The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes git rev-parse using subprocess.run to identify repository roots and validate project paths. These executions are constrained to a fixed list of arguments and do not utilize a shell, mitigating command injection risks.
[DATA_EXFILTRATION]: The skill accesses sensitive local directories containing agent conversation histories, including ~/.claude, ~/.codex, and Warp Terminal's internal SQLite databases. This access is required for the skill's core functionality of analyzing past interactions. The implementation is restricted to local file operations, and the instructions explicitly forbid uploading transcripts or excerpts.
[INDIRECT_PROMPT_INJECTION]: The skill processes conversation transcripts which are treated as untrusted data. This establishes a surface where instructions embedded in logs could theoretically influence the agent during the scoring phase.
Ingestion points: scripts/collect_sessions.py ingests local JSONL and SQLite session files.
Boundary markers: Transcripts are structured using role-based labels (e.g., [user], [assistant], [tool:name]) in the render_transcript function.
Capability inventory: The skill possesses file system read/write capabilities, repository discovery via git, and local HTML report generation.
Sanitization: Input content is truncated via the truncate function to limit the volume of data processed in a single turn.

skill-doctor