Testing Livepeer

Artifact Discovery

Source: prebuilt artifacts from the Upload artifacts to google bucket step in the Build binaries workflow for PR #3884. This PR (ja/serverless) is still WIP — once it merges, update this skill to pull from a release or master build instead.

Always resolve the latest head SHA dynamically — do not hardcode a SHA. Run:

SHA=$(gh pr view 3884 --repo livepeer/go-livepeer --json headRefOid -q '.headRefOid')
echo "Latest SHA: $SHA"

Then construct the download URL for the current platform:

# Detect platform
ARCH=$(uname -m)
OS=$(uname -s | tr '[:upper:]' '[:lower:]')
case "$ARCH" in
  x86_64)  ARCH="amd64" ;;
  aarch64|arm64) ARCH="arm64" ;;
esac
URL="https://build.livepeer.live/go-livepeer/${SHA}/livepeer-${OS}-${ARCH}.tar.gz"
echo "Download URL: $URL"

Full platform matrix (all use https://build.livepeer.live/go-livepeer/${SHA}/ as base):

• livepeer-darwin-amd64.tar.gz
• livepeer-darwin-arm64.tar.gz
• livepeer-linux-amd64.tar.gz
• livepeer-linux-arm64.tar.gz
• livepeer-linux-gpu-amd64.tar.gz
• livepeer-linux-gpu-arm64.tar.gz
• livepeer-windows-amd64.zip

If the artifact is missing, the PR head may have changed or CI hasn't finished. Verify with gh pr view 3884 --repo livepeer/go-livepeer --json headRefOid,statusCheckRollup. Never build go-livepeer from source.

serverless.json

[
  {
    "pipeline": "live-video-to-video",
    "model_id": "scope",
    "url": "ws://localhost:8001/ws",
    "warm": false,
    "capacity": 10
  }
]

The url points to the local livepeer-runner websocket endpoint, not the Scope server.

Start The Local Stack

Kill old processes on the relevant ports first:

lsof -ti:8001 -ti:8022 -ti:8935 | xargs kill -9 2>/dev/null
lsof -tiUDP:9001 | xargs kill -9 2>/dev/null

Orchestrator:

./livepeer -orchestrator -aiWorker -aiServerless -aiModels serverless.json -serviceAddr localhost:8935

Runner (SCOPE_PORT=9001 prevents OSC port collision with the main Scope server; LIVEPEER_DEV_MODE=1 skips auth for local testing):

# CRITICAL: The runner needs the websockets library for its /ws endpoint.
# Without it, the orchestrator gets "websocket: bad handshake" when connecting.
uv pip install 'uvicorn[standard]'

LIVEPEER_DEV_MODE=1 SCOPE_PORT=9001 UV_NO_SYNC=1 uv run --extra livepeer livepeer-runner

The runner has no /health endpoint; verify by checking logs and that port 8001 is listening.

Scope (LIVEPEER_SIGNER=off is intentional for local dev; do not add signing unless the user asks):

LIVEPEER_DEV_MODE=1 \
LIVEPEER_SIGNER=off \
LIVEPEER_ORCH_URL=localhost:8935 \
SCOPE_CLOUD_MODE=livepeer \
UV_NO_SYNC=1 \
uv run daydream-scope --no-browser --port 8022

Wait for health: curl -s http://localhost:8022/health

For MCP-driven testing after Scope is healthy, follow the testing-scope-mcp skill to start the MCP server and connect to port 8022.

Verification

Verify in this order, and do not skip ahead:

1. go-livepeer runs and accepts serverless.json
2. livepeer-runner starts on ws://localhost:8001/ws and stays running
3. Scope starts without import, config, or OSC port-collision errors
4. Scope operates in SCOPE_CLOUD_MODE=livepeer
5. The specific user-facing behavior works end to end

Do not declare success after partial recovery. The final check must cover the full Livepeer-mode scenario the user cares about. A rerun after any fix must behave consistently.

Interpretation note: a successful connect_to_cloud + pipeline/load proves the control path. A headless start_stream + capture_frame may exercise local media paths only. If the user cares about true remote media/WebRTC behavior, finish with a browser or UI-driven verification.

Headless Session in Livepeer Mode

The headless session (/api/v1/session/start) runs locally on the Scope instance — it is not proxied to cloud. However, when Scope is connected to cloud, POST /api/v1/pipeline/load is proxied to the runner. This means the pipeline gets loaded only on the runner side, and the local headless session fails with "Pipeline passthrough not loaded".

Workaround — load the pipeline locally before starting the session:

1. POST /api/v1/cloud/disconnect — disconnect from cloud
2. POST /api/v1/pipeline/load body: {"pipeline_ids": ["passthrough"]} — loads locally (no proxy)
3. POST /api/v1/cloud/connect body: {} — reconnect to cloud
4. Wait for cloud status to show connected: true
5. POST /api/v1/session/start body: {"pipeline_id": "passthrough", ...} — now works

This applies to all pipelines in Livepeer mode, not just passthrough.

Video input for passthrough: Use the bundled test video at frontend/public/assets/test.mp4 as input_source with source_type: "video_file" and input_mode: "video". Do not create a synthetic test video with ffmpeg.

Troubleshooting

When something fails:

1. Identify the failing boundary: orchestrator / runner / Scope startup / cloud connect / websocket handshake / stream behavior.
2. Read logs from the failing process first.
3. Make the narrowest fix and restart only the affected processes.
4. Rerun the failing step, then the full end-to-end flow.

HTTP API fallbacks useful during Livepeer testing (when MCP is unavailable):

Practical Notes

• If uv run --extra livepeer ... fails resolving a platform-inapplicable wheel (e.g., flash-attn on macOS), retry with UV_NO_SYNC=1 uv run --extra livepeer ....
• If Scope cannot open ~/.daydream-scope/logs in a sandboxed environment, set DAYDREAM_SCOPE_LOGS_DIR to a writable directory.
• If using the Docker artifact path instead of the native binary, ensure the container can reach the host runner on port 8001.

Updating This SKILL.md

Artifact URLs are resolved dynamically from the PR head SHA (see "Artifact Discovery"), so this file does not need manual SHA updates. If the PR merges, update the artifact discovery section to pull from a master build or release instead of the PR.