orchestra-monitor

name	orchestra-monitor
description	Monitor an orchestra run — team status, unread messages, costs, failures, and live activity. Designed to run with /loop for continuous monitoring as a human coordinator.
argument-hint	[path-to-.orchestra] — e.g., ~/dev/my-project/.orchestra
user_invocable	true

/orchestra-monitor -- Orchestra Run Monitor

Single-pass status check for an active orchestra run. Designed to be called via /loop 1m /orchestra-monitor <path>.

Workspace path

If <user_argument> is provided, use it as the path to the .orchestra/ directory. Otherwise default to .orchestra/ in the current working directory.

Store the resolved path as $WS for the rest of this skill. Messages are at $WS/messages/.

Steps

Perform ALL of the following checks in a single pass. Be concise — this runs every minute.

1. Team Status

Read $WS/state.json and $WS/registry.json. state.json uses the v2 run-state shape: top-level project, optional backend/run_id/started_at, and a teams map keyed by team name. Local runs still use status: "done" for completed teams.

Count teams by status and show a one-line summary:

Teams: 2/7 done | 2 running | 3 pending | 0 failed

If any team has status "failed", flag it prominently:

!! FAILED: data-engine — check .orchestra/logs/data-engine.log

2. Cost & Duration

From state.json, sum cost_usd across all completed teams. From registry.json, compute wall clock for running teams (now - started_at).

Cost: $4.23 total | Running: full-stack (12m), security-caching (8m)

3. Live Activity (per running team)

For each running team, provide a SHORT activity snapshot. Use these methods (in order of preference):

a) Files created/modified — Run find on the project directory for files modified since the team started. Show counts and notable recent files:

  phoenix-ideation (4m): 3 files written (research-notes.md, brainstorm.md, architecture.md)
  titan-ideation (4m): 1 file written (research-notes.md), 2 teammates active

b) Latest log line — Extract the most recent meaningful assistant text from $WS/logs/<team>.log. Parse the JSON log for the last "type":"assistant" message with "type":"text" content. Show a truncated snippet (max 80 chars):

  hydra-ideation (4m): "Evaluating swarm consensus patterns for multi-agent coor..."

c) Teammate count — Check if the team has spawned subagents by looking for TeamCreate/SendMessage tool calls in the log:

  cipher-ideation (4m): 3 teammates spawned, waiting on results

Pick whichever signals are available. Show 1 line per running team. Skip teams with nothing interesting.

4. Deliverables Progress

For each running team, check what deliverable files exist vs what's expected. Use the orchestra.yaml deliverables or verify fields as a guide, or simply count files in the team's expected output directory.

Deliverables: phoenix 2/6, titan 4/6, hydra 1/6, nova 0/6, cipher 3/6

If deliverables aren't clearly defined or no files exist yet, omit this section.

5. Unread Messages

Check ALL inboxes under $WS/messages/ for unread messages (JSON files where "read": false).

Focus on 0-human/inbox/ and 1-coordinator/inbox/ first (messages directed at you), then scan team inboxes.

If unread messages exist, list them:

Unread messages (3):
  -> 0-human: [gate] "Need approval for API key rotation" from 2-data-engine
  -> 1-coordinator: [blocking-issue] "Missing dep" from 3-full-stack
  -> 1-coordinator: [status-update] "API routes done" from 3-full-stack

If no unread messages: No unread messages.

6. Shared Artifacts

List any files in .orchestra/messages/shared/:

Shared artifacts: 2 (api-contract-v1.json, db-schema-v1.json)

If none: omit this section entirely.

Output Format

Keep it tight but informative — this is a dashboard glance. Target ~8-15 lines per tick. Example of a full output:

[orchestra] cf-open-estimate — 14m elapsed
Teams: 2/7 done | 2 running | 3 pending | 0 failed
Cost: $4.23 | Running: full-stack (12m), security-caching (8m)

Activity:
  full-stack (12m): 5 files written (routes.go, handlers.go, main.go...), "Wiring up health endpoint..."
  security-caching (8m): 3 teammates spawned, 2 files written (ratelimit.go, safeguard.go)

Deliverables: full-stack 3/5, security-caching 2/4

Unread messages (1):
  -> 1-coordinator: [blocking-issue] "Missing dep" from 3-full-stack

Shared: api-contract-v1.json

Quiet tick (no messages, early in run):

[orchestra] Agent Hackathon — 4m elapsed
Teams: 0/11 done | 5 running | 6 pending | 0 failed
Cost: $0.00

Activity:
  phoenix-ideation (4m): 3 files written (research-notes.md, brainstorm.md, architecture.md)
  hydra-ideation (4m): 2 teammates active, "Designing swarm consensus protocol..."
  titan-ideation (4m): concept picked: SentinelAgent, 1 file written
  nova-ideation (4m): 3 teammates spawned, researching creative AI patterns
  cipher-ideation (4m): 2 files written (research-notes.md, threat-model.md)

No unread messages

Important

• Do NOT read full message contents — just show sender, type, and subject
• If the user wants to read or respond to a message, tell them to use /orchestra-inbox or /orchestra-msg
• If a team has failed, suggest checking the log file at $WS/logs/<team>.log
• The Activity section is the most valuable part — make it count. Show what's actually happening, not just "running".
• When extracting log snippets, skip tool calls and system messages — only show assistant text that reveals intent or progress.