What problem does it solve? When multiple agent systems are available β Orca for lifecycle, the harness's SendMessage for content between Claude Code sessions, and ecp for code-graph visibility β it is easy to send a message down the wrong channel, collide with another agent's edits, or hand-write a Playwright script when an embedded browser already exists. This Skill decides which runner, channel, or browser each piece of work belongs to. ## Core Features & Use Cases - Channel routing: Lifecycle messages (worker_done, heartbeat, escalation) stay on Orca; follow-ups and content between Claude Code sessions go over SendMessage; codex, opencode, gemini, and bare shells go through Orca terminals. - Collision-aware dispatch: Runs ecp peers plan before fanning work out so each dispatched task owns a disjoint surface of the code graph, and chains overlapping targets with task-create --deps. - Browser routing: Any task that needs to see a page β rendering a URL, screenshotting a dev server, checking layout β goes to Orca's embedded browser via orca-cli, never a hand-written Playwright script; desktop windows go to computer-use. - Guide drift detection: The bundled check-anchors.sh script verifies that three quoted passages from the Orca orchestration guide still match the installed Orca version, warning when the overrides in SKILL.md go stale. - Use Case: Before dispatching three workers to refactor separate modules, run ecp peers plan --targets to confirm their blast radii do not intersect, then brief each Claude worker with SendMessage while a codex worker gets its task through an Orca terminal. ## Quick Start Ask the agent to decide which runner and channel should handle a task, for example whether to dispatch a code change through Orca or message an existing Claude session directly.