The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the gh (GitHub CLI) and git command-line tools across multiple scripts (e.g., heartbeat-state.sh, pr-readiness.sh, request-review.sh) to list pull requests, read comments, post reactions, merge code, and manage local worktrees.
[REMOTE_CODE_EXECUTION]: In scripts/repair-runner.sh, the skill executes AI-generated code using the codex exec tool with the --dangerously-bypass-approvals-and-sandbox flag. This configuration allows for the execution of arbitrary system commands without human oversight or a security sandbox.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it incorporates external, untrusted content into its decision-making and code-generation process.
Ingestion points: Pull request comments, review bodies, and thread content are fetched in scripts/pr-readiness.sh and stored in a snapshot used for subsequent repair tasks.
Boundary markers: Absent. The instructions generated in scripts/start-repair.sh do not use delimiters or safety instructions to prevent the model from following malicious commands embedded in pull request feedback.
Capability inventory: The repair-runner.sh script possesses the capability to modify the filesystem and execute arbitrary code via codex exec without sandbox restrictions, and run-heartbeat.sh can merge branches to the main repository.
Sanitization: Absent. No filtering or validation is performed on the content retrieved from the GitHub API before it is processed by the AI model.
[EXTERNAL_DOWNLOADS]: The skill clones external repositories to a local workspace using gh repo clone in scripts/start-repair.sh and scripts/start-fallback-review.sh to facilitate the repair and review loops.

pr-comment-sentinel