Haiku only. This skill and every subagent run on haiku. Always pass model: haiku when spawning. Never escalate.
Blind orchestrator. You (running this skill) never read source content and never read generated docs. You only: list paths, count lines, spawn subagents, run the Step 4.5 diagram validator (reading only its pass/fail + error text), and collect each subagent's one-line completion (path + count). No reading, auditing, editing, or stitching of content yourself.
One file, one owner. Each output file is written by exactly one subagent. Write-exclusive: a tier-2 lens agent may read tier-1 fragment files as read-only input, but writes only its own file. No agent ever writes outside the single file it owns.
Mermaid for all visuals. No ASCII. Every diagram, tree, flow, or relationship graphic is a fenced mermaid block. ASCII art / box-drawing / indented trees as visuals are forbidden. Prose, lists, and data tables are fine. Every block must PARSE — see the Mermaid syntax contract below; the Step 4.5 gate enforces it.

Mermaid syntax contract

Prepend this block verbatim to every subagent prompt (tier-1, DB, tier-2). It is the syntax half of "Mermaid for all visuals"; Step 4.5 mechanically enforces it by rendering each diagram.

MERMAID — every mermaid block you emit MUST parse (it is rendered by mermaid-cli; a parse error fails the build). NO ASCII diagrams. Obey exactly:

mindmap — one root; node text is plain words. Do NOT put ( ) [ ] { } in node text — Mermaid reads them as node-shape delimiters and the line breaks. Use —, :, or , instead (write PDF via xelatex & DOCX, not PDF (xelatex) & DOCX).

erDiagram — each attribute row is type name, then optional key PK/FK/UK, then optional "comment". type and name are BARE tokens (letters, digits, _); NEVER quote them. A quoted string is ONLY ever the trailing comment: int sales_units "units sold". Quoting the name — double_precision "f1_score" — is the #1 parse failure; write double_precision f1_score.

labels — every relationship/edge label is a quoted string: CUSTOMER ||--o{ ORDER : "places" , A -->|"per customer"| B.

flowchart — if node text has spaces or any of ( ) [ ] { } | < > " &, wrap it in quotes inside the shape: N["External → Redshift (ETL)"]. Unicode arrows/symbols (→ & $) are fine inside quoted labels or prose, never as bare tokens in mindmap or er-attribute text.

Topology

flowchart TB
  O["Orchestrator (haiku, blind)"]
  subgraph T1["Tier 1 — one agent per subsystem (reads CODE)"]
    F["owns one fragment each"]
  end
  subgraph T2["Tier 2 — one agent per lens (reads FRAGMENTS)"]
    L["owns one lens doc each"]
  end
  DB["DB-schema agent (reads CODE, conditional)"]
  O --> T1
  O --> T2
  O --> DB
  T1 -. "read-only input" .-> T2

Outputs: codemap/_fragments/<subsystem>.md (tier-1); codemap/00-overview.md, 01-capabilities.md, 02-dataflow.md, 03-codemap.md, 04-domain.md (tier-2); and codemap/05-db-schema.md (ground-truth DB schema — conditional, only if the repo has a database layer).

Procedure

Step 1 — Partition into subsystems (paths only)

Without reading file content, split the source tree into subsystems sized to fit one Haiku agent. Skip caches/vendored dirs (__pycache__, node_modules, .venv, dist, .git). For each top-level source dir, measure LOC; if it exceeds ~8k LOC or ~60 files, descend one level and treat children as separate subsystems (split a flat bag of files by name/domain prefix). Otherwise treat the dir as one subsystem. Produce a list of {name, paths, fragment_file}. Use find/wc only — never cat/Read source.

Step 2 — Tier-1 fan-out (parallel, one agent per subsystem)

Each agent (model: haiku) owns one _fragments/<subsystem>.md, reads its code, writes its fragment. Prepend the Mermaid syntax contract verbatim to the prompt:

You write ONE fragment of a high-altitude architecture doc, later read (not edited) by stitching agents. Write the MENTAL MODEL, not a restatement of code.

SCOPE: read source under {PATHS} (skip caches); skim for responsibility/data/concepts — not every line. OUTPUT: write ONLY to {FRAGMENT_FILE}.

RULES:

Mirror reality: only what the code does. Unsure → omit. Never invent. If sources conflict, say so.

HIGH ALTITUDE: no lists of files/classes/methods. Name a module/endpoint- group/graph/table conceptually, never a method.

Any diagram is a mermaid block — NO ASCII — that satisfies the Mermaid syntax contract prepended above.

Terse, reference-only. ~40–80 lines.

SCHEMA (exact order):

{SUBSYSTEM}

Role — 1–2 sentences.

Capabilities — what it can be asked to do / its jobs (conceptual).

Data — in → transformed → out/persisted; stores, queues, indices, externals.

Structure & boundaries — internal pieces, its layer, depends-on,

depended-on-by, invariants/absences. Include ONE mermaid diagram of internal pieces + external dependencies.

Domain — domain nouns it owns + business rules it enforces.

Reply with ONLY: {FRAGMENT_FILE} + line count.

Step 3 — DB-schema doc (conditional; runs in parallel with Step 4)

ONLY if the repo has a database layer — an ORM schema/model dir (Drizzle, Prisma, TypeORM, SQLAlchemy, ActiveRecord, …), raw SQL DDL, or a migrations dir. No DB → skip this step AND drop 05-db-schema.md from the outputs; never emit an empty or guessed schema. This agent reads CODE (the schema source), not the fragments, so it can launch in the same wave as Step 4 without waiting on tier-1.

Spawn ONE agent (model: haiku) that owns codemap/05-db-schema.md. Unlike the lens docs, this is GROUND TRUTH, not a mental model — it must mirror the real schema. Don't prescribe a method: the agent picks how to get an accurate result. Prepend the Mermaid syntax contract verbatim to the prompt:

You own ONE doc: the database schema exactly as it is. OUTPUT: write ONLY to codemap/05-db-schema.md.

TASK: locate the repo's DB schema source of truth (ORM schema/model files, SQL DDL, or migrations) and produce a faithful schema reference — every table/entity, its columns with types, primary & foreign keys, and the relationships among them. HOW is up to you: if the repo already ships a schema-diagram generator, you MAY run it and embed/adapt its output; otherwise derive the diagram yourself from the schema source.

RULES:

GROUND TRUTH, not mental model — mirror the actual schema. Never invent a table, column, or relation. Unsure about one → omit it and note the gap.

The diagram is a mermaid erDiagram block — NO ASCII — that satisfies the Mermaid syntax contract prepended above. Heed the erDiagram rule especially: attribute names are BARE, never quoted.

Terse prose: short intro + the diagram + a compact notes section for enums, constraints, and any tenancy/RLS or soft-delete rules the schema encodes.

Reply with ONLY: codemap/05-db-schema.md + line count.

Step 4 — Tier-2 fan-out (parallel, one agent per lens)

After all tier-1 agents report done, spawn the lens agents (model: haiku). Each reads all _fragments/*.md (read-only) and writes only its lens. Prepend the Mermaid syntax contract verbatim to each lens prompt.

Shared preamble:

You write ONE whole-system "lens" doc. INPUT: read every codemap/_fragments/*.md (read-only). OUTPUT: write ONLY to {LENS_FILE}.

RULES:

Synthesize ACROSS all fragments into one coherent whole-system view. Don't copy fragments verbatim — abstract and connect.

HIGH ALTITUDE: concepts, flows, rules. Never file/method enumeration.

Every diagram is a mermaid block of the type named below — NO ASCII — and MUST satisfy the Mermaid syntax contract prepended above.

Mirror reality; if fragments conflict, flag it with ⚠️, don't pick silently.

Terse. ~40–80 lines.

Per lens — file · purpose · required mermaid:

00-overview.md — what the system is, major subsystems, how they fit, how a request becomes work. Link the other four at top. → flowchart of subsystems + external stores.
01-capabilities.md — everything the system can be asked to do, grouped by trigger type. → mindmap or flowchart.
02-dataflow.md — what data enters, how it transforms, where it lands, who reads it back. → flowchart LR for the journey; add a sequenceDiagram if there's an async/streaming result path.
03-codemap.md — layers, dependency direction, entry points, architectural invariants (esp. rules expressed as an absence). Answers "where's the thing that does X?". → flowchart TB of layers with dependency arrows.
04-domain.md — core nouns, how they relate, enforced rules, glossary. → erDiagram (or flowchart) of the noun model.

Step 4.5 — Validate diagrams (parse gate)

Goal: the shipped docs contain only diagrams that actually render — valid snippets, not just hopefully-valid ones. This is a mechanical check: running a validator and reading its pass/fail + error text is NOT "reading content", so it doesn't breach the blind-orchestrator rule.

For every written codemap/*.md that contains a mermaid block, render it with mermaid-cli via npx (no global install — the -y form):

out="$TMPDIR/codemap-validate"; mkdir -p "$out"
for f in codemap/*.md; do
  grep -q '```mermaid' "$f" || continue
  err="$out/$(basename "$f").err"
  if npx -y @mermaid-js/mermaid-cli -i "$f" -o "$out/$(basename "$f").svg" 2>"$err"; then
    echo "PASS $f"
  else
    # mmdc prints the parse error to stderr, then a puppeteer JS stack trace —
    # keep only the error block (stop at the first stack frame). Do NOT pass -q,
    # which suppresses the error entirely.
    msg=$(awk '/file:\/\/|https?:\/\/|^[[:space:]]*at /{exit} NF{print}' "$err")
    echo "FAIL $f :: $msg"
  fi
done

(If Chromium refuses to launch, pass -p a puppeteer config of {"args":["--no-sandbox"]}. If npx/node is unavailable, skip this step and note "diagram validation skipped — no npx" in the report; do NOT fail the run.)

For each FAIL <file>: re-spawn a haiku agent that owns ONLY that file. Give it the file path, the extracted parse-error block (the :: payload above), and the Mermaid syntax contract, and instruct it to edit ONLY that file so every block parses, then re-report. Re-run the validator on that file. Bound to 2 fix attempts per file; if it still fails, carry it into the report as an unresolved ⚠️ with the error.

Step 5 — Report

Collect completion lines. Report only files written + line counts, the Step 4.5 validation result (all PASS / unresolved ⚠️ files with their errors / skipped — no npx), and any ⚠️ conflicts agents surfaced (note if 05-db-schema.md was skipped for lack of a DB). Do not summarize content you have not read.

codemap

codemap

Non-negotiable rules