EPS Dashboard

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a gap the orchestrator hit directly on task #1739 (2026-08-02): ~12 consecutive dispatches passed --backend gcp explicitly, silently bypassing the free fellows (charmander H200) lane that the auto chain tries FIRST.

Goal

Add a "do NOT pass --backend gcp out of habit" caution (the free-lane sibling of the existing runpod caution) and, if the spawned session's planner agrees it is mechanizable, a lint/WARN that flags an explicit non-auto --backend flag with no stated residual-gap reason.

Workflow gap

Bug observed: Every #1739 dispatch this session (gap-fill boxes, 7 partial-item boxes, the Gap 2 OOM relaunch, 3 ladder shards) passed --backend gcp explicitly. That flag bypasses the auto chain, so the FREE fellows lane was never attempted on any of them, and ~40+ GPU-h of GCP credits were spent while charmander sat with free capacity (sinfo at 2026-08-02: general 14 nodes mixed, overflow 16 nodes mixed, dev 2 nodes mixed). The flag was originally added to dodge FLEX_START preemption, but the correct remedy for that (--provisioning-model STANDARD) was ALREADY being passed alongside it — so the --backend gcp pin was pure habit with no residual-gap justification.
Why it is a workflow gap: CLAUDE.md § "Compute backends" states the fellows-first default only for the ABSENT/EMPTY-frontmatter (auto) case, and carries an explicit habit-guard for exactly ONE lane: "Do NOT pass --backend runpod out of habit" (line 279). There is no equivalent guidance for --backend gcp, and no mechanical check anywhere. The runpod guard exists because runpod costs money; the same reasoning applies to gcp (credits) vs fellows (free), but was never written down. Nothing in the surface tells an orchestrator that an explicit --backend gcp silently forfeits the free lane, and nothing warns at dispatch time.
Confidence (emitter): high (directly observed; grep-verified below)
verified-at-filing: grep -n "out of habit" CLAUDE.md -> 1 hit, line 279, runpod-only; grep -noE '[^.]*--backend gcp[^.]*\.' CLAUDE.md -> 0 hits (no gcp caution exists); grep -rn "check-.*backend\|backend.*habit" scripts/workflow_lint.py -> 0 hits (no lint exists); grep -n "fellows FIRST" CLAUDE.md -> lines 274, 278 (fellows-first documented for auto only). (2026-08-02)

Proposed change (candidate diff sketch — refine in planning)

In CLAUDE.md § "Compute backends — multi-lane router", alongside the existing runpod habit-guard:

- Do NOT pass --backend gcp out of habit either. An explicit --backend gcp bypasses the auto
chain and silently forfeits the FREE fellows lane that auto tries first. Prefer the auto chain
(omit --backend) and let it walk fellows -> gcp; --provisioning-model STANDARD is the correct
remedy for FLEX_START preemption on the GCP rung, NOT a gcp pin. Reserve an explicit `--backend
gcp for a named residual gap (e.g. the SLURM venv-extras mismatch, a sentinel//workspace`
requirement the target lane cannot meet) and name that gap in the launch marker note.

Optionally (planner's call, if judged mechanizable and low-noise): a workflow_lint.py check, or a dispatch_issue.py stderr WARN, when an explicit non-auto --backend is passed without a stated reason. A WARN is likely the right severity — explicit pins are legitimate for the named gaps.

Scope / surfaces

Primary target: CLAUDE.md (the § "Compute backends — multi-lane router" bullet list, near line 279)
Possible secondary (planner's call): scripts/workflow_lint.py (a new check) and/or scripts/dispatch_issue.py (a dispatch-time WARN).
Grep the workflow surface for sibling guidance that should stay consistent (grep -rn "out of habit\|auto_fallback\|DEFAULT_AUTO_LANE_ORDER" .claude/ CLAUDE.md scripts/) and update every hit that would otherwise contradict the new caution; list them in the plan.

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
Do NOT change DEFAULT_AUTO_LANE_ORDER or any router precedence: the lane order is already correct; this is a GUIDANCE + (optional) WARN gap, not a routing bug. The durability pin tests/test_router.py::test_default_auto_lane_order_is_gcp_first must stay green and untouched.
Keep the existing runpod habit-guard intact; the new text is its free-lane sibling, not a replacement.
Any new lint/WARN must be WARN-severity (explicit pins are legitimate for the named residual gaps) and must not fail existing dispatches.
scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
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: CLAUDE.md
fingerprint: (computed by the filer at creation)

Origin: orchestrator's own observation on #1739, 2026-08-02. The user asked "can you run it faster by using runpod/afp?" (afp = charmander/fellows), which surfaced that every dispatch this session had pinned --backend gcp and never tried the free lane; the user then asked "isn't that in the workflow already?" — it is, but only as the auto default, with a habit-guard written for runpod alone.

Supersession — task re-scoped 2026-08-08

The filed ask above is superseded and is NOT being implemented as written. #2028 (GCP provisioning disabled) and #2054 (runpod-first auto order) both landed after this task was auto-filed on 2026-08-02, and between them they falsified every clause of it:

The requested caution already exists, stronger, at .claude/rules/compute-backends.md:15 — including a verbatim citation of #1739, this task's own founding incident.
#2028 replaced prose persuasion with a hard mechanical refusal: an explicit backend: gcp raises the typed GcpDisabledError. A habit pin can no longer silently forfeit the free lane — it cannot provision at all.
The durability pin this body names as a constraint (test_default_auto_lane_order_is_gcp_first) was deliberately INVERTED by #2028 to test_default_auto_lane_order_has_no_gcp. The constraint is unimplementable.
The --backend runpod habit-guard this body proposes to mirror was deleted by #2054 and replaced with "Prefer bare auto".
The optional dispatch-time WARN leg already exists as the 3-way override classifier at scripts/dispatch_issue.py:1410-1493.

Re-scoped deliverable. The residual this body's own scope line authorizes ("Grep the workflow surface for sibling guidance that should stay consistent … and update every hit that would otherwise contradict the new caution"): 9 live sites that still direct a gcp pin or assert GCP is the auto default, plus a WARN-level workflow_lint.py regression check. Highest-severity site is scripts/dispatch_issue.py:1425-1445 — a live warning that fires on --backend runpod and tells the operator "the standing default is GCP FIRST (credits before real money)", pushing them off what is now the first auto rung.

Full evidence: the epm:clarify v1 marker. Plan: plans/v6.md (critic-APPROVED at round 4). 0 GPU-h.