/init-deep

Generate hierarchical AGENTS.md files: root + complexity-scored subdirectories, produced by a size-formula-driven dag map-reduce (quick scanners -> unspecified-high writers) so the main session's context stays flat at any repo size.

Usage

/init-deep                      # Update mode: modify existing + create new where warranted
/init-deep --create-new         # Read existing → remove all → regenerate from scratch
/init-deep --max-depth=2        # Limit directory depth (default: 3)

---

Workflow (High-Level)

1. Size & route (main session) - ONE eval cell measures the repo and computes the node formula.
2. Map (dag) - quick scanner nodes extract per-chunk facts into bounded file reports.
3. Reduce (dag) - unspecified-high writer nodes own disjoint subtrees: score, write AGENTS.md files, emit digests.
4. Root & verify (dag) - one node writes root AGENTS.md from digests only; one node verifies every file.
5. Snapshot & mode (main session) - snapshot contract unchanged.

<critical> **THE ALWAYS-REDUCE RULE.** The main session NEVER reads chunk reports or raw node outputs - only the verify node's verdict and, when a repair needs it, one digest. Context protection is structural (bounded fan-in at every stage), never a runtime "how much context is left" guess.

todo init the five phases; start/done each transition in real time. </critical>

---

Phase 1: Size & Route

Measure and compute in ONE eval cell - code, not mental arithmetic:

# Measure (tracked files minus vendored/generated: node_modules, .git, dist,
# build, out, vendor, target, coverage, lockfiles, minified and binary files)
S        = total source bytes after exclusions
per_dir  = source bytes per directory            # bin-packing input
depth    = max directory depth                   # respect --max-depth (default 3)
existing = every AGENTS.md / CLAUDE.md path      # read the ROOT one now

# Formula
CHUNK   = 400 * 1024          # ~100k tokens of source; a quick worker's usable window
                              # is ~150k over its fallback chain - leave room for its
                              # prompt and report
N_quick = ceil(S / CHUNK)     # bin-pack WHOLE directories into chunks; a directory
                              # larger than one chunk splits at its children
N_high  = ceil(N_quick / 12)  # one reducer absorbs ~12 reports (~60k tokens) and
                              # still has room to spot-check real code

Route:

• **N_quick < 4** -> inline path below; a dag costs more than it saves.
• N_quick > task.dag.max_nodes_per_run (default 64) -> raise the knob in omo config, or run one chained dag per top-level directory (multi-run composition, mass-ulw skill).
• Otherwise -> dag path: emit CHUNKS = [{id, dirs, bytes}], assign each writer a directory SUBTREE (disjoint - no two writers own the same directory), and mkdir -p .omo/init-deep/reports .omo/init-deep/digests.

--create-new: read every existing AGENTS.md FIRST (still-true facts survive as scanner input), then delete all, then regenerate.

Inline path (N_quick < 4)

Small repo - skip the dag. Fire 2-4 parallel explore agents (structure, entry points, conventions, anti-patterns), for example:

task(subagent_type="explore", run_in_background=true, prompt="Project structure: map real layout via ast-grep structural search (sg/ast_grep MCP) + rg --files -> REPORT deviations from standard patterns")

Run the LSP/ast-grep code map yourself (lsp_symbols outlines + workspace inventory, lsp_find_references on top exports, ast-grep import/call shapes; when neither resolves, mark centrality unmeasured). Then score with the matrix below and write every file per the templates yourself. Phase 5 applies unchanged.

---

Phase 2: Map Wave - quick Scanners (dag)

Build and start the run in one eval JS cell with the dag SDK (OMO_DAG_SDK_ROOT); wave doctrine, the node prompt contract, and the failure playbook come from the mass-ulw skill's references/planning.md. One scanner node per chunk, category: "quick", no load_skills - scanners stay lean and their prompt is a rigid numbered extraction template. Quick workers extract; they never judge and never write AGENTS.md:

TASK: Extract knowledge-base facts for chunk <id> (<dirs>) of <repo-root>.
Steps, in order:
1. Inventory each directory in scope: file count, LOC, languages, entry files.

…