The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses local shell commands to interact with development tools including git, gh (GitHub CLI), origin, and gt (Graphite). These commands are used for pull request management, worktree auditing, and orchestration tasks.- [EXTERNAL_DOWNLOADS]: The watch-pr utility downloads logs and test artifacts from GitHub Actions runs. The implementation includes strict security controls, such as limiting download sizes to 8MiB per archive and enforcing total budget limits to prevent denial-of-service.- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from pull request comments and review bodies. It mitigates injection risks by explicitly instructing the agent to treat this data as untrusted information for triage rather than as valid instructions.
Ingestion points: Pull request comments and reviews fetched in scripts/watch-pr/github.ts and diagnostic artifacts in scripts/watch-pr/failed-check-evidence.ts.
Boundary markers: Verdicts are expected in a specific four-line block format to help distinguish them from other prose.
Capability inventory: Shell command execution (git, gh, origin, gt, bun, python, node) and local file system access across all scripts.
Sanitization: Implementation includes HTML tag stripping in scripts/watch-pr/review-bodies.ts and explicit rejection of DTD/entities in the XML parser in scripts/watch-pr/junit-archive.py.- [DYNAMIC_EXECUTION]: A bootstrap script (scripts/bootstrap.ts) manages local dependencies by running bun install with a frozen lockfile and restarting the current process. This mechanism is used to ensure the orchestration tools have their required environment before execution.

poteto-mode