Source gh-aw.effective_tokens from agent_usage.json fallback in conclusion spans

[Copilot]

sendJobConclusionSpan only read GH_AW_EFFECTIVE_TOKENS, which is not reliably present in post-step environments, so gh-aw.effective_tokens was silently missing from conclusion spans. This change makes the metric durable by sourcing it from /tmp/gh-aw/agent_usage.json when the env var is unavailable.

• Token sourcing reliability

  • In actions/setup/js/send_otlp_span.cjs, conclusion-span token sourcing now:
    • prefers process.env.GH_AW_EFFECTIVE_TOKENS
    • falls back to /tmp/gh-aw/agent_usage.json → effective_tokens when env is absent
  • Keeps existing emission behavior (gh-aw.effective_tokens only when parsed value is a positive integer).

• Precedence + fallback behavior codified in tests

  • In actions/setup/js/send_otlp_span.test.cjs, added focused coverage for:
    • artifact fallback when env var is absent
    • env var precedence when both env and artifact are present
    • omission case when neither source provides a usable value
  • Ensures the env-precedence path does not read agent_usage.json.

• Implementation sketch

  const rawETFromEnv = process.env.GH_AW_EFFECTIVE_TOKENS || "";
  const usageFromArtifact = rawETFromEnv ? null : readJSONIfExists("/tmp/gh-aw/agent_usage.json");
  const rawET = rawETFromEnv || String((usageFromArtifact && usageFromArtifact.effective_tokens) || "");
  const effectiveTokens = rawET ? parseInt(rawET, 10) : NaN;