The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill orchestrates multiple sub-agents and consolidates their reports. These reports, along with external memory entries and user directives, represent untrusted data sources that could potentially contain malicious instructions intended to influence the orchestrating agent's behavior.
Ingestion points: Processes reports returned by dispatched agents, entries from project memories via list_memories, and CI failure logs fetched from remote repositories.
Boundary markers: The skill requires reports to cite checkable evidence, such as specific file:line references or command output, to prevent blind obedience to agent claims. It also warns that a dispatch-prompt override is not a guarantee of compliance.
Capability inventory: The agent can modify the file system (using the jj tool and direct edits), execute shell commands for verification and build processes, and dispatch additional autonomous agents.
Sanitization: The instructions explicitly mandate that the orchestrator must synthesize findings itself and verify any fixes prescribed by an agent before adoption, which serves as a robust semantic sanitization process.
[COMMAND_EXECUTION]: The workflow relies on the execution of shell commands through the jj (Jujutsu) and gh (GitHub) CLI tools to manage workspaces, fetch repository state, and interact with the Git history. It also facilitates the execution of arbitrary test and build commands as part of its verification gates.

execution-workflow