Surface OTel token usage from agent-stdio.log when proxy usage logs are missing

[Copilot]

Most gh-aw.agent.conclusion spans were missing gen_ai.usage.* attributes because token enrichment depended solely on /tmp/gh-aw/agent_usage.json, which is only written when proxy token logs are present. Engines already emit equivalent usage data in agent-stdio.log result events, but that path was not being consumed for OTel enrichment.

• Add runtime token-usage fallback

  • Extend readAgentRuntimeMetrics() to read result.usage from agent-stdio.log
  • Capture input, output, cache-read, and cache-write token counts from runtime result events
  • Preserve existing agent_usage.json behavior as the primary source when available

• Normalize engine usage shapes

  • Map runtime usage fields onto the existing OTel enrichment shape used by agent_usage.json
  • Support both runtime cache field variants already seen in engine logs:
    • cache_read_input_tokens → cache_read_tokens
    • cache_creation_input_tokens → cache_write_tokens

• Keep span semantics unchanged

  • Continue emitting gen_ai.usage.total_tokens as input + output
  • Continue attaching usage attributes only where the current agent/conclusion span routing expects them
  • Avoid changing downstream-job behavior or double-counting rules

• Cover fallback and precedence

  • Add focused tests for:
    • fallback to agent-stdio.log when agent_usage.json is absent
    • precedence of agent_usage.json when both sources exist

Example of the runtime event now used as fallback:

{
  "type": "result",
  "num_turns": 3,
  "usage": {
    "input_tokens": 48200,
    "output_tokens": 1350,
    "cache_read_input_tokens": 41000,
    "cache_creation_input_tokens": 3100
  }
}

[Copilot]

Pull request overview

Adds a runtime fallback so OTLP gh-aw.agent.conclusion spans still get gen_ai.usage.* token attributes when /tmp/gh-aw/agent_usage.json is missing, by extracting and normalizing usage from agent-stdio.log result events (while preserving agent_usage.json as the preferred source).

Changes:

• Extend readAgentRuntimeMetrics() to parse result.usage from agent-stdio.log and normalize cache token field variants.
• Prefer /tmp/gh-aw/agent_usage.json when present; otherwise fall back to runtime metrics token usage for span enrichment.
• Add tests covering fallback behavior and precedence when both sources exist.

Show a summary per file

File	Description
actions/setup/js/send_otlp_span.cjs	Adds runtime token-usage normalization + fallback path for GenAI OTLP usage enrichment.
actions/setup/js/send_otlp_span.test.cjs	Adds regression tests for fallback to agent-stdio.log and precedence of agent_usage.json.

Copilot's findings

Tip

Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

• Files reviewed: 2/2 changed files
• Comments generated: 0