The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple shell commands to manage the repository state and analyze changes.
Evidence: Uses git rev-parse, git diff, git log, and git worktree add in SKILL.md to analyze branch differences and prepare temporary work environments.
Evidence: Executes a vendor-provided Python script skills/track/map-features/scripts/reconcile.py during the initial phase of the review.
[DYNAMIC_EXECUTION]: The skill includes a "Blast prove" workflow that requires generating and running executable code to verify runtime safety.
Evidence: The instructions in blast-prove.md direct the agent to "Prove the one fact: smallest script or test that imports the real code / drives the real consumer... Paste the command and output."
Evidence: This behavior involves creating and executing scripts that interact directly with untrusted code changes from the repository being reviewed.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data as part of its core logic, which could be used to manipulate the agent's behavior.
Ingestion points: Reads repository diffs via git diff and requirement files such as requirements.md in SKILL.md (Steps 1 and 2).
Boundary markers: The skill uses structured report headings (## Standards, ## Spec) to separate outputs but lacks explicit instructions to ignore instructions embedded within the content of the diff or requirements files.
Capability inventory: The skill has access to shell command execution, repository manipulation (git), and the ability to generate and run scripts locally.
Sanitization: There is no evidence of sanitization, escaping, or explicit filtering of the ingested code or requirement text before it is processed by the subagents.

inspect-change