The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external plan files and repository instructions to define its execution logic, representing a vulnerability surface where malicious instructions could attempt to subvert the orchestration process (e.g., bypassing gates or overriding boundaries).
Ingestion points: The skill reads the 'plan', 'repository instructions', and 'resumes' from persisted state files in Phase 1 and Phase 5.
Boundary markers: The skill explicitly establishes 'write boundaries', 'write isolation', and 'authorization boundaries' to mitigate scope creep.
Capability inventory: The skill invokes delegation tools (flow-work, flow-verify), writes state records to the .plans/ directory, and reads repository source code.
Sanitization: The skill mandates that verification occurs in a 'separate agent or fresh context' to prevent worker implementations from influencing the ground truth of the verification phase.
[COMMAND_EXECUTION]: The orchestrator coordinates the execution of tasks and verification commands defined within the external plan.
Evidence: Phase 3 instructs the agent to dispatch 'exact verification commands' to workers, and Phase 4 executes the flow-verify skill based on the plan's contract.

flow-orchestrate