The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests review findings, which may originate from external sources such as teammate review comments. These findings are passed as verbatim instructions to sub-agents (Sonnet models) that possess the capability to modify the codebase and execute project-specific shell commands.
Ingestion points: Finding text sourced from the terminal, a sidecar-stored Review Disposition Ledger, or teammate review threads in references/state-backend.md.
Boundary markers: The skill does not implement explicit delimiters or sanitization for the finding text when constructing sub-agent prompts, relying on the model to only address what the finding names.
Capability inventory: Sub-agents have permissions to write to implementation files, perform git commit operations, and execute the project's configured test command.
Sanitization: No explicit sanitization of the external finding text is performed.
[DYNAMIC_CONTEXT_INJECTION]: The skill utilizes a dynamic execution block in the SKILL.md header to manage workflow state.
Evidence: A shell command is executed using the exclamation-backtick syntax to create a .fix-findings-active marker file in the project's .claude directory at load time.
Context: This is a benign use of project-specific state management used to signal a post-review edit lock to other tools in the suite.
[COMMAND_EXECUTION]: The skill and its bundled sidecar.sh script execute local shell commands for repository operations and verification.
Evidence: The workflow involves commands such as git archive, tar, git commit --fixup, and the project's test suite as defined in the sidecar profile.
Safety: The skill implements an isolation mechanism by archiving the target commit into a throwaway directory before running mutation tests, ensuring that destructive testing (the 'breaker' pass) does not affect the main working tree.

fix-findings