The Agent Skills Directory

[COMMAND_EXECUTION]: The relay.mjs script utilizes the node:child_process module (specifically spawn and execFileSync) to run the aider CLI and git commands. This is the core functionality required to delegate tasks and track repository changes.
[COMMAND_EXECUTION]: The applyFleetLane function in relay.mjs attempts to execute a local JavaScript file located at ../../delegate-setup/scripts/lane.mjs. This facilitates integration between sibling skills within the delegate ecosystem but relies on the local availability of that script.
[SAFE]: The skill implementation significantly reduces the attack surface by pinning restrictive Aider configuration flags by default, including --no-suggest-shell-commands (to prevent execution of unreviewed shell code), --no-detect-urls (to prevent unauthorized outbound network fetches), and --no-analytics (to disable telemetry).
[SAFE]: The orchestrator explicitly disables Aider's autonomous committing features (--no-auto-commits and --no-dirty-commits). This ensures that the tool cannot unilaterally modify the repository history and that all changes remain in the working tree for a mandatory review-and-land workflow.
[PROMPT_INJECTION]: The skill processes a 'brief' containing instructions that are passed to an LLM-driven agent (Aider). While this creates a surface for indirect prompt injection, the skill mitigates this through extensive documentation on defining file-access boundaries and establishing a manual verification process where all implementation work must pass local project gates before being committed.

aider-delegate