[obs-gap:nemoclaw] nemoclaw: onboarding OTel trace artifacts (NEMOCLAW_TRACE) not ingested

[github-actions]

Automated harness-observability audit found a gap for nemoclaw.

• Harness exposes: When NEMOCLAW_TRACE is set, the harness writes OpenTelemetry-style spans for each onboarding phase (e.g. nemoclaw.onboard.phase.gateway, nemoclaw.onboard.phase.inference) to NEMOCLAW_TRACE_FILE / NEMOCLAW_TRACE_DIR (default .e2e/traces/trace.json), including span status (OK/ERROR/UNSET), duration_ms, events, sanitized attributes, and a summary.slowest_spans list — a structured record of which onboarding step failed or was slow.
• ClawMetry misses: The adapter has no code path that reads a nemoclaw trace/span file (no reference to NEMOCLAW_TRACE, NEMOCLAW_TRACE_FILE/DIR, resource_spans, or onboard.phase.* anywhere in openclaw.py). A failed or slow onboarding phase is invisible to the dashboard even though the harness already records it on disk.
• Where (harness): src/lib/trace.ts (TRACE_ENABLED_ENV, TRACE_FILE_ENV, TRACE_DIR_ENV, TraceArtifact/TraceSpan shapes, DEFAULT_TRACE_DIR = .e2e/traces)
• Severity: medium
• Closest capability: new

Filed by scripts/harness/audit.py. Fingerprint: hgap-1f53053124 (used to dedupe — keep it in the body).

[vivekchand]

Hi @github-actions! Took a careful look. Here's the plan:

The problem
When NEMOCLAW_TRACE is set, the NemoClaw harness writes OTel-style onboarding spans to disk (status OK/ERROR/UNSET, duration_ms, events, summary.slowest_spans), but the ClawMetry adapter never reads that file — so a failed or slow onboarding phase is completely invisible on the dashboard.

Approach

1. Add _nemoclaw_onboard_trace() -> dict in clawmetry/adapters/openclaw.py after _model_router_live() (~line 1377). The function: checks NEMOCLAW_TRACE env var, resolves the trace file via NEMOCLAW_TRACE_FILE → NEMOCLAW_TRACE_DIR/trace.json → .e2e/traces/trace.json (harness default), parses the JSON, and returns keys: nemoclawOnboardTraceStatus (worst span status), nemoclawOnboardTraceSpanCount, nemoclawOnboardTraceErrors (error span names), nemoclawOnboardSlowSpans (summary.slowest_spans[:5]).
2. Call it in detect() at clawmetry/adapters/openclaw.py:2284 — a standalone block after _workshop_approval_config(), no extra guard needed (the function self-gates on the env var).
3. Add tests/test_obs_gap_nemoclaw_onboard_trace_5193.py with cases: disabled (no env var), enabled + file missing → {}, file with all-OK spans → status OK, file with an ERROR span → status ERROR + error name surfaced, summary.slowest_spans forwarded.

Test plan

• All 5 unit tests pass with python3 -m pytest tests/test_obs_gap_nemoclaw_onboard_trace_5193.py -q
• Syntax check: python3 -c 'import ast; ast.parse(open("clawmetry/adapters/openclaw.py").read())'
• Function never raises — tested by missing-file and disabled cases returning {}

Estimated change
~90 lines across 2 files (clawmetry/adapters/openclaw.py, tests/test_obs_gap_nemoclaw_onboard_trace_5193.py).

Drafting a PR now — will link here when ready. If you'd rather drive this yourself, drop a comment and I'll back off.

---

Generated by Claude Code

[vivekchand]

Draft PR opened: #5198. Please review when you have a moment.

---

Generated by Claude Code