agentleFS
desplega-aidesplega-ai/agent-swarm/.claude/internal-skills/swarm-local-e2e/SKILL.md

swarm-local-e2e

Skill1.1k starsChanged 2 months agoMITView on GitHub

Guide for running local E2E tests with API server, Docker lead/worker containers, task creation, log verification, UI dashboard, and cleanup

  • Reads credentials
  • Sends data out

Adopted by

Steady: 0 new projects took it in the last 4 weeks, none the 4 weeks before.thin data

Week of Sep 20: 0 new projectsWeek of Sep 27: 0 new projectsWeek of Oct 4: 0 new projectsWeek of Oct 11: 0 new projectsWeek of Oct 18: 0 new projectsWeek of Oct 25: 0 new projectsWeek of Nov 1: 0 new projectsWeek of Nov 8: 0 new projects
Sep 2026Oct 2026Nov 2026
New projects taking a version of this file each week, over the 8 weeks our latest run read: only 0 in all, so this is a hint, not a trend.

Who's talking about desplega-ai/agent-swarm

The biggest 5 of 57 mentions linking to the repository that our crawl found. Each line is their own headline.

See all 57

Talked about on GitHub, Hacker News and Bluesky (57 times).

On GitHub · 50

45 more on GitHub

On Hacker News · 4

The file

Show its rules and text

We split this skill into the rules it gives an agent

25 rules. About 0 are common, 12 are distinctive, and 13 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈12

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 13

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Run full end-to-end tests of the agent swarm locally with a real API server and Docker containers.

    line 8
  2. Distinctive:

    Automated change verification: After implementing changes that touch the API, runner, polling, task lifecycle, session logs, Docker …

    line 16
  3. Distinctive:

    Session log changes: Run two sequential tasks on the same agent, verify log isolation (unique sessionIds, no cross-contamination)

    line 18
  4. Distinctive:

    UI changes: Start the dashboard, use agent-browser to verify rendering

    line 20
  5. Distinctive:

    API endpoint changes: Call the endpoint directly and verify the response

    line 21
  6. Very specific:

    Check .env for the configured port — do not assume 3013:

    line 34
  7. Very specific:

    Use this value as $PORT throughout.

    line 40
  8. Distinctive:

    If they don't match, update them before starting containers.

    line 48
  9. Very specific:

    Keep DESPLEGA_TELEMETRY_ENV=test on every API and container you boot for a test.

    line 64
  10. Distinctive:

    Use a unique container name to avoid conflicts with other worktrees (e.g.

    line 76
  11. Distinctive:

    Wait ~15s, then verify:

    line 89
  12. Very specific:

    If port 3201 is taken by another worktree, pick a different host port (e.g.

    line 95
  13. Very specific:

    Should show both lead and worker registered as idle.

    line 122
  14. Very specific:

    Important: Use agentId (not assignedTo) to assign tasks.

    line 136
  15. Distinctive:

    Start the dashboard to visually verify tasks, logs, and agent status:

    line 187
  16. Distinctive:

    If the UI port is taken by another worktree, start on an alternate:

    line 194
  17. Very specific:

    Use agent-browser to automate UI checks (never qa-use unless explicitly asked):

    line 203
  18. Very specific:

    Verify task was created with agentId (not assignedTo) — wrong param silently creates an unassigned task

    line 247
  19. Very specific:

    Check task status isn't already in_progress (e.g.

    line 248
  20. Very specific:

    Check worker has capacity: docker logs <container> 2>&1 | grep "capacity"

    line 253
  21. Distinctive:

    If "At capacity" — a previous task is still running.

    line 254
  22. Very specific:

    If another worktree is running, set a different PORT in .env and update MCP_BASE_URL in .env.docker* to …

    line 264
  23. Very specific:

    Check claudeSessionId is set on the task: GET /api/tasks/<id> should show it

    line 268
  24. Very specific:

    If logs were stored under wrong taskId, check the session_logs table directly

    line 269
  25. Distinctive:

    Use simple tasks like "Say hello" for E2E tests.

    line 275

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Local E2E Testing Guide

Run full end-to-end tests of the agent swarm locally with a real API server and Docker containers.

When to Use This Skill

This skill should be invoked in two modes:

  1. User-requested QA: The user asks you to run E2E tests, verify a feature, or QA a specific flow. Follow the steps below targeting what they asked for.
  2. Automated change verification: After implementing changes that touch the API, runner, polling, task lifecycle, session logs, Docker entrypoint, or worker/lead behavior — use this skill proactively to verify the changes work end-to-end. Determine what's testable based on the diff:
    • Task lifecycle changes (poll, runner, store-progress): Create assigned + pool tasks, verify they complete and have correct logs
    • Session log changes: Run two sequential tasks on the same agent, verify log isolation (unique sessionIds, no cross-contamination)
    • Docker / entrypoint changes: Build image, start containers, verify boot logs and registration
    • UI changes: Start the dashboard, use agent-browser to verify rendering
    • API endpoint changes: Call the endpoint directly and verify the response

You do not need to run every step — pick the subset relevant to the changes being tested.

Prerequisites

  • OrbStack or Docker Desktop running (open -a OrbStack if needed)
  • .env with API_KEY and PORT configured
  • .env.docker-lead with lead config (AGENT_ID, CLAUDE_CODE_OAUTH_TOKEN, MCP_BASE_URL)
  • .env.docker with worker config (AGENT_ID, CLAUDE_CODE_OAUTH_TOKEN or OPENROUTER_API_KEY, MCP_BASE_URL)

Step 1: Determine Your Port

Check .env for the configured port — do not assume 3013:

grep ^PORT= .env

Use this value as $PORT throughout. In worktrees, each worktree may have a different port. Always verify and use the value from .env.

Also verify the Docker env files match:

grep MCP_BASE_URL .env.docker-lead .env.docker
# Both should point to http://host.docker.internal:$PORT

If they don't match, update them before starting containers.

Step 2: Clean DB + Start API Server

# Kill any existing API process on your port
lsof -ti :$PORT | xargs kill 2>/dev/null

# Clean DB for fresh state
rm -f agent-swarm-db.sqlite agent-swarm-db.sqlite-wal agent-swarm-db.sqlite-shm

# Start API server (tagged as a test install so it stays out of production telemetry)
DESPLEGA_TELEMETRY_ENV=test bun run start:http &
# Wait ~3s for startup, confirm "MCP HTTP server running on http://localhost:$PORT/mcp"

Keep DESPLEGA_TELEMETRY_ENV=test on every API and container you boot for a test. Without it, the boot reports as a production install and inherits the swarm's SWARM_ORG_NAME.

Step 3: Build Docker Image

bun run docker:build:worker:slim

This builds agent-swarm-worker:slim from the current code — the slim target is faster and sufficient for E2E (all four harnesses included). Use bun run docker:build:worker + agent-swarm-worker:latest instead only when the test needs playwright/agent-browser, local postgres/redis, or glab. Rebuild after every code change.

Step 4: Start Lead Container

Use a unique container name to avoid conflicts with other worktrees (e.g. include branch name or feature):

docker run --rm -d \
  --name e2e-lead-$(git branch --show-current | tr '/' '-') \
  --env-file .env.docker-lead \
  -e DESPLEGA_TELEMETRY_ENV=test \
  -e AGENT_ROLE=lead \
  -e MAX_CONCURRENT_TASKS=1 \
  -p 3201:3000 \
  agent-swarm-worker:slim

Wait ~15s, then verify:

docker logs e2e-lead-$(git branch --show-current | tr '/' '-') 2>&1 | tail -5
# Should see: "[lead] Polling for triggers (0/1 active)..."

If port 3201 is taken by another worktree, pick a different host port (e.g. -p 3211:3000).

Step 5: Start Worker Container

docker run --rm -d \
  --name e2e-worker-$(git branch --show-current | tr '/' '-') \
  --env-file .env.docker \
  -e DESPLEGA_TELEMETRY_ENV=test \
  -e MAX_CONCURRENT_TASKS=1 \
  -p 3203:3000 \
  agent-swarm-worker:slim

Wait ~15s, then verify:

docker logs e2e-worker-$(git branch --show-current | tr '/' '-') 2>&1 | tail -5
# Should see: "[worker] Polling for triggers (0/1 active)..."

Step 6: Verify Registration

curl -s -H "Authorization: Bearer $API_KEY" "http://localhost:$PORT/api/agents" \
  | jq -r '.agents[] | "\(.name) | isLead: \(.isLead) | status: \(.status) | id: \(.id)"'

Should show both lead and worker registered as idle. Save the agent IDs for task creation.

Step 7: Create Tasks

Assigned task (picked up by lead)

const t = await (await fetch('http://localhost:$PORT/api/tasks', {
  method: 'POST', headers,
  body: JSON.stringify({ task: 'Say hello. Call store-progress with status completed.', agentId: LEAD_ID })
})).json();
console.log('Task:', t.id, '| status:', t.status);

Important: Use agentId (not assignedTo) to assign tasks. Wrong param silently creates an unassigned task.

Pool task (auto-claimed by worker)

const t = await (await fetch('http://localhost:$PORT/api/tasks', {
  method: 'POST', headers,
  body: JSON.stringify({ task: 'Say hello. Call store-progress with status completed.' })
})).json();
console.log('Pool task:', t.id, '| status:', t.status);

Workers auto-claim unassigned tasks at poll time. Leads do not auto-claim pool tasks.

Step 8: Monitor Progress

# Watch lead logs (use your container name)
docker logs -f e2e-lead-$(git branch --show-current | tr '/' '-') 2>&1 | tail -20

# Watch worker logs
docker logs -f e2e-worker-$(git branch --show-current | tr '/' '-') 2>&1 | tail -20

Poll task status:

const t = await (await fetch('http://localhost:$PORT/api/tasks/<task-id>', { headers })).json();
console.log(t.status);  // pending → in_progress → completed/failed

Step 9: Verify Session Logs

const logs = await (await fetch('http://localhost:$PORT/api/tasks/<task-id>/session-logs', { headers })).json();
console.log('Log count:', logs.logs.length);
// Should be > 0 for completed tasks

For log isolation verification (multiple sequential tasks from same agent):

const [l1, l2] = await Promise.all([
  fetch('http://localhost:$PORT/api/tasks/<task1>/session-logs', { headers }).then(r => r.json()),
  fetch('http://localhost:$PORT/api/tasks/<task2>/session-logs', { headers }).then(r => r.json()),
]);
const s1 = [...new Set(l1.logs.map(l => l.sessionId))];
const s2 = [...new Set(l2.logs.map(l => l.sessionId))];
console.log('Unique sessionIds:', s1[0] !== s2[0]);  // Should be true

Step 10: Test the Dashboard UI

Start the dashboard to visually verify tasks, logs, and agent status:

cd apps/ui && bun run dev &
# Defaults to port from APP_URL in .env (check with: grep APP_URL ../.env)

If the UI port is taken by another worktree, start on an alternate:

cd apps/ui && bun run dev --port 5276

The UI connects to the API via VITE_API_URL (check apps/ui/.env or defaults to http://localhost:$PORT).

Visual verification with agent-browser

Use agent-browser to automate UI checks (never qa-use unless explicitly asked):

# Quick visual gut-check with agent-browser
agent-browser open http://localhost:5175
agent-browser snapshot
agent-browser screenshot /tmp/e2e-dashboard.png

Things to verify in the UI:

  • Agents page: Lead and worker both show as registered with correct status
  • Tasks page: Tasks appear with correct status, assigned agent, and timestamps
  • Task detail → Logs tab: Session logs render in the conversation viewer (not "No session data available")
  • Task detail → Outcome tab: Completed tasks show output
  • Costs: Session costs appear for completed tasks

Step 11: Cleanup

# Stop containers (use your branch-specific names)
docker stop e2e-lead-$(git branch --show-current | tr '/' '-') e2e-worker-$(git branch --show-current | tr '/' '-') 2>/dev/null

# Stop API server
lsof -ti :$PORT | xargs kill 2>/dev/null

# Stop UI dev server (if started)
lsof -ti :5175 | xargs kill 2>/dev/null

Troubleshooting

Docker daemon not running

ERROR: Cannot connect to the Docker daemon

Fix: open -a OrbStack and wait ~5s.

Container name conflict

docker: Error response from daemon: Conflict. The container name "..." is already in use

Another worktree has a container with the same name. Either stop it (docker stop <name>) or use a different name suffix.

Lead not picking up tasks

  • Verify task was created with agentId (not assignedTo) — wrong param silently creates an unassigned task
  • Check task status isn't already in_progress (e.g. from a manual poll call that consumed the trigger)
  • Restart container if stuck: docker restart <container-name>

Worker not picking up pool tasks

  • Workers auto-claim via poll. Leads do not claim pool tasks.
  • Check worker has capacity: docker logs <container> 2>&1 | grep "capacity"
  • If "At capacity" — a previous task is still running. Wait or restart.

Poll returns 404

  • Poll endpoint is GET /api/poll (not POST)
  • Requires X-Agent-ID header with a valid agent UUID

Port conflicts (worktrees)

lsof -i :3013  # Check what's using the port

If another worktree is running, set a different PORT in .env and update MCP_BASE_URL in .env.docker* to http://host.docker.internal:<new-port>.

Session logs show 0 entries

  • Task must have actually run (status completed or failed, not just in_progress)
  • Check claudeSessionId is set on the task: GET /api/tasks/<id> should show it
  • If logs were stored under wrong taskId, check the session_logs table directly

Task cancellation doesn't stop Claude

Direct API cancellation (POST /api/tasks/<id>/cancel) updates the DB but doesn't kill the Claude process inside Docker. Use docker restart <container> to force-stop.

Keep tasks trivial

Use simple tasks like "Say hello" for E2E tests. Complex tasks waste time and API credits.

UI shows stale data

The dashboard auto-polls every 5 seconds. If data looks stale, hard-refresh (Cmd+Shift+R) or check VITE_API_URL points to the correct API port.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.