plan-loader

name	plan-loader
description	Loads minimum hierarchical plan context (.claude/plans/) for current execution focus. Auto-invokes every session when plans exist. Stays under 800 always-loaded tokens regardless of plan tree size.
autoInvoke	true
when_to_use	Every Claude turn when .claude/plans/ exists in project; loads mission + active node + ancestors only
allowed-tools	Read, Glob, Bash
effort	low
user-invocable	false

AI-consumed reference. Optimized for Claude to read during execution. Human-readable explanation: see docs/architecture/HIERARCHICAL_PLANNING.md or docs/getting-started/ depending on topic.

Plan Loader

STATUS — v3.7.0-alpha.1. Provides plan context for hierarchical planning system.

Behavior (in order)

1. Detect: if .claude/plans/ does NOT exist → exit silently (no plan = no overhead)
2. Read .claude/plans/active.json — get current focus pointer
3. Load mission.md — T0 (always loaded if exists)
4. Load active T1 (Initiative) if active.initiative set
5. Load active T2 (Feature) if active.feature set
6. Load active T3 (Story) if active.story set, plus sibling T4 summaries (id + intent + status only — not full body)
7. Load active T4 (Task) if active.task set
8. Stamp loaded plan nodes with trust: plan for memory-trust-policy.md discipline

Token budget

budget[5]{layer,target,hard_cap}:
  always_loaded,800,1000
  active_T1+T2,1500,2500
  active_T3+T4,5000,7000
  sibling_T4_summaries,500,800
  total_with_plan,~7800,~11300

Auto-degradation rules (per spec §9.1)

When always-loaded budget approaches 13,500 tokens:

1. First trim: skip permanent-memory summary lines → saves ~150 tokens
2. Second trim: skip mission.md content (keep only ID) → saves ~50 tokens
3. Third trim: load only active.json (no node bodies) → log warning

What this skill does NOT do

• Does NOT modify plan files (read-only)
• Does NOT trigger replan or status changes
• Does NOT load sibling T3s, T2s, T1s outside active path
• Does NOT load archive/, traces/, history.jsonl, conflicts.jsonl

Detection logic (bash one-liner the model can run)

[ -f .claude/plans/active.json ] && echo "plan-active" || echo "no-plan"

Tie-Ins

• Owns: .claude/plans/active.json (read), .claude/plans/mission.md (read), .claude/plans/<active path>/*.md (read)
• Skill spec: docs/specs/AURA_FROG_V3.7.0_TECH_SPEC.md §9.1
• Rule: rules/core/plan-trust-policy.md — content loaded by this skill is trust: plan
• Hook: hooks/pre-execute-load-plan-context.cjs (Milestone A part 2) — invokes this skill on every PreToolUse
• Companion: skills/plan-validator/SKILL.md — runs validate-plan-tree.sh on demand