Bounded, read-only codebase exploration. Do not write files or propose implementation — return triaged, structured findings for downstream skills.

Workflow

• Frame — state a concrete exploration question or hypothesis before searching.

• Bounded discovery — find relevant modules, symbols, and entry points. MUST complete within 3 refinement cycles (search → read → refine). If unanswered after 3 cycles, return best partial results with explicit risks_or_unknowns .

• Context triage — label findings: core (must include), related (include if space), peripheral (mention only if risk signal). Drop peripheral from handoff by default.

• Map flow — trace control flow, data flow, and cross-file dependencies.

• Extract surface — identify files to modify and expected side effects.

• Handoff — return concise structured output (see below).

Discover project-specific terminology and naming conventions in cycle 1. Use discovered terms in subsequent searches instead of generic names.

Output Format

• key_findings — confirmed observations with file paths

• files_to_touch — with triage priority per file (core / related / peripheral)

• risks_or_unknowns — explicit gaps and unexplored areas

• terminology_discovered — non-obvious project naming (omit when not applicable)

External Dependencies

When exploring external libraries or APIs, use this source reliability hierarchy: official docs > versioned examples > changelog > blog posts > community forums > Stack Overflow.

• Never mix guidance from different major versions

• Check maintenance status before recommending a library

• Capture version and context assumptions for every recommendation

Anti-Patterns

• Speculating about behavior without repository evidence

• Returning file lists without explaining each file's role in the flow

• Assuming behavior from file names alone without verifying call sites

• Proposing code edits or implementation during exploration

• Overloading output with low-value file dumps instead of triaged signal