OpenClaw COLO interaction benchmark (2026-07-07)

Publication redaction: The operator-specific network hostname was replaced with a public-safe placeholder. The observed workflow and result are unchanged.

STATUS: LIVE PASS CAPTURED. From the Companion Node OpenClaw gateway host, the COLO runner reached the Primary Node anvil-serving router and completed the repeatable direct-router interaction benchmark without HTTP failures, truncation, or warnings. Separate OpenClaw config/plugin checks in the same artifact verified that the gateway had the Anvil provider and intent plugin installed.

This note is the site/blog-ready benchmark citation for the OpenClaw gateway to anvil-serving router path. It summarizes one bounded live run of direct router probes launched from the gateway host; it is not proof of OpenClaw's full agent attempt loop, not a model-card maximum throughput claim, and should not be used as a standalone promotion decision.

Environment

• Capture time: 2026-07-07T01:25:03Z
• Gateway: Companion Node running OpenClaw
• Router: Primary Node anvil-serving front door at http://100.64.0.10:8000/v1
• Runner: examples/openclaw/colo_smoke.py --run-generations --run-interaction-benchmark
• Artifact filename: openclaw-colo-live-interactions-repeatable.json in the operator evidence root
• Artifact SHA-256: 6e108cb68fa9b28600f3854406ffd51302900db267ed0d9a314ba0614768239f
• Router recipe source: examples/primary-node/anvil-router.live.toml
• Router recipe SHA-256: e03c6684b4262ca10753a698494f5e3f930202e5f7956d2bce060055477269bf
• Verdict: pass
• Proof counts: 8 pass, 0 warn, 0 fail

Recipe

The benchmark reads its measurement recipe from router tier params. These values travel with the model/tier recipe, so a future heavy-model swap updates config metadata rather than the runner or skill prompt.

Tier	Model	Exact max tokens	Stream max tokens	Benchmark reasoning effort	Intent overrides
fast-local	qwen36-27b	192	128	none	none
heavy-local	gpt-oss-120b	1024	512	low	planning: 2048 exact, 1024 stream

The router does not forward params upstream. They are repeatable smoke/eval metadata. Runtime defaults for ordinary callers remain in fields such as extra_body_defaults.

Results

Metric	Result
Interaction requests	10
Completed requests	10
HTTP status counts	200: 10
Finish reasons	stop: 10
Benchmark warnings	0
Latency p50	407.6 ms
Latency p95	1041.9 ms
Streaming TTFT p50	411.9 ms
Streaming TTFT p95	1049.4 ms
Exact output tokens	350
Exact tokens/sec p50	85.37
Exact tokens/sec p95	170.82

Intent coverage:

Intent	Requests	Completed	Finish reasons
chat-fast	2	2	stop: 2
quick-edit	2	2	stop: 2
review	2	2	stop: 2
planning	2	2	stop: 2
long-context	2	2	stop: 2

The completion requests use the router URL directly. Route evidence below is from companion /v1/route probes in the same artifact, not from OpenClaw's provider dispatch logs.

Route evidence:

Intent	Observed route
chat-fast	fast-local / qwen36-27b
quick-edit	heavy-local / gpt-oss-120b
review	heavy-local / gpt-oss-120b
planning	heavy-local / gpt-oss-120b
long-context	heavy-local / gpt-oss-120b

Site Citation

Suggested copy:

In a live Companion Node to Primary Node COLO smoke test, direct router probes launched from the OpenClaw gateway host completed 10/10 repeatable intent benchmark requests through anvil-serving with all responses finishing by stop, no truncation warnings, p50 end-to-end latency of 407.6 ms, p95 latency of 1041.9 ms, streaming TTFT p50 of 411.9 ms, and exact-generation throughput p50 of 85.37 tokens/sec across the bounded benchmark prompts. The run covered chat-fast, quick-edit, review, planning, and long-context, with companion route probes showing chat-fast on qwen36-27b and heavier intents on gpt-oss-120b.

Use the caveat with the citation:

These numbers are from a bounded OpenClaw-gateway-host smoke/eval on fixed prompts. They validate gateway-to-router reachability, router routing, recipe wiring, and direct router interaction behavior; they are not a general maximum-throughput benchmark for the models and do not by themselves prove OpenClaw's full provider attempt loop.

Reproduce

python examples/openclaw/colo_smoke.py \
  --live \
  --gateway-host mini-host.example \
  --router-base-url http://100.64.0.10:8000/v1 \
  --run-generations \
  --run-interaction-benchmark \
  --artifact .anvil/evidence/openclaw-colo-live-interactions-repeatable.json \
  --pretty

If a new recipe changes model family, context window, reasoning controls, serving engine, quantization, or throughput, update the tier params in the router config and rerun this command before citing new numbers.