Repository navigation
[model-resolution] Model resolution audit findings #66468
Description
🔬 Sub-Agent Model Resolution Audit — 2026-10-07
Period: last 2 days · Runs analyzed: 18 · With sub-agents: 9
⚠️ Mismatches Found (24 total)
| Workflow | Sub-agent | Declared | Actual Model | Status |
|---|---|---|---|---|
| [aw] Failure Investigator (6h) | failure-classifier | small |
claude-sonnet-5 |
|
| [aw] Failure Investigator (6h) | issue-matcher | small |
claude-sonnet-5 |
|
| [aw] Failure Investigator (6h) | cluster-evidence-extractor | small |
claude-sonnet-5 |
Mismatch details
All 8 sampled runs of [aw] Failure Investigator (6h) resolved every declared small sub-agent to claude-sonnet-5 (large):
- run
37574472659 - run
37572644919 - run
37569726943 - run
37565285885 - run
37563931251 - run
37561000989 - run
37559283703 - run
37556039445
Workflow source confirms each sub-agent is declared model: small in .github/workflows/aw-failure-investigator.md (lines 469-506), so this is a runtime resolution mismatch, not a declaration mismatch.
✅ Correct Resolutions (1 sub-agent verified)
| Workflow | Sub-agent | Declared | Actual Model |
|---|---|---|---|
| PR Code Quality Reviewer | grumpy-coder | small |
gpt-5.4-mini-2026-03-17 |
📋 Analysis Notes
- I used
agenticworkflows auditfor authoritative model usage becauseagent-stdio.logheuristics can misclassify shell/tool names as model requests. - The sampled set also included several workflows with only top-level
model:declarations; those were not counted as sub-agent resolution checks. - No other sampled sub-agent workflows showed a size-class mismatch in the audited runs.
Generated by 🔬 Daily Sub-Agent Model Resolution Audit · copilot · gpt54mini · 66.5 AIC · ⌖ 2.48 AIC · ⊞ 14.7K · ◷
- expires on Oct 9, 2026, 9:54 PM UTC-08:00
Activity
github-actions commentedon Oct 8, 2026
This issue is being closed as outdated. A newer issue has been created: #66801
View newer issue
This action was performed automatically by the
Daily Sub-Agent Model Resolution Auditworkflow.