Executive Summary
Investigation of the last 6 hours (2026-05-05 ~19:30–01:17 UTC) found 2 distinct failure clusters across 4 runs out of 29 total. A fifth run (Smoke CI §25351381782) was cancelled due to push supersession and is not actionable.
Critical (P0): APM unpack failures are blocking 3 runs across 2 different PRs, preventing agents from ever launching. This is systemic — present on both firewall v0.25.35 and v0.25.38. Separate P1: Smoke Gemini fails because the Gemini CLI can't reach localhost:8080 (MCP gateway) or the Gemini API via play.googleapis.com, resulting in 94% firewall block rate and exhausted retries.
Failure Clusters
| Priority |
Cluster |
Affected Runs |
Workflows |
Root Cause |
| P0 |
APM Unpack Failure |
3 |
Matt Pocock Skills Reviewer (×2), Smoke Claude |
apm unpack failed for bundle 1 of 1 — exit code 1 on apm-default.tar.gz |
| P1 |
Gemini API Connectivity |
1 |
Smoke Gemini |
TypeError: fetch failed — MCP gateway at localhost:8080 blocked (267/301 requests blocked) |
| N/A |
Smoke CI Cancelled |
1 |
Smoke CI |
Superseded by newer push to main; next run succeeded immediately |
Evidence
Cluster A: APM Unpack Failures (P0) — 3 runs
Affected runs:
| Run |
Workflow |
Branch |
Firewall |
| §25350390860 |
Matt Pocock Skills Reviewer |
copilot/fix-ai-moderator-workflow-allowlist |
v0.25.35 |
| §25351475330 |
Smoke Claude |
copilot/bump-firewall-to-v0-25-38-and-mcpg-to-v0-3-6 |
v0.25.38 |
| §25351482890 |
Matt Pocock Skills Reviewer |
copilot/bump-firewall-to-v0-25-38-and-mcpg-to-v0-3-6 |
v0.25.38 |
Consistent error (all 3 runs):
##[error]APM action failed: apm unpack failed for bundle 1 of 1
(path: /tmp/gh-aw/apm-bundles/apm-default.tar.gz, exit code: 1)
Failure point: agent job → step Restore APM packages (all bundles). APM download succeeds (bundles present: 3–78 KB), but unpack exits non-zero.
Audit-diff evidence (Smoke Claude vs Smoke Codex on same PR):
- Smoke Claude: 0 network requests — agent never activated
- Smoke Codex (success): 16+ API calls, 8 MCP tool invocations,
api.openai.com:443 allowed
- Duration delta: −3m39s confirming early failure, not a slow run
Scope: Two distinct PRs + two distinct firewall versions rule out the firewall bump as the sole cause.
Cluster B: Smoke Gemini — API Connectivity Failure (P1)
Affected run: §25351475292 — Smoke Gemini on copilot/bump-firewall-to-v0-25-38-and-mcpg-to-v0-3-6
Repeated error (4 retries before termination):
Attempt N failed. Retrying with backoff...
Error: exception TypeError: fetch failed sending request
at Models.generateContent → NumericalClassifierStrategy.route
Firewall analysis:
- 301 total requests; 283 blocked (94% block rate)
localhost:8080 — 267 blocked (MCP gateway)
172.30.0.30:10003 — 15 blocked (MCP sidecar/proxy)
play.googleapis.com:443 — 18 allowed (Gemini API attempts that still fail upstream)
Audit-diff vs Smoke Pi (success, same PR):
- 3 new anomaly domains in Gemini run vs Pi:
localhost:8080, 172.30.0.30:10003, (unknown) — all newly blocked
- Pi had
api-proxy:10000 and api-proxy:10002 blocked (expected internal proxy traffic) — absent in Gemini
- Duration: +3m3s longer than Pi (Gemini exhausted 4 retries before giving up)
- GitHub API rate limit usage: +627% vs Pi (retry amplification)
Root cause: The Gemini model router (NumericalClassifierStrategy) cannot complete its API classification call; MCP tools are inaccessible because the MCP gateway is blocked. This is likely a firewall allowlist gap for the Gemini engine setup.
Existing Issue Correlation
⚠️ GitHub API unavailable in this environment (HTTP 403) — existing open issue correlation could not be performed. Sub-issue created for the P0 APM cluster pending deduplication review.
Proposed Fix Roadmap
| Priority |
Fix |
Area |
| P0 |
Investigate APM unpack failure — check bundle integrity, APM action version, and filesystem permissions |
APM / Platform |
| P1 |
Add localhost:8080 and 172.30.0.30:10003 to Gemini allowed domains in firewall config |
Firewall / Gemini engine |
| N/A |
Smoke CI cancellation expected (superseded push) |
— |
Sub-issues Created
References:
Executive Summary
Investigation of the last 6 hours (2026-05-05 ~19:30–01:17 UTC) found 2 distinct failure clusters across 4 runs out of 29 total. A fifth run (Smoke CI §25351381782) was cancelled due to push supersession and is not actionable.
Critical (P0): APM unpack failures are blocking 3 runs across 2 different PRs, preventing agents from ever launching. This is systemic — present on both firewall v0.25.35 and v0.25.38. Separate P1: Smoke Gemini fails because the Gemini CLI can't reach
localhost:8080(MCP gateway) or the Gemini API viaplay.googleapis.com, resulting in 94% firewall block rate and exhausted retries.Failure Clusters
apm unpack failed for bundle 1 of 1— exit code 1 onapm-default.tar.gzTypeError: fetch failed— MCP gateway atlocalhost:8080blocked (267/301 requests blocked)main; next run succeeded immediatelyEvidence
Cluster A: APM Unpack Failures (P0) — 3 runs
Affected runs:
copilot/fix-ai-moderator-workflow-allowlistcopilot/bump-firewall-to-v0-25-38-and-mcpg-to-v0-3-6copilot/bump-firewall-to-v0-25-38-and-mcpg-to-v0-3-6Consistent error (all 3 runs):
Failure point:
agentjob → stepRestore APM packages (all bundles). APM download succeeds (bundles present: 3–78 KB), but unpack exits non-zero.Audit-diff evidence (Smoke Claude vs Smoke Codex on same PR):
api.openai.com:443allowedScope: Two distinct PRs + two distinct firewall versions rule out the firewall bump as the sole cause.
Cluster B: Smoke Gemini — API Connectivity Failure (P1)
Affected run: §25351475292 — Smoke Gemini on
copilot/bump-firewall-to-v0-25-38-and-mcpg-to-v0-3-6Repeated error (4 retries before termination):
Firewall analysis:
localhost:8080— 267 blocked (MCP gateway)172.30.0.30:10003— 15 blocked (MCP sidecar/proxy)play.googleapis.com:443— 18 allowed (Gemini API attempts that still fail upstream)Audit-diff vs Smoke Pi (success, same PR):
localhost:8080,172.30.0.30:10003,(unknown)— all newly blockedapi-proxy:10000andapi-proxy:10002blocked (expected internal proxy traffic) — absent in GeminiRoot cause: The Gemini model router (
NumericalClassifierStrategy) cannot complete its API classification call; MCP tools are inaccessible because the MCP gateway is blocked. This is likely a firewall allowlist gap for the Gemini engine setup.Existing Issue Correlation
Proposed Fix Roadmap
localhost:8080and172.30.0.30:10003to Gemini allowed domains in firewall configSub-issues Created
References: