plan-loader

> **AI-consumed reference.** Optimized for Claude to read during execution. > Human-readable explanation: see [docs/architecture/HIERARCHICAL_PLANNING.md](../../../docs/architecture/HIERARCHICAL_PLANNING.md) > or [docs/getting-started/](../../../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 ```toon 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` (...