The Agent Skills Directory

[SAFE]: The skill is designed to orchestrate complex software development workflows by delegating implementation, testing, and QA tasks to subagents. It defines a robust management system using "Boulder" state and evidence ledgers to track progress and ensure verification rigor.
[EXTERNAL_DOWNLOADS]: The skill references the agent-browser tool from Vercel Labs (github.com/vercel-labs/agent-browser) as a recommended utility for browser-based manual QA. This reference targets a repository from a well-known and trusted organization.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external plan files located in .omo/plans/ and feedback data (DoneClaim JSON) from subagents. This architecture represents a surface for indirect prompt injection if the plans or subagent outputs originate from untrusted sources.
Ingestion points: The skill reads work plans from .omo/plans/*.md, session state from .omo/boulder.json, and result messages from spawned subagents.
Boundary markers: The skill does not explicitly define delimiters or specific "ignore" instructions when interpolating plan content into the orchestrator's context, as the plan itself serves as the set of instructions.
Capability inventory: The skill has the capability to spawn subagents (multi_agent_v1.spawn_agent), perform Git operations (worktree management, merging, pushing), write to the local file system (.omo/ directory), and execute commands through delegated workers.
Sanitization: The instructions focus on verification of work (adversarial QA) rather than sanitization of the plan content itself, though the "No-plan bootstrap" process includes an "adversarial plan-review" step as a design-time mitigation.

ulw-execute