• MTTR: For hard crashes (OOM, runner termination, segfault) the trace is the first place an on-call engineer looks. Currently those traces show status=ERROR with no exception event — engineers must context-switch to GitHub Actions logs to learn anything.
• Alerting: Exception-based alerts ("alert when exception.type =~ \"gh-aw.*\" rate > X") miss this entire class of hard-failure runs because no exception event is emitted.
• Dashboards: A common pattern is grouping failures by exception.type to see top failure modes. Right now gh-aw.AgentTimedOut and gh-aw.AgentCancelled are visible, but raw failure outcomes silently drop out of the breakdown.
• Span hygiene: An OTel span with status.code = ERROR and no exception event is a known anti-pattern flagged by linters like the OTel Collector's spanstatusprocessor and breaks parity with the OTel exception semantic conventions.

In actions/setup/js/send_otlp_span.cjs (around lines 1387-1399):

const buildSpanEvents = eventTimeMs => {
  const shouldEmitSyntheticException = hasNoReadableAgentOutput && (isAgentTimedOut || isAgentCancelled);
  if (outputErrors.length === 0) {
    if (shouldEmitSyntheticException) {
      const exceptionType = isAgentTimedOut ? "gh-aw.AgentTimedOut" : "gh-aw.AgentCancelled";
      const exceptionMessage = (statusMessage || `agent ${agentConclusion}`).slice(0, MAX_ATTR_VALUE_LENGTH);
      return [{ timeUnixNano: toNanoString(eventTimeMs), name: "exception", attributes: [buildAttr("exception.type", exceptionType), buildAttr("exception.message", exceptionMessage)] }];
    }
    return [];
  }
  // ... per-error events when outputErrors.length > 0 ...
};

The condition (isAgentTimedOut || isAgentCancelled) excludes the most common hard-failure mode: agentConclusion === "failure" with no readable agent_output.json. Those runs return [] (no events) even though statusCode === 2 is set later in the same function.

Note that isAgentNonOK = isAgentFailure || isAgentCancelled is already computed nearby and is used to drive the span status — but it is not used to drive the synthetic-exception emission.

Reuse the existing isAgentNonOK predicate so any non-OK agent outcome (failure, timeout, cancellation) triggers a synthetic exception event when no agent_output.json is available. Add a third exception type (gh-aw.AgentFailureWithoutOutput) so backends can distinguish hard crashes from clean timeouts/cancellations.

// Proposed replacement in actions/setup/js/send_otlp_span.cjs
const buildSpanEvents = eventTimeMs => {
  // Emit a synthetic exception event whenever the agent did not finish OK and
  // there is no agent_output.json to extract concrete errors from. This covers
  // hard crashes (OOM, runner termination, signal kill) where the agent process
  // is killed before it can write structured output — not just clean timeouts
  // and cancellations.
  const shouldEmitSyntheticException = hasNoReadableAgentOutput && isAgentNonOK;
  if (outputErrors.length === 0) {
    if (shouldEmitSyntheticException) {
      let exceptionType;
      if (isAgentTimedOut) {
        exceptionType = "gh-aw.AgentTimedOut";
      } else if (isAgentCancelled) {
        exceptionType = "gh-aw.AgentCancelled";
      } else {
        exceptionType = "gh-aw.AgentFailureWithoutOutput";
      }
      const exceptionMessage = (statusMessage || `agent ${agentConclusion}`).slice(0, MAX_ATTR_VALUE_LENGTH);
      return [{
        timeUnixNano: toNanoString(eventTimeMs),
        name: "exception",
        attributes: [
          buildAttr("exception.type", exceptionType),
          buildAttr("exception.message", exceptionMessage),
        ],
      }];
    }
    return [];
  }
  // ... existing per-error events branch unchanged ...
};

This is a minimal change: one predicate swap + one extra branch in the type selection. No new attributes, no schema changes, no impact on non-failure paths.

After this change:

• In Grafana / Honeycomb / Datadog: Every red conclusion span (and dedicated agent span when emitted) carries an exception event. Backends that auto-detect exception events show "❗ exception" markers in the trace UI for hard-crashed agent runs, just as they do today for timeouts and cancellations. TraceQL / Honeycomb queries like { status = error && event:name = "exception" } cover all failed runs instead of only the timeout/cancellation subset.
• In the local JSONL mirror (/tmp/gh-aw/otel.jsonl): The conclusion span entry now has a non-empty events: [...] array on hard-crash runs, making post-hoc artifact-based debugging viable without a live collector.
• For on-call engineers: A new exception.type = "gh-aw.AgentFailureWithoutOutput" value becomes alertable and groupable, surfacing the runs that previously had status=ERROR with no diagnostic context.
• Dashboards: Stacked-bar "failures by exception type" panels start showing a gh-aw.AgentFailureWithoutOutput slice. Currently those runs are invisible to that breakdown.

• Edit buildSpanEvents in actions/setup/js/send_otlp_span.cjs (around lines 1390-1397) to use isAgentNonOK and the three-way type selection shown above.
• Add a unit test in actions/setup/js/action_conclusion_otlp.test.cjs (or action_otlp.test.cjs) that asserts: when GH_AW_AGENT_CONCLUSION=failure and agent_output.json is missing, the conclusion span carries exactly one exception event with exception.type = "gh-aw.AgentFailureWithoutOutput".
• Extend the existing timed_out / cancelled synthetic-exception tests to also cover the new failure branch (parametrized table-driven test if convenient).
• Run make test-unit (or cd actions/setup/js && npx vitest run) to confirm the new and existing tests pass.
• Run make fmt to ensure formatting.
• Open a PR referencing this issue.

The Grafana Tempo datasource (grafanacloud-traces) attached to this workflow returned zero traces for any {} TraceQL query across the last 30 days, and /api/v2/search/tags returned only the intrinsic scope with no resource.* or span.* tags populated:

GET /api/datasources/proxy/uid/grafanacloud-traces/api/search?q=%7B%7D&limit=10
→ {"metrics":{"completedJobs":3,"totalJobs":3},"traces":[]}

GET /api/datasources/proxy/uid/grafanacloud-traces/api/v2/search/tags
→ {"metrics":{},"scopes":[{"name":"intrinsic","tags":["duration","event:name",...]}]}

The Prometheus and Loki datasources also returned no gh-aw.*, gen_ai.*, or traces_spanmetrics_* series. So the recommendation in this issue is grounded in static analysis of the instrumentation code rather than live telemetry. The gap above is verifiable by running the existing action_otlp.test.cjs harness with GH_AW_AGENT_CONCLUSION=failure and no agent_output.json on disk — the resulting conclusion span has status.code = 2 and events = [], confirming the missing exception event.

• actions/setup/js/send_otlp_span.cjs (primary change site — buildSpanEvents inside sendJobConclusionSpan)
• actions/setup/js/action_conclusion_otlp.cjs (no change; documents the env vars used)
• actions/setup/js/action_otlp.test.cjs (test for the new exception-event emission path)
• actions/setup/js/action_conclusion_otlp.test.cjs (parallel test surface)