[obs-gap:nemoclaw] nemoclaw: onboarding OTel trace artifacts (per-phase spans/status/duration) not ingested

[github-actions]

Automated harness-observability audit found a gap for nemoclaw.

• Harness exposes: The harness onboarding CLI (src/lib/trace.ts) emits OpenTelemetry-shaped trace artifacts when NEMOCLAW_TRACE=1: a TraceArtifact with resource_spans/scope_spans/spans, each span carrying name (e.g. 'nemoclaw.onboard.phase.gateway', 'nemoclaw.onboard.phase.inference'), start/end time, duration_ms, a status.code of OK/ERROR/UNSET (+ message), sanitized attributes (provider, credential_env, curl-probe results from runCurlProbe), child TraceEvents, and a top-level summary with total_duration_ms and slowest_spans. Output path is controlled by NEMOCLAW_TRACE_FILE or NEMOCLAW_TRACE_DIR (default .e2e/traces), confirmed exercised in src/lib/trace.test.ts (onboard trace artifacts, curl-probe sanitization, secret redaction).
• ClawMetry misses: No code path in the OpenClaw/NemoClaw adapter reads NEMOCLAW_TRACE_FILE / the .e2e/traces directory or parses this trace artifact. Per-phase onboarding failures (status=ERROR on the gateway or inference phase, with duration and a redacted-but-informative message) and the curl-probe reachability checks against inference endpoints are invisible to ClawMetry today — the adapter only surfaces sandbox routing config and post-hoc doctor/model-router fingerprint files, never a timeline of what onboarding actually did or where it failed.
• Where (harness): src/lib/trace.ts (TraceArtifact, TRACE_ENABLED_ENV, TRACE_FILE_ENV, TRACE_DIR_ENV, DEFAULT_TRACE_DIR), src/lib/trace.test.ts (onboard span names, redaction, curl probe)
• Severity: high
• Closest capability: new

Filed by scripts/harness/audit.py. Fingerprint: hgap-e62ba1a084 (used to dedupe — keep it in the body).

[vivekchand]

Closing: this was a false positive, and the auditor that filed it is fixed

The capability exists and shipped. clawmetry/adapters/openclaw.py:1705 reads NEMOCLAW_TRACE_FILE, falls back to NEMOCLAW_TRACE_DIR, then the harness default, and puts nemoclawOnboardTraceStatus, nemoclawOnboardTraceSpanCount, nemoclawOnboardTraceErrors and nemoclawOnboardSlowSpans on the detection record. clawmetry/adapters/nemo.py:1040 reads it too. REQ-OBS-RSO-034 specifies the whole capability, delivered by PR #5198.

Why the audit could not see it

scripts/harness/audit.py::_adapter_source read the first 60,000 characters. clawmetry/adapters/openclaw.py is 193,340. The reader sits at line 1690, roughly 18k characters past the cut — so 69% of the adapter was invisible, and both runtimes this OSS audit covers map to that same file, so every run judged a two-thirds-clipped adapter.

The prompt then told the model: "verify against the FULL adapter above (it is provided in full)". For a task that is entirely about reporting absence, that sentence converts "I did not see it" into "it is not there".

Second occurrence. The docstring records the first: aider's conditional COST at line ~527 was cut, the audit flagged "no COST", and the cap was raised to 60k. Raising a number is not a fix for a file that grows.

Fixed in #5764, released in 0.12.846

• the budget is far above any adapter here, and _adapter_source reports whether it trimmed rather than trimming silently;
• when it does trim, the prompt says so instead of claiming completeness;
• _adapter_index always carries every def and every UPPER_CASE string from the whole file, so an absence claim stays checkable even under trimming.

Verified on a clean origin/main checkout at the release commit (4fb077a4db):

openclaw  chars=193104  trimmed=False
   NEMOCLAW_TRACE_FILE in body : True
   NEMOCLAW_TRACE_FILE in index: True
nemoclaw  chars=193104  trimmed=False
   NEMOCLAW_TRACE_FILE in body : True

tests/test_harness_audit_reads_whole_adapter.py auto-discovers from manifest.json, so an adapter that outgrows the budget tomorrow fails there rather than quietly filing fiction; one case pins the NEMOCLAW_TRACE_FILE witness past the old 60k mark so the test keeps proving something. Restoring the cap and the completeness claim reds 5 of 6.

No wheel check applies here: scripts/harness/ is repo tooling and is not packaged (confirmed absent from the 0.12.846 wheel). The artifact that matters is origin/main, which harness-observability-audit.yml checks out on its daily 06:00 UTC run, and that is what was verified above.

The four siblings are real, and were checked rather than assumed

#5746 to #5749 came from the same truncated runs, so I verified each against the full adapter. All four reason about functions inside the old window (_openclaw_doctor_findings char 4,899; _clawrouter_detect 6,053; _sandbox_inference_configs 46,194), and the full-file grep confirms every claim: no share/publish reader, no meeting-library reader, no connected-accounts reader, and only _gateway_migration_warning rather than update candidate state. This issue was the only fiction of the five.

---

Generated by Claude Code