How Probe Works ​

Probe combines fast text scanning with structure-aware parsing to return grounded, reliable context from real codebases. This page covers Probe internals (not Visor workflows).

The retrieval pipeline ​

  1. Scan: ripgrep finds candidate files and matches while honoring ignore rules.
  2. Parse: tree-sitter builds an AST for structural understanding.
  3. Match: query terms are matched against code and structure.
  4. Rank: results are scored for relevance and stability.
  5. Extract: full functions, classes, or modules are returned.
  6. Shape output: JSON/SARIF or human output, with token limits when needed.

Why results are reliable ​

  • Complete context: full code blocks, not partial snippets.
  • Deterministic ranking: consistent ordering across runs.
  • Local-first processing: your code stays in your environment.

Agent bootstrap (optional) ​

If you run Probe in agent mode (MCP/ACP), it will also load repo guidance files when present: AGENTS.md, ARCHITECTURE.md, and the AgentSkills standard.
Raw CLI/MCP usage stays unchanged and only exposes retrieval tools.

Where it runs ​

  • CLI in local dev or CI
  • GitHub Actions for reviews and automation
  • MCP + SDKs for agents and custom tooling
  • Web UI for non-engineers