The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements an 'Evolution Agent' loop that ingests execution traces from quality_reports/traces/. This creates an indirect prompt injection surface where malicious content within processed traces could attempt to influence the agent's logic to propose insecure modifications to harness rules or permission boundaries.
Ingestion points: Reads .jsonl trace files from quality_reports/traces/ and .tsv baselines.
Boundary markers: The instructions do not define specific prompt delimiters for the trace content, though they emphasize evidence-based analysis.
Capability inventory: Can modify files in hooks/, rules/, skills/, and agents/ directories.
Sanitization: Relies on mandatory human-in-the-loop (HITL) review for all changes as a primary safety mechanism.- [COMMAND_EXECUTION]: The skill facilitates the modification of project source code and configuration files. It specifies a 'Promote' stage where edits are applied to hooks/, rules/, skills/, and agents/ after human approval. This is the primary intended function of the skill for harness optimization.

harness-mechanic