EPS
← All tasks·#1503Completed

daily-fix: windowed-read duty for first-pass guard briefs

kind: infra#wf-fix#wf-fix-fp:5dfa21294202#daily-auto-filed
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 /daily 2026-07-17 (route 2) from a transcript-mined problem (chunk-2 miner): 4 Usage-Policy refusal kills across 2 review sessions on guard/security-artifact targets — #1436's fact-checker (04:41 and 04:45Z, the second on an already-thinned brief) and #1443's Alternatives critic (09:35 and 09:43Z, killed mid-read of a trigger-dense guard file). All recovered via rung-(b2) sonnet pins, at ~35+ min total cost.

Goal

Extend .claude/rules/trigger-dense-review.md (and the Phase-2 brief-composition guidance it anchors) so FIRST-PASS critic/fact-checker briefs targeting guard/security artifacts carry the windowed-read + findings-by-reference discipline — not only review/reconcile roles and revision-round briefs (#1413's scope).

Workflow gap

  • Bug observed: first-pass Phase-2 critic and fact-checker briefs for guard-script targets do not carry the windowed-read discipline; the agents paged trigger-dense guard-file content and were refusal-killed, twice each, before the per-model-pin recovery rung fired.
  • Why it is a workflow gap: the trigger-dense discipline exists but its stated triggers are "reviewing/reconciling a guard/security artifact" and "composing a revision-round/bounce brief" (#1413) — plan-time Phase-2 FIRST-PASS briefs on guard targets fall between the covered cases, so each new guard-file task re-pays the refusal-ladder cost.
  • Confidence: medium
  • verified-at-filing: .claude/rules/LESSONS.md trigger row for trigger-dense-review.md reads "reviewing/reconciling a guard/security artifact or refusal corpus (…), or composing a revision-round/bounce brief from such verdicts (#1413)" — first-pass plan-review briefs absent from the trigger set (presence-of-gap read on the live index, 2026-07-18 UTC). CLAUDE.md refusal rung (e) covers brief VOCABULARY neutralization first-pass (#1073/#1398) but not the windowed-READ discipline for critic briefs. Incident evidence: transcript-mined, 4 kills as described (chunk-2 findings file).

Proposed change (candidate diff sketch — refine in planning)

# .claude/rules/trigger-dense-review.md — widen the fires-when + add a section:
+ First-pass plan-review / fact-check briefs on guard or security artifacts
+ carry the same duties as review rounds: windowed reads (grep -n + ≤60-line
+ windows, never paging the whole guard file), findings by reference, and
+ neutral gate vocabulary in the brief itself (CLAUDE.md rung (e)).

Plus the matching LESSONS.md index-row update (and the --check-lessons-index lint + ratchet bump if needed).

Scope / surfaces

  • Primary target: .claude/rules/trigger-dense-review.md, .claude/rules/LESSONS.md
  • Possibly one sentence in .claude/skills/adversarial-planner/SKILL.md Phase-2 brief composition.

Constraints / invariants

  • Workflow-surface only. workflow_lint.py no-flags run green (LESSONS index + ratchet).
  • This session runs under the recursion guard — it MUST NOT auto-route its own subagents' workflow-fix candidates.

Provenance

  • fingerprint: 5dfa21294202

  • workflow_fix_target: .claude/rules/trigger-dense-review.md

source: /daily 2026-07-17 transcript sweep (chunk-2 miner) — 4 refusal kills on guard-target first-pass review briefs (#1436 fact-checker ×2, #1443 Alternatives critic ×2), all recovered via rung-(b2).

Activity