The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the execution of the fallow CLI and git commands to generate review briefs and validate judgments. These operations are core to the skill's functionality and are performed on local repositories.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code changes (diffs). To mitigate risks, it uses a deterministic 'guide' derived solely from the project's module graph, intentionally excluding potentially malicious prose from pull requests. Findings are also clearly labeled with deterministic: false when generated by a model.
Ingestion points: git diffs, JSON walkthrough guides, and production runtime data from Fallow Cloud.
Boundary markers: Model-generated trade-offs are explicitly fenced with deterministic: false and kept separate from graph-grounded facts.
Capability inventory: Execution of fallow and git commands via the shell; network requests to Fallow Cloud for runtime coverage data.
Sanitization: The skill's architectural design separates the untrusted input (code diffs) from the control mechanism (graph-based signals), reducing the ability for injected code to influence the agent's logic.

fallow-review