The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes build and test commands as part of its fix verification loop.
Commands are derived from user parameters (verify_cmd) or auto-detected from standard repository files like package.json, Makefile, and common test frameworks.
Execution is performed through a wrapper (run-reviewer.sh) that enforces wall-clock timeouts and captures exit statuses to prevent silent hangs.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the repository under review, including code and Pull Request history.
Ingestion points: The skill fetches PR diffs, bodies, and review history (SKILL.md Section 1).
Boundary markers: Review history is encapsulated in a dedicated block in the prompt, and specific instructions are provided to identify the source of each entry.
Capability inventory: The skill can execute shell commands, read and write files, and create Git commits.
Sanitization: The instructions explicitly mandate that the agent must treat author replies as untrusted technical claims to be validated against the code, rather than instructions to follow (SKILL.md Section 1, Section 3).
[DYNAMIC_EXECUTION]: The skill supports an opt-in isolated proof runner for executing probes to verify findings.
Execution requires an absolute path to a trusted runner executable located outside the reviewed worktree (SKILL.md Parsing section).
The execution environment is heavily sanitized by scripts/run-proof.sh, which disables network access, provides read-only access to source and proof inputs, and bounds system resources.
Fingerprinting logic is implemented to detect and reject any mutations to the source tree or proof inputs during execution.

review-anvil