- Open https://renderops-director-w6mw3t2ita-uc.a.run.app.
- Switch between Original plate, Failed render, and Canary fix and verify that all three previews play.
- Confirm that the frame timeline marks failed frames and explains why editorial review is blocked.
- Keep the default shot
SH-042and mission, then click Investigate shot. - Wait up to 90 seconds for the live Gemini and Grafana MCP workflow.
- Verify that the result includes:
- runtime
Gemini + MCP; - at least three evidence items;
- failed-frame, GPU-memory, log, or trace evidence;
- recovery cost
$35.90; - avoided cost
$150.50 · 81%; - exactly three recovery actions;
- a human approval requirement.
- runtime
- Click Approve canary and verify:
5 / 5frames;- GPU memory
96% → 72%; - successful Prometheus, Loki, and Tempo checks;
- a real FFmpeg check with exit code, duration, frames, and output bytes;
- the 38-frame recovery becomes enabled.
- Click Approve 38 frames and verify
38 / 38,EDITORIAL READY, and the recovered video. - Open Runtime evidence and verify MCP tool names such as
query_prometheus,query_loki_logs, andtempo_traceql-search.
The initial cinematic GPU failure is synthetic and reproducible. Approved canary and recovery actions run real FFmpeg processing; their output files and actual execution telemetry are then verified through Grafana Cloud and official MCP queries.
curl -fsS https://renderops-director-w6mw3t2ita-uc.a.run.app/api/healthExpected indicators:
{
"status": "ok",
"version": "0.8.0",
"agent_runtime": "live",
"grafana_transport": "stdio"
}python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[dev]'
python -m ruff check .
python -m pytest
python -m compileall app
docker build -t renderops-director:test .- No Grafana write tools are exposed.
mcp-grafanaruns with--disable-write.- Tokens are injected through Google Secret Manager.
- Render mutations require human approval.