The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external, potentially attacker-influenced data (failure logs) to dynamically rewrite the instructions of other agent skills.
Ingestion points: Reads data from metrics/analytics.jsonl and various Task Failure Logs.
Boundary markers: The documentation explicitly acknowledges the risk of prompt injection via errors and states the system 'must sanitize' input, but no specific boundary markers or escaping logic are implemented in this file.
Capability inventory: The skill has the capability to read log files and perform file-write/overwrite operations on .skill.md files throughout the system.
Sanitization: Although the skill instructions mandate sanitization of incoming error text, the lack of defined validation patterns for log content poses a risk where malicious input could influence the 'Optimization' (Rewrite) step.
[DYNAMIC_EXECUTION]: The skill generates and modifies executable instructions (skills) at runtime. By rewriting .skill.md files, it performs a form of self-modification that alters the logic of the agentic system. The inclusion of 'Optimization Sandboxing' and 'Regression Testing' as described in the security section provides some mitigation against the execution of degraded or malicious instructions.

meta-optimize