You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Fix the Smoke Copilot build fleet now — Go toolchain pin is broken and blocking all 3 smoke variants
One-sentence rationale:go.mod/go tool pin go1.26.5 no longer matches the toolchain image's go1.26.6, so make build fails before any smoke test can run.
This report covers 21 failed runs across the last 6 hours (2026-08-20 13:26–18:41 UTC). 8 distinct root-cause clusters were identified; 5 already have tracking issues (updated below with fresh evidence), 1 new P0 gap gets a linked sub-issue, and 2 are low-priority/non-infra and need no new issue.
Act on this first (P0)
Merge a toolchain pin fix immediately.go1.26.6 does not match go tool version go1.26.5 breaks make: *** [Makefile:42: build] Error 1 on all 3 Smoke Copilot variants (apikey, AOAI apikey, AOAI Entra). Tracked in the new sub-issue below — fix is a one-line go.mod/toolchain bump.
Not a bug — linter caught real MD013/MD009/MD060/MD034/MD049 violations in specs/security-architecture-spec.md; fix the doc, not the workflow
Evidence detail per cluster
Cluster 1 — Cloud Hypervisor guest-network down (P0): All 9 runs show Cloud Hypervisor guest connectivity probe failed with exit code 4, preceded in most cases by MCP gateway health check failed (attempt N/120): read ECONNRESET. Confirmed via full job logs on §32395535034, §32394421163, §32375962917.
Cluster 2 — Codex invalid_project 401 (P1):401 Unauthorized ... auth error code: invalid_project returned by the internal API proxy (`(172.30.0.30/redacted) on every Codex CLI call. This is a NEW concrete signature — #54242 previously had no captured error. Daily Cache Strategy Analyzer is a newly-affected workflow not previously listed on that issue.
Cluster 3 — Pi CLI segfault (P1): Auto-Triage Issues run §32402675238 exits 139 (SIGSEGV) inside the Pi CLI engine — matches #54186 exactly.
Cluster 4 — Claude Code CLI rate-limit cap (P1): Design Decision Gate run §32394421229 fails with Maximum LLM invocations exceeded (30 / 30), a non-retryable 429 guard — NOT a silent hang as #53619 currently states. No commit since #53619 opened has touched this workflow's .md/.lock.yml with a targeted fix, so the issue stays open per the tool-denial-limit rule; corrected characterization posted as a comment.
Cluster 5 — Go toolchain mismatch (P0, NEW): All 3 Smoke Copilot variants fail identically: compile: version "go1.26.6" does not match go tool version "go1.26.5" repeated across every package, terminating in make: *** [Makefile:42: build] Error 1. No existing issue covers a build-time toolchain mismatch (#54116 and #53262 cover artifact-upload/naming symptoms, not this). Sub-issue created below.
Cluster 6 — approve_workflow_run unhandled (P1): PR Sous Chef runs §32389691755 and §32381373575 both hard-fail the safe_outputs job with "No handler loaded" for approve_workflow_run — matches #53263 exactly.
audit-diff regression check (Cluster 1, highest severity): Comparing failed run 32404370749 against baseline success 32387174411 shows the failed run never reached o205451.ingest.us.sentry.io or otlp-gateway-prod-eu-west-2.grafana.net (both new/absent) and made only 2 GitHub API calls vs. 34 in the successful baseline — the sandbox never obtained working outbound network access at all, consistent with the guest-connectivity-probe failure occurring before telemetry/API setup.
Existing issue correlation
All 5 matched issues remain open and accurately track active, recurring problems — none are fixed or stale, so none are closed. Each gets a comment with today's confirming run IDs; #53619 additionally gets a root-cause correction (see above).
Sub-issues created
Go toolchain version mismatch breaks Smoke Copilot build fleet (P0) — linked below.
Follow-up — 2026-08-21 19:01 UTC (6h) — no new parent needed, three items updated
One-sentence rationale: all 14 failures in this window map to clusters already covered by open trackers or by a same-day merged fix — no P0 gap, so no new report issue was created.
Linter Miner's Go/Serena failure (§32508657133, 17:32 UTC) needs no action — already fixed.audit-diff against the prior successful run showed the Serena/Go error is a pre-existing, non-fatal red herring present in both runs; the real cause was model_not_supported for the linter-writer sub-agent, fixed by #54579 which merged at 18:55 UTC — after this run failed. No recurrence expected on the next scheduled run.
Cluster table (this window)
Cluster
Severity
Runs
Workflows
Status
Codex CLI invalid_project 401
P1
6
AI Moderator ×5, Daily Cache Strategy Analyzer ×1
Existing #54242 — reopened with confirmed root cause
Serena Go-LS red herring / real cause model_not_supported
P0→fixed
1
Linter Miner
Existing #54571, fixed by #54579 same day — no action
Isolated single-run failures, no shared signature
P2
6
Avenger, Daily Security Observability Report, Daily Fact, Super Linter Report, CI Optimization Coach, Code Scanning Fixer
Not tracked — monitor only
Evidence detail
Codex CLI invalid_project 401:audit on §32515609498 shows firewall clean (74/74 allowed, 0 blocked) and the actual fault: unexpected status 401 Unauthorized ... auth error code: invalid_project from the AWF api-proxy. codex-harness doesn't classify this as an auth failure, so it retries 3x then surfaces bare exit code 1 — explaining why this was previously filed as "no signature captured." Full detail and affected run IDs in #54242.
Linter Miner:audit-diff between failed §32508657133 and last success §32398150992 shows Setup Go/Go toolchain identical and successful in both; the Serena Go-LS error fires identically in both runs (ephemeral container never bundles Go — cosmetic noise). The actual differentiator is failureClass=model_not_supported on the Copilot backend, absent from the success run, and already fixed same-day by #54579.
Isolated cluster: six workflows each failed exactly once with no retrievable detailed logs and no shared error signature between them — consistent with this report's prior isolated/no-action items (KVM module, lint content violations). No new issue opened; will re-cluster if any of these repeats in the next window.
Fix the Smoke Copilot build fleet now — Go toolchain pin is broken and blocking all 3 smoke variants
One-sentence rationale:
go.mod/go toolpingo1.26.5no longer matches the toolchain image'sgo1.26.6, somake buildfails before any smoke test can run.This report covers 21 failed runs across the last 6 hours (2026-08-20 13:26–18:41 UTC). 8 distinct root-cause clusters were identified; 5 already have tracking issues (updated below with fresh evidence), 1 new P0 gap gets a linked sub-issue, and 2 are low-priority/non-infra and need no new issue.
Act on this first (P0)
go1.26.6 does not match go tool version go1.26.5breaksmake: *** [Makefile:42: build] Error 1on all 3 Smoke Copilot variants (apikey, AOAI apikey, AOAI Entra). Tracked in the new sub-issue below — fix is a one-linego.mod/toolchain bump.Cluster table
invalid_project401approve_workflow_rununhandled in safe_outputsspecs/security-architecture-spec.md; fix the doc, not the workflowEvidence detail per cluster
Cluster 1 — Cloud Hypervisor guest-network down (P0): All 9 runs show
Cloud Hypervisor guest connectivity probe failed with exit code 4, preceded in most cases byMCP gateway health check failed (attempt N/120): read ECONNRESET. Confirmed via full job logs on §32395535034, §32394421163, §32375962917.Cluster 2 — Codex
invalid_project401 (P1):401 Unauthorized ... auth error code: invalid_projectreturned by the internal API proxy (`(172.30.0.30/redacted) on every Codex CLI call. This is a NEW concrete signature — #54242 previously had no captured error. Daily Cache Strategy Analyzer is a newly-affected workflow not previously listed on that issue.Cluster 3 — Pi CLI segfault (P1): Auto-Triage Issues run §32402675238 exits 139 (SIGSEGV) inside the Pi CLI engine — matches #54186 exactly.
Cluster 4 — Claude Code CLI rate-limit cap (P1): Design Decision Gate run §32394421229 fails with
Maximum LLM invocations exceeded (30 / 30), a non-retryable 429 guard — NOT a silent hang as #53619 currently states. No commit since #53619 opened has touched this workflow's.md/.lock.ymlwith a targeted fix, so the issue stays open per the tool-denial-limit rule; corrected characterization posted as a comment.Cluster 5 — Go toolchain mismatch (P0, NEW): All 3 Smoke Copilot variants fail identically:
compile: version "go1.26.6" does not match go tool version "go1.26.5"repeated across every package, terminating inmake: *** [Makefile:42: build] Error 1. No existing issue covers a build-time toolchain mismatch (#54116 and #53262 cover artifact-upload/naming symptoms, not this). Sub-issue created below.Cluster 6 —
approve_workflow_rununhandled (P1): PR Sous Chef runs §32389691755 and §32381373575 both hard-fail the safe_outputs job with "No handler loaded" forapprove_workflow_run— matches #53263 exactly.audit-diff regression check (Cluster 1, highest severity): Comparing failed run 32404370749 against baseline success 32387174411 shows the failed run never reached
o205451.ingest.us.sentry.ioorotlp-gateway-prod-eu-west-2.grafana.net(both new/absent) and made only 2 GitHub API calls vs. 34 in the successful baseline — the sandbox never obtained working outbound network access at all, consistent with the guest-connectivity-probe failure occurring before telemetry/API setup.Existing issue correlation
All 5 matched issues remain open and accurately track active, recurring problems — none are fixed or stale, so none are closed. Each gets a comment with today's confirming run IDs; #53619 additionally gets a root-cause correction (see above).
Sub-issues created
References:
Follow-up — 2026-08-21 19:01 UTC (6h) — no new parent needed, three items updated
One-sentence rationale: all 14 failures in this window map to clusters already covered by open trackers or by a same-day merged fix — no P0 gap, so no new report issue was created.
Act on this first
invalid_project401 auth error on the AWF api-proxy was confirmed viaauditand recurred 5 more times today (AI Moderator ×4 new + 1 more, Daily Cache Strategy Analyzer ×1) after being closed with no linked fix. Root cause and fix are now documented there.audit-diffagainst the prior successful run showed the Serena/Go error is a pre-existing, non-fatal red herring present in both runs; the real cause wasmodel_not_supportedfor thelinter-writersub-agent, fixed by #54579 which merged at 18:55 UTC — after this run failed. No recurrence expected on the next scheduled run.Cluster table (this window)
invalid_project401model_not_supportedEvidence detail
Codex CLI
invalid_project401:auditon §32515609498 shows firewall clean (74/74 allowed, 0 blocked) and the actual fault:unexpected status 401 Unauthorized ... auth error code: invalid_projectfrom the AWF api-proxy. codex-harness doesn't classify this as an auth failure, so it retries 3x then surfaces bareexit code 1— explaining why this was previously filed as "no signature captured." Full detail and affected run IDs in #54242.Linter Miner:
audit-diffbetween failed §32508657133 and last success §32398150992 showsSetup Go/Go toolchain identical and successful in both; the Serena Go-LS error fires identically in both runs (ephemeral container never bundles Go — cosmetic noise). The actual differentiator isfailureClass=model_not_supportedon the Copilot backend, absent from the success run, and already fixed same-day by #54579.Isolated cluster: six workflows each failed exactly once with no retrievable detailed logs and no shared error signature between them — consistent with this report's prior isolated/no-action items (KVM module, lint content violations). No new issue opened; will re-cluster if any of these repeats in the next window.
References: