The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data like specifications, stories, and issues, which could potentially contain malicious instructions targeting the agent.
Ingestion points: The orchestrator reads specs, issues, and repository rules (SKILL.md, Section 1).
Boundary markers: The skill includes a defensive instruction in references/briefs.md that mandates sub-agents treat all input as data and ignore any embedded instructions.
Capability inventory: The skill utilizes file system access, git operations, and sub-agent execution.
Sanitization: Instructions require agents to flag instruction-shaped text rather than obeying it.
[DYNAMIC_EXECUTION]: The skill generates JavaScript orchestration scripts to be executed by the Workflow tool in Claude Code based on the task graph.
Evidence: references/dynamic-workflows.md provides rules and templates for generating these scripts.
Context: These scripts are constructed dynamically to manage the parallel or serial execution of sub-agents.
[PERSISTENCE]: The skill implements a mechanism to maintain state and resume execution across sessions using system-level schedulers to manage usage limits.
Evidence: SKILL.md (Section 6) and references/dynamic-workflows.md describe a 'Heartbeat' mechanism that monitors limits and schedules a resume via cron jobs or scheduled tasks.
Purpose: This persistence is used to ensure long-running multi-agent tasks can survive platform-imposed token or rate limit resets.
[COMMAND_EXECUTION]: The workflow relies on executing shell commands for repository management and verification tasks.
Evidence: The instructions involve automated use of git (fetch, merge, worktree) and package managers (npm install, npm test) to prepare isolated worker environments.

graph-engineering