Cut per-run AI-credit (AIC) burn and raise/reset the shared budget — the account blew past its 1000-AIC cap (1006.273/1000) at 2026-07-19T00:11Z, and every agentic run that touches the pool after that point 403s at authentication before doing useful work.
This is a new signature not covered by any open agentic-workflows issue. It is distinct from #46481 (Copilot watchdog / retry-exhaustion), #46326 (CAPI 400 model), #46301 (Codex 404 model), #46172 (EACCES sandbox), and #44298 (CLI hang-on-exit) — those are per-engine defects; this is an account-wide budget wall that hard-fails runs regardless of engine.
Failure cluster
| Cluster |
Sev |
Representative run |
Signature |
| AIC budget exceeded |
P0 |
§29666222310 |
403 Maximum AI credits exceeded (1006.273000 / 1000) |
| Zero-log scheduled batch |
P2 (unconfirmed) |
§29663541268 |
no retrievable logs — cause not established |
Evidence (confirmed cluster)
Workflow: Semantic Function Refactoring (.github/workflows/semantic-function-refactor.lock.yml), Claude engine, model claude-opus-4-8.
Run §29666222310 — started 00:00:54Z, died 00:11:12Z (2026-07-19).
The run fanned out multiple general-purpose subagents ("Find duplicates in pkg/cli", "...pkg/workflow") and hit the cap mid-flight:
{"type":"text","text":"Failed to authenticate. API Error: 403 Maximum AI credits exceeded (1006.273000 / 1000)."} ... "error":"authentication_failed"
{"type":"system","subtype":"task_updated","patch":{"status":"failed","error":"Agent terminated early due to an API error: Failed to authenticate. API Error: 403 Maximum AI credits exceeded (1006.273000 / 1000)."}}
Once the shared AIC pool is over budget, the 403 is returned at auth for any subsequent run, so the impact is account-scoped even though only one run is directly evidenced here. The heavy parallel subagent pattern in this workflow is a primary burn driver.
Existing-issue correlation
No open agentic-workflows tracking issue matches the Maximum AI credits exceeded signature. Checked: #46481, #46326, #46301, #46172, #44298 — all distinct. No existing issue was found to be fixed/stale on fresh evidence, so none are being closed in this pass.
Fix roadmap
P0 — stop the account-wide 403 wall:
- Raise or reset the shared AIC budget (cap is 1000; the account overshot to 1006.273). Confirm the reset window so runs stop 403-ing.
- Add a pre-flight AIC-budget guard: on insufficient budget, emit
missing_data/noop and exit clean instead of hard-failing at 403 mid-run.
P1 — cut per-run burn so the cap isn't hit:
- Cap subagent fan-out/concurrency in Semantic Function Refactoring — the duplicate-finding sweep spawns many
general-purpose subagents per run.
- Stagger the 22:29 scheduled batch (4+ workflows fire within ~2s) so concurrent runs don't race the shared pool.
**P2 — unresolved / needs (redacted)
- 4 runs at 2026-07-18T22:29Z (Auto-Triage Issues, Duplicate Code Detector, Detection Analysis Report, Documentation Unbloat) and 5 further run IDs (29663540300, 29663540049, 29662788042, 29660744841, 29658071589) failed with zero retrievable logs in the prefetch payload; the logs API returns 0 runs for them. Root cause is not established — possibly the same budget wall (they emitted no safe outputs), possibly a distinct early-setup failure. Do not assume; re-pull logs next window before acting.
Investigation notes
audit-diff was intentionally skipped: the root cause is an unambiguous billing/budget 403, for which run-to-run behavioral deltas add no fix-relevant signal.- Log retrieval for the zero-log batch failed (GitHub API intermittency +
logs tool returned total_runs: 0).
- Prefetch:
failed_run_ids=10; only 1 carried an error signature.
Success criteria / verification
- No
403 Maximum AI credits exceeded in agentic runs across a rolling 48h window.
- Semantic Function Refactoring completes end-to-end within budget and emits its intended safe outputs.
- The 22:29 scheduled batch stops failing en masse (or its true cause is identified from fresh logs).
References: §29666222310, §29663541268
Generated by 🔍 [aw] Failure Investigator (6h) · 140.3 AIC · ⌖ 38 AIC · ⊞ 6.2K · ◷
Cut per-run AI-credit (AIC) burn and raise/reset the shared budget — the account blew past its 1000-AIC cap (1006.273/1000) at 2026-07-19T00:11Z, and every agentic run that touches the pool after that point 403s at authentication before doing useful work.
This is a new signature not covered by any open
agentic-workflowsissue. It is distinct from #46481 (Copilot watchdog / retry-exhaustion), #46326 (CAPI 400 model), #46301 (Codex 404 model), #46172 (EACCES sandbox), and #44298 (CLI hang-on-exit) — those are per-engine defects; this is an account-wide budget wall that hard-fails runs regardless of engine.Failure cluster
403 Maximum AI credits exceeded (1006.273000 / 1000)Evidence (confirmed cluster)
Workflow: Semantic Function Refactoring (
.github/workflows/semantic-function-refactor.lock.yml), Claude engine, modelclaude-opus-4-8.Run §29666222310 — started
00:00:54Z, died00:11:12Z(2026-07-19).The run fanned out multiple
general-purposesubagents ("Find duplicates in pkg/cli", "...pkg/workflow") and hit the cap mid-flight:Once the shared AIC pool is over budget, the 403 is returned at auth for any subsequent run, so the impact is account-scoped even though only one run is directly evidenced here. The heavy parallel subagent pattern in this workflow is a primary burn driver.
Existing-issue correlation
No open
agentic-workflowstracking issue matches theMaximum AI credits exceededsignature. Checked: #46481, #46326, #46301, #46172, #44298 — all distinct. No existing issue was found to be fixed/stale on fresh evidence, so none are being closed in this pass.Fix roadmap
P0 — stop the account-wide 403 wall:
missing_data/noopand exit clean instead of hard-failing at 403 mid-run.P1 — cut per-run burn so the cap isn't hit:
general-purposesubagents per run.**P2 — unresolved / needs (redacted)
Investigation notes
audit-diffwas intentionally skipped: the root cause is an unambiguous billing/budget 403, for which run-to-run behavioral deltas add no fix-relevant signal.logstool returnedtotal_runs: 0).failed_run_ids=10; only 1 carried an error signature.Success criteria / verification
403 Maximum AI credits exceededin agentic runs across a rolling 48h window.References: §29666222310, §29663541268