Executive Summary
Fix the Gemini model router first — it routes text tasks to the unpriced gemini-3.1-flash-tts-preview model and has broken every Smoke Gemini run in the last 6h (P0, sub-issue #38558). Two content workflows are exhausting the 1000-AIC per-run budget; a Windows integration test is timing out; and the Git Simulator memory-branch push is blocked by a signed-commit rule. 9 recovered/stale auto-trackers were closed; 4 active trackers are linked below.
- Window: 2026-06-11 02:25Z–08:25Z (6h)
- Failed agentic runs analyzed: 12
- Clusters: 5 active + 1 recovered
- Status: ❌ 1 P0 active · ⚠️ 3 P1 active · 🟡 1 P2 active
Why a new parent report: No open Failure Investigator parent existed (all prior [aw-failures] reports are closed) and the P0 Gemini failure had only a thin per-run auto-tracker (#38515). This report consolidates coverage and surfaces 3 previously-untracked failures.
Failure Cluster Table
| # |
Cluster |
Workflows |
Runs |
Failure class |
Root cause |
Priority |
Tracking |
| C1 |
Gemini router → unpriced TTS model |
Smoke Gemini |
4 |
engine/API pricing misconfig |
Router selects gemini-3.1-flash-tts-preview (no AIC pricing) → gemini-cli status 400 unknown_model_ai_credits on first call |
P0 |
#38515 + fix #38558 |
| C2 |
AIC per-run budget exhaustion |
Code Simplifier, Workflow Skill Extractor |
2 |
engine/API 429 |
Per-run AI-credit cap (1000) exceeded (~1014–1025) after 5 retries → harness budget abort |
P1 |
#38499, #38501 |
| C3 |
Windows integration timeout |
Daily Windows CLI Integration |
1 |
build/integration timeout |
gh-aw.exe compile --help exceeded 30000ms in Windows scenario-matrix step (build+verify passed) |
P1 |
untracked |
| C4 |
Memory-branch push blocked |
Daily Safe Outputs Git Simulator |
1 |
permissions / repo policy |
Orphan branch memory/git-simulator push declined — GH013 signed-commit ruleset (3-day streak; agent succeeded) |
P1 |
untracked |
| C5 |
Safe-output validation |
LintMonster, Smoke Copilot |
2 |
safe-output validation |
update_issue "Target is 'triggering' but not running in issue context" (digest mismatch); add_labels/remove_labels "No issue/PR number available" on workflow_dispatch |
P2 |
recovered |
| C6 |
Recovered / stale |
Codex, Pi, Antigravity, Copilot, Copilot-AOAI, jsweep, PR Sous Chef, Conformance Checker |
— |
mixed |
Latest runs green |
— |
9 trackers closed |
Evidence
C1 — Gemini router → unpriced TTS model (P0)
Audited runs §27327422881 and §27331325693 — identical signature, both fail on the first API call (0 tokens, 0 tool calls):
status: 400
[API Error: {"type":"unknown_model_ai_credits","message":"Model \"gemini-3.1-flash-tts-preview\" has no AI credits pricing and no default pricing is configured. Set apiProxy.defaultAiCreditsPricing in the AWF config ... or add the model to the pricing table.","model":"gemini-3.1-flash-tts-preview"}]
Anomaly: the router (NumericalClassifierStrategy.route → ModelRouterService.route) selected a TTS preview model for a text task; configured text models were gemini-2.5-flash-lite / gemini-3.1-pro-preview. All 4 window runs (27327422881, 27331325693, 27329688408, 27320398264) share this signature.
C2 — AIC per-run budget exhaustion (P1)
- Code Simplifier §27324513461:
CAPIError: 429 Maximum AI credits exceeded (1014.39 / 1000), "Max AI credits exceeded (harness budget abort): true". Failing 3 consecutive daily runs.
- Workflow Skill Extractor §27325185979:
CAPIError: 429 Maximum AI credits exceeded (1025.08 / 1000).
No tool/permission/firewall blocks; agent aborts after 5 failed model-request retries.
C3 / C4 / C5 — Untracked + validation failures
- C3 Daily Windows CLI Integration §27332389236:
TIMEOUT: 'D:\a\gh-aw\gh-aw\gh-aw.exe compile --help' exceeded 30000ms. Cross-compile + binary verify passed — this is a CI/integration matrix failure, not an agentic-engine failure.
- C4 Daily Safe Outputs Git Simulator §27327037693:
remote: error: GH013 ... refs/heads/memory/git-simulator ... requires signed commit ... push declined after 4 attempts. Agent succeeded (all 4 sim configs PASS). Failing 3 days straight.
- C5 LintMonster §27322319441: agent succeeded; safe_outputs failed on digest-mismatch + 3×
update_issue "Target is 'triggering' but not running in issue context". Smoke Copilot §27320477001: add_labels/remove_labels "No issue/PR number available" on workflow_dispatch (now recovered).
Existing Issue Correlation
Fix Roadmap
Issues Closed (9)
Recovered/stale auto-trackers closed with evidence: #38535, #38517 (dup), #38512, #38513, #38519, #38542, #38530, #38540, #38505.
Sub-Issues
References: §27327422881 · §27324513461 · §27332389236
Executive Summary
Fix the Gemini model router first — it routes text tasks to the unpriced
gemini-3.1-flash-tts-previewmodel and has broken every Smoke Gemini run in the last 6h (P0, sub-issue #38558). Two content workflows are exhausting the 1000-AIC per-run budget; a Windows integration test is timing out; and the Git Simulator memory-branch push is blocked by a signed-commit rule. 9 recovered/stale auto-trackers were closed; 4 active trackers are linked below.Failure Cluster Table
gemini-3.1-flash-tts-preview(no AIC pricing) → gemini-clistatus 400 unknown_model_ai_creditson first callgh-aw.exe compile --helpexceeded 30000ms in Windows scenario-matrix step (build+verify passed)memory/git-simulatorpush declined — GH013 signed-commit ruleset (3-day streak; agent succeeded)update_issue"Target is 'triggering' but not running in issue context" (digest mismatch);add_labels/remove_labels"No issue/PR number available" on workflow_dispatchEvidence
C1 — Gemini router → unpriced TTS model (P0)
Audited runs §27327422881 and §27331325693 — identical signature, both fail on the first API call (0 tokens, 0 tool calls):
Anomaly: the router (
NumericalClassifierStrategy.route→ModelRouterService.route) selected a TTS preview model for a text task; configured text models weregemini-2.5-flash-lite/gemini-3.1-pro-preview. All 4 window runs (27327422881, 27331325693, 27329688408, 27320398264) share this signature.C2 — AIC per-run budget exhaustion (P1)
CAPIError: 429 Maximum AI credits exceeded (1014.39 / 1000), "Max AI credits exceeded (harness budget abort): true". Failing 3 consecutive daily runs.CAPIError: 429 Maximum AI credits exceeded (1025.08 / 1000).No tool/permission/firewall blocks; agent aborts after 5 failed model-request retries.
C3 / C4 / C5 — Untracked + validation failures
TIMEOUT: 'D:\a\gh-aw\gh-aw\gh-aw.exe compile --help' exceeded 30000ms. Cross-compile + binary verify passed — this is a CI/integration matrix failure, not an agentic-engine failure.remote: error: GH013 ... refs/heads/memory/git-simulator ... requires signed commit ... push declinedafter 4 attempts. Agent succeeded (all 4 sim configs PASS). Failing 3 days straight.update_issue"Target is 'triggering' but not running in issue context". Smoke Copilot §27320477001:add_labels/remove_labels"No issue/PR number available" on workflow_dispatch (now recovered).Existing Issue Correlation
Fix Roadmap
gemini-3.1-flash-tts-previewand add it (orapiProxy.defaultAiCreditsPricing) to the pricing table. Unblocks all Smoke Gemini runs.gh-aw.exe compile --helpon Windows or warm the binary before timing — the binary builds fine.memory/*branch commit or exemptmemory/*refs from the signed-commit ruleset.update_issue/label outputs when no triggering issue/PR context exists (LintMonster, workflow_dispatch smoke runs).Issues Closed (9)
Recovered/stale auto-trackers closed with evidence: #38535, #38517 (dup), #38512, #38513, #38519, #38542, #38530, #38540, #38505.
Sub-Issues
References: §27327422881 · §27324513461 · §27332389236