The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (logs, feedback, artifacts) to propose and apply structural upgrades to other skills, creating a surface for indirect prompt injection.
Ingestion points: Untrusted data enters the context through run logs, screenshots, and user feedback as described in SKILL.md and references/evidence-and-scope.md.
Boundary markers: The skill enforces a mandatory two-stage approval process (plan_only followed by apply_after_approval) in SKILL.md to prevent automatic execution of malicious instructions.
Capability inventory: The skill possesses the capability to modify SKILL.md, reference files, and Python scripts within the target skill directory.
Sanitization: Uses logical gates such as the 'Generalization Gate' in references/upgrade-decision-protocol.md to filter incident-specific data from permanent rules.
[COMMAND_EXECUTION]: The skill instructs the agent to execute local diagnostic and validation commands.
Evidence: SKILL.md and references/validation-and-regression.md detail the execution of python -m py_compile, python -m json.tool, and the local script scripts/validate_upgrade_artifacts.py for ensuring the integrity of upgraded artifacts.

run-history-skill-upgrader