The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains strong instructional directives such as 'IMPORTANT: READ VERY CLOSELY' and sections labeled 'Never' and 'Always'. These are used to define the operational constraints and safety boundaries of the orchestration process, rather than attempting to bypass the agent's core safety guidelines.
[DATA_EXFILTRATION]: No patterns of unauthorized data exposure or exfiltration were found. The skill interacts with local project files and configuration, maintaining a local execution scope.
[REMOTE_CODE_EXECUTION]: The skill does not execute remote code from external sources. It manages local subprocesses and other internal agent skills using established routing logic.
[INDIRECT_PROMPT_INJECTION]: The skill manages the risk of malicious input from sub-agent workers by implementing a multi-stage verification funnel. 1. Ingestion points: The orchestrator ingests code deltas (diffs) produced by workers and reads from local doctrine.toml files. 2. Boundary markers: The skill mandates the use of 'doctrine check prove' for linting and 'doctrine check regression' to capture and compare baselines before and after worker execution. 3. Capability inventory: The skill utilizes git for worktree management, local doctrine CLI tools, and internal routing to sub-agent skills. 4. Sanitization: There is an explicit 'R-5 belt' rule that rejects any modifications to the .doctrine/ or .claude/ directories, preventing sub-agents from altering the coordination framework's configuration.

dispatch