The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute several CLI tools provided by the author's own package, @ericrisco/rsc, to manage the review process. These include commands for freezing the current code state (sello freeze), approving reviews (sello approve), blocking merges (sello block), and consolidating multi-agent reports (review consolidate).
[EXTERNAL_DOWNLOADS]: The skill utilizes npx to download and run the @ericrisco/rsc package from the NPM registry. As this package belongs to the skill's author (ericrisco), it is considered a legitimate and expected vendor resource for the skill's operation.
[INDIRECT_PROMPT_INJECTION]: The skill has an inherent attack surface as it processes external, potentially untrusted data including code diffs, human PR comments, and machine-generated feedback. However, it includes a mandatory security protocol for receiving reviews: the agent must restate the claim, verify it against the actual code, and reproduce the defect before accepting it. This 'Verify Before Acting' requirement, combined with a lack of dangerous automated capabilities on unverified input, effectively mitigates injection risks.
Ingestion points: External PR comments, review feedback, and code diffs processed in SKILL.md.
Boundary markers: The skill employs a logical boundary by requiring a 'Restate' and 'Verify' step for every finding before it enters the action phase.
Capability inventory: File system writing to 02-DOCS/wiki/sdd/decisions.md and .rsc/sello-log.jsonl, and execution of vendor CLI tools.
Sanitization: Strict protocol requiring code-based traces or reproduction cases (repro) for all findings ensures that external instructions cannot bypass verification.
[DYNAMIC_EXECUTION]: The skill orchestrates parallel analysis by dispatching specialized subagents (refuter-correctness, refuter-security, refuter-tests, and finding-verifier) to challenge the readiness of a code change. This multi-agent verification process is a structured implementation of the skill's adversarial review logic.

review