EPS Dashboard

Overview / Motivation

Auto-filed by the /daily 2026-07-26 problem sweep (route 2). Surfaced by 1 independent miner group(s) over the 2026-07-26 session transcripts.

Goal

Add a code-review check that every *_smoke / *_full variable pair in a dispatcher diff is gated on the same smoke predicate, so a smoke-scoped list can never ship as the production default.

Workflow gap

Bug observed: scripts/issue1689_dispatch.sh::run_phase_capture assigned local conds="$conds_smoke" unconditionally — the $SMOKE gate was applied to the models variable but never to conds — so a 21-condition lattice silently captured 1 condition on a real experiment.
Why it is a workflow gap: no review lens anywhere in the workflow surface checks that a smoke-scoped variable is gated, so eight implementer / code-review rounds and a smoke-run gate all passed over a one-line scope collapse.
Confidence (emitter): high
verified-at-filing: grep -c '_smoke' .claude/agents/code-reviewer.md → 0; grep -c '_smoke' .claude/agents/code-correctness-critic.md → 0; grep -rn '_smoke"' .claude/agents .claude/rules .claude/skills CLAUDE.md | wc -l → 0 hits across the entire workflow surface — the absence-of-guard evidence; semantic sibling probe grep -n 'Step 0.69\|Step 0.55' .claude/agents/code-reviewer.md → L198 (smoke-architecture marker presence) and L943 (phase-idempotency + inter-phase contract, landed ad3549bc2a 2026-07-26T02:49:20-07:00 for this same issue #1689), both read in context and neither covering smoke-variable gating; git log --oneline --since='7 days ago' -- .claude/agents/code-reviewer.md .claude/agents/code-correctness-critic.md → 3 commits, none landing this check (2026-07-26)

Evidence

Session 5c5a89e8, task #1689, bug present since round 1 and discovered roughly 14 hours into the run at 2026-07-26T20:26:40Z. The dispatcher source: "local conds_smoke=\"assistant_chat\"\n local conds=\"$conds_smoke\" # full path drives via a per-cell loop in production\n ... local models=\"$models_full\"\n [ -n \"$SMOKE\" ] && models=\"$models_smoke\"". The $SMOKE guard is present on models and absent on conds; run_phase_fit_cells was hardcoded to one cell the same way.
The run produced 1 percell JSON and then crashed: FileNotFoundError: [Errno 2] No such file or directory: 'analysis_tensors/issue_1689/store/Qwen_Qwen2.5-7B/assistant_naturalistic/L14.pt' — a store path that was never written because the condition was never captured.
Rounds R1 through R8 of implementer and code review, plus a smoke-run gate, all PASSed over the unconditional assignment.
Measured cost: roughly 50 minutes of wall clock directly (the R12 relaunch ran 30 minutes producing 1 cell, then R13 implementer plus code review plus relaunch). The larger cost is a science-integrity one: Phase D's whole result would have been a silent 1-of-21 scope collapse had the orchestrator not manually inspected the store tree.
The pattern is mechanically greppable from the dispatcher diff alone: a X="$X_smoke" assignment with no sibling [ -n "$SMOKE" ] guard on the same variable. The existing Step 0.69 gate already establishes that dispatcher-shape greps are an accepted review-time instrument in this file.

Proposed change

In .claude/agents/code-reviewer.md, add a sub-check to the dispatcher review steps (alongside the Step 0.69 phase-idempotency gate at L943, which already greps dispatcher shapes): for every dispatcher file in the diff, run grep -nE '=\"\$[a-z_]+_smoke\"' <dispatcher> and, for each hit, require a sibling gate on the SAME variable — a [ -n "$SMOKE" ] && <var>= line, a ${SMOKE:+...} expansion, or an equivalent conditional.
Treat an ungated X="$X_smoke" assignment as a blocker, not a CONCERNS: the failure is silent scope collapse on a real run, with no crash at the assignment site and no smoke-gate coverage.
Check the converse direction too — a *_full variable defined but never assigned to the live variable on any path is the same defect seen from the other side.
Give the check a stable blocker tag consistent with the existing Step 0.5 / 0.55 / 0.6 contract-tag vocabulary, so the Step 5c-bis strip classifies it correctly; this is a SUBSTANTIVE finding and must not be strippable.
Record an explicit N/A verdict line when the diff carries no dispatcher, matching the Step 0.69 convention (Step 0.69: N/A — diff carries no multi-phase dispatcher).
Mirror the check into .claude/agents/code-correctness-critic.md, the v2 successor lens, so workflow: v2 tasks are covered identically.
Add the pattern to the smoke-gate material in .claude/rules/gotchas.md so the implementer side sees it before the reviewer does.

Scope / surfaces

Primary target: .claude/agents/code-reviewer.md
.claude/agents/code-correctness-critic.md (v2 mirror)
.claude/rules/gotchas.md (implementer-side smoke-gate entry)

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
uv run python scripts/workflow_lint.py passes (no-flags); ruff clean on touched files.
This session runs under a workflow_fix_target: Provenance line — it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

fingerprint: b1706a8c8135
workflow_fix_target: .claude/agents/code-reviewer.md
fingerprint: PENDING

/daily 2026-07-26 route-2 filing. Miner refs: A-P1.