The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (PR comments and code diffs) and acts on high-privilege keywords.
Ingestion points: The skill reads human comments via cmux comments list (referenced in references/review-receipts.md) and file diffs using git commands.
Boundary markers: Instructions recommend independent discovery to separate the author's reasoning from the review, but the agent still treats comments as sources of intent.
Capability inventory: The agent is empowered to perform merges to the main branch when it identifies a 'merge directive'.
Sanitization: There are no explicit instructions to verify that a merge keyword originated from an authorized user rather than the data being reviewed.
[COMMAND_EXECUTION]: The skill requires the agent to run various shell commands and platform-specific tools.
Evidence: SKILL.md contains the command cmux workspace status set review. references/review-receipts.md contains a shell script snippet for cryptographic tree snapshotting using git read-tree, git add -A, and git write-tree.
[DYNAMIC_EXECUTION]: The skill involves the generation and execution of temporary shell sequences.
Evidence: The procedure in references/review-receipts.md for generating a tree_sha requires the agent to generate a temporary index file using mktemp and execute a multi-line sequence of Git operations against it.

cmux-review