[deep-report] Cross-engine zero-token 'driver_exit' crash: PR Triage Agent (copilot) + Avenger (codex) failing 100% of sampled runs #65189

Description

Two fleet-health reports from 2026-10-02 independently flagged the Avenger workflow as unhealthy — Detection Analysis Report (#65128) recorded 0/6 successful runs (0 tokens each), and Agent Job Health Monitor (#65133) recorded Avenger failing its "Execute Codex CLI" step on 6/6 observed runs, below the auto-file threshold so left for manual triage. I re-sampled fleet logs today (2026-10-03) via agenticworkflows logs to confirm and extend this:

• Avenger (Codex engine): 3/3 freshly sampled runs (37073003325, 37067359802, 37061250616) are classified failure_kind: driver_exit, token_usage: 0 — the CLI driver process exits before producing any agent output.
• PR Triage Agent (Copilot engine): 3/3 freshly sampled runs (37084143668 — as recent as today — 37047749014, 37007549310) show the identical signature: driver_exit, token_usage: 0.

By contrast, I also sampled the other workflows #65133 grouped into the same "Execute GitHub Copilot CLI step failure" cluster by step name — Code Scanning Fixer (11.5–12.1K tokens/run) and Matt Pocock Skills Reviewer/Delight (12.5–12.7K tokens/run) — and all of those are classified failure_kind: agent_logic with real turns and token usage. That means the step-name-based clustering in #65133 conflated two distinct failure classes: a handful of genuine agent-logic errors, and a narrower, 100%-reproducible, zero-token CLI driver crash affecting at least Avenger (Codex) and PR Triage Agent (Copilot) across two different engines.

The historical #54186 "cross-engine segfault (exit 139)" issue (closed not_planned, 2026-08-27) described a similarly-shaped cross-engine driver crash but a different confirmed root cause (Bun-based bridge segfault) — it was never fixed, just closed stale. This new signature may or may not be the same root cause; that needs to be checked explicitly with raw step logs/exit codes, not assumed.

Expected Impact

Root-causing this restores 100% of currently-wasted scheduled runs for both workflows, and because the failure spans two different agent engines (Codex and Copilot) with an identical 0-token driver-exit shape, fixing the likely shared infra/wrapper component would prevent recurrence across other workflows too, rather than patching one engine at a time.

Suggested Agent

[aw] Failure Investigator (or manual engineering triage) — pull raw step logs/exit codes for run IDs above to confirm the exact driver exit signal before assuming it matches or differs from #54186's Bun segfault hypothesis.

Estimated Effort

Medium (1-4 hours) — mostly log retrieval and root-cause confirmation, not yet a code fix.

Data Source

DeepReport Intelligence analysis, 2026-10-03 cycle. Cross-referenced discussions #65128 and #65133 (2026-10-02), plus fresh agenticworkflows logs samples taken this cycle for Avenger, PR Triage Agent, Code Scanning Fixer, Matt Pocock Skills Reviewer, and Delight.

Generated by 🔬 Deep Report · claude · agent · 294.3 AIC · ⌖ 7.82 AIC · ⊞ 7.3K · ◷

• expires on Oct 4, 2026, 5:12 PM UTC-08:00