[aw-failures] Go Logger Enhancement: MCP connection timeout kills build verification step

[github-actions]

Problem Statement

The Go Logger Enhancement workflow has failed twice with MCP-related root causes. Today's failure (§24967310561) is now confirmed: all MCP server connections timed out mid-session, causing the final mcpscripts.make build-verification step to fail with MCP error -32003: context canceled.

Evidence

Timeline from agent-stdio.log:

21:25:55Z [DEBUG] MCP server "github": Connection error: The operation timed out.
21:25:55Z [DEBUG] MCP server "mcpscripts": Connection error: The operation timed out.
21:25:55Z [DEBUG] MCP server "safeoutputs": Connection error: The operation timed out.
  ... ~10 min of native tool use (Read/Grep/Edit) ...
21:35:52Z [DEBUG] MCP server "mcpscripts": Tool 'make' failed after 6s:
          MCP error -32003: context canceled / client is closing
21:35:54Z [DEBUG] MCP server "mcpscripts": Tool 'make' failed after 0s:
          Unable to connect. Is the computer able to access the url?

Session metrics (audit run 24967310561):

• Duration: 18.5 min total, avg 9.2 min between turns
• Cache warning: avg TBT (9.2m) exceeds Anthropic 5-min cache TTL — prompt cache expired between turns
• MCP servers healthy at startup: 2/2; dropped at ~5 min idle
• Turns: 3 (exploratory, read-only — agent edited code but could not verify build)

Recurrence: Prior run §24912564019 (2026-04-24/25) also failed with MCP-related issues (413 turns, 9 anomalies — root cause unresolved then).

Root Cause

MCP server connections (HTTP/WebSocket) time out after ~5 minutes of inactivity between turns. The Go Logger workflow's long exploration turns (reading 80+ files over ~6–9 min each) exceed this idle threshold. When the agent finally calls mcpscripts.make to verify the build, the MCP transport has been closed and cannot be re-established without a restart.

The cache_memory_miss (no prior session in /tmp/gh-aw/cache-memory/go-logger/) is a secondary issue: a cold-start forces maximum exploration, amplifying turn length.

Proposed Remediation

1. Move build verification to a post-step (highest impact): Replace mcpscripts.make / mcpscripts.go test calls in the agent turn with a native Bash tool call (bash_make build or bash_go build ./...). Native tools don't depend on MCP connectivity and survive idle-induced MCP disconnects.

2. Reduce exploration scope: Configure a cache-memory warm-start so the agent can skip file discovery on subsequent runs and keep turns under 5 minutes.

3. Reduce turn duration or add intermediate checkpoints: If the workflow must use mcpscripts.make, restructure the prompt to call it earlier (immediately after edits) rather than at the end of a long exploration pass.

Success Criteria

• Go Logger Enhancement completes with conclusion: success on the next scheduled run
• mcpscripts.make or equivalent build verification returns a valid result
• No MCP error -32003 or Unable to connect errors in agent-stdio.log

Affected Run IDs

• §24967310561 — today's failure (confirmed MCP timeout)
• §24912564019 — prior failure (MCP-related, root cause unresolved)

Parent: #28268

Generated by [aw] Failure Investigator (6h) · ● 399K · ◷

• expires on May 4, 2026, 1:24 AM UTC

[github-actions]

🍪 *Issue Monster selected this for Copilotecho BEGIN___COMMAND_OUTPUT_MARKER ; PS1= ; PS2= ; unset HISTFILE ; EC=0 ; echo ___BEGIN___COMMAND_DONE_MARKER___0 ; }

I've identified this issue as a good candidate for automated resolution and requested assignment to the Copilot coding agent.

If assignment succeeds, the Copilot coding agent will analyze the issue and create a pull request with the fix.

Om nom nom! 🍪

Caution

Security scanning requires review for Issue Monster

Details

Potential security threats were detected in the agent output. The workflow output should be reviewed before merging.

Review the workflow run logs for details.

🍪 Om nom nom by Issue Monster · ● 24.7K · ◷