The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill uses a feedback loop that processes execution traces to generate new agent instructions, creating a surface where malicious external data could influence the agent's long-term behavior.
Ingestion points: The scripts/analyze_traces.py script reads trace files from .tmp/learning/traces/, which capture raw task interactions and tool outputs that may include untrusted content.
Boundary markers: The trace extraction logic does not implement specific delimiters or instructions to ignore embedded commands within the analyzed text.
Capability inventory: The skill uses scripts/apply_adaptation.py to modify the agent's own markdown instruction files, granting the learning process write access to the agent's core logic.
Sanitization: There is no validation or escaping of the extracted 'insights' before they are appended to the agent's definition files.
[DYNAMIC_EXECUTION]: The skill implements self-modifying instruction patterns by programmatically updating the agent's markdown configuration files at runtime.
The apply_agent_adaptation function in scripts/apply_adaptation.py searches for specific headers in agent files and inserts new rules, which can persistently alter the agent's behavior across all future sessions.

self-learning