From fc7d161f9093ca9c4c8be150f87ac5b5d436e335 Mon Sep 17 00:00:00 2001
From: Daniel Bejarano <dbejarano820@gmail.com>
Date: Tue, 2 Jun 2026 13:23:05 -0600
Subject: [PATCH 1/2] DOJ-4835: optional per-module fan-out orchestration for
 workbook-generate
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Size-gated subagent orchestration (discover -> map per module -> assemble) so
per-course workbooks scale without exhausting one context window. Portable: uses
the standard Task/Agent tool, NOT the session-specific Workflow tool.

- agents/workbook-module-composer.md (NEW): composes ONE module into a
  namespaced .wb-module fragment + manifest (no global chrome; m{N}- ID
  namespace; design-token-only; voice-neutral).
- commands/workbook-generate.md: Phase 2.5 execution-strategy gate (small ->
  linear; large -> fan out per module); Phase 4 assembly builds TOC/nav/
  progress/footers from manifests and dedupes namespaced IDs.
- skills/workbook-generate/SKILL.md: documents the map->reduce orchestration,
  the fragment contract, and the consistency safeguards (shared spec + assembly
  normalization pass).
- CLAUDE.md: corrected stale surface counts to match disk (commands 27->30,
  agents 19->20, skills 18->20) — drive-by accuracy fix.

No workbook-base.html / invariant change. Voice-neutral. CI lint green.

Created by Claude Code on behalf of @daniel

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 CLAUDE.md                          |   6 +-
 agents/workbook-module-composer.md | 123 +++++++++++++++++++++++++++++
 commands/workbook-generate.md      |  38 ++++++++-
 skills/workbook-generate/SKILL.md  |  40 ++++++++++
 4 files changed, 202 insertions(+), 5 deletions(-)
 create mode 100644 agents/workbook-module-composer.md

diff --git a/CLAUDE.md b/CLAUDE.md
index ca9bb11..742be0f 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -73,11 +73,11 @@ order until it carries the consumer's voice.
 
 ### Surface counts (current)
 
-- **27 commands** under `commands/` (plus 2 internal `_translation-pipeline.md`
+- **30 commands** under `commands/` (plus 2 internal `_translation-pipeline.md`
   / `_translation-strategy.md` includes — underscore prefix marks them as
   helpers reused by `translate-content`).
-- **19 agents** under `agents/`.
-- **18 skills** under `skills/<skill-name>/SKILL.md`.
+- **20 agents** under `agents/`.
+- **20 skills** under `skills/<skill-name>/SKILL.md`.
 - **1 runtime** under `assets/runtime/overlay-protocol.md` (the executable
   spec of the Base + Overlay loop).
 - **1 CI workflow** under `.github/workflows/lint.yml` with 3 lint scripts
diff --git a/agents/workbook-module-composer.md b/agents/workbook-module-composer.md
new file mode 100644
index 0000000..7f51b91
--- /dev/null
+++ b/agents/workbook-module-composer.md
@@ -0,0 +1,123 @@
+---
+name: workbook-module-composer
+description: >
+  Composes ONE module ("lesson") of an interactive workbook into a self-contained HTML
+  fragment + manifest, for the `workbook-generate` fan-out (DOJ-4835). Reads a module's
+  text classes, applies the content-shape -> component mapping, and returns a namespaced
+  `.wb-module` fragment — NOT a full document. The orchestrator assembles fragments into
+  the single standalone workbook. Voice-neutral; the shared spec is the consistency
+  contract.
+model: sonnet
+tools:
+  - Read
+  - Grep
+  - Glob
+---
+
+# Workbook Module Composer Agent
+
+Compose a single module of a course workbook into a **content fragment** the
+`workbook-generate` orchestrator can concatenate into one self-contained HTML
+file. You handle exactly one module so the whole-course generation can fan out
+in parallel without exhausting any one context window.
+
+Read the design system first:
+`${CLAUDE_PLUGIN_ROOT}/skills/workbook-generate/SKILL.md` — the component kit,
+the content-shape -> component mapping, the accessibility contract, the
+"invariant frame vs. creative payload" split, and "the kit is a floor, not a
+ceiling". This agent **follows that skill**; it does not redefine it.
+
+## Inputs (passed by the orchestrator)
+
+- `module_index` (required): the module's 1-based position in the course (the
+  `N` used for the ID namespace and the `mod-{N}` id).
+- `module_path` (required): absolute path to the module directory.
+- `text_classes` (required): the ordered list of `text-*.md` paths for this
+  module (already ordered by the orchestrator).
+- `module_title` (required): the display title for the module.
+- `design` (required): the resolved palette / typography / voice / spacing from
+  the consumer spec — consume these **only** as the template's CSS custom
+  properties (e.g. `var(--wb-accent)`). Never hardcode hex values.
+- `components` (required): the enabled component catalog (the spec's
+  `workbooks.components`, or IDT's full built-in catalog when absent).
+- `sibling_artifacts` (optional): `quiz-*.md` to fold in as predict-and-reveal
+  checks; `video-*.md` / `slides-*.html` / `challenge-*.md` to reference (never
+  embed).
+
+## Process
+
+1. Read the module's `text-*.md` (frontmatter + H1->H2->H3 / lists / tables /
+   code / prose) and any `quiz-*.md`.
+2. Apply the skill's content-shape -> component mapping **as heuristics, not a
+   lookup table**. Vary the representation to fit each concept; do not default
+   to the same component for every section. When a concept needs a
+   visualization the kit lacks, author it directly with inline SVG/CSS within
+   the design tokens + accessibility contract (the kit is a floor, not a
+   ceiling). Reserve "compose from the tested kit" for fragile stateful JS.
+3. Build the module fragment per the contract below. Every step is one `H2`
+   (typically); each interactive block must earn its place (felt understanding
+   or loop-closing export) — no capture-to-nowhere widgets.
+
+## Output contract (return EXACTLY these two blocks)
+
+Return a `MANIFEST` then a `FRAGMENT`, nothing else. Your final message IS the
+data the orchestrator parses — no preamble.
+
+**1. `MANIFEST` — a fenced `json` block:**
+
+```json
+{
+  "moduleId": "mod-{N}",
+  "title": "{module_title}",
+  "steps": [
+    { "id": "m{N}-step-1", "title": "..." },
+    { "id": "m{N}-step-2", "title": "..." }
+  ],
+  "components_used": ["callout", "flow-diagram", "annotated-code", "predict-reveal"]
+}
+```
+
+**2. `FRAGMENT` — a fenced `html` block** containing a single module chunk:
+
+```html
+<div class="wb-module" id="mod-{N}" role="group" aria-label="{module_title}">
+  <section class="wb-step" id="m{N}-step-1" role="region" aria-label="...">
+    <div class="wb-chapter"><div class="wb-eyebrow">{module_title}</div></div>
+    <h2>...</h2>
+    <!-- kit components / bespoke inline SVG, all theme-variable-driven -->
+  </section>
+  <!-- more steps... -->
+</div>
+```
+
+## Fragment rules (non-negotiable — the assembler depends on them)
+
+- **ID namespacing:** every `id`, `aria-controls`, and `aria-labelledby` you
+  emit is prefixed `m{N}-` (e.g. `m3-acc-1-panel`). This guarantees global
+  uniqueness when fragments are concatenated. The module wrapper id is
+  `mod-{N}`.
+- **Fragment only — no global chrome.** Emit the single `.wb-module` div and its
+  `.wb-step` sections. Do **NOT** emit `<html>`, `<head>`, `<style>`,
+  `<script>`, the `.wb-progress` bar, the `.wb-nav` table of contents, or the
+  `.wb-module__nav` footer. Those are global and the orchestrator owns them
+  (the footer is added by the assembler so the "Module N of M" arc is correct).
+- **No hardcoded brand.** Use the CSS custom properties (`var(--wb-accent)`,
+  `var(--wb-text)`, `var(--wb-font-body)`, etc.) — never literal hex or font
+  names. Brand lives in the assembled document's `:root`.
+- **Accessibility:** each step is `role="region"` with an `aria-label`;
+  interactive controls are focusable; quiz/diff/comparison state is conveyed by
+  text + icon, not color alone; honour `prefers-reduced-motion` patterns from
+  the kit.
+- **In-memory only / standalone:** no `localStorage`, no `postMessage`, no CDN
+  runtime deps. Any image is inlined or relative.
+- **Voice-neutral:** no brand voice, no "Dojo"-isms, no momentum copy. Leave the
+  forward-hook teaser to the overlay (the assembler/overlay owns it).
+
+## Anti-patterns
+
+- Emitting a full HTML document, a `<style>`/`<script>` block, the nav, the
+  progress bar, or the module footer — all assembler-owned.
+- Un-namespaced ids (collide on concatenation).
+- Hardcoded colors/fonts instead of the design tokens.
+- Mapping every section to the same component (monotony) — vary to fit.
+- Capture-to-nowhere widgets; freeform stateful JS.
diff --git a/commands/workbook-generate.md b/commands/workbook-generate.md
index 646223b..584b5cb 100644
--- a/commands/workbook-generate.md
+++ b/commands/workbook-generate.md
@@ -76,9 +76,36 @@ and `workbooks.step_patterns`. Read:
 6. Read each reading's frontmatter (`title`, `order`, `duration_min`, `module`,
    `course`, `tags`) and body (H1 -> H2 -> H3, lists, tables, code, prose).
 
+## Phase 2.5 — Execution strategy (linear vs. per-module fan-out)
+
+Choose how to compose the workbook based on course size:
+
+- **Small course** (≈ 2 modules or fewer, or little total reading) — compose
+  **linearly** in this session (Phases 3–4 inline). Simplest; no drift risk.
+- **Larger course** — **fan out**: dispatch one
+  `${CLAUDE_PLUGIN_ROOT}/agents/workbook-module-composer.md` subagent **per
+  module, in parallel** (via the standard Task/Agent tool). This keeps each
+  agent's context focused on one module, handles big courses without exhausting
+  the window, and tends to vary visualizations across modules. This uses
+  ordinary subagents — **NOT** any session-specific orchestration tool — so the
+  command stays portable across consumer installs.
+
+When fanning out, pass each composer: `module_index` (its 1-based position, used
+as the `m{N}-` ID namespace), `module_path`, the ordered `text_classes` list,
+`module_title`, the resolved `design` tokens + enabled `components`, and any
+`sibling_artifacts`. Each returns a **MANIFEST + a namespaced `.wb-module`
+FRAGMENT** per that agent's output contract — not a full HTML. Collect all
+fragments + manifests for Phase 4 assembly.
+
+Both paths produce the same thing — composed module chunks. Phase 3 is the
+composition ruleset (applied inline when linear, or by each composer when fanned
+out); Phase 4 assembles.
+
 ## Phase 3 — Map content to components
 
-Apply the **content-shape -> component mapping** from the skill. Summary:
+These rules govern composing **each module's chunk** — applied inline (linear)
+or by each `workbook-module-composer` subagent (fan-out). Apply the
+**content-shape -> component mapping** from the skill. Summary:
 
 - Course -> module **chapters** + a **table of contents** + course **progress**.
 - Each reading -> a run of steps in its module chapter; each `H2` -> a **step**.
@@ -119,7 +146,14 @@ payload" and "The kit is a floor, not a ceiling".
    properties (`:root` variables). Under graceful-degrade, keep the neutral
    defaults already in the template.
 3. Replace the content region between the WORKBOOK CONTENT markers with the
-   composed module chapters, steps, and components from Phase 3.
+   composed module chunks, in module order. **If you fanned out (Phase 2.5):**
+   drop in each composer's `FRAGMENT` verbatim, ordered by `module_index`;
+   verify every `id` is `m{N}-`-namespaced and **globally unique** across
+   fragments (fix or re-request any collision); the composers deliberately omit
+   the TOC, progress bar, and module footers — **you own those** and build them
+   here (the omitted module footer is why the "Module N of M" arc stays correct
+   at assembly). Build the `.wb-nav` table of contents from the `MANIFEST`s
+   (module links `#mod-{N}` + step links `#m{N}-step-{k}`).
 4. Honour `workbooks.scroll_behavior` — `doc` (navigable, module-chunked;
    **default**) or `stepped` (one screen at a time) — and `workbooks.animation`
    (enter transition + `prefers-reduced-motion`). In `doc` mode the TOC is a
diff --git a/skills/workbook-generate/SKILL.md b/skills/workbook-generate/SKILL.md
index 7c96a85..edf38d1 100644
--- a/skills/workbook-generate/SKILL.md
+++ b/skills/workbook-generate/SKILL.md
@@ -250,6 +250,46 @@ visualization is encouraged. The only thing to avoid is fragile, untested
 **stateful JS** widgets — those should be added to the kit deliberately (and
 tested), not improvised per page.
 
+## Generation at scale: per-module fan-out (orchestration)
+
+A per-course workbook can span many modules × text classes — more than one
+context window composes well in a single pass. So generation is **size-gated**:
+
+- **Small course** (≈ 2 modules or fewer) → compose **linearly** in one pass.
+- **Larger course** → **map → reduce**: dispatch one `workbook-module-composer`
+  subagent **per module** (in parallel, via the standard Task/Agent tool — never
+  a session-specific orchestration tool; the command must stay portable), then
+  the orchestrator **assembles** the returned fragments into the single
+  self-contained workbook.
+
+This is also a quiet win for variety: independent per-module composition resists
+the "I already used a chart, reuse it" monotony a single pass can fall into.
+
+### The fragment contract (what binds it together)
+
+Each composer returns a **MANIFEST** (`{ moduleId, title, steps[], components_used }`)
+plus a **FRAGMENT**: one `<div class="wb-module" id="mod-{N}">` of namespaced
+`.wb-step`s and kit components — **not** a full document. Rules:
+
+- **ID namespacing:** every `id` / `aria-controls` / `aria-labelledby` is
+  prefixed `m{N}-` (module index) so fragments concatenate without collisions.
+- **Fragment only:** no `<html>/<head>/<style>/<script>`, no `.wb-progress`, no
+  `.wb-nav`, **no `.wb-module__nav` footer**. Those are global and the
+  orchestrator owns them — the footer is added at assembly so the "Module N of M"
+  arc is correct.
+- **No hardcoded brand:** consume design tokens (`var(--wb-accent)`, …), never
+  literal hex/fonts. Voice-neutral; the forward-hook teaser stays an overlay
+  slot.
+
+### Why it stays consistent (no stitched-together feel)
+
+The shared `instruction-bundle-spec.yaml` (brand / a11y / kit) is the binding
+contract handed to every composer, so fragments are visually uniform by
+construction. The **assembler** then runs a normalization pass — verify each
+fragment uses only allowed components, IDs are namespaced + globally unique,
+a11y roles are present — and runs the variety/fit self-check on the assembled
+whole. Drift is bounded by the frame, not left to chance.
+
 ## Consuming the consumer spec (`instruction-bundle-spec.yaml`)
 
 The workbook's brand + structural vocabulary comes from the consumer's

From dc70c33163e37826a42273e97d7475899fda9130 Mon Sep 17 00:00:00 2001
From: Daniel Bejarano <dbejarano820@gmail.com>
Date: Tue, 2 Jun 2026 13:26:25 -0600
Subject: [PATCH 2/2] =?UTF-8?q?DOJ-4835:=20address=20review=20=E2=80=94=20?=
 =?UTF-8?q?namespace=20label=20for=20/=20input=20name=20(P3)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Expand the fragment-contract namespacing rule (agent + skill) so each composer
also prefixes label `for` and form-control `name` (radio/checkbox groups) with
`m{N}-`, not just id/aria-* — prevents broken label associations and
cross-module radio-group bleed when fragments are concatenated.

(The 2 P4 nits about escaped quotes do not apply — the source uses clean quotes
inside backticks; the backslashes appeared only in the diff JSON.)

CI lint green.

Created by Claude Code on behalf of @daniel

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 agents/workbook-module-composer.md | 10 ++++++----
 skills/workbook-generate/SKILL.md  |  6 ++++--
 2 files changed, 10 insertions(+), 6 deletions(-)

diff --git a/agents/workbook-module-composer.md b/agents/workbook-module-composer.md
index 7f51b91..30c370e 100644
--- a/agents/workbook-module-composer.md
+++ b/agents/workbook-module-composer.md
@@ -92,10 +92,12 @@ data the orchestrator parses — no preamble.
 
 ## Fragment rules (non-negotiable — the assembler depends on them)
 
-- **ID namespacing:** every `id`, `aria-controls`, and `aria-labelledby` you
-  emit is prefixed `m{N}-` (e.g. `m3-acc-1-panel`). This guarantees global
-  uniqueness when fragments are concatenated. The module wrapper id is
-  `mod-{N}`.
+- **ID & attribute namespacing:** every `id`, `aria-controls`,
+  `aria-labelledby`, label `for`, and form-control `name` (radio/checkbox
+  groups) you emit is prefixed `m{N}-` (e.g. `m3-acc-1-panel`,
+  `name="m3-quiz-1"`). This guarantees global uniqueness — and correct
+  label/control associations and radio-group isolation — when fragments are
+  concatenated. The module wrapper id is `mod-{N}`.
 - **Fragment only — no global chrome.** Emit the single `.wb-module` div and its
   `.wb-step` sections. Do **NOT** emit `<html>`, `<head>`, `<style>`,
   `<script>`, the `.wb-progress` bar, the `.wb-nav` table of contents, or the
diff --git a/skills/workbook-generate/SKILL.md b/skills/workbook-generate/SKILL.md
index edf38d1..e9cc077 100644
--- a/skills/workbook-generate/SKILL.md
+++ b/skills/workbook-generate/SKILL.md
@@ -271,8 +271,10 @@ Each composer returns a **MANIFEST** (`{ moduleId, title, steps[], components_us
 plus a **FRAGMENT**: one `<div class="wb-module" id="mod-{N}">` of namespaced
 `.wb-step`s and kit components — **not** a full document. Rules:
 
-- **ID namespacing:** every `id` / `aria-controls` / `aria-labelledby` is
-  prefixed `m{N}-` (module index) so fragments concatenate without collisions.
+- **ID & attribute namespacing:** every `id` / `aria-controls` /
+  `aria-labelledby` / label `for` / form-control `name` is prefixed `m{N}-`
+  (module index) so fragments concatenate without id collisions, broken label
+  associations, or cross-module radio-group bleed.
 - **Fragment only:** no `<html>/<head>/<style>/<script>`, no `.wb-progress`, no
   `.wb-nav`, **no `.wb-module__nav` footer**. Those are global and the
   orchestrator owns them — the footer is added at assembly so the "Module N of M"