plan-review

plan-review — Human-in-the-loop plan/diff review sublane

OPTIONAL HITL surface. Capture an agent plan, let a human annotate it, ship structured feedback back, and compute a Plan-Diff across resubmissions. Clean-room re-imagining of the plan-annotation idea (github-adoption 2026-06-13, human-review-sublane) — no external code copied.

Posture (read this first)

• AUTO-BLOCK (owner decision 2026-06-13, Marcel "auto block with a reason provided"). When plan-review mode is ON, the PreToolUse guard DENIES the next post-ExitPlanMode mutation, with a reason, until the plan is reviewed and approved (a changes-requested verdict keeps the block). Hard gate; requires CLAUDE_SESSION_ID (else degrades to WARN); a long TTL failsafe prevents a forgotten review-mode from wedging the session. Fires ONLY in review mode (default OFF), so autonomous sessions are unaffected.
• Mutual exclusion (load-bearing safety invariant). Plan-review mode and the autonomous plan_dispatch_gate NEVER both fire on the same ExitPlanMode event. The mode toggle is the single source of truth; the PostToolUse arm site + the two PreToolUse gate-eval sites all exclude off isPlanReviewMode().
• Fail-loud toggle. on/off MUST fail loud on null/corrupt/missing session-state (exit nonzero + stderr + read-back verify) — never a silent no-op.

CLI

# Toggle HITL plan-review mode (fail-loud; exits nonzero on bad state)
node _SYSTEM/Scripts/plan-review.mjs on
node _SYSTEM/Scripts/plan-review.mjs off
node _SYSTEM/Scripts/plan-review.mjs status

# Capture a plan revision (auto-increments revision number)
node _SYSTEM/Scripts/plan-review.mjs capture "<plan-id>" "plan text here"
echo "plan text" | node _SYSTEM/Scripts/plan-review.mjs capture "<plan-id>"

# Plan-Diff across two stored revisions (LCS, scale-capped)
node _SYSTEM/Scripts/plan-review.mjs diff "<plan-id>" 0 1

Programmatic: import { isPlanReviewMode, setPlanReviewMode, capturePlan, annotatePlan, computePlanDiff, buildResultLabel } from '_SYSTEM/Scripts/plan-review.mjs'.

When to reach for it

• A human wants to review/annotate an agent plan before any mutation is dispatched.
• You want a recorded Plan-Diff showing what actually changed between plan resubmissions.
• You want to GATE the autonomous dispatch loop with a human checkpoint (blocks the next mutation until you approve) instead of the route-plan auto-dispatch.

Do NOT reach for it for routine autonomous work — leave the mode OFF and the existing plan_dispatch_gate flow runs unchanged.

RESULT_LABEL

Emits a BARE Lane Result Grammar label: 09PR_PLAN_REVIEW_<verb>_<X|P|F>_<PASS_COMMITTED|BLOCKED>. X/P use PASS_COMMITTED; F uses BLOCKED.

Session Notes

2026-06-13

• session: 24m | peak ctx: 54% | compacts: 0
• tools: Bash×129, Write×56, Read×56, Edit×39, StructuredOutput×6, Agent×5, TodoWrite×5, TaskUpdate×2, Workflow×1
• corrections: Base directory for this skill: /Users/marcelspatz/.claude/skills/cross-reference-navigation

Cross-Reference Navigation (XREF)

The GROUND step of the work loop, made reflexive. One question asked a | Base directory for this skill: /Users/marcelspatz/.claude/skills/quantum-hypothesis-simulation

Quantum Hypothesis Simulation

The quantum-probability layer for YURI's claim/pulse machinery. It mode

• errors: none