The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the orca command-line utility and local Node.js scripts (mo-harness-screen.mjs, mo-review-resource.mjs) to interact with terminal sessions, read screen content, and manage development worktrees.
[INDIRECT_PROMPT_INJECTION]: The skill gathers status information by querying other active agents and reviewers for messages regarding blocking issues. These external responses are processed by the agent to generate diagnostic recommendations, creating an attack surface where untrusted input from other agents could influence behavior.
Ingestion points: SKILL.md (Gathering positions by querying messages from other agent terminals).
Boundary markers: The skill utilizes scripts/mo-harness-screen.mjs to verify terminal states and ensure task bytes are only injected into a confirmed agent_prompt state, reducing the risk of accidental instruction following.
Capability inventory: Terminal read/write access, worktree management, and local script execution (SKILL.md, references/orca-mechanics.md).
Sanitization: The methodology references a redaction mechanism for credentials in references/methodology.md, but no specific filters are described for instructions embedded within the ingested agent responses.
[PRIVILEGE_ESCALATION]: The architecture defined in references/backend-contract.md and references/orca-mechanics.md requires that agent harnesses (Codex, Claude Code, OpenCode) be launched with an "unsandboxed posture." This configuration grants the agents unrestricted access to the host system and filesystem, which is a significant security stance for the execution environment.

mo-convergence