Skill: Orchestrate

CRITICAL: Run HARNESS_DIR=$(epic-harness path) first. NEVER use .harness/ in the project directory.

Process

Each mode follows its own step sequence below. In auto mode: check state → read inbox → check control → check deps → execute → hand off. In status mode: check state → read agents → render dashboard. In intervene mode: check state → parse command → write directive → confirm.

Mode: auto (default)

Triggered automatically when an active multi-agent orchestration is detected.

Step 1: Check orchestration state

1. Run HARNESS_DIR=$(epic-harness path)
2. Read $HARNESS_DIR/orchestrator/run.json
3. If no active run, this skill does not apply — return

Step 2: Read agent inbox

1. Read $HARNESS_DIR/orchestrator/agents/{my_id}/inbox.jsonl
2. Process any pending messages from other agents or user interventions
3. Messages from other agents: integrate into current task context
4. Messages from user (via intervene redirect): adjust task accordingly

Step 3: Check control directives

1. Read $HARNESS_DIR/orchestrator/control.json
2. If a directive targets this agent (or "all"), act on it:
   • pause: halt work after current tool call, wait for resume
   • cancel: stop immediately, set status to "cancelled"
   • redirect: adopt the new instruction, discard current task
   • resume: resume from paused state
3. After acting on a directive, increment generation in control.json to acknowledge

Step 4: Check dependencies

1. Read $HARNESS_DIR/orchestrator/run.json dependency graph
2. Identify which agents this agent depends on
3. Read their status.json — if any dependency is not DONE, wait
4. If all dependencies are DONE, proceed with task

Step 5: Execute with progress reporting

While working on the assigned task:

1. The Rust orchestrate hook automatically appends events to stream.jsonl on each tool call
2. Status updates happen automatically via the hook
3. No manual progress reporting needed — the hook handles it

Step 6: Complete and hand off

When task is complete:

1. Report status via structured output format (DONE/DONE_WITH_CONCERNS/BLOCKED/NEEDS_CONTEXT)
2. The orchestrate hook will evaluate dependencies and notify downstream agents
3. If results should be shared with specific agents, include in the output summary

---

Mode: status — Orchestration Dashboard

Read-only view of the current orchestration run. Displayed when user runs /status or asks about orchestration state.

Step 1: Check orchestration state

1. Check if $HARNESS_DIR/orchestrator/run.json exists
2. If the file does not exist, output:

   No active orchestration. Run /go or /orbit to start.

3. If status is not "running", output:

   Last orchestration ended with status: {status}. No active run.

4. If status is "running", proceed to Step 2.

Step 2: Read all agent states

Extract the agent list from run.json.agents. For each agent {id}:

1. Read agent status from $HARNESS_DIR/orchestrator/agents/{id}/status.json (mark unknown if missing)
2. Read the last 5 events from $HARNESS_DIR/orchestrator/agents/{id}/stream.jsonl (show "no events yet" if missing)
3. Calculate elapsed time:
   • started_at set, completed_at not set: elapsed = now - started_at
   • Both set: elapsed = completed_at - started_at
   • Format as Xm Ys or Xh Ym
4. Read the dependency graph from run.json.dependencies

Step 3: Render dashboard

Output a formatted dashboard (keep total under 30 lines):

## Orchestration Dashboard

- **Run ID**: {run.id}
- **Status**: {run.status}
- **Elapsed**: {total_elapsed since run.started_at}
- **Agents**: {count where status=running} / {total} active

| Agent | Role | Status | Elapsed | Last Event |
|-------|------|--------|---------|------------|
| {id} | {role} | {status} | {elapsed} | {summary of last event} |

### Dependency Graph
{render with ASCII arrows, e.g.:}
  builder ──→ reviewer ──→ integrator
  builder ──→ tester

### Recent Events (last 5 across all agents)
- [{timestamp}] {agent_id}: {event_summary}

### User Interventions
{read control.json; show directives if any, else "none"}

Dependency graph rendering rules:

• Read edges from run.json.dependencies (format: {"agent_a": ["agent_b", "agent_c"]})
• Render using ──→ arrows
• If no dependencies exist, output: "No inter-agent dependencies."

Event summary formatting:

• Each event is one line from stream.jsonl with fields: timestamp, type, message
• Summarize message to at most 60 characters
• Sort by timestamp descending, take top 5

Step 4: Optional — file change summary

git diff --stat

If changes exist, append:

### File Changes
- {count} files modified ({insertions} insertions, {deletions} deletions)
- Key changes: {list up to 5 file paths}

Note: This mode is read-only — never modify orchestration state.

---

Mode: intervene — Agent Intervention

Write control directives that the orchestrate hook reads on the next agent tool call. Triggered when user runs /intervene.

Step 1: Check active orchestration

1. Read $HARNESS_DIR/orchestrator/run.json
2. If no active run (status != "running"), output: "No active orchestration to intervene in."
3. Show current agent list with statuses

Step 2: Parse user command

Invocation format:

• intervene pause {agent_id} — pause a specific agent (blocks next tool call)
• intervene pause all — pause all agents
• intervene cancel {agent_id} — cancel a specific agent
• intervene cancel all — cancel entire orchestration
• intervene redirect {agent_id} {new_instruction} — change what an agent is doing
• intervene resume {agent_id} — resume a paused agent

If the user provides no arguments, ask which action and target they want.

Validate the agent_id against the agent list in run.json. If not found, output: "Agent '{agent_id}' not found in active orchestration. Available agents: {list}" and stop.

Step 3: Write control directive

Read current generation from $HARNESS_DIR/orchestrator/control.json (or 0 if file doesn't exist). Increment by 1.

Write to $HARNESS_DIR/orchestrator/control.json:

{
  "action": "pause|cancel|redirect|resume",
  "target": "agent_id or 'all'",
  "message": "optional user message",
  "generation": <current_generation + 1>
}

For redirect, the message field contains the new instruction. For pause/cancel, the message field is optional — use an empty string if none provided. For resume, set action to "resume" and target to the agent_id.

Additional actions by type:

• cancel all: Also update $HARNESS_DIR/orchestrator/run.json status to "aborted"
• redirect: Also append the new instruction to the target agent's $HARNESS_DIR/orchestrator/agents/{agent_id}/inbox.jsonl as a new line: {"type": "redirect", "instruction": "{new_instruction}", "at": "{ISO-8601}"}
• cancel {agent_id}: Also update the agent's status in run.json to "cancelled"

Step 4: Confirm

Output confirmation:

Intervention recorded: {action} {target}
The {target} agent will respond on its next tool call.
Use /status to monitor.

Notes:

• Multiple interventions can be queued (generation number increases each time)
• The orchestrate hook checks control.json before every agent tool call, so interventions take effect within one tool call cycle
• "cancel all" is destructive — confirm with the user before executing
• "redirect" requires a new instruction — if none is provided, ask the user

---

Message Format Convention

When agents need to communicate (via SendMessage or future inbox/outbox):

{
  "from": "agent_id",
  "to": "agent_id",
  "type": "handoff|question|blocked|result",
  "body": "message content",
  "timestamp": "ISO-8601"
}

Anti-Rationalization

Excuse	Rebuttal	What to do instead
"I'll just work independently"	Orchestration exists to prevent conflicts and duplication	Check dependencies first, report status
"Progress reporting slows me down"	The hook does it automatically on every tool call	No manual action needed
"I'll read all agents' state"	Only read your own inbox and dependencies	Respect isolation boundaries
"I'll intervene directly in agent files"	Direct file edits bypass the generation protocol	Always write to control.json
"I'll skip the status check"	Stale state leads to incorrect interventions	Always read run.json first

Evidence Required

• Inbox checked before starting work
• Control directives checked (auto mode)
• Dependencies verified (all upstream agents DONE)
• Structured output format used for completion report
• Dashboard shows all agent states accurately (status mode)
• Intervention written to control.json with incremented generation (intervene mode)

Red Flags

• Starting work before dependencies are met
• Ignoring inbox messages from other agents
• Not reporting BLOCKED state when stuck
• Modifying orchestration state in status mode (read-only)
• Cancelling all agents without user confirmation
• Redirecting an agent without a concrete new instruction
• Writing to control.json when run.json status is "aborted" or "complete"
• Intervening without an active orchestration (check run.json first)
• Showing raw JSON instead of a formatted dashboard (status mode)