[aw-failures] [aw] Failure Report 2026-05-10 (6h): Go Logger toolchain mismatch + Schema lock-file drift (recurs #31178, #31221)

[github-actions]

Executive summary

Two workflow runs failed in the 2026-05-10 ~01:35–07:36 UTC (6h) window; both already have auto-generated tracking issues. Each is a recurrence of an open parent failure cluster — no new P0 emerges, but the Go Logger Enhancement run revealed a new, actionable Go toolchain mismatch (go.mod requires Go 1.25.8, runner has 1.24.13 with GOTOOLCHAIN=local) that is worth tracking on top of the existing #31178 max-turns cluster.

• 42 runs, 2 failures (4.8% failure rate).
• 0 P0 new — both clusters are already tracked.
• 1 P1 sub-pattern (Go toolchain mismatch) is new and should be addressed inside [aw-failures] Claude workflows hit max-turns burning every turn on Bash permission denials (Step Name Alignment, Design Decision Gate) #31178 or as a follow-up to that cluster.

Failure clusters

#	Workflow	Run	Engine	Conclusion	Cluster	Existing tracking
1	Schema Consistency Checker	§25621677191	claude	failure (activation, 36s)	Lock-file drift	#31307 (auto), parent #31221
2	Go Logger Enhancement	§25619824735	claude (Opus 4.7)	failure (agent timeout, 17.4m)	Bash permission denials + Go toolchain mismatch	#31292 (auto), parent #31178

Evidence

Cluster 1 — Schema Consistency Checker (lock-file drift)

audit of run 25621677191:

• activation job failed in 13s; agent was skipped.
• Single error from the activation step:

##[error]ERR_CONFIG: Lock file '.github/workflows/schema-consistency-checker.lock.yml' is outdated! The workflow file '.github/workflows/schema-consistency-checker.md' frontmatter has changed. Run 'gh aw compile' to regenerate the lock file.

audit-diff vs the most recent successful baseline (§25593663016, 2026-05-09 06:05 UTC):

• 0 new domains, 0 status changes; the only delta is removal of api.anthropic.com:443 traffic (because the agent never started).
• No firewall, MCP, or quota anomalies — purely a stale-lock failure that prevented the agent from booting.

This is a recurrence of the #31221 "agentic workflows out of sync" pattern. The auto-generated #31307 already provides the exact remediation (gh aw compile).

Cluster 2 — Go Logger Enhancement (agent timeout, mixed root cause)

audit of run 25619824735:

• activation succeeded (15s); agent ran 16.8m and was terminated as failure (no result event emitted before kill).
• 247 turns across 56 tool types with tool_breadth=broad and resource_profile=heavy.
• 0 firewall blocks (143 requests, all allowed).
• mcp_server_health clean (1 server mcpscripts, 6 calls, 0 errors).
• 11 anomalous event patterns flagged by template clustering.

Burned-turn pattern from agent-stdio.log (matches #31178):

This Bash command contains multiple operations. The following parts require approval: ls -la /tmp/gh-aw/cache-memory/go-logger/, cat /tmp/gh-aw/cache-memory/go-logger/processed-files.json, ...
ls in '/tmp/gh-aw/cache-memory' was blocked. For security, Claude Code may only list files in the allowed working directories: '/home/runner/work/gh-aw/gh-aw'.
mkdir in '/tmp/gh-aw/cache-memory/go-logger' was blocked.
Output redirection to '/tmp/gh-aw/agent/all_go.txt' was blocked.
find with '-exec' executes commands or modifies files — cannot be auto-allowed by a Bash(find:*) prefix rule
This Bash command contains multiple operations. The following part requires approval: go version

New finding — Go toolchain mismatch (not present in #31178):

go: go.mod requires go >= 1.25.8 (running go 1.24.13; GOTOOLCHAIN=local)
make: *** [Makefile:23: build] Error 1

The runner has Go 1.24.13 and GOTOOLCHAIN=local is set, suppressing the auto-upgrade path that go.mod's 1.25.8 directive would normally trigger. The agent eventually pivoted to the mcpscripts go MCP tool (which has the correct toolchain), but by then most of its time/turn budget was gone — and mcpscripts then hit a hard 60s subprocess timeout for go test ./pkg/workflow/:

Error [-32603]: calling "tools/call": Command failed: /home/runner/work/_temp/gh-aw/mcp-scripts/go.sh (signal: SIGTERM)
stdout:
go test ./pkg/workflow/ -count=1 -timeout 300s

The agent passed -timeout 300s to go test, but the mcpscripts go wrapper enforces a 60s wall-clock kill, so the test was always going to be SIGTERM'd. The wallclock job limit hit before recovery.

Existing issue correlation

Existing issue	State	Today's evidence	Recommendation
#31178 (Claude max-turns / Bash denials, Step Name + Design Decision Gate)	open	Go Logger Enhancement is a fresh 3rd recurrence with the same /tmp/gh-aw/cache-memory/* denial signature	Keep open; Go Logger Enhancement should be added to its "affected workflows" table
#31221 (workflows out of sync)	open	Schema Consistency Checker hit it again on 2026-05-10 06:19 UTC	Keep open until lock files are regenerated and committed; auto-issue #31307 already linked
#31292 (Go Logger Enhancement failed, auto)	open, expires 2026-05-10 16:47 UTC	Same evidence as Cluster 2 above	Link as sub-issue of #30961; let auto-expiry handle close
#31307 (Schema Consistency Checker failed, auto)	open, expires 2026-05-10 18:20 UTC	Same evidence as Cluster 1 above	Link as sub-issue of #30961; let auto-expiry handle close
#31285 / #31286 (printf allow-list)	closed	Today's failures unrelated; fix landed	No action
#31309 (GitHub Remote MCP Auth, "no safe outputs")	open	The referenced run §25621704254 reports success in the audit; this is a known false-positive class	No action; auto-expires
#31297 (Step Name Alignment, cache_memory_miss)	open	Run §25620382907 reports success; missing-data on cache key	Distinct from #31178; leave open for the cache-key bug
#31287 (Daily Firewall Logs Collector, "no safe outputs")	open	Same pattern as #31309	No action; auto-expires

Proposed fix roadmap

P0

None. Both clusters have existing trackers and neither caused a service-wide outage.

P1 (this 6h window introduces new evidence)

• Add Go toolchain handling to [aw-failures] Claude workflows hit max-turns burning every turn on Bash permission denials (Step Name Alignment, Design Decision Gate) #31178 follow-up. Either (a) bump runner Go to ≥ 1.25.8 in the Go Logger Enhancement workflow runtimes.go block, or (b) drop GOTOOLCHAIN=local so go.mod triggers an auto-download. Without this fix, every Go-touching agent workflow will keep failing on make build until the agent finds the mcpscripts go wrapper. (success criteria: a fresh Go Logger Enhancement run completes make build on the host without falling back to mcpscripts)
• Raise the mcpscripts go.sh subprocess timeout above 60s for go test invocations, or document the cap in the workflow prompt so the agent doesn't pass -timeout 300s it can never use. (success criteria: a go test ./pkg/workflow/ -timeout 300s invocation via mcpscripts completes without SIGTERM)

P2

• Continue the [aw-failures] Claude workflows hit max-turns burning every turn on Bash permission denials (Step Name Alignment, Design Decision Gate) #31178 broader fix (allow-list canonical bash forms for /tmp/gh-aw/cache-memory/* and chained commands) — Go Logger Enhancement is now the 3rd workflow caught by this.

Sub-issues linked

• [aw] Go Logger Enhancement failed #31292 — Go Logger Enhancement failed (auto-issue from this run)
• [aw] Schema Consistency Checker failed #31307 — Schema Consistency Checker failed (auto-issue from this run)

References

• §25621677191 — Schema Consistency Checker (failed, lock drift)
• §25619824735 — Go Logger Enhancement (failed, agent timeout)
• §25593663016 — Schema Consistency Checker (last successful baseline used in audit-diff)
  Related to [aw-failures] [aw] Failure Investigator (6h) - Issue Group #30961

Generated by [aw] Failure Investigator (6h) · ● 15.4M · ◷

• expires on May 17, 2026, 7:45 AM UTC