EP000 · IMPLEMENTED RUNTIME · AUTHENTICATE CLAUDE AND GEMINI TO COMPLETE FOUR-AGENT LAUNCH Cooperative CLI Workflow Claude, Gemini, Hermes and Codex coordinate through a standalone runtime and the Agent Message Board while retaining independent tasks, cursors and file ownership. Native WSL profiles are configured; Codex and Hermes passed a real smoke launch.
Allocation and live-state runtime implemented Codex + Hermes smoke passed Claude + Gemini authentication pending
All stages Live state Launch Execute Coordinate Optional watcher
0 · Continuous workflow truth — mandatory
STATE.0 Allocate all agent nodes Before execution, agree the complete node-to-agent map and reject overlapping assignments.
0% planned
STATE.1 Assign node owner Before work starts, bind the node to agent, model, task ID and execution ID.
0% planned
STATE.2 Stream node activity Continuously publish current action, percent, heartbeat time, blockers and evidence.
0% planned
STATE.3 Render live workflow Update the workflow display from structured events so everyone sees who is doing what.
0% planned
STATE.4 Detect stale ownership Mark silent or expired agents stale; never leave abandoned nodes looking active.
0% planned
1 · Launch and preflight
LAUNCH.1 Invoke wrapper User starts one named CLI with task ID, prompt, files and topics.
0% planned
LAUNCH.2 Resolve agent profile Load executable, model, working directory, timeout and cooperation policy.
0% planned
LAUNCH.3 Board preflight Read relevant unread events and active claims without moving the cursor.
0% planned
2 · Safety and context gate
GATE.1 Filter relevance Match exact task, normalized files, dependencies, topics and broadcasts.
0% planned
GATE.2 Check edit claims Surface overlaps before shared-file editing; advisory claims never become locks.
0% planned
GATE.3 Build safe prompt Inject board context as untrusted coordination data plus explicit task boundaries.
0% planned
3 · Execute individual task
EXEC.1 Assess impact Model classifies each relevant event before beginning task work.
0% planned
EXEC.2 Post safe responses Auto-post low-risk assessments; pause for blockers, changes or conflicts.
0% planned
EXEC.3 Run CLI task The selected model works only within its assigned scope and ownership.
0% planned
4 · Cooperate during work
COOP.1 Publish discoveries Material findings are timestamped and scoped to affected tasks, files and topics.
0% planned
COOP.2 Checkpoint re-read Re-read before shared edits and before completion; debounce related events.
0% planned
COOP.3 Conflict stop gate Pause overlapping edits, post conflict, and await an explicit ownership decision.
0% planned
5 · Complete and checkpoint
CLOSE.1 Final board read Check events arriving during execution before claiming the task is complete.
0% planned
CLOSE.2 Post outcome Publish completion, handoff or blocker with concrete validation evidence.
0% planned
CLOSE.3 Advance own cursor Advance only after successful assessment and preserve other agents' positions.
0% planned
6 · Optional idle watcher
WATCH.1 Poll unread events A separate lightweight process checks only events newer than the agent cursor.
0% planned
WATCH.2 Debounce and budget Group related events and enforce invocation, cost and retry limits.
0% planned
WATCH.3 Request assessment turn Invoke a short model turn only for relevant events; never edit task files itself.
0% planned
Runtime: skills/cooperative-cli/scripts/cooperative_cli.py. Use create, allocate, activate, update, handoff, reassign, not-required, stale and render. Reassign only queued, released, blocked or stale nodes; active work must use a recorded handoff. Mark a superseded node or workflow not-required with its reason; active work must first be released or handed off. Generated live views are stored under agent_board/workflows/<workflow_id>/index.html.