Bug Report: gh aw does not correctly count sub-agent model usage

Summary

gh aw appears to attribute all token usage to the orchestrator model and does not break out usage for sub-agents invoked with explicit model overrides.

In my workflow Test Sub-Agent Models, three sub-agents were invoked with claude-haiku-4.5, but the audit output and token summary only report claude-sonnet-4.6.

Environment

• gh-aw CLI version: v0.77.5

Reproduction

1. Use a workflow with engine copilot and inline sub-agents configured with:
   • model: claude-haiku-4.5
2. Dispatch three sub-agents (agent-alpha, agent-beta, agent-gamma) with explicit model override claude-haiku-4.5.
3. Run the workflow.
4. Inspect audit output and artifacts:
   • run_summary.json -> token_usage_summary.by_model
   • sandbox/firewall/logs/api-proxy-logs/token-usage.jsonl
   • agent-stdio.log

Expected Behavior

• Either:
  • token/model accounting includes separate entries for sub-agent model usage (claude-haiku-4.5), or
  • audit clearly indicates that sub-agent model overrides were ignored/fell back and why.
• by_model should not imply only one model was used when sub-agents were dispatched with explicit different model overrides.

Actual Behavior

• Audit output reports only orchestrator model usage:
  • by_model contains only claude-sonnet-4.6
  • token-usage.jsonl entries all show model:"claude-sonnet-4.6"
• At the same time, sub-agent dispatch logs show requests with explicit override claude-haiku-4.5.

This creates a counting/attribution gap for sub-agent usage and makes model-level cost/perf analysis unreliable.

Evidence

1) Sub-agent dispatch indicates haiku override

From agent-stdio.log:

● Agent-alpha(claude-haiku-4.5) Get model name
● Agent-beta(claude-haiku-4.5) Get model name
● Agent-gamma(claude-haiku-4.5) Get model name

2) Audit by_model only has sonnet

From run_summary.json:

"token_usage_summary": {
  "by_model": {
    "claude-sonnet-4.6": {
      "provider": "copilot",
      "requests": 15
    }
  }
}

3) Raw token usage also only has sonnet

From sandbox/firewall/logs/api-proxy-logs/token-usage.jsonl:

{"event":"token_usage","provider":"copilot","model":"claude-sonnet-4.6", ...}

Impact

• Incorrect or incomplete model attribution in audit results.
• Misleading cost optimization decisions when sub-agents are intended to use cheaper models.
• Difficult to validate whether sub-agent model overrides are actually honored.

Requested Fix

1. Include sub-agent model attribution in token accounting and by_model aggregation.
2. If fallback/override is ignored, emit explicit machine-readable reason per sub-agent invocation.
3. Add an audit section like:
   • subagent_model_requests (requested/effective)
   • subagent_model_actuals (observed from token logs)
   • mismatch_count and reason codes.

Notes

• It may indicate an underlying runtime override problem, but even then the audit output should surface that mismatch explicitly.
• Potential regression signal: this behavior appears inconsistent with recent gh-aw guidance that sub-agent model: pins are honored in v0.77.4+.

test-subagent-models.md