EPS Dashboard

Overview / Motivation

Auto-filed by the /daily Step-C parked-candidate sweep (2026-07-08) from a candidate parked on task #748 (recursion-guarded workflow-fix session).

Goal

Add a workflow_lint check or hub-upload-time assert flagging a workload uploading >10000 files into a single repo directory (or a single upload_folder commit staging >10000 sibling files into one dir), mirroring the --check-upload-as-file backstop pattern.

Workflow gap

Bug observed: The HF Hub 10000-files-per-directory commit limit (#658 r2) is documented in gotchas.md but has no mechanical enforcement — unlike its sibling HF-upload traps (--check-upload-as-file; the poller gpu-idle escalation) — so a workload emitting >10k sibling files into one HF directory silently re-hits a non-retriable BadRequestError after the full compute has already run.
Why it is a workflow gap: Documentation alone is the same 'documented but not enforced' pattern that let #664/#658 recur; only a human who reads gotchas.md first is protected.
Confidence (emitter): medium

Proposed change (candidate diff sketch — refine in planning)

Option A (runtime, preferred): in the shared upload helper (orchestrate/hub.py), before upload_folder, count staged files per top-level dir and raise/shard when any dir >10000. Option B (lint): flag upload_folder call sites lacking a documented per-dir sharding strategy. Planner picks.

Scope / surfaces

Primary target: scripts/workflow_lint.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: scripts/workflow_lint.py
origin: parked candidate on task #748 at 2026-06-30T04:18:26Z

Verbatim parked note:

parked — running under EPM_WORKFLOW_FIX_SESSION / workflow_fix_target (.claude/rules/gotchas.md), see .claude/rules/workflow-fix-on-bug.md § Recursion guard. Surfaced by methodology critic on plan v1 (NOT a redundant re-raise of #748; distinct fingerprint).

Candidate source: methodology critic, round 1. target_file: scripts/workflow_lint.py bug_observed: the HF Hub 10000-files-per-directory commit limit (#658 r2) is being documented in gotchas.md but has no mechanical enforcement, unlike its two sibling HF-upload traps (line 101 has --check-upload-as-file, line 102 has the poller gpu-idle escalation); documentation alone is the same "documented but not enforced" pattern that let #664/#658 recur. why_workflow_gap: a future workload emitting >10k sibling files into one HF directory will silently re-hit a non-retriable BadRequestError after the full compute has already run — the doc bullet only helps a human who reads gotchas.md first. proposed_change: add a workflow_lint or hub-upload-time assert that flags a workload uploading >10000 files into a single repo directory (or a single upload_folder commit staging >10000 files into one dir), mirroring the --check-upload-as-file backstop pattern. confidence: medium related_task: #748 (this task documents; that one enforces)

Routing: PARKED under recursion guard. Distinct fingerprint from #748 (target_file: scripts/workflow_lint.py vs .claude/rules/gotchas.md; alternatives critic also weighed and ruled this NOT a blocker on the current doc fix). Recovered after this workflow-fix session completes: an interactive orchestrator (or a /daily pass) re-raising the candidate outside the recursion-guarded session will auto-file + spawn under the normal workflow-fix-on-bug default; alternatives critic noted "premature at N=1 recurrence; lint check is the right deferred follow-up, not a reason to bounce the doc bullet."