Layered Context Loading Protocol

Inspired by OpenViking (ByteDance) L0/L1/L2 architecture. Adapted for MUSE's on-demand Markdown context loading.

Why

Full-loading all .muse/*.md files during /resume wastes tokens when the Agent only needs one role's context. The layered approach loads minimum context first, then deepens on demand.

Three Layers

Layer	Token Budget	Content	When to Load
L0	~100 tokens	One-line HTML comment at top of each .muse/*.md	Always — scan ALL role files
L1	~2K tokens	Full role file content	On demand — only the CURRENT role's file
L2	Unbounded	memory/*.md + code files + docs	On demand — grep search when needed

L0 Format

Every .muse/*.md file MUST have an L0 comment as the first line:

<!-- L0: v2.10.1 | P0=竞品技术吸收, P1/P2全清, QA PASS, S036已接收 -->

L0 Content Rules

1. Max 120 characters (excluding <!-- L0: and -->)
2. Must include: current version + top priority + blocking issues
3. Pipe-separated sections: version | priorities | status
4. Updated every /bye — when the role file is synced

L0 Examples

<!-- L0: v2.10.1 | P0=竞品技术吸收(mem0/OpenViking), P1全清, QA 10/10 PASS -->
<!-- L0: 9/9渠道已发, Show HN暂缓, S040梗图排期中, Stars=2 -->
<!-- L0: 最近QA全PASS(10/10 v2.3), 无待修FAIL, QA清洁状态 -->

Boot Sequence with Layered Loading

/resume [role]
  │
  ├─① Read CLAUDE.md + MEMORIES.md (constitutional layer, always)
  │
  ├─② Scan ALL .muse/*.md L0 lines (grep "<!-- L0:" .muse/*.md)
  │   → Get one-liner status of every role in ~400 tokens total
  │
  ├─③ Deep-read CURRENT role's .muse/*.md (L1, full file)
  │   → Only the file matching /resume [role]
  │
  ├─④ Scan memory/ for unfinished items (L2, on demand)
  │   → grep 🔲 and [ ] in recent memory files
  │
  └─⑤ grep strategy.md for 🟡 directives (L2, on demand)
      → Only if non-strategy role

Decision Tree: When to Upgrade

Agent receives a question/task
  │
  ├─ Can answer from L0? → Answer immediately
  │   (e.g., "What version is MUSE?" → L0 has it)
  │
  ├─ Need role details? → Load L1 (full role file)
  │   (e.g., "What's the P0 task?" → need build.md details)
  │
  └─ Need historical context? → Load L2 (memory/grep)
      (e.g., "What did we decide about X last week?" → grep memory/)

Maintaining L0

Who Updates L0

The /bye workflow updates L0 as part of Step 3.5 (role file sync):

1. After syncing the role file content
2. Rewrite the L0 comment to reflect current state
3. Keep within 120 char limit

L0 Staleness Detection

If the L0 comment's version doesn't match the latest tag, the /resume boot should flag it:

⚠️ L0 stale: build.md says v2.10.1 but latest tag is v2.11.0