EPS
← All tasks·#1093Completed

workflow-fix: verify_plan check for 7B capture booked on eval/debug intent

kind: infra#wf-fix#wf-fix-fp:f7436ef0e447
track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #825 (emitting agent: critic, Methodology lens, round 1 of the base-separator-control amendment review).

Goal

Add a verify_plan.py check that flags a plan whose design/§9 text indicates hidden-state/activation extraction on a ≥7B model while the launch command books --intent eval or --intent debug, pointing at capture-7b.

Workflow gap

  • Bug observed: Plan v17 booked --intent eval (L4) for a 7B all-layer activation-capture phase — the exact #666/#744 OOM configuration — and sailed through verify_plan PASS (0 FAIL/0 WARN); only manual critic review caught it.
  • Why it is a workflow gap: .claude/rules/plan-compute-sizing.md § Activation-capture HBM sizing is MUST-level but has no mechanical check; c26 covers machine-costing WARN only, not the HBM-intent mismatch.
  • Confidence (emitter): medium

Proposed change (candidate diff sketch — refine in planning)

+ CAPTURE_PAT = re.compile(r"(hidden.state|activation (store|capture|extract)|extract_store)", re.I)
+ INTENT_PAT = re.compile(r"--intent\s+(eval|debug)\b")
+ if CAPTURE_PAT.search(plan_text) and INTENT_PAT.search(plan_text) and re.search(r"7B", plan_text):
+     fail("c<N>: 7B activation-capture phase booked on --intent eval/debug (L4); use capture-7b (>=40 GB HBM) per plan-compute-sizing")

Scope / surfaces

  • Primary target: scripts/verify_plan.py
  • Grep the workflow surface for the pattern before editing (grep -rln 'capture-7b\|Activation-capture HBM' .claude/ CLAUDE.md scripts/) and update every hit; list them in the plan. Consider a canonical N/A escape phrase for plans whose eval-intent booking is legitimate (no capture phase, or a <7B model).

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).
  • Beware false positives: the check must not FAIL a plan that merely MENTIONS extraction while booking eval-intent for a non-capture phase — key the check on the launch-command fence + a capture-phase design section, and provide a canonical escape phrase (see verify_plan.py's existing N/A conventions).

Provenance

  • workflow_fix_target: scripts/verify_plan.py
  • fingerprint: f7436ef0e447

target_file: scripts/verify_plan.py bug_observed: Plan v17 booked --intent eval (L4) for a 7B all-layer activation-capture phase — the exact #666/#744 OOM configuration — and sailed through verify_plan PASS (0 FAIL/0 WARN); only manual critic review caught it. why_workflow_gap: plan-compute-sizing § Activation-capture HBM sizing is MUST-level but has no mechanical check; c26 covers machine-costing WARN only, not the HBM-intent mismatch. proposed_change: Add a check that flags a plan whose design/§9 text indicates hidden-state/activation extraction on a ≥7B model while the launch command books --intent eval or --intent debug, pointing at capture-7b. diff_sketch: |

  • CAPTURE_PAT = re.compile(r"(hidden.state|activation (store|capture|extract)|extract_store)", re.I)
  • INTENT_PAT = re.compile(r"--intent\s+(eval|debug)\b")
  • if CAPTURE_PAT.search(plan_text) and INTENT_PAT.search(plan_text) and re.search(r"7B", plan_text):
  • fail("c<N>: 7B activation-capture phase booked on --intent eval/debug (L4); use capture-7b (>=40 GB HBM) per plan-compute-sizing")
    

confidence: medium related_task: #825

Activity