The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The governance rules define a pipeline for 'self-improvement' triggered by external inputs ('sanitized signals'). This ingestion pattern creates a vector where malicious instructions embedded in input data could influence the 'Analyst' or 'Coder' steps to produce undesirable modifications to the agent's scaffold. * Ingestion points: The Candidate Lifecycle (Step 1) involves observing a 'sanitized signal' from the environment. * Boundary markers: While the skill mentions 'sanitized' signals and 'schema validation' (Step 3), it does not define specific prompt delimiters or instructions to ignore embedded commands within the signal data. * Capability inventory: The 'Coder' (Step 5) and 'Promotion' (Step 8) mechanisms possess the capability to write files to the '.agents/skills/self-improvement-overlays/' directory. * Sanitization: The policy explicitly requires 'sanitized signals' and 'clean schema validation' as prerequisites for promotion, providing basic protection against malformed inputs.

self-improvement-governance