The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from GitHub issue and PR titles, bodies, and comments, which could contain malicious instructions designed to influence the agent's autonomous decisions.
Ingestion points: Data is ingested via gh issue view and gh pr view which retrieve bodies, comments, and review threads (referenced in SKILL.md, agents/cluster-evidence-agent.md, and scripts/alias.sh).
Boundary markers: No delimiters or explicit instructions are provided to the agent to ignore potentially adversarial instructions embedded within the ingested data.
Capability inventory: The skill can execute gh CLI commands to mutate repository state (merge, close, label), perform git operations (branch, patch), and execute local tests (referenced in SKILL.md and scripts/alias.sh).
Sanitization: No sanitization or validation of the external content is performed before it is passed to the sub-agent decision chain.
[COMMAND_EXECUTION]: The skill relies on a local bash script (scripts/alias.sh) and the GitHub CLI (gh) to perform repository management tasks. In autonomous mode, the agent executes these commands without mandatory per-action user confirmation, relying on internal confidence scores and repo policy.
[DYNAMIC_EXECUTION]: The skill's autonomous mode includes instructions to "patch the smallest implicated surface" and "run targeted local gates," which involves the generation and execution of code and tests at runtime based on external bug reports and PR descriptions (referenced in SKILL.md).

openclaw-github-dedupe