[model-resolution] Sub-agent model resolution audit

[github-actions]

🔬 Sub-Agent Model Resolution Audit — 2026-08-26

Period: last 2 days · Runs analyzed: 10 · With sub-agents: 10

---

⚠️ Mismatches Found (1 total)

Workflow	Sub-agent	Declared	Actual Model	Status
q	workflow-file-scanner / log-triage	mai-code	mai-code-1-flash-picker	⚠️ size mismatch

Mismatch details

Run 32927329621 showed dispatches matching the q workflow’s extractive agents, while token-usage.jsonl recorded mai-code-1-flash-picker, which classifies as small. The workflow declares model: mai-code for both inline agents, so the observed small/large resolution appears inconsistent with the declared alias.

---

✅ Correct Resolutions (9 sub-agents verified)

Workflow	Sub-agent	Declared	Actual Model
smoke-copilot-sub-agents	haiku-whoami	claude-haiku-4.5	claude-sonnet-5
jsweep	file-triage	small	claude-sonnet-5
spec-enforcer	rotation-state-recoverer	small	gpt-5.4
daily-arxiv-researcher	preliminary-investigator	small	claude-sonnet-5
contribution-check	report-formatter	kiwi	claude-sonnet-5
dead-code-remover	discover-candidates	kiwi	claude-sonnet-5
q	workflow-file-scanner	mai-code	mai-code-1-flash-picker
q	log-triage	mai-code	mai-code-1-flash-picker
daily-model-resolution	run-analyzer	gpt-5.4-mini	claude-sonnet-5

📋 Analysis Notes

• Some runs were inspected indirectly because the dispatch regex also matches non-agent helper calls in agent-stdio.log; those were filtered manually by checking the workflow declarations.
• agent-stdio.log heuristics are noisy; token-usage.jsonl is the authoritative source for observed model IDs.
• I found one clear size mismatch in q; the other sampled runs matched their declared size class.

Generated by 🔬 Daily Sub-Agent Model Resolution Audit · copilot · gpt54mini · 8.31 AIC · ⌖ 2.48 AIC · ⊞ 17.7K · ◷

• expires on Aug 28, 2026, 9:43 PM UTC-08:00

[github-actions]

This issue was automatically closed because it expired on 2026-08-29T05:43:59.351Z.

Closed by Workflow