tool calls select the registered repository identity; the current repository is the default when the server starts inside it. Catalog mode also provides `cross_repository_search` for bounded SQLite FTS5 literal search across all registered repositories. Results include each repository's generation and coverage; matching text is candidate evidence, not an inferred dependency relationship. The catalog defaults to the shared user state location (`$XDG_STATE_HOME/repowitness`, or `~/.local/state/repowitness`) so any local AI client can use the same catalog. Use `--catalog-state-dir` only to deliberately select another state root. MCP also loads the optional shared user configuration from `$XDG_STATE_HOME/repowitness/config.toml` (or `~/.local/state/repowitness/config.toml`); `--user-config` overrides it. The server is read-only by default. Add `--enable-memory-writes --memory-actor <validated-actor>` only when the local process should manage team memory. Stdout is reserved for MCP JSON-RPC; startup help and errors go to stderr. There is no daemon, connected-workspace CLI, SCIP import surface, personal-memory surface, or durable task surface. The catalog is read-only by default; explicitly enabling memory writes adds the same fixed-actor `memory_manage` capability to each selected repository. It reloads its bounded control file at MCP request boundaries; run `onboard` and the next request sees the updated repository set without an MCP restart. A malformed later catalog leaves the last valid snapshot in place. ## Development Read [`AGENTS.md`](AGENTS.md), then: ```text cargo fmt --all -- --check cargo check --workspace --all-targets --all-features --locked cargo test --workspace --locked ./scripts/check-docs ``` The repository uses the MIT License. See [`CONTRIBUTING.md`](CONTRIBUTING.md) for provenance and clean-room requirements.