<!-- [INPUT]: Depends on local project materials and $DECK_DIR room artifacts. [OUTPUT]: Provides the /codeck entry protocol, routing rules, and Decision Ask semantics. [POS]: skills/codeck entry lane; orchestrates fixed lanes and owns room state rules. [PROTOCOL]: 变更时更新此头部，然后检查 CLAUDE.md -->

codeck — Deck Room Entry

Open the deck room, scan materials, diagnose project state, route work to role lanes, and write the role channel as work moves.

Flag anomalies proactively: stale stages, upstream changes not reflected downstream, unresolved threads, and tasks without a handoff.

Deck room model

codeck follows a slock-style collaboration model in one skill runtime.

A skill is a channel: an addressable role with a clear write boundary, durable room files, and a handoff protocol. /codeck is the entry channel. The sub-skills are internal channels that own one part of the room.

Fixed role lanes:

Dynamic role persona:

• Read diagnosis.md for the recommended person behind @outline, @design, and @review.
• The fixed handle owns the work. The dynamic person shapes the judgment.
• Example: @outline may work with Feynman's instinct; @design may use Ravel's formal logic.

Default user-facing output is compact: judgment, artifact, next action. The expanded role channel is written to channel/YYYY-MM-DD.md; show it only when the user asks to see the channel.

Compact response shape:

codeck: {state}

I read this as {task}. I {did/will do} {action}.

Artifact: `{path}`
Next: `{next command or next lane}`

When the user asks to see the channel, show the handoff:

@orchestrator
I read the current request as: {task}. Owner: @outline.

@outline
I will change `deck.md` first and leave visual files alone.

@design
I will rebuild `slides.html` and `custom.css` after the content changes.

@review
I will check whether the change lands for the target audience.

Room document contract

Room documents are not equal. The room has one current truth layer, one work-state layer, and one audit layer.

Read order for every lane:

1. MEMORY.md Active Context, Role Registry, Open Threads, Task Index, Artifacts, Assumed Defaults, Decision Log.
2. tasks/tasks.md open or active rows only.
3. threads/threads.md open rows and their detail blocks only; closed rows are audit evidence.
4. diagnosis.md for material interpretation and dynamic persona, when present.
5. roles/{lane}.md for current lane persona and lane-local rules.
6. Owned source: deck.md; DESIGN.md before custom.css / slides.html; assembled HTML only after build-html.sh.
7. channel/YYYY-MM-DD.md only for audit/debug. Channel text must not override current truth.

Legacy artifacts:

• outline.md is audit-only if found from an older room. Do not read, write, or recreate it during normal generation.
• design-notes.md is a legacy design scratchpad from main. Do not recreate it; migrate useful facts into DESIGN.md, roles/design.md, or latest valid review.md.
• review.md is the latest QA report only when it matches the current assembled HTML. If HTML changed after review, treat review as stale.
• Old project-root *-rN.html or *-deck.css files are historical unless they were produced by the current build-html.sh path.

Memory protocol

MEMORY.md is the deck room index and current-state map. It replaces PROJECT.md.

Keep it short. Store current decisions, state, and indexes. Do not paste the full deck, channel transcript, or stale preview history.

MEMORY.md must include:

# Memory

## Active Context
- Current request:
- Active lane:
- Next:
- Blockers:

## Room Truth Contract
- Current truth: MEMORY.md, deck.md, DESIGN.md, custom.css, slides.html, latest assembled HTML, speech.md when present.
- Work state: diagnosis.md current material reading, tasks/tasks.md active tickets, threads/threads.md open decisions and decision ledger, roles/*.md lane memory, latest valid review.md.
- Audit only: channel/YYYY-MM-DD.md, legacy PROJECT.md, legacy outline.md, legacy design-notes.md, superseded reviews, old previews, project-root sibling CSS.
- Rule: audit text never overrides current truth; legacy outline.md is never a generation source.

## Role Registry
| Handle | Dynamic persona | Owns | Writes |
|--------|-----------------|------|--------|
| @orchestrator | codeck room lead | room state, routing, handoffs | MEMORY.md, tasks, threads, channel, roles |
| @outline | {from diagnosis.md or fallback} | narrative and deck content | deck.md |
| @design | {from diagnosis.md or fallback} | design skeleton, visual system, HTML source | DESIGN.md, custom.css, slides.html |
| @review | {from diagnosis.md or fallback} | audience resistance and QA | review.md, scoped source fixes |
| @speech | {coach if needed} | speech script and presenter notes | speech.md, HTML data-notes |
| @export | publisher | PDF/PPTX output and QA | PDF/PPTX, export notes |

## Latest Channel Summary
{short summary of the latest role exchange}

## Open Threads
| ID | Kind | Owner | Status | Blocking | Writes To | Decision |
|----|------|-------|--------|----------|-----------|----------|

## Task Index
| Task | Owner | Status | Artifact | Handoff |
|------|-------|--------|----------|---------|

## Artifacts
- Diagnosis:
- Content source:
- Design source:
- Current preview:
- Latest review:
- Speech:
- Final exports:

## Assumed Defaults
- {field}: {value} — {reason}

## Decision Log
| Time | ID | Moment | Resolution | Source |
|------|----|--------|------------|--------|

Workspace directories:

$DECK_DIR/MEMORY.md
$DECK_DIR/channel/YYYY-MM-DD.md
$DECK_DIR/tasks/tasks.md
$DECK_DIR/threads/threads.md
$DECK_DIR/roles/orchestrator.md
$DECK_DIR/roles/outline.md
$DECK_DIR/roles/design.md
$DECK_DIR/roles/review.md
$DECK_DIR/roles/speech.md
$DECK_DIR/roles/export.md

channel/YYYY-MM-DD.md stores expanded role turns. Append one block per handoff:

Channel is an audit trail. Do not read it during normal generation, and do not let old channel facts override MEMORY.md, deck.md, or DESIGN.md.

## {ISO time} — {request}

@orchestrator
{routing read}

@{owner}
{claim and action}

@{next-owner}
{handoff if any}

tasks/tasks.md stores live claim tickets. Do not use hard locks. Each ticket has owner, status, artifact, and handoff. Closed history belongs in MEMORY.md summaries or channel, not as a growing task backlog.

threads/threads.md stores unresolved decisions, content conflicts, design disputes, cross-owner proposals, and the Decision Ask ledger. A user-facing question is never free-floating; it must first exist here as a decision record. Once resolved, copy the outcome into the owned source and MEMORY.md Decision Log, then mark the thread answered/defaulted/superseded. Normal reads treat only open rows as work state.

For Decision Ask entries, keep an index row and a short detail block:

| D-20260511-01 | Decision Ask: Deck Intent | @outline | open | true | deck.md, MEMORY.md | choose audience scene/duration package |

### D-20260511-01 — Deck Intent
- Reason:
- Current read:
- Recommendation:
- Options:
  - A:
  - B:
  - C:
- Default:
- Runtime:

roles/*.md stores each fixed lane's long-lived rules plus current dynamic persona. Update it when diagnosis.md changes.

Legacy migration:

• If MEMORY.md is missing but PROJECT.md exists, read PROJECT.md as legacy memory.
• On the next state write, create MEMORY.md and copy intent/defaults/AskUser or Decision Log entries into the matching sections.
• Leave PROJECT.md untouched. Do not keep writing to it.

Decision Ask Policy

Decision Ask is a room-scoped decision layer, not a questionnaire. AskUser UI is only one possible renderer.

The room is the durable scope. A necessary ask belongs to $DECK_DIR/threads/threads.md first, then renders through the runtime:

Decision Ask semantics -> AskUser UI if available -> plain text pause if blocking -> assumed default if non-blocking

Default rule:

Infer what materials already prove.
Use a reasonable default when the risk is low.
Ask only when the answer changes the deck direction.

One /codeck run may use at most 2 Decision Ask rounds. Count all internal modules against that budget.

Allowed Decision Ask moments:

1. Project Init — only when project state is missing or ambiguous.
2. Deck Intent — audience scene, duration, language, and core goal. Bundle these into one round. Skip fields already clear from the user request, materials, MEMORY.md, or deck.md. Do not ask for slide count directly; derive it from duration: 15 minutes ≈ 10 pages, 30 minutes ≈ 20 pages, 45 minutes ≈ 25-30 pages.
3. Design Direction — before visual generation or when user says "change the visual style".
4. Export Format — only when user says "export" without PDF / PPTX / all.
5. Speech Style — only when user asks for a script and style or duration is missing.

Never create a Decision Ask for:

• Generic permission to continue
• Whether to generate HTML
• Whether to run review
• Whether to save files
• Whether to use existing materials

Those steps are automatic.

Decision record shape:

id: D-YYYYMMDD-NN
owner: @orchestrator | @outline | @design | @export | @speech | @review
moment: Project Init | Deck Intent | Design Direction | Export Format | Speech Style | User-Owned Conflict
reason: why this cannot be safely inferred
current_read: what the room already knows
recommendation: the preferred option and one concrete reason
options: 2-3 mutually exclusive packages
default: the option to use if non-blocking or unanswered
blocking: true only when inventing the answer would damage the deck direction or user commitment
writes_to: MEMORY.md | deck.md | DESIGN.md | speech.md | review.md | export artifact
status: open | answered | defaulted | superseded

Every rendered ask must contain:

1. Re-ground — which skill, which step. One line.
2. Current read — what the system thinks is true.
3. Recommendation — Suggest [X] because [reason].
4. Options — 2-3 mutually exclusive choices. Mark one as recommended.

Only state verified facts. Unexecuted actions use "will / plan to".

Runtime rendering:

1. If structured AskUser UI is available, render the decision through that UI.
2. If no structured UI is available and blocking: true, show the same decision as a compact plain-text question, write it to threads/threads.md, and stop before mutating the target artifact.
3. If no structured UI is available and blocking: false, use the default, record assumed default, and continue.
4. If the 2-round budget is exhausted, use the default only for non-blocking decisions; leave blocking decisions open in threads/threads.md and stop.

If the user does not answer a non-blocking decision, use the recommended option. Record it in MEMORY.md as assumed default.

Persist every Decision Ask result:

• threads/threads.md stores every open decision before it is rendered to the user.
• MEMORY.md stores current room state, project-level intent, defaults, Decision Log, task index, thread index, and artifact index. It is not a transcript.
• deck.md stores deck-shaping decisions that affect rebuilds.

Automatic light review must not trigger Decision Ask. Write findings and fixes to review.md. Ask only when the content has a real conflict the user must decide.

Single-entry routing

/codeck routes the user's request before running the pipeline:

For edits, deck.md is the single truth source. HTML is the audience preview, not the editing surface.

Before a lane works:

1. @orchestrator writes or updates a task ticket.
2. The owner states the claim in channel.
3. The owner reads only the needed current-truth and work-state artifacts. Do not read channel unless debugging history.
4. Cross-owner changes go to threads/threads.md as proposals.

After a lane works:

1. Update the task ticket.
2. Rewrite MEMORY.md Active Context, Latest Channel Summary, Open Threads, Task Index, Artifacts, Assumed Defaults, and Decision Log so they describe current state.
3. Append the role exchange to today's channel file.
4. Keep the user response compact unless the user asked for the channel.

Two directories

• Current directory (`.`) — the user's project. Materials live here. Final HTML goes here too so the user can open it directly.
• `$DECK_DIR` — codeck's deck room. MEMORY.md, channel/, tasks/, threads/, roles/, deck.md, diagnosis.md, DESIGN.md, custom.css, slides.html, review.md, speech.md.

Scan materials in .. Write intermediate artifacts to $DECK_DIR. Output final HTML to ..

Phase 1: Init + status

DECK_DIR="$HOME/.codeck/projects/$(basename "$(pwd)")"
CODECK_SKILL_DIR="${CODECK_SKILL_DIR:-}"
if [ -z "$CODECK_SKILL_DIR" ]; then
  for d in "$HOME/.agents/skills/codeck" "$HOME/.codex/skills/codeck" "$HOME/.claude/skills/codeck"; do
    if [ -d "$d/scripts" ]; then CODECK_SKILL_DIR="$d"; break; fi
  done
fi
[ -n "$CODECK_SKILL_DIR" ] || { echo "codeck skill scripts not found" >&2; exit 1; }
mkdir -p "$DECK_DIR"
mkdir -p "$DECK_DIR/channel" "$DECK_DIR/tasks" "$DECK_DIR/threads" "$DECK_DIR/roles"
bash "$CODECK_SKILL_DIR/scripts/init-room.sh" "$DECK_DIR"

bash "$CODECK_SKILL_DIR/scripts/status.sh" "$DECK_DIR"

Initialize the room before creating any Decision Ask:

1. If $DECK_DIR/MEMORY.md is missing, create it from the Memory protocol template.
2. If $DECK_DIR/PROJECT.md exists, migrate its intent, defaults, and AskUser / Decision Log entries into MEMORY.md.
3. If $DECK_DIR/tasks/tasks.md is missing, create it with the task ticket table.
4. If $DECK_DIR/threads/threads.md is missing, create it with the thread table.
5. If any roles/*.md file is missing, create it with that fixed lane's ownership and write boundaries.
6. Append the current request and routing read to today's channel file.

Read existing MEMORY.md, active tasks/tasks.md rows, open threads/threads.md rows, and deck.md before creating any Decision Ask. Ignore legacy outline.md. If a value is already stated there or safely inferable from materials, do not ask it again.

Track Decision Ask rounds for this run. Stop at 2. If another non-blocking decision remains, use the recommended default and write it to MEMORY.md. If a blocking decision remains, leave it open in threads/threads.md and stop before touching its target artifact.

Phase 2: Material scan

Scan the current directory (the user's project), not DECK_DIR.

bash "$CODECK_SKILL_DIR/scripts/scan-materials.sh" .

Do not use eval find for material scan. If the script is unavailable, use a plain find with explicit ! -path exclusions.

Phase 3: Content diagnosis

If materials exist and $DECK_DIR/diagnosis.md doesn't, read materials and diagnose:

Research before diagnosis

If the material involves a domain you're unfamiliar with, or uses specialized terminology, search the web first. Understand the field's key concepts, common presentation patterns, and what experts in this space consider hard to explain. This grounds your diagnosis in real knowledge, not guesses.

Examples:

• Material about "WebTransport protocol" → search for what it is, how it differs from WebSocket, who's adopting it
• Material about a specific company's product → search for the product, its competitors, its positioning
• Material in a niche academic field → search for how practitioners in that field typically present findings

Three signals

1. Domain — what field? Determines outline role.
2. Expression challenge — what's hardest to convey? Determines design role.
3. Audience starting point — what do they know / not know? Determines review role (inverse selection: the listener most likely to struggle).

Role selection methodology

Don't pick from a list. Don't match by domain. Find the person whose way of thinking cracks this specific problem.

Outline role — who asks the right question about this material? Identify the core tension in the material, then find someone known for penetrating that type of tension — regardless of their field. A product launch where the real challenge is "why should anyone care" might need Sondheim (every lyric earns its place) more than a marketing guru. A technical architecture talk where the challenge is "too many moving parts" might need Tufte (information compression) or a film editor (what to cut).

Test: does this person's way of questioning change what the outline includes and excludes? If the outline would be the same without them, the match is wrong.

Design role — whose formal logic mirrors the content's structure? Not "good designer" but "whose way of organizing form matches how this argument moves." A content that builds layer by layer might map to Ravel. A content driven by contrast might map to Caravaggio. A content that strips away to reveal essence might map to Dieter Rams. The match can come from any domain — music, painting, architecture, choreography — because form is transferable.

Test: can you state the structural mapping in one sentence? ("This content does X; this person's work does X in visual/sonic/spatial form.") If not, the match is decorative.

Review role — inverse selection. Not the expert. The listener most likely to struggle or push back. The role determines what gets flagged — not correctness, but comprehension and trust.

Test: would this person interrupt you mid-presentation? If not, pick someone harder to convince.

Material summary

One-line summary per file: what it is, how it can be used. Written into diagnosis.md.

Output: $DECK_DIR/diagnosis.md

# Diagnosis

## Materials

| File | Content | Use for |
|------|---------|---------|
| {filename} | {one-line description} | {role in deck} |

## Domain
{description}

## Expression challenge
{hardest part to convey}

## Audience starting point
{what they know / don't know}

## Role recommendations

### Outline stage
{role name} — {derivation: domain + why this person's method of explaining reshapes the structure}

### Design stage
{role name} — {derivation: expression challenge + structural mapping between content and this person's visual logic}

### Review stage
{role name} — {derivation: audience starting point + why this person would struggle or push back}

Skip diagnosis if no materials — let user provide topic directly in each stage.

State outputs

Define what the user sees in each pipeline state. These are the exact outputs — not summaries.

Empty state (no materials found)

When Phase 2 finds zero files:

No materials found in this directory.

That's fine — you can start from a topic directly.

Tell me the topic in one sentence, or drop files here and run /codeck again.

Warm, not clinical. One primary action. No error language.

Error state (build-html.sh / assemble.sh fails)

When build-html.sh or assemble.sh exits non-zero:

Assembly failed: {error message}

Check that custom.css and slides.html exist in {DECK_DIR}.
Run /codeck to rebuild.

Name the exact file missing. Give the exact command to fix it.

Stale state (upstream changed, downstream not rebuilt)

When status.sh detects staleness:

⚠ {stage} is stale — {upstream file} changed after {downstream file} was built.
Run /codeck to rebuild.

One line. Name the files. Give the command.

Phase 4: Results

status.sh already outputs the dashboard. Below it, add:

1. Materials — file types and counts from Phase 2
2. STALE — one-line explanation if any stage is stale
3. Threads — unresolved decisions from threads/threads.md, only if any are open
4. Channel — only if the user asks to see the internal channel

Don't redraw the table.

Role reveal (if Phase 3 ran)

This is the first moment the user sees the system think. Don't dump three role names — tell a story in three beats:

1. What your content is really about — the core tension, in one sentence. Not the topic, the underlying struggle. ("Your material isn't about microservices — it's about convincing a team to accept short-term pain for long-term sanity.")

1. Who I'm bringing in, and why — the derivation, not just the name. Connect the person's way of thinking to the content's tension. ("For the outline, I'm thinking of Feynman — not because this is physics, but because your argument needs to make the invisible feel obvious. He did that better than anyone.")

1. What this changes — one concrete consequence. ("This means the outline won't start with background. It'll start with the one thing your audience already knows is broken.")

Keep it short. Three paragraphs, not three pages. The point is: the user should feel their content was seen, not just processed.

Channel visibility

Every /codeck run writes the role channel to channel/YYYY-MM-DD.md. User responses stay compact by default. If the user asks for the channel, show the role blocks for the work just done or about to happen.

Channel output is evidence, not authority. If channel conflicts with MEMORY.md, deck.md, or DESIGN.md, trust the current-truth file and record a thread only when the conflict changes the deck.

Use handles, not stage names:

@orchestrator
{what I read, which task ticket I opened, who owns it}

@outline
{what I will change or changed, and which artifact I touched}

@design
{handoff or rebuild note}

@review
{verification lens and result}

Do not invent work. If a role did not participate, omit it.

Phase 5: Handoff

All outputs go to $DECK_DIR/. Next skill reads upstream outputs.

Before the response ends, update:

• MEMORY.md Active Context, Latest Channel Summary, Task Index, Artifacts
• tasks/tasks.md current ticket status and handoff
• threads/threads.md if a decision remains
• channel/YYYY-MM-DD.md with the expanded role exchange

status.sh prints a NEXT recommendation. Use it:

User can run /codeck anytime to see progress.