The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The skill uses !command`` syntax in SKILL.md to execute shell commands automatically when the skill is loaded. This is used to run skill-toc and list the skill's internal references.
[PERSISTENCE]: Instructions in the skill direct the agent to establish a heartbeat poll using the CronCreate tool to maintain progress across session breaks.
[DYNAMIC_EXECUTION]: The work-stage.mjs script uses the AsyncFunction constructor to dynamically load and execute JavaScript logic from the workflow.js file at runtime.
[COMMAND_EXECUTION]: The skill is built to execute arbitrary shell commands defined in a 'plan' (e.g., redCommand, acceptanceCmd, and mechanical checks). This allows for extensive system interaction based on the configuration provided in Markdown plans.
[DATA_EXFILTRATION]: The rule-judging functionality in rule-check.ts sends data to the OpenRouter API. This data includes repository file paths and extracts used as evidence for compliance checks.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from command outputs and repository files into a review lens. The following evidence chain applies:
Ingestion points: Command results and file content are read into the agent context via scripts like work-checks.sh.
Boundary markers: Prompt templates use explicit headers like AUTHORITY: and GOAL: to demarcate instructions.
Capability inventory: The skill uses Bash, Write, and Agent tools across various scripts.
Sanitization: The red_probe_gate in work-dispatch.sh includes a regex-based filter to avoid processing sensitive files like .env or SSH keys.

work