[model-resolution] 🔬 Sub-Agent Model Resolution Audit — 2026-07-10

[github-actions]

Period: last 2 days · Runs analyzed: 18 downloaded locally · With sub-agents: 18 heuristic matches · Verified mismatches: 2

---

⚠️ Mismatches Found (2 total)

Workflow	Sub-agent	Declared	Actual Model	Status
PR description caveman	Chunk-analyzer	small	claude-sonnet-4.6	⚠️ size mismatch
Contribution Check	Contribution-checker	small	claude-sonnet-4.6	⚠️ size mismatch

Mismatch details

• Run 29071597406: agent-stdio.log records Chunk-analyzer(small) and Pr-description-synthesizer(large). The authoritative token-usage.jsonl contains 13 requests for claude-sonnet-4.6 (large). The workflow source .github/workflows/pr-description-caveman.md declares chunk-analyzer as model: small and pr-description-synthesizer as model: large; the small sub-agent therefore resolved to a large model. The large declaration is consistent with the observed size.
• Run 29073176775: agent-stdio.log records nine Contribution-checker(small) dispatches. The authoritative token-usage.jsonl contains 19 requests for claude-sonnet-4.6 (large). .github/workflows/contribution-check.md explicitly requires model: small for every contribution-checker call; these calls therefore resolved to a large model.

---

✅ Correct Resolutions

Workflow	Sub-agent	Declared	Actual Model
PR description caveman	Pr-description-synthesizer	large	claude-sonnet-4.6
Matt Pocock Skills Reviewer	Pr-triage	claude-haiku-4.5	claude-haiku-4.5
Issue Monster	dispatches with large-class requests	large-class	gpt-5.4

The Pr-description-synthesizer(large) and Pr-triage(claude-haiku-4.5) resolutions matched their declared size. Other runs with no classifiable declared sub-agent model were not counted as correct or mismatched.

---

❓ Unknown / Not Counted

• The downloaded artifact sample included incomplete metadata for several runs (aw_info.json and/or events.jsonl absent), and some agent-stdio.log matches were shell/tool text such as container(s), json.load(sys.stdin), or CSS :lang(fa). These heuristic matches were excluded from verified mismatch counts unless a recognizable workflow sub-agent declaration was present.
• Runs with no authoritative sub-agent model association in token-usage.jsonl were classified unknown rather than inferred from the parent model.
• The log download request for 30 runs exceeded the service outer timeout after leaving a usable local sample; the report therefore covers the 18 qualifying downloaded runs available locally.

📋 Analysis Notes

• Size classification used /(mini|haiku|flash|nano)/i; all other observed model IDs were large.
• token-usage.jsonl was treated as authoritative for observed models; agent-stdio.log was used to identify requested aliases and provide dispatch evidence.
• No workflow source declaration was changed by this audit.

Generated by 🔬 Daily Sub-Agent Model Resolution Audit · 92.1 AIC · ⌖ 8.85 AIC · ⊞ 17.8K · ◷

• expires on Jul 12, 2026, 10:51 PM UTC-08:00

[github-actions]

This issue is being closed as outdated. A newer issue has been created: #44882

View newer issue

---

This action was performed automatically by the Daily Sub-Agent Model Resolution Audit workflow.