EPS Dashboard

Overview / Motivation

Auto-filed by the /daily Step-C parked-candidate sweep from a candidate parked on task #1055 (park_form: recursion-guard).

Goal

Investigate why the #854-hardened _eps_persist_diagnostics left no issue811_partial// prefix on either 2026-07-04 #811 GCE crash, and harden the crash-persist path so a persist no-fire is itself observable off-VM.

Workflow gap

Bug observed: Both 2026-07-04 #811 GCE crashes left NO issue811_partial/att-20260703-172624/ prefix on the HF data repo despite the #854 300s-bounded, transcript-uploaded-last _eps_persist_diagnostics — the promised diagnosability was absent and the #811 diagnosis cycles were manual.
Why it is a workflow gap: The crash-persist trap is the ONLY diagnostics channel surviving the GCE --instance-termination-action=DELETE; a silent persist no-fire reproduces the exact blindness #854 was built to close.
Confidence (emitter): medium (alternatives critic, Phase 2)

Proposed change (candidate diff sketch — refine in planning)

Audit the EXIT-trap wiring for the #811 attempts' crash modes (trap not installed? persist guarded-out? pre-trap poweroff?); add an off-VM persist-attempted breadcrumb (e.g. an early tiny marker upload at trap entry) so absence distinguishes 'persist never ran' from 'persist ran and failed'.

Scope / surfaces

Primary target: src/explore_persona_space/backends/gcp.py
Grep the workflow surface for the pattern before editing (grep -rln '<pattern>' .claude/ CLAUDE.md scripts/) and update every hit; list them in the plan.

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; if workflow.yaml or CLAUDE.md change, they stay consistent with the rule file.
This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).

Provenance

workflow_fix_target: src/explore_persona_space/backends/gcp.py
origin: parked candidate on task #1055 at 2026-07-05T18:48:16Z

parked: EPM_WORKFLOW_FIX_SESSION / workflow_fix_target — see workflow-fix-on-bug § Recursion guard. source: prose-followup (alternatives critic, Phase 2). NOT auto-routed (this session is itself a workflow-fix session); logged for the next orchestrator/human pass.

Candidate (synthesized from critic prose): the #854-hardened GCE crash-persist (_eps_persist_diagnostics) left NO issue811_partial/att-20260703-172624/ prefix on either 2026-07-04 #811 crash despite the 300s bound + transcript-uploaded-last design — the very diagnosability #854 promised was absent, which is half of why the #811 diagnosis cycles were manual. Suggested target: src/explore_persona_space/backends/gcp.py (_eps_persist_diagnostics) — investigate why the persist died/skipped on those attempts (serial evidence unreadable post-DELETE; the #854 fix-engaged signal never fired). Confidence: medium.