The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted data from repository files (specifically under docs/solutions/ and MEMORY.md) to guide its maintenance logic. This creates a vulnerability where malicious content or instructions embedded in these files could influence the agent to perform unintended documentation replacements, deletions, or metadata changes.
[COMMAND_EXECUTION]: The mode:autonomous argument enables a fully automated workflow that skips all human checkpoints. The agent is empowered to perform file system modifications (updates, archiving) and repository state changes (branching, committing, and opening PRs) without any human oversight or approval.
[COMMAND_EXECUTION]: The instructions explicitly direct subagents to use internal platform tools (Glob, Grep, Read) instead of standard shell commands specifically to "avoid permission prompts." This design choice intentionally circumvents security controls intended to notify the user of file system access and requires them to grant permission.
[COMMAND_EXECUTION]: In autonomous mode, the Phase 5 (Commit Changes) logic automatically executes Git operations. Depending on the environment context (such as being on the main branch), the skill may create new branches and attempt to open Pull Requests based on the results of the automated, potentially compromised, analysis.

compound-refresh