chore(release): 0.7.3

- PAPERCRAFT_HOME also holds the chat settings (/.config),
  so a test copy never touches the installed app's.
- get_computed_styles reports a text colour bound to a variable as
  var(--name). Closes #13.
- Composer chips wrap instead of running under Stop and Add; the agent pill
  drops its text when it would run under the top bar's buttons.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Cargo.lock

@@ -3128,7 +3128,7 @@ dependencies = [
 
 [[package]]
 name = "papercraft"
-version = "0.6.0"
+version = "0.7.3"
 dependencies = [
  "arboard",
  "axum",

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "papercraft"
-version = "0.6.0"
+version = "0.7.3"
 edition = "2024"
 rust-version = "1.95"
 description = "A native infinite-canvas app for planning, working and sketching"

PROGRESS.md

@@ -1,6 +1,6 @@
 # Progress
 
-Last updated: 2026-10-21.
+Last updated: 2026-10-24.
 
 ## Phase 3: local-first design and whiteboard tool
 
@@ -76,6 +76,55 @@ are credited in `THIRD_PARTY.md`.
 * **Not carried over**: comments, design tokens, AI image generation, video export, teams;
   moving layers between pages; remote image URLs (see README Known gaps).
 
+### 2026-10-25: 0.7.3 (QA sub-agent, helper agents, auto mode, mid-turn messages, isolated dev settings)
+
+* QA sub-agent: Claude Code chats always get a Sonnet `qa` agent that lints and screenshots built frames and returns only findings, so screenshots stay out of the main context.
+* Helper agents (`src/mcp/delegate.rs`, Settings, Chat, Helper): `delegate` hands `qa`, `read` or `build` tasks to a cheaper agent CLI (pi with MiniMax M3.1 Flash, opencode, an Ollama-backed agent, any command taking `{prompt}`). It drives PaperCraft through `papercraft-cli`, runs off the UI thread, and only its last 6,000 characters come back. With a helper set, the chat agent orchestrates: keeps the brief and taste, delegates review, bulk reading and fully specified edits.
+* Dev config isolation: `PAPERCRAFT_HOME` also isolates the chat settings (`$PAPERCRAFT_HOME/.config`), so a second copy for testing never touches the installed app's settings.
+* Live check in an isolated copy: `delegate` ran the pi helper (MiniMax M3.1 Flash) as QA, a message added mid-turn went Waiting then Seen, and the agent drafted #13.
+* Fixes #10, #11, #12, #13:
+  * #10: `move_nodes` to another page keeps mains' component keys and relinks every instance to moved nodes; new `rename_file` renames a project.
+  * #11: swapped nested instances persist as swap overrides (`InstanceLink.swap`), preserving component identity across outer component updates and copies.
+  * #12: `move_nodes` to another page takes the moved nodes' annotations along; detached notes resolve as `Detached #n`.
+  * #13: `get_computed_styles` reports a text colour bound to a variable as `var(--name)`.
+* Auto mode (the composer's Auto chip): when a turn ends on `ask_user` questions, the first (recommended) option of each is sent and the chat goes on, until Stop or a turn hits its budget. The agent is told to put its recommendation first, default what only the user can know, and end each turn with next steps.
+* The Present button uses a play icon drawn like the icons beside it. Composer chips wrap instead of running under Stop and Add. Top bar: the agent pill drops its text when it would run under the buttons on the right.
+* While a chat turn works, the composer's Add (or Enter) queues a message (`src/mcp/steer.rs`). It rides on the agent's next PaperCraft tool result for that chat, so every backend reads it mid-turn; the composer shows each as Waiting or Seen. What no tool call picked up is sent as the next message when the turn ends, or goes back into the box after Stop.
+
+### 2026-10-24: 0.7.2 (UI kits, code links, issues from the app, design playbook)
+
+* Code links (Figma Code Connect style): a main component carries its library import and a React and an egui template (`set_code_link`). The React, Tailwind, inline JSX and egui exports write linked instances as that code, with overrides filled in and imports gathered once; the AI spec lists them as Library Components with a "use it, do not rebuild it" rule and stops describing their inner layers.
+* UI kits (`src/ui_kits.rs`, `apply_ui_kit`): shadcn/ui, Ant Design, React Bits and egui, each with Light and Dark variables, tokens and components in its look and density, all linked. The chat brief asks for the code library and starts from its kit.
+* Issues from the app (`src/issues.rs`): the agent's `report_issue` draft card (Send or Discard) and a Report a PaperCraft issue form, filed on the repository in Settings (token in the keychain, else the `tea` login), de-duplicated by title against open and closed issues.
+* design-playbook in the chat: with the `playbook` CLI installed, the brief asks whether to follow the playbook and the agent looks rules up with `design_rules`, citing ids in its report.
+* Top bar: project name and page picker on one centred row; the two-row title had outgrown the bar and covered the canvas.
+* Evidence: `src/ui_kits.rs`, `src/issues.rs`, `src/mcp/codegen.rs` (`linked_components_export_as_library_code`), `specs/ui-kits/spec.md`.
+
+### 2026-10-24: 0.7.1 (rename fix)
+
+* Enter, Escape and clicking away end a rename again (projects on the home screen, library assets, versions). The fields took focus back before checking whether they had lost it.
+
+### 2026-10-24: 0.7.0 (images, research, plugins, fewer tokens)
+
+* Images in chat: paste from the clipboard (Cmd/Ctrl+V with an image-only clipboard now reaches the composer), thumbnails in the composer and in sent messages, a lightbox, image-only messages. Attachments are sent at most 1568 px (originals kept); `zoom_image` reads a region of the original.
+* Brief and finish: `ask_user` draws questions as a card of clickable answers (palette swatches, reference image tiles, the recommended answer marked). The chat rules ask for a brief before the first design, variables and components first, sizes from the agreed density, a short report and next steps as a card.
+* Web research (the composer's Web chip): `fetch_reference_images` (public addresses only, icons, logos and backdrops dropped, social cards last), a References gallery that dims what the agent did not keep; on Claude Code a Haiku `researcher` sub-agent does the searching.
+* Image generation, opt-in (Settings, Chat, Images): Codex (offered once when signed in with a ChatGPT plan), Google Flow through `gf`, or any command (`{prompt}`, `{aspect}`, `{ratio}`, `{out}`, passed as data).
+* Plugins (Settings, Plugins): Avataaars, DiceBear, Iconify, Lorem Picsum, Flags, QR codes, placed with `<x-plugin use="…">`, cached on disk. SVG import draws masked groups as images (Avataaars).
+* Tokens: lean Claude Code sessions (42.7k to 20.6k prompt tokens), deferred tool schemas and context editing on the API, a fresh session with a recap past 120k (or on demand from the meter), per-turn cost for resumed Claude Code sessions, a 1M context meter, a turn budget. Model and effort pickers for Claude Code.
+* Also: page switcher in the Chat top bar, steps stay folded while a turn runs (no flicker), `lint_design` asks 24 px targets on desktop screens, the installer keeps `papercraft-mcp` current.
+* Evidence: `src/mcp/{refs,imagegen}.rs`, `src/plugins.rs`, `src/io/svg_import.rs`, `src/ui/{chat,chat_msgs,chat_agents,settings}.rs`, `src/chat/{anthropic,claude_code}.rs`, `tests/chat_tools.rs`, `specs/chat-tokens/spec.md`.
+
+### 2026-10-22: 0.6.1 (agents via ACP, annotations)
+
+* Agents are found through the login shell's PATH (read once at startup, 3 s limit, merged with the app PATH and the usual folders), so an app opened from the Dock finds Claude Code, Codex and the rest. No install instructions anywhere; a missing agent shows Not found with Set path… and Docs.
+* ACP backend (`src/chat/acp.rs`): any agent that speaks the Agent Client Protocol over stdio, with PaperCraft's MCP server, streamed updates, permission flow, cancel and load. Registry as data in `src/chat/agents.rs`: native agents found on PATH, npx agents off until enabled, custom agents by name, command and arguments. Verified here: OpenCode works; Gemini CLI's server currently refuses ACP clients; Cursor needs `cursor-agent login`; Pi works through `pi-acp` after enabling.
+* Agent panel over the chat column (Detected, Enable via npx, Not found, API endpoints, Custom, search), a Connect an agent setup card, a sign-in card, and multiple `@ Selection` chips.
+* Settings → Chat redesigned: one agents list with details, endpoints, defaults, scrolling page on one label grid.
+* Send bug: a turn that cannot start keeps the typed text and shows an error row with Retry, never a silent no-op (`a_failed_start_keeps_the_text_and_shows_an_error`).
+* Annotations: notes with leader lines on a page, six categories with keys (style comes only from the category), project and per-page templates with built-ins and a Customize modal, view modes (⌘⇧A), Annotate tool (N), side view, inspector section, Annotate with AI, six MCP tools (`annotate` batches), and comments in `get_design_spec` and the code exports. Exports and Present leave annotations out.
+* Evidence: `src/chat/{path,agents,acp}.rs`, `src/ui/{chat_agents,annotations,annotations_canvas}.rs`, `src/model/annotation.rs`, `src/editor/annotations.rs`, `src/mcp/tools.rs`, `tests/annotations*.rs`, `specs/chat-agents/`, `specs/annotations/`, `docs/chat.md`, `docs/annotations.md`.
+
 ### 2026-10-21: Chat mode
 
 * Chat layout next to the canvas (⌘J, `Design | Chat`), with three backends: Claude Code and Codex through their CLIs (`claude -p` stream-json, `codex exec --json`) against the app's MCP server, and the Anthropic API with a key in the OS credential store. One event stream for all three; tool calls run through the same MCP dispatch, so runs, review and revert behave as before.

README.md

@@ -90,7 +90,7 @@ Projects live in `~/Documents/papercraft/<name>/` (`project.json`, one JSON file
 
 ## Agents
 
-PaperCraft exposes **61 tools** (`src/mcp/tools.rs`) three ways. They run the same code inside the app.
+PaperCraft exposes **70 tools** (`src/mcp/tools.rs`) three ways. They run the same code inside the app.
 
 | Way | Notes |
 |---|---|
@@ -138,7 +138,7 @@ EOF
 
 ### Chat
 
-Chat mode (⌘J, or `Design | Chat` in the top bar) puts a conversation next to the canvas. It runs Claude Code or Codex through their CLIs with PaperCraft's MCP server, or the Anthropic API with your own key, and it shows tool calls, checkpoints and run review in the thread. Chats are saved in the project's `chats/` folder. See [docs/chat.md](docs/chat.md).
+Chat mode (⌘J, or `Design | Chat` in the top bar) puts a conversation next to the canvas. It runs Claude Code and Codex through their CLIs, any agent that speaks the Agent Client Protocol (ACP, such as OpenCode, Gemini CLI or Cursor) found on your PATH, or the Anthropic API with your own key. It shows tool calls, checkpoints and run review in the thread. Chats are saved in the project's `chats/` folder. See [docs/chat.md](docs/chat.md).
 
 <p align="center">
   <img src="docs/screenshots/readme/chat.png" alt="Chat mode: the chat column on the left shows a finished agent run with its Review, Revert and Accept actions, and the canvas on the right" width="100%">
@@ -146,6 +146,10 @@ Chat mode (⌘J, or `Design | Chat` in the top bar) puts a conversation next to
   <sub>Chat mode with a finished agent run: review, revert or accept from the thread.</sub>
 </p>
 
+### Annotations
+
+Annotations are notes that explain a design in place: business rules, behaviour, data sources, accessibility and dev notes, each a numbered card with a leader line to its layer. Categories set the look, templates say which notes a page wants (product spec, dev handoff, accessibility review and more), and the view cycles between full notes, badges and hidden with ⌘⇧A. Exports and Present mode leave them out. Agents add them in one `annotate` call. See [docs/annotations.md](docs/annotations.md).
+
 ### Skills
 
 [`skills/`](skills) holds agent skills: `papercraft` (the build, screenshot, lint, fix loop and tool cheat sheet), `papercraft-design` (design-quality rules) and `papercraft-critique`. [docs/agents.md](docs/agents.md) is the full guide: install, registering with agents, scripts, runs and review, safety. [docs/design-system.md](docs/design-system.md) covers variables, components and libraries.
@@ -223,6 +227,7 @@ The scenario tests drive the editor with synthetic pointer and keyboard events;
 | ⌘= / ⌘- / ⇧0 | Zoom in / out / 100% | ⇧1 / ⇧2 | Zoom to fit / selection |
 | PageUp / PageDown | Previous / next page | ⌘\\ | Show / hide UI |
 | ⌘J | Design / Chat mode | ⌘L | Add selection to chat (Chat mode) |
+| N | Annotate tool | ⌘⇧A | Annotations: full, badges, hidden |
 
 Draw a frame around existing shapes and it adopts them; dragging a layer into or out of a frame re-parents it. Hold ⌘ while dragging to suspend snapping.
 
@@ -237,7 +242,7 @@ Draw a frame around existing shapes and it adopts them; dragging a layer into or
 * **Variables:** gradients keep fixed colors; there are no text styles and no component variants or properties.
 * **Booleans are destructive:** the result is a new vector path with flattened curves.
 * **`write_html`** ignores margin, tables and transforms (it reports them as `ignoredStyles`) and does not fetch remote images; pass a local path or a `data:` URI.
-* **Agents:** one project is open at a time in the app, and `move_nodes` stays on one page. The in-app Chat mode runs Claude Code, Codex or your own Anthropic key; other agents run outside and connect over MCP or the CLI.
+* **Agents:** one project is open at a time in the app, and `move_nodes` stays on one page. The in-app Chat mode runs Claude Code, Codex, ACP agents or your own Anthropic key. Agents that are not in Chat connect over MCP or the CLI.
 * **Sharing** has no live presence (cursors, selections) yet.
 * **Pen pressure** is simulated from stroke velocity on macOS tablets, because winit 0.30 does not expose it for mouse-like pen events.
 * **Releases** cover macOS (Apple silicon) and Linux x86_64. Windows packaging scripts exist but are not part of releases.

ROADMAP.md

@@ -2,7 +2,7 @@
 
 PaperCraft is an early alpha: a local-first infinite-canvas design app that agents can drive. This
 file lists what is still missing, in priority order, with an honest status. Done items are checked
-against `src/mcp/tools.rs` and the UI code (2026-10-12); rows 7, 8, 12, 15, 23, 25, 32, 34, 39, 40, 44, 45, 46 and section 6 were updated 2026-10-21. Milestone history is in
+against `src/mcp/tools.rs` and the UI code (2026-10-12); rows 7, 8, 12, 15, 23, 25, 32, 34, 39, 40, 44, 45, 46 and section 6 were updated 2026-10-21; rows 15, 47 and 48 on 2026-10-22. Milestone history is in
 [PROGRESS.md](PROGRESS.md); known rendering and import limits are in the README.
 
 Status: **done** (works today), **partial** (some of it works), **todo**.
@@ -32,7 +32,7 @@ Agents can already build a design system through the tools; people cannot yet ma
 
 | # | Item | Status | Notes |
 |---|---|---|---|
-| 15 | MCP and CLI tool surface | **done** | 61 tools, one undo step per call, `papercraft-cli list / schema / call`. |
+| 15 | MCP and CLI tool surface | **done** | 67 tools, one undo step per call, `papercraft-cli list / schema / call`. |
 | 16 | Agent skills | **done** | `skills/papercraft` and `skills/papercraft-design`. |
 | 17 | Accessibility annotations | todo | Reading order, roles and alt text on layers, checked by the linter. |
 | 18 | Code export from frames | **done** | `get_jsx` supports `format: "inline"`, `"tailwind"` (Tailwind utility classes), and `"html"` (HTML + CSS rules with `:root` variables). |
@@ -86,3 +86,10 @@ Spec: [specs/agent-runs/spec.md](specs/agent-runs/spec.md), tasks: [specs/agent-
 | 43 | Comments assigned to agents | todo | `@agent` in a thread triggers an edit and a reply. Needs threaded comments. |
 | 44 | In-app prompt bar | **done** | Chat mode: Claude Code, Codex or your own Anthropic key; see `docs/chat.md`. |
 | 45 | Correct tool annotations | **done** | `lint_design` no longer sends `readOnlyHint`; `fix: true` edits are classified. |
+
+## 7. Annotations and chat agents
+
+| # | Item | Status | Notes |
+|---|---|---|---|
+| 47 | Design annotations | **done** | Notes with leader lines and categories, templates per project and page, view modes (⌘⇧A), Annotate tool (N), side view, inspector section, six MCP tools; see `docs/annotations.md`. |
+| 48 | ACP agents in Chat | **done** | Any ACP agent on the PATH (OpenCode works; Gemini CLI refuses ACP clients for now; Cursor needs `cursor-agent login`), npx agents after enabling, custom agents; see `docs/chat.md`. |

deploy/Dockerfile

@@ -6,7 +6,7 @@ FROM rust:slim AS build
 # Must match the wasm-bindgen crate in Cargo.lock.
 ARG WASM_BINDGEN=0.2.129
 RUN rustup target add wasm32-unknown-unknown \
- && apt-get update && apt-get install -y --no-install-recommends binaryen pkg-config libssl-dev \
+ && apt-get update && apt-get install -y --no-install-recommends binaryen pkg-config libssl-dev libdbus-1-dev \
  && rm -rf /var/lib/apt/lists/* \
  && cargo install wasm-bindgen-cli --version ${WASM_BINDGEN} --locked
 COPY --from=vendor crates /vendor
@@ -21,7 +21,7 @@ COPY . .
 RUN sh web/build.sh && cargo build --release --locked --bin papercraft-relay
 
 FROM debian:stable-slim
-RUN apt-get update && apt-get install -y --no-install-recommends libssl3 ca-certificates \
+RUN apt-get update && apt-get install -y --no-install-recommends libssl3 libdbus-1-3 ca-certificates \
  && rm -rf /var/lib/apt/lists/* \
  && useradd --system --uid 10001 papercraft
 COPY --from=build /src/target/release/papercraft-relay /usr/local/bin/papercraft-relay

docs/agents.md

@@ -1,14 +1,14 @@
 # Driving PaperCraft with agents
 
-PaperCraft exposes one set of 61 tools three ways. They all run the same code inside the app, so an agent sees what you see and every call is one undo step.
+PaperCraft exposes one set of 70 tools three ways. They all run the same code inside the app, so an agent sees what you see and every call is one undo step.
 
 | Way | Use it when | Needs the app running |
 |---|---|---|
 | MCP over HTTP, `http://127.0.0.1:29980/mcp` | the agent client speaks MCP (Claude Code, Codex, pi) | yes |
 | MCP over stdio, `papercraft-mcp` | the client only starts stdio servers | for calls; `tools/list` works offline |
-| Shell, `papercraft-cli` | the agent has a terminal and you want no tool schemas in its context | for `call`; `list`, `schema` and `icons` work offline |
+| Shell, `papercraft-cli` | the agent has a terminal and you want no tool schemas in its context | for `call`; `list`, `schema` and `icons` work offline (with the app running, `list` and `schema` show the app's tools) |
 
-If you would rather talk to the agent inside the app, Chat mode (⌘J, or `Design | Chat` in the top bar) is the in-app alternative. It runs Claude Code or Codex through their CLIs against the same MCP server, or the Anthropic API with your own key, and keeps the conversation next to the canvas, with the same runs, review and revert. The setup and the thread are in [chat.md](chat.md). The rest of this guide covers driving the app from your own terminal or agent client.
+If you would rather talk to the agent inside the app, Chat mode (⌘J, or `Design | Chat` in the top bar) is the in-app alternative. It runs Claude Code or Codex through their CLIs against the same MCP server, ACP agents such as OpenCode or Gemini CLI, or the Anthropic API with your own key, and keeps the conversation next to the canvas, with the same runs, review and revert. The setup and the thread are in [chat.md](chat.md). The rest of this guide covers driving the app from your own terminal or agent client.
 
 ## Install
 
@@ -91,6 +91,8 @@ papercraft-cli call get_screenshot '{"nodeId":"abc123:1"}' --project Demo -o car
 
 ### Shell and scripts
 
+When `call` or `shell` reaches the app and the app is another version, the CLI prints once on stderr, for example `papercraft-cli 0.6.1 ≠ app 0.6.2; reinstall the CLI`. `list` and `schema` read the running app's tool list, so tools newer than the CLI still show; with the app closed they use the CLI's built-in list.
+
 `papercraft-cli shell [--project X] [--offline]` keeps one session open (offline, one in-process project; otherwise the app). A line is `<tool> <json>` (JSON may span lines until its braces close) or a command: `:help :tools :schema <tool> :open <project> :pages :page <name|id> :shot [nodeId] [file] :undo :redo :save :watch on|off :quit`. `:watch on` renders the artboard last edited to `preview.png` after each editing call, shown inline on iTerm2, WezTerm and kitty. Input is plain stdin (no line editing or history; use `rlwrap` for that). Piped input prints no prompt and exits 1 if any line failed, listing the line numbers on stderr.
 
 ```sh
@@ -112,6 +114,24 @@ EOF
 * Components: `get_components` lists mains of every page and enabled library with a `componentKey`; `create_instance` and `swap_instance` accept a main from any page or library (`componentId` or `componentKey`).
 * Libraries: `publish_library`, `list_libraries`, `enable_library`, `update_library`, `disable_library`. They work offline too, on the project store.
 
+## Seeing what the user sees
+
+* `get_basic_info` and `get_selection` include `viewport: {pageId, rect: [x, y, w, h], zoom}`: the page the user is viewing and the world rect on screen.
+* `set_viewport {pageId?, nodeIds? | rect?, zoom?, padding?, force?}` switches the user's page (as the page list does) and fits the nodes or rect, or centres them at an explicit `zoom` (0.01 to 256; `select_nodes {zoom: true}` stops at 100%). While the user is panning or zooming (10 s after their last camera move) it is refused with "user is moving the canvas", unless `force: true`.
+* `find_nodes {textValue}` finds text layers containing the phrase on every page (or `pageId`), ignoring case and line breaks; `*` wildcards match the whole value instead. `get_tree_summary {nodeId, fullText: true}` prints every text layer's complete text, for copy review across screens.
+* `get_screenshot {nodeId, annotations: true}` draws the notes on the captured layers (cards, number badges, leader lines), so an agent can check note placement.
+
+## Annotation tools
+
+* `annotate {pageId?, notes: [{node, text, type?}]}` adds many notes in one undo step. `node` is a node id or an exact layer name; the first line of `text` is the label and the rest the body; `type` is a category key (default `rule`) that the page's template must enable. Placement is automatic. `annotate {tidy: true, frame?}` with no notes re-lays-out the existing notes and returns `{"tidied": n}`, the number of notes it moved.
+* `get_annotation_template {pageId?}` returns the allowed types with their guidance and fields. Call it once before `annotate`.
+* `set_annotation_template {template, scope?: "page" | "project", pageId?}` switches the template by id or name.
+* `upsert_annotation_template {template: {id?, name, display?, categories: [{key, enabled?, guidance?, fields?}]}, categories?: [{key, name, color, icon}], setAs?: "project" | "page", pageId?}` creates or replaces a template and its category styles in one undo step. `delete_annotation_template {template}` removes one (the last stays).
+* `get_annotations {nodeId?, pageId?, full?}` reads notes as `[number, node, type, text]` rows.
+* `update_annotation {id | ref, text?, type?}` and `delete_annotations {ids}` take ids or references such as `Sign in #3`.
+
+See [annotations.md](annotations.md) for categories, templates and what the notes look like.
+
 ## Safety
 
 The server listens on loopback only, rejects requests that carry an `Origin` header (browsers) and foreign `Host` headers. Anything on your machine that can reach `127.0.0.1:29980` can edit the open project, so do not forward the port. Undo (⌘Z, or the `undo` tool) reverts any agent edit.

docs/annotations.md

@@ -0,0 +1,148 @@
+# Annotations
+
+An annotation is a note that explains a design in place: a business rule, a behaviour, where the data comes from, an accessibility gap. Each note is a small card with a number, a category, a label and an optional body, joined to its layer by a leader line. Teams use them for review and handoff, and turn them off for presenting.
+
+Annotations live on a page and are saved with it (the `annotations` list in the page file). Pages without any load and save exactly as before. Categories and templates are project-wide and live in `design.json`.
+
+## Categories
+
+Every note has one category. The category is the only thing that sets how a note looks: its icon, colour and badge come from the category, and a note stores none of them. Change a category's colour or icon and every note in it changes.
+
+Each category has a short type key. Agents and the Annotate popover use the key; the UI shows the name.
+
+| Key | Name | Icon | Colour | Use for |
+|---|---|---|---|---|
+| `rule` | Business rule | scale | violet `#7c5cff` | the rule or constraint the part enforces |
+| `behaviour` | Behaviour | zap | blue `#2f80ed` | states, transitions, edge cases |
+| `content` | Content | type | green `#1fa87a` | copy, tone, length |
+| `data` | Data | database | amber `#e08a1e` | where a value comes from |
+| `a11y` | Accessibility | accessibility | pink `#d94f8a` | focus, labels, contrast, reading order |
+| `dev` | Dev note | code | grey `#6b7280` | implementation notes for engineers |
+
+You can add categories, rename them, change their key, colour and icon, and write guidance for each one in the Customize modal. A custom category gets a key too.
+
+## Templates
+
+A template says which kinds of notes a page wants, and what each note should contain. The project has a default template; a page can use its own instead.
+
+Built-in templates are editable copies, not locked. Each category in a template has enabled or not, guidance (a short prompt for what the note should say) and optional fields (structured lines such as `Owner`).
+
+| Template | Categories | Fields |
+|---|---|---|
+| Product spec | rule, behaviour, content, data | none |
+| Dev handoff | behaviour, data, dev, a11y | none |
+| Accessibility review | a11y | Issue, WCAG, Fix |
+| Content review | content | Tone, Max length |
+| Minimal | rule | none |
+
+Pick the project default or "This page" from the template chips at the top of the Annotations side view. Built-ins can be reset; custom templates can be deleted. Duplicate copies any template so you can change it.
+
+**Customize…** (side view) opens the template editor: name, display (full notes or badges), the enabled categories, the guidance and fields of each. Changes to categories apply to the whole project.
+
+The Annotate popover offers only the enabled categories of the active template, and shows each field as a labelled line. A field is stored as a `Name: value` line at the top of the note's body. Agents get the same rule: `annotate` refuses a type the template disables and lists the allowed keys.
+
+## View modes
+
+Three levels, all view state. They are not saved in the design and not undoable.
+
+* **Full**: every note shows its label and the first lines of its body. Below 60% zoom notes show as badges, so cards never spill over the design; zoom in to read them.
+* **Badges**: only the number badges are shown.
+* **Hidden**: nothing is drawn.
+
+`⌘⇧A` (View → Annotations) cycles Full → Badges → Hidden → Full. The starting mode follows the template's display setting until you change it.
+
+Two finer controls:
+
+* **Per note**: collapse a note to its number badge; click the badge to expand it again.
+* **Per category**: show or hide a category from the category filter in the side view.
+
+Annotations are drawn as an overlay on the canvas only. Exports (PNG, SVG, PDF) and Present mode never include them. There is no option to include them in an export yet.
+
+## Annotate tool
+
+Press `N` for the Annotate tool, then click a layer. The New annotation popover opens beside it with:
+
+* Category, from the template's enabled categories.
+* Label, up to 40 characters. The popover shows the count.
+* The template's fields, if any.
+* A note, in Markdown.
+
+Save creates the note with its leader line to the layer's side; Cancel discards it. You can also select layers, right-click, and choose **Annotate**.
+
+Each note's leader line starts at the note card and ends at the anchor on the target. The card and the anchor are stored relative to the target, so moving or resizing the layer keeps the note attached. Drag a card to move it; the line re-routes.
+
+If the target is deleted, the note stays where it was in the side view's **Detached** group, with a warning badge.
+
+## Annotations side view
+
+The side view lists the page's notes:
+
+* Grouped by top-level frame, plus **Page** (notes on no frame) and **Detached**.
+* Each row shows the number, the category chip and the label. Click a row to select the layer and zoom to it.
+* The category filter shows or hides categories.
+* Edit a note inline, or delete it. Edits are one undo step each.
+* **Add annotation** opens the popover for the current selection.
+* **Annotate with AI** puts a prompt in the Chat composer, unsent, with the selected frame (or the page) as a chip. The prompt asks the agent to read the template with `get_annotation_template` and add all the notes in one `annotate` call. It needs a saved project and Chat mode.
+
+Numbers run per frame in reading order (top to bottom, then left to right), so a frame reads 1, 2, 3. Notes on no frame are numbered after that, from 1.
+
+## Inspector
+
+When a single layer is selected, the inspector has an Annotations section. It lists that layer's notes, with edit and delete, and an Add annotation button. The note's anchor is one of the four sides of the target (Top, Right, Bottom, Left) with a position along it.
+
+## Agents
+
+Agents work with notes through six MCP tools, all in `src/mcp/tools.rs`. They are one undo step per call and appear in `edits_doc`, so a run that only adds notes still gets a Review card (a "Notes" group, counted as added, changed or removed), and Revert removes them.
+
+| Tool | What it does |
+|---|---|
+| `get_annotation_template` | Reads the template: `[type key, guidance, field names]` for each enabled type. Call it once before `annotate`. |
+| `set_annotation_template` | Switches the template by id or name, for one page (default) or the project (`scope: "project"`). |
+| `upsert_annotation_template` | Creates or replaces a template, and creates or restyles categories in the same call (`categories: [{key, name, color, icon}]`). `setAs: "project"` or `"page"` makes it the default or the page override. Returns `{id, name, types}`. |
+| `delete_annotation_template` | Deletes a template by id or name. The last template stays. |
+| `annotate` | Adds many notes at once. Each note names a node id or an exact layer name on the page, a `text` and a `type` (default `rule`). |
+| `get_annotations` | Reads notes as `[number, nodeName, type, text]` rows (text capped at 120 characters). `full: true` returns complete objects; `nodeId` limits to a subtree; `pageId` picks another page. |
+| `update_annotation` | Changes the text or type of one note, by `id` or by reference such as `Sign in #3`. |
+| `delete_annotations` | Deletes notes by ids or references such as `Sign in #3`. |
+
+Custom template with a new category, used on the current page:
+
+```json
+{
+  "template": {"name": "Legal review", "categories": [
+    {"key": "legal", "guidance": "Legal risk this part carries.", "fields": ["Owner", "Law"]},
+    {"key": "rule"}]},
+  "categories": [{"key": "legal", "name": "Legal", "color": "#aa3355", "icon": "scale"}],
+  "setAs": "page"
+}
+```
+
+`icon` must be one of the built-in icon names (an unknown name lists them). Every key in the template must exist, either already or in `categories`. The template and category edit is one undo step; `setAs: "page"` adds a second.
+
+For `annotate` and `update_annotation`, the first line of `text` is the label and the rest is the body. There are no positions, icons or colours: PaperCraft places each note beside its target, outside the frame, stacked so none overlap. `annotate` is all or nothing: if one note is refused (an unknown layer, or a type the template disables), none is added and the error lists the allowed keys.
+
+Batched call, on a page whose template is Product spec:
+
+```json
+{
+  "notes": [
+    {"node": "Sign in", "type": "rule", "text": "Lock after 5 failed tries\nLocks for 15 minutes, then resets."},
+    {"node": "Email field", "type": "data", "text": "Source: users.email\nValidated on blur"},
+    {"node": "Submit", "type": "behaviour", "text": "Disabled until both fields are valid\nSpinner while the request runs"}
+  ]
+}
+```
+
+The reply is one row per note, `[number, node, type, label]`. Then `get_annotations {}` reads them back, and `delete_annotations {"ids":["Sign in #3"]}` removes one by its frame reference.
+
+`annotate {"tidy": true}` with no notes re-runs placement for every note on the page (or on `frame`) in one undo step and returns `{"tidied": n}`, the number of notes it moved; `0` means the layout was already tidy.
+
+To see the notes, `get_screenshot {"nodeId": "<frame>", "annotations": true}` draws the cards, number badges and leader lines of the notes whose target is inside the captured layers, placed and routed as on the canvas (full cards, scale 1). The image grows to include the cards. Without `annotations`, or with `false`, the screenshot is the same as before.
+
+**Handoff.** `get_design_spec` and the code exports (`get_jsx`) include the notes of the frame as comments. Each note is one line, `// 1 Business rule: Lock after 5 failed tries`, with its body lines as more `//` lines, so code generated from a frame carries the rules next to the markup.
+
+## Storage
+
+* The page file has an `annotations` list. Each note holds its target (or a point, when it has none), its anchor, label, body, category, position, author (`user` or `agent`) and creation time. A note whose target is gone is marked `detached`.
+* `design.json` holds the categories and the templates, with the project default. A page may name its own template.
+* Old projects load unchanged. A page without annotations saves byte-identical to before.

docs/chat.md

@@ -8,18 +8,35 @@ Chat mode is native only. The web build shows the switch disabled with a note.
 
 ## Backends
 
-Chat drives the same tools as an agent connected over MCP, so the edits, runs, undo and revert behave the same whichever backend you use. There are three kinds of backend.
+Chat drives the same tools as an agent connected over MCP, so the edits, runs, undo and revert behave the same whichever backend you use. There are three ways in: two CLIs that PaperCraft talks to directly, any agent that speaks ACP, and API endpoints.
 
 **Claude Code** runs `claude -p` with streaming JSON in and out. PaperCraft writes an MCP config that points the CLI at the app's own server on `127.0.0.1`, allows the `mcp__papercraft__*` tools, and passes `--strict-mcp-config`, so the CLI sees only PaperCraft's server. The PaperCraft skill text goes in through `--append-system-prompt-file`. The working directory is the project folder. Follow-up messages resume the CLI session by its id, so the agent keeps its context.
 
 **Codex** runs `codex exec --json` in the project folder. The PaperCraft server is added with a `-c mcp_servers.papercraft.url=...` override, and its tools are pre-approved (`default_tools_approval_mode="approve"`), so Codex does not stop to ask for each one. The PaperCraft instructions are prepended to the first prompt of a chat; resumed sessions do not repeat them. Follow-ups resume the Codex session the same way.
 
+**ACP agents** run any agent that speaks the Agent Client Protocol (https://agentclientprotocol.com) over stdio. PaperCraft starts the agent once per turn, sends `initialize` and `session/new` with the project folder as the working directory, and passes PaperCraft's MCP server: over HTTP when the agent supports it, otherwise through the bundled `papercraft-mcp` stdio bridge. The prompt goes out as `session/prompt`, and the agent's streamed updates become the same thread as the other backends: text, thinking, tool cards and the plan. Stop sends `session/cancel`. Follow-ups resume with `session/load` when the agent supports it. ACP has no system prompt field, so the PaperCraft instructions go in front of the first prompt.
+
+The ACP agents PaperCraft knows about, and what was checked on the development Mac:
+
+* **OpenCode** (`opencode acp`): works.
+* **Gemini CLI** (`gemini --acp`): found and started, but its server currently refuses ACP clients, so turns fail with the agent's error until Google changes that.
+* **Cursor** (`cursor-agent acp`): sign in first with `cursor-agent login`.
+* **Pi** (`pi-acp`, an adapter run through `npx`): enable it once in the agent panel or Settings, then it works like the others.
+* **Goose, Kimi, Devin, Junie, Mistral Vibe, Hermes, OpenHands, Stakpak, VT Code, Poolside, Droid**: started with the command their own docs give. They are not all verified here.
+* **Claude Code and Codex ACP adapters** (`npx`): listed under Enable via npx. Claude Code and Codex themselves use the direct backends above.
+
+Agents that run through `npx` are off until you enable them, because enabling runs code that npx fetches from npm.
+
+ACP permissions: PaperCraft's own tools run without asking, except the destructive ones listed under Approval below. Any other tool the agent asks to run (its own shell or file tools, for example) shows an approval card with Approve or Deny in the thread, and Deny is sent back to the agent.
+
+**Custom agents** are your own ACP agents: a name, a command (a program name on the resolved `PATH`, or a full path) and arguments. The arguments are split on spaces, with no quoting. Custom agents appear under Custom in the agent panel and in Settings → Chat.
+
 **API endpoints** call a model over HTTPS with your own key. Settings → Chat holds a list of endpoints, each with a name, a format, a base URL, a model and an optional key. The app's tool list is sent as the API tools, and every tool call runs through the same code path as an MCP call on the UI thread. The loop continues until the model ends its turn, for up to 40 tool rounds per message.
 
 * **Format**: `Anthropic` is the Messages API (`POST {base}/v1/messages`, `x-api-key`). `OpenAI` is Chat Completions (`POST {base}/chat/completions`, `Authorization: Bearer`, streaming `choices[].delta`, tools as `{"type":"function",...}`, tool results as `tool` messages). Both give the same thread: text, tool cards, approvals, Stop, the usage line.
 * **Base URL and model** are free text, for example `https://api.anthropic.com`, any Anthropic-compatible URL, or `http://127.0.0.1:11434/v1`. The key is optional, so local servers need none.
 * A default endpoint named `Anthropic` (`https://api.anthropic.com`, `claude-opus-5-5`) exists from the start, so the first run needs only a key.
-* The agent picker in the chat header lists Claude Code, Codex, then every endpoint by name. Each chat remembers its endpoint.
+* The agent panel lists Detected agents, API endpoints and Custom agents (see Setup). Each chat remembers its backend.
 * **Claude-only fields** (effort, the thinking summary, the refusal fallback and its beta header, eager tool-input streaming, prompt-cache breakpoints) are sent only when the base URL host is `api.anthropic.com`. For that host the model menu offers `claude-opus-5-5`, `claude-sonnet-5-5` and `claude-haiku-5-5`, effort is `medium` by default (`low` and `high` in the menu), and the Thinking row shows the summary. Every other endpoint gets a plain request. Missing `usage`, thinking or cache fields in the reply are fine.
 * **OpenAI format and images**: attached images are sent as `image_url` data URIs. A tool message in this format carries text only, so a tool result with an image (such as `get_screenshot`) goes back as the text `[image omitted]`, and the image follows as an extra user message in the same round.
 * **Usage line**: input, output and cached tokens, as the server reports them; nothing when it reports none. A cost estimate shows only for the Claude models in the small price table in `src/chat/endpoint.rs`. It is not a bill.
@@ -27,66 +44,133 @@ Chat drives the same tools as an agent connected over MCP, so the edits, runs, u
 
 ## Setup
 
-Chat looks for `claude` and `codex` on `PATH`. A missing CLI is not an error: its row says `Not found` and shows the install command:
+### Finding agents
 
-* Claude Code: `npm install -g @anthropic-ai/claude-code`
-* Codex: `npm install -g @openai/codex`
+Chat finds agents the way your terminal does. At startup PaperCraft reads your login shell's `PATH` once, off the UI thread (`$SHELL -l`, 3 s limit), and merges it with the app's own `PATH` and the usual folders (`~/.local/bin`, `/opt/homebrew/bin`, `/usr/local/bin`, `~/.cargo/bin`, `~/.bun/bin`, `~/.npm-global/bin`, Volta and nvm). Detection and the agent processes both use that merged `PATH`. This is why an agent installed under `~/.local/bin` or Homebrew is found even when PaperCraft was opened from the Dock.
 
-Install one of them, or add a key, then press Detect again on the Chat page. Settings → Chat shows each CLI found with its version.
+A row shows a version when the agent reports one. A program with the right name that is not the agent (for example the `goose` database migration tool, which shares its name with Block's Goose agent) shows `Wrong program` with its path, not a false Ready.
 
-The Chat page also holds:
+PaperCraft never shows install commands. An agent that is not found shows `Not found`, with `Set path…` (opens Settings) and `Docs` (the agent's own documentation). Install the agent with its own installer, then press Detect again.
 
-* **Paths**: an override for the `claude` and `codex` binaries, for when they are not on `PATH` or you want a specific install. Empty means the first one on `PATH`.
-* **Default backend**: the backend a new chat starts with. The agent picker in the chat header changes it per chat.
+### Agent panel
+
+Click the agent chip in the chat header. The panel opens over the chat column with a shadow and a dimmed backdrop, sized like the column. It has a search field (name or id), and these sections:
+
+* **Detected**: agents found on your `PATH`, with the version and path. Click one to use it for this chat.
+* **Enable via npx**: agents that start through `npx` (Pi, Qwen Code, Kilo, Cline, Auggie, Copilot CLI, Qoder, CodeBuddy and the Claude Code and Codex ACP adapters). Each row has an `Enable` chip. The section says that it runs code fetched by npx.
+* **Not found**: registry agents that are not on this computer, folded by default.
+* **API endpoints**: your endpoints, with an `Add` action that opens Settings.
+* **Custom**: `Add custom agent` opens Settings.
+
+Each row has a status chip: `Ready`, `Enable`, `Not found`, `Wrong program`, `No key` (an endpoint without a key) or `Set up` (an endpoint without a URL or model). Arrow keys move through the rows, Enter picks one, Escape closes the panel.
+
+Settings → Chat has the same list, so an agent can also be hidden from the panel there (the `Show in the chat agent picker` switch on its row).
+
+### Setup card
+
+A chat with no backend chosen shows one card, "Connect an agent". It lists the agents detected on this computer, up to four, with `Show all agents` for the rest. Below them are `Add an API endpoint` and `Add custom agent`. The card never lists install steps.
+
+### Sign-in
+
+When an agent says it needs you to sign in (ACP's "Authentication required" and similar messages), the failed turn shows "<Agent> needs you to sign in" with the command from the agent's docs, then Retry. For Cursor that is `cursor-agent login`, for OpenCode `opencode auth login`, for Codex `codex login`, for Claude Code's ACP adapter `claude /login`. For agents without a known command the card says to sign in with the agent's own app and retry. Chat does not sign in for you.
+
+### Settings → Chat
+
+The page scrolls and keeps one grid for its labels. Its sections:
+
+* **Agents**: one row per agent, with its status, version and path. A details disclosure holds the **Path override** (for a binary that is not on `PATH`, or a specific install; empty means the first match on the merged `PATH`), the **Arguments override** (replaces the registry's ACP arguments, for example `--acp --debug`), and for npx agents the **Enable** button. The **Show in the chat agent picker** switch hides an agent from the panel. Custom agents have Name, Command and Arguments fields and a Remove button. `Detect again` re-runs detection.
 * **API endpoints**: add, edit and remove endpoints (name, format, base URL, model) and set the default one. Each has a key field with Save and Remove. Keys are stored in the macOS Keychain, Windows Credential Manager or the Linux Secret Service, under the service name `PaperCraft` and the user `api-key:<base URL>`. An older single key (`anthropic-api-key`) is still read for `api.anthropic.com` and moves to the new entry the next time you save it. Removing a key makes that endpoint stop until a key is saved again (for `api.anthropic.com`; other endpoints work without one).
 * **Defaults for new chats**: the backend, the endpoint, and the effort for `api.anthropic.com`.
 
-Settings from before endpoints existed (`api_model`) move into the default endpoint on first load. These defaults and the endpoint list (never the keys) are saved in `chat_settings.json` in your config folder (`~/Library/Application Support/papercraft/` on macOS). They are not part of any project.
+Settings from before endpoints existed (`api_model`) move into the default endpoint on first load. The `claude_path` and `codex_path` settings from earlier versions still win for their agent until you save the Agents list. These defaults, the agent settings and the endpoint list (never the keys) are saved in `chat_settings.json` in your config folder (`~/Library/Application Support/papercraft/` on macOS). They are not part of any project.
 
 ## The thread
 
-The header has the chat title (rename), the chat list (new chat, switch, delete), the agent picker (for example `Claude Code · Opus 5.5`) and the context meter. The meter shows the tokens in the context at the last model call against a 200k window, and turns amber past 75%. It is hidden for Codex, which reports no per-call usage.
+The header has the chat title (rename), the chat list (new chat, switch, delete), the agent chip (for example `Claude Code · Opus 5.5`) and the context meter. The meter shows the tokens in the context at the last model call against a 200k window, and turns amber past 75%. It is hidden for Codex, which reports no per-call usage.
 
 Replies are markdown: paragraphs, lists, code blocks with a copy button, inline code and links. While a turn runs, the reply streams in.
 
 * **Thinking**: when the backend reports reasoning, a row shows "Thinking…" while it runs and collapses to a summary when it ends. Expand it to read the text.
 * **Tool cards**: one card per tool call with the verb and target (for example "Drawing layout · Card"), a spinner while it runs, and a failed state with the error. Args and result are collapsible. `get_screenshot` results show a thumbnail.
 * **Step groups**: consecutive tool calls in one reply fold into one line ("Built the card · 5 steps") that expands to the cards.
-* **Plan**: when the agent keeps a todo list (Claude Code's `TodoWrite`), the latest list is pinned above the composer with a "done of total" count.
-* **Approval**: on the API backend, the destructive tools (`delete_pages`, `disable_library`, `import_design_system`, and `delete_nodes` on more than 20 layers) wait for Approve or Deny before they run. Deny returns an error to the model, which can then carry on. The CLI backends do not show this card. They use the CLI's own permission rules, and in non-interactive mode a tool that needs permission is denied. The papercraft tools are pre-allowed on both CLIs.
+* **Plan**: when the agent keeps a todo list (Claude Code's `TodoWrite`, or an ACP plan), the latest list is pinned above the composer with a "done of total" count.
+* **Approval**: on the API backend and for ACP agents, the destructive tools (`delete_pages`, `disable_library`, `import_design_system`, and `delete_nodes` on more than 20 layers) wait for Approve or Deny before they run. Deny returns an error to the model, which can then carry on. For ACP agents, any non-PaperCraft tool the agent asks for shows the same card. The Claude Code and Codex CLIs do not show this card. They use the CLI's own permission rules, and in non-interactive mode a tool that needs permission is denied. The papercraft tools are pre-allowed on both CLIs.
 * **Checkpoint**: when an agent run starts, the thread shows "Checkpoint · before run N". "Restore to here" reverts that run and every run after it, through the same revert as the Agent view.
-* **Run card**: when a run ends, "Agent finished, N changes" with Review, Revert and Accept. These are the same actions as the run review on the canvas.
-* **Errors**: a failed turn shows a red row with Retry. A missing CLI or key shows a setup row instead.
+* **Run card**: when a run ends, "Agent finished, N changes" with Review, Revert and Accept. These are the same actions as the run review on the canvas. The card names the client that made the run ("Chat", "Claude Code · 2") and refreshes after an undo.
+* **Progress**: while a run is open, the step group and the Working pill count calls and touched layers ("Styling · 212 calls · 1,340 changes"). The pill clears when the turn ends.
+* **Errors**: a failed turn shows a red row with Retry. A sign-in problem shows the sign-in card (see Setup).
 * **Canvas follow**: when a tool call targets a node, the canvas pans to it, unless you moved the camera in the last 10 seconds.
+* **Agent presence**: after each agent call, the layers it pointed at or changed get an accent outline that fades over 3 seconds, and the newest carries a label with the verb and the client ("Styling · Claude Code"). This works for chat turns and for any MCP client. When the agent works on another page, a chip "Agent on page X →" switches to it.
+
+## Design-system context
+
+Every turn carries a summary of the project's design system, so the agent reuses what exists instead of drawing look-alikes:
+
+* variables by collection and group with their modes (colors as `name=#light/#dark`), tokens as scales;
+* main components on every page and in enabled libraries, as `componentKey · name · w×h · uses`, most used first;
+* enabled libraries and their update status, and the page's annotation template;
+* a reuse rule: place components with `create_instance` / `swap_instance`, colors as `var(--…)`, spacing on tokens.
+
+The summary is capped at about 1,500 tokens. A large system keeps its most used components and collapses variable groups to counts, with a pointer to `get_components` / `get_variables` for the rest. It is rebuilt only when the design system changes. The API backends get it as a separate system block (cached on `api.anthropic.com`); CLI and ACP agents get it in their first prompt, and again as `[Design system changed] …` on the next message after a change. The context meter includes it. Turn it off in Settings → Chat → **Send design-system summary**.
+
+`lint_design` has a matching rule, `not-an-instance`: a frame built by hand that matches a component (same structure, size within 4 px, same fill, radius and stroke) is flagged with the component to use.
+
+## Pointing at the canvas
+
+* **Pick**: the Pick chip next to `@ Selection` turns the canvas into a picker without changing your selection. Click a layer to add it, or Shift+drag an area; Esc cancels. Each pick types `@Name` at the composer cursor and adds a chip.
+* **⌥-click**: in Chat mode, ⌥-click a layer on the canvas to add it the same way.
+* **Area chips** are sent as `[Area] page P, rect x,y,w,h, contains: <ids…>` (the top layers wholly inside the area, up to 40).
+* **Crops**: layer, area and `@ Selection` chips carry an image of that part of the page with its notes drawn (1024 px max). The API backend gets it as an image block; CLI and ACP agents get a file under `.papercraft-chat-tmp/`.
+* **Ask (⌘K)**: with a selection, ⌘K opens a prompt box anchored above it. Enter sends to the chat with the selection chip and its crop, and the reply streams in the chat column without leaving Design mode. With nothing selected, ⌘K opens the command palette (also ⌘⇧/).
+
+## Actions on a selection
+
+The **Actions** menu opens from the canvas right-click menu, from the Actions chip in the composer and from the command palette ("Actions: …"). It works in Design and Chat modes.
+
+Instant actions run in the app with no model. Each one is a single undo step and its own run (client "Actions"), and a toast reports what it did with Review and Revert:
+
+* **Fix lint here**: `lint_design {fix: true}` on the selection.
+* **Bind to tokens**: binds hard-coded colors to the nearest variable (within 12 RGB) and gaps, padding and radii to the nearest spacing or radius token (within 2 px). Values with no close match are listed.
+* **Use components**: replaces each `not-an-instance` hit with an instance of its component in the same place, keeping the texts.
+* **Rename layers**: names default layers (`Frame 12`, `Text`) from their text or role.
+* **Annotate**: opens the annotation popover.
+
+Agent actions send a prompt with a selection chip to the chat and switch to Chat mode: **Optimize** (tighten layout and spacing to the token scale, fix contrast, keep the content, prefer components), **Make consistent with the design system**, **Write states** (empty, loading, error), **Variations ×3** and **Annotate with AI**.
+
+**Send open notes to chat**, in the inspector's Annotations section for a top-level frame, sends the frame's open feedback and its `todo` / `question` notes as one prompt with a chip.
 
 ## Composer and attachments
 
 Type in the composer, ⏎ sends and ⇧⏎ adds a line. The empty state offers suggestions built from the project, such as "Design a sign-in screen using this project's variables" or "Lint this page and fix warnings".
 
-* **Selection**: the `@ Selection` chip adds the current selection to the message (also ⌘L). It is sent as a short line with the layer name and keys.
+* **Selection**: the `@ Selection` chip adds the current selection to the message as a chip (also ⌘L). Each click adds another chip, labelled with the layer name or `<first> · N layers`; an identical selection is added once, and each chip is removed on its own. Every chip is sent as a short line with the layer name and node ids.
+* **Start failures**: if a turn cannot start (agent missing, bad path), an error row with Retry shows in the thread and your text stays in the composer.
 * **Images**: the image button opens a file picker. You can also paste a clipboard image or drop image files on the column (PNG, JPEG, GIF, WebP).
 
-For the API backend, images are sent as image blocks, so the model sees them. For the CLIs, each image is saved under `.papercraft-chat-tmp/` in the project and its path is added to the prompt for the agent to read.
+For the API backend, images are sent as image blocks, so the model sees them. For the CLIs and ACP agents, each image is saved under `.papercraft-chat-tmp/` in the project and its path is added to the prompt for the agent to read.
 
-While a turn runs, the send button becomes **Stop**. Stop asks the backend to interrupt: the CLIs get an interrupt, or SIGINT, then a kill after 3 seconds. Every backend ends within 3 seconds. Tools that already ran stay applied, and the partial run can be reviewed or reverted like any other.
+While a turn runs, the send button becomes **Stop**. Stop asks the backend to interrupt: the CLIs get an interrupt, or SIGINT, then a kill after 3 seconds; ACP agents get `session/cancel`, then a kill if they do not exit. Every backend ends within 3 seconds. The thread shows a muted "Stopped" line, not an error. Tools that already ran stay applied, and the partial run can be reviewed or reverted like any other.
 
 ## Storage
 
-A chat is one file in the project: `chats/<id>.json`, holding the title, the backend (kind, model, effort, CLI session id) and the messages with their tool calls, usage and run ids. The API backend also keeps `chats/<id>.api.json`, the raw messages array (Messages format with thinking blocks, or Chat Completions format, whichever the endpoint speaks), so a follow-up replays exactly. The chat records its endpoint id. Deleting a chat asks first and removes both files, and its saved images.
+A chat is one file in the project: `chats/<id>.json`, holding the title, the backend (kind, model, effort, CLI or ACP session id, agent) and the messages with their tool calls, usage and run ids. The API backend also keeps `chats/<id>.api.json`, the raw messages array (Messages format with thinking blocks, or Chat Completions format, whichever the endpoint speaks), so a follow-up replays exactly. The chat records its endpoint id. Deleting a chat asks first and removes both files, and its saved images.
 
 Projects without a `chats/` folder load unchanged. The newest chat opens when you enter Chat mode.
 
 ## Privacy
 
-PaperCraft never writes an API key into a project file, a chat or a log. The key lives in the OS credential store. The CLIs run with your own logins and do not get the API keys from the environment: `ANTHROPIC_API_KEY`, `ANTHROPIC_AUTH_TOKEN`, `OPENAI_API_KEY` and `CODEX_API_KEY` are removed from the CLI's environment, so Claude Code and Codex bill your account, not a key you set for PaperCraft. The MCP server listens on loopback only, as it does for any other client.
+PaperCraft never writes an API key into a project file, a chat or a log. The key lives in the OS credential store. The CLIs and ACP agents run with your own logins and do not get the API keys from the environment: `ANTHROPIC_API_KEY`, `ANTHROPIC_AUTH_TOKEN`, `OPENAI_API_KEY` and `CODEX_API_KEY` are removed from their environment, so Claude Code and Codex bill your account, not a key you set for PaperCraft. The MCP server listens on loopback only, as it does for any other client.
 
-Prompts, project content and tool results go to whichever backend you chose: the CLI's model service, or the base URL of the endpoint you picked. Nothing goes to any other service. If a server echoes your key in an error message, the app hides it before showing the message.
+Prompts, project content and tool results go to whichever backend you chose: the CLI's or agent's model service, or the base URL of the endpoint you picked. Nothing goes to any other service. An npx agent downloads its package from npm the first time it runs. If a server echoes your key in an error message, the app hides it before showing the message.
 
 ## Limits
 
 * One turn per project at a time. Sending while a turn runs is refused with "A turn is already running"; use Stop first. Switching to another project stops a running turn.
 * The web build has no Chat mode.
-* The canvas shows the run's outlines after the turn ends, when the run is reviewed. They are not drawn live while the agent works.
+* While the agent works, the canvas shows fading outlines of the layers it touches; the full change review (added, changed, removed) comes after the turn, when the run is reviewed.
+* Variations ×3 is a prompt for now: the agent makes the variants; there is no side-by-side Keep tag yet.
+* Over the ACP stdio bridge the chat id does not reach the server, so those runs show under the agent's client name and close after 30 seconds idle instead of at turn end.
 * The API backend stops after 40 tool rounds in one message and says so.
 * Codex reports no per-call usage, so the context meter is hidden for it.
-* The CLI backends need the CLI installed and logged in. Chat does not install or log in for you.
+* Agents need to be installed and signed in with their own tools. Chat does not install them or sign them in for you.
+* ACP terminal and file-system requests from an agent are answered as not supported. Agents use their own file tools.

docs/design-system.md

@@ -60,7 +60,7 @@ papercraft-cli call set_text_content '{"updates":[{"nodeId":"<label layer id fro
 * `create_instance` returns the instance id and its layers, so you can set text and styles on them. Changes made through `set_text_content` or `update_styles` on an instance are recorded as overrides; everything else follows the main component.
 * In `write_html`, `<x-paper-instance component-id="<id>" layer-name="Button / Save"/>` places an instance inline.
 * `get_components` lists the main components of every page (and library components) with `componentKey`, page and how many instances each has.
-* **Across pages.** Instances can be placed from a main on any page: `create_instance` and `swap_instance` take a `componentId` (`pageId:num`) or a `componentKey` (stable, from `get_components`). Editing a main on page A updates its instances on every other page, one undo step per page. If the main is deleted its instances stay as they are, marked "Main missing" in Assets, where you can detach them. `import_components` still copies a main onto a page, but reuse no longer needs it.
+* **Across pages.** Instances can be placed from a main on any page: `create_instance` and `swap_instance` take a `componentId` (`pageId:num`) or a `componentKey` (stable, from `get_components`). Editing a main on page A updates its instances on every other page, one undo step per page. If the main is deleted its instances stay as they are, marked "Main missing" in Assets, where you can detach them. `<x-paper-instance component="Name"/>` in `write_html` links the same way when the main is on another page (it used to copy the main onto the page). `import_components` still copies mains onto a page, detached from the source; it is for moving a design system, not for using components.
 * **Limits.** No variants or properties: model variants as separate components (`Button / Primary`, `Button / Secondary`). `detach_instance` and `swap_instance` exist as tools; in the editor detach with ⌘⌥B.
 
 ## Icons

packaging/windows/package.ps1

@@ -59,7 +59,7 @@ function Get-PeHeader([string] $Path) {
   return @{ Machine = [BitConverter]::ToUInt16($bytes, $pe + 4); Subsystem = [BitConverter]::ToUInt16($bytes, $pe + 0x5C) }
 }
 $Machine = switch ($Arch) { 'x64' { 0x8664 } 'x86' { 0x14C } 'arm64' { 0xAA64 } }
-foreach ($name in @('papercraft.exe', 'papercraft-cli.exe', 'papercraft-mcp.exe')) {
+foreach ($name in @('papercraft.exe', 'papercraft-cli.exe', 'papercraft-mcp.exe', 'papercraft-relay.exe')) {
   $exePath = Join-Path $Bin $name
   if (Test-Path $exePath) {
     $h = Get-PeHeader $exePath
@@ -71,13 +71,13 @@ foreach ($name in @('papercraft.exe', 'papercraft-cli.exe', 'papercraft-mcp.exe'
 $Stage = Join-Path $TargetDir "windows-package\$Arch"
 Remove-Item -Recurse -Force $Stage -ErrorAction SilentlyContinue
 New-Item -ItemType Directory -Force -Path $Stage | Out-Null
-Copy-Item (Join-Path $Bin 'papercraft.exe'), (Join-Path $Bin 'papercraft-cli.exe'), (Join-Path $Bin 'papercraft-mcp.exe') $Stage
+Copy-Item (Join-Path $Bin 'papercraft.exe'), (Join-Path $Bin 'papercraft-cli.exe'), (Join-Path $Bin 'papercraft-mcp.exe'), (Join-Path $Bin 'papercraft-relay.exe') $Stage
 foreach ($f in 'LICENSE-MIT', 'LICENSE-APACHE', 'ATTRIBUTION.md', 'README.md') {
   $p = Join-Path $Root $f
   if (Test-Path $p) { Copy-Item $p $Stage }
 }
 
-& (Join-Path $PSScriptRoot 'sign.ps1') (Join-Path $Stage 'papercraft.exe') (Join-Path $Stage 'papercraft-cli.exe') (Join-Path $Stage 'papercraft-mcp.exe')
+& (Join-Path $PSScriptRoot 'sign.ps1') (Join-Path $Stage 'papercraft.exe') (Join-Path $Stage 'papercraft-cli.exe') (Join-Path $Stage 'papercraft-mcp.exe') (Join-Path $Stage 'papercraft-relay.exe')
 
 # ---- MSI ---------------------------------------------------------------------------------------
 $Msi = Join-Path $Dist "papercraft-$Version-windows-$Arch.msi"
@@ -119,9 +119,22 @@ $CliZip = Join-Path $Dist "papercraft-cli-$Version-windows-$Arch.zip"
 Remove-Item -Force $CliZip -ErrorAction SilentlyContinue
 Compress-Archive -Path $CliPortable -DestinationPath $CliZip
 
+# ---- relay zip ---------------------------------------------------------------------------------
+$RelayPortable = Join-Path $TargetDir "windows-package\papercraft-relay-$Version-windows-$Arch"
+Remove-Item -Recurse -Force $RelayPortable -ErrorAction SilentlyContinue
+New-Item -ItemType Directory -Force -Path $RelayPortable | Out-Null
+Copy-Item (Join-Path $Stage 'papercraft-relay.exe') $RelayPortable
+foreach ($f in 'README.md', 'LICENSE', 'LICENSE-MIT', 'LICENSE-APACHE', 'ATTRIBUTION.md') {
+  $p = Join-Path $Root $f
+  if (Test-Path $p) { Copy-Item $p $RelayPortable }
+}
+$RelayZip = Join-Path $Dist "papercraft-relay-$Version-windows-$Arch.zip"
+Remove-Item -Force $RelayZip -ErrorAction SilentlyContinue
+Compress-Archive -Path $RelayPortable -DestinationPath $RelayZip
+
 # Smoke test
 Invoke-Native 'papercraft-cli --help' { & (Join-Path $Stage 'papercraft-cli.exe') --help | Out-Null }
 
-$artifacts = @($Zip, $CliZip)
+$artifacts = @($Zip, $CliZip, $RelayZip)
 if (Test-Path $Msi) { $artifacts += $Msi }
 Get-Item $artifacts | Format-Table Name, Length

skills/papercraft-design/SKILL.md

@@ -78,7 +78,7 @@ Sizes: **12, 14, 16, 20, 24, 32, 40, 48, 64.** `lint_design` flags anything else
 ## 6. Touch targets and density
 
 * Interactive targets are **at least 44px** tall and wide on touch (mobile, tablet). Make them 44 with padding (12 + 20 + 12), not just a bigger font. On dense desktop UIs 32 to 40 is acceptable; say so, and keep spacing between targets at 8 or more.
-* `lint_design` checks frames named `Button…`, `Tab /…` and `Tab Bar /…` for height under 44, so **name interactive layers that way**.
+* `lint_design` checks frames named `Button…`, `Tab /…` and `Tab Bar /…`: under 44px tall on a touch screen (or a component on its own), under 24px on a screen 1024px or wider (WCAG 2.2 AA for pointers). **Name interactive layers that way.**
 * Mobile: body 16px or larger, side padding 16 to 24, bottom navigation within thumb reach, safe areas (status bar 44, home indicator 34 at the bottom).
 
 ## 7. Components and states

skills/papercraft/SKILL.md

@@ -21,6 +21,10 @@ papercraft-cli call get_basic_info      # fails with exit 1 if the app is not ru
 * `pageId` defaults to the page the user is viewing, which can change between calls: pass it explicitly. Node ids look like `3fa9c1:14`; never show them to the user.
 * App closed (server, CI)? `papercraft-cli call --offline ... --project <name>` runs the same tools on the project store and saves after every editing call; `papercraft-cli shell` keeps one session for many calls and reads scripts from stdin. It is refused while the app has that project open. See `docs/agents.md`.
 * With MCP tools you do not need the CLI; with a shell, `papercraft-cli list` and `papercraft-cli schema <tool>` give the schemas on demand. Pass big JSON on stdin: `... | papercraft-cli call write_html -`.
+* **Working in the user's open project: use `papercraft-cli --online call …`** (or `PAPERCRAFT_CLI_NO_OFFLINE=1`). Without it, a call made while the app is starting or stuck silently falls back to offline writes on the files. `stderr` warns when the CLI and app versions differ: tell the user, don't ignore it.
+* **One call at a time per agent.** No background loops or parallel batches against the app; a backlog of screenshot calls makes the app unresponsive for every agent.
+* **App stops answering** ("did not answer in time", 0% CPU): usually a macOS dialog (a privacy prompt such as "access files on a removable volume") or a modal in the app. Tell the user and wait; never kill, quit or restart their PaperCraft.
+* **See what the user sees**: `get_basic_info` / `get_selection` include `viewport {pageId, rect, zoom}`. `set_viewport {pageId, nodeIds | rect, zoom?}` switches page and frames things for the user; it refuses while they are moving the canvas (pass `force` only when they asked you to show something).
 
 ## 2. The loop: build, look, lint, fix
 
@@ -34,6 +38,23 @@ papercraft-cli call get_basic_info      # fails with exit 1 if the app is not ru
 
 Lint passes are necessary, not sufficient: contrast and grids say nothing about composition. The screenshot is the real test.
 
+## 2b. Big jobs and several agents on one project
+
+Lessons from redesigning a 3-page, 130-screen project with up to six agents at once:
+
+* **Split by section, not by page.** Give each agent named sections (frame ids) and forbid the rest. New sections go below the current lowest one; re-read positions right before creating, since another agent may have just added one.
+* **Section layout**: a section frame holds a title and rows of screen slots (a route label above a fixed-size screen). Keep 640 px between screens and rows and 1600 px between sections, so notes fit beside screens. Stack sections in one column.
+* **Recreate, then delete.** Build the new screen next to the old one, finish it, then delete the old one, and record the deleted ids and names in the report. A cleanup agent once deleted buttons and lost icons by editing in place.
+* **Reuse the design system across pages**: `create_instance` / `swap_instance` / `<x-paper-instance>` take a main from any page and stay linked. Never `import_components` to use a component: that copies mains onto the page, and instances made from the copies stop following the design system. `lint_design` flags hand-made look-alikes as `not-an-instance` (with the component key to use).
+* **QA pass after a build** (screen by screen at scale 1, one call at a time): text layers with stale fixed widths clip labels ("In prog", "New sess"): set `width: fit-content`. Sidebars and panels fixed at 900 px stop short on taller screens: make them fill the height and let the screen hug with a 900 px minimum. Tags and chips overflowing their box, cards overlapping buttons, footers not pinned, leftover `Text` / `Frame` names, gendered copy about people whose gender is unknown (use their name or neutral wording).
+* **Report per agent** in `specs/<feature>/<agent>-report.md`: what changed (ids), why (playbook rule ids from `playbook search`), open questions. Unknown product facts are written as questions for the team, never invented.
+
+## 2c. Annotations
+
+* Read the page's template first (`get_annotation_template {pageId}`), then one batched `annotate {pageId, notes:[{node, text, type}]}` per section, then `annotate {pageId, frame: <section id>, tidy: true}` (returns `{tidied: n}`; 0 means already placed).
+* Content: the first line is the complete rule in 60 characters or fewer; the body is condition → outcome with values as `Field: value` lines; attach to the exact element, not the screen; few high-value notes; unknowns as `? (team)`. Rules: `playbook rule ANN`.
+* Check placement with `get_screenshot {nodeId: <screen slot>, annotations: true, scale: 0.5}`; plain screenshots do not show notes. Cards must sit beside their target off any screen, lines must not cross other screens. Below 60% zoom the canvas shows notes as badges.
+
 ## 3. Naming
 
 * Layers: `Noun` or `Component / Variant` (`Button / Primary`, `Tab / Today`, `Card / Task`, `Icon / gear`). Slashes group the layers panel.
@@ -79,17 +100,28 @@ Full flows: [recipes.md](references/recipes.md).
 | Know the project, pages, artboards | `get_basic_info`, `get_tree_summary`, `get_children` |
 | Know what is selected | `get_selection` |
 | Inspect one node | `get_node_info`, `get_computed_styles`, `get_jsx` |
-| Find nodes by text or style | `find_nodes` |
-| See it | `get_screenshot` (`nodeId`, `scale` 1 or 2) |
+| Find nodes by text or style | `find_nodes` (text contains the phrase, any case, every page) |
+| Read all copy in a subtree | `get_tree_summary {nodeId, fullText: true}` (depth counts from the node: rows under a "Screens" frame need depth 3) |
+| See it | `get_screenshot` (`nodeId`, `scale` 1 or 2, `annotations: true` to include notes) |
+| Show the user something | `set_viewport` (page switch, fit, zoom) |
+| Document rules on the design | `annotate`, `get_annotation_template`, `get_annotations`, `update_annotation`, `delete_annotations` |
 | New screen | `create_artboard` (`width` and `height` required) |
 | Add content | `write_html` (`insert-children` or `replace`) |
 | Change styles or text | `update_styles`, `set_text_content`, `rename_nodes` |
 | Whiteboard shapes, sticky notes, arrows | `create_shape` |
 | Copy, move, delete | `duplicate_nodes`, `move_nodes`, `delete_nodes` |
 | Tokens | `get_variables` (`format` css, json or tailwind), `set_variables` (`collection`, `type`, `mode`, `alias`), `create_collection`, `delete_variables`, `set_variable_mode`, `bind_variable` |
-| Reuse | `create_component`, `create_instance`, `swap_instance`, `get_components`, `<x-paper-instance>` |
+| Reuse | `create_component`, `create_instance`, `swap_instance` (also a nested badge inside a row instance: a swap that survives main edits), `get_components`, `<x-paper-instance>` |
+| Project | `rename_file {name}` (open project stays open), `move_nodes {toPageId}` keeps instances linked and notes attached |
 | Shared libraries | `publish_library`, `list_libraries`, `enable_library`, `update_library`, `disable_library` |
 | Icons | `search_icons`, `<x-paper-icon name="set:name">`; CLI: `papercraft-cli icons add\|list\|search` |
+| Design in a library's look, hand off its code | `apply_ui_kit` (shadcn, antd, reactbits, egui), then `<x-paper-instance>` of its components; `set_code_link` for a component of your own; `get_jsx` / `get_design_spec` then write `<Button variant="outline">` instead of divs |
+| Avatars, more icons, photos, flags, QR | `<x-plugin use="avataaars\|dicebear\|iconify\|picsum\|flags\|qr" …/>` inside `write_html` (plugins the user turned on) |
+| References and art | `fetch_reference_images`, `zoom_image`, `generate_image` (when the user turned image generation on) |
+| Ask the user | `ask_user` (clickable answers; end your turn after it) |
+| Hand work to the helper agent | `delegate` (`role` qa, read or build, a self-contained `task`; only when Settings, Chat, Helper is set) |
+| Design rules (design-playbook) | `design_rules` (`query` or `rule` id; needs the `playbook` CLI; cite ids in the report) |
+| Report a PaperCraft problem | `report_issue` (a draft the user sends; synthetic repro only) |
 | Check quality | `lint_design` |
 | Export | `export` (PNG, SVG, PDF to files), `export_combined_pdf` |
 | Point at things for the user | `select_nodes` (`zoom: true`) |
@@ -113,8 +145,9 @@ Full flows: [recipes.md](references/recipes.md).
 * **Text inside nested frames**: `set_text_content` needs the Text node id. Walk `get_children` down from the instance; instances copy the main's structure (`Meta > Label`, not a flat list).
 * **No icon-colour follow for stroke sets** (Lucide, Tabler): the colour is frozen at write time, so a mode switch (Mocha to Latte) leaves icons as they were. Use a fill set (Phosphor, Material) when a design must switch modes.
 * **Pages**: `create_page`, `rename_pages`, `delete_pages` (needs `confirm: true` when the page has layers; the last page stays). `move_nodes` accepts `toPageId` to copy between pages.
-* **`get_screenshot` of `root` on a wide page** is scaled to fit and unreadable. Screenshot one artboard at a time. Hoisted components are not inside an artboard screenshot.
-* **CLI saves images to the cwd** as `<tool>.png`; newer builds take `-o FILE`. Run the CLI from a scratch directory, or pass `-o`, and rename after.
+* **`get_screenshot` of `root` on a wide page** is scaled to fit and unreadable. Screenshot one artboard (or one screen slot) at a time. Hoisted components are not inside an artboard screenshot.
+* **CLI saves images to the cwd** as `<tool>.png`, or to `-o FILE`. The CLI writes the file itself; on older builds an `out` path made the app write it and could freeze the app behind a macOS privacy prompt, so update the CLI with the app (`tools/install-local.sh` installs both).
+* **`swap_instance` renames the instance root to the new main's name** unless the name was overridden. Re-apply custom layer names after a batch swap (`rename_nodes`).
 
 ## 7. Troubleshooting
 
@@ -125,7 +158,10 @@ Full flows: [recipes.md](references/recipes.md).
 | `write_html` reports `ignoredStyles` | You used a CSS property that has no equivalent (`margin`, `box-shadow`, `transform`...). Replace it, see [html-subset.md](references/html-subset.md). |
 | Content is cut off | The artboard height is fixed. `update_styles` with `height: "fit-content"` or a larger number. |
 | A color did not change with the theme | It was written as hex, or it is a stroke or gradient. Use `var(--name)` on fills and text; strokes need `update_styles` after a mode switch. |
-| `create_instance` fails | The component is on another page. Instances must be on the component's page. |
+| `create_instance` fails | Wrong `componentId` / `componentKey` (read `get_components`). Mains on other pages and enabled libraries work; no import needed. |
+| A page is full of copied components at y=0 | Something imported mains onto the page (`import_components`, or old `write_html` builds). Swap their instances to the design-system mains (`swap_instance {componentKey}`), check a screenshot, then delete the unused copies. |
+| `papercraft-cli --online` says "did not answer in time" | The app is blocked: look for a macOS dialog or an in-app modal and ask the user to answer it. Don't retry in a loop and don't restart the app. |
+| A tool name is unknown, or the docs list tools the CLI lacks | The CLI is stale. `papercraft-cli list` reads the app's live list; reinstall the CLI with the app. |
 | Icon does not appear | The set is not installed: `papercraft-cli icons add phosphor`, then `papercraft-cli icons search <words>` for the exact `set:name`. |
 | Layer panel is a pile of `Frame` | You skipped `layer-name`. Fix with `rename_nodes` (batch) before finishing. |
 | Text looks tiny in the screenshot | Screenshots default to scale 1. Use `scale: 2`. |

skills/papercraft/references/compact-workflow.md

@@ -13,7 +13,7 @@ How to keep designs small, keep design and code in step, and spend few tokens do
 
 * Build every repeated thing once: `create_component`, then `create_instance` or `<x-paper-instance>`. Never paste markup twice. A screen is mostly instances plus layout frames.
 * No raw colours, radii, spacing or shadows in `write_html`: `var(--token)` only. Changing a token then updates every page, and `get_design_spec` reports token names, which is what code needs.
-* Reuse across pages with `import_components` (instances stay linked) or `<x-paper-clone node-id="page:id">`. Do not rebuild a component per page.
+* Reuse across pages with `create_instance {componentKey}` or `<x-paper-instance component="Name"/>` in `write_html`: a main on any page (or an enabled library, for `create_instance`) works directly and the instance stays linked to it. Do not `import_components` for that (it makes detached copies; it is for moving a design system). Do not rebuild a component per page.
 * One design system per brand: `export_design_system` writes an OpenDesign package (variables, modes, tokens); `import_design_system` loads it into another project. That is the sync path between files and apps.
 * Write one visual group per `write_html` (a header, a row), with `layer-name` on every box. Fewer, larger writes cost fewer round trips; groups that are too big make one bad style hard to find.
 

skills/papercraft/references/tools.md

@@ -60,14 +60,14 @@ Delete one or more pages. The last page of a project cannot be deleted. A page t
 
 ### `get_basic_info` (read-only)
 
-Essential context about the open project: project name, the page, node count, artboards (top-level frames) with ids, sizes and world positions, other top-level layers, fonts, and `pages` with the active one marked isActive.
+Essential context about the open project: project name, the page, node count, artboards (top-level frames) with ids, sizes and world positions, other top-level layers, fonts, `pages` with the active one marked isActive, and `viewport` ({pageId, rect: [x, y, w, h] world, zoom}: what the user sees).
 - worldX/worldY are world positions; x/y are relative to the parent.
 - Pass pageId to inspect another page without switching the user's view.
 
 
 ### `get_selection` (read-only)
 
-The selected layers on the page the user is viewing: ids, names, types, sizes, positions and the artboard each belongs to.
+The selected layers on the page the user is viewing: ids, names, types, sizes, positions and the artboard each belongs to, plus `viewport` ({pageId, rect: [x, y, w, h] world, zoom}).
 
 
 ### `get_node_info` (read-only)
@@ -94,9 +94,20 @@ Compact indented text tree of a subtree: type, name, id and size per node. Much
 
 | Argument | Type | Required | Notes |
 |---|---|---|---|
-| `depth` | number |  | Maximum depth (default 3, max 10). |
+| `depth` | number |  | Levels shown, counting nodeId (or each top-level layer for "root") as level 1 (default 3, max 10). Deeper levels collapse to "… N children". |
+| `fullText` | boolean |  | Show each text layer's complete text instead of its first 40 characters (for copy review). |
 | `nodeId` | string | yes | Root of the summary, or "root". |
 
+Depth counts from the node you pass, not from the page. For an artboard `Home` with a `Screens` frame holding rows, `{"nodeId": "<Home id>", "depth": 2}` stops at `Screens`:
+
+```
+frame "Home" [3fa9c1:1] 1440×900
+  frame "Screens" [3fa9c1:2] fill(1440)×hug(820)
+    … 3 children
+```
+
+The rows inside `Screens` are level 3, so they need `depth: 3` from `Home` (the default), or `depth: 2` from `Screens` itself. With `nodeId: "root"` every artboard is level 1, so the same rows also need `depth: 3`.
+
 
 ### `get_jsx` (read-only)
 
@@ -151,7 +162,7 @@ CSS styles (camelCase) of one or more nodes, as {nodeId: styles}.
 
 ### `find_nodes` (read-only)
 
-Find nodes by text content and/or computed style. Searches the whole project by default; pass pageId for one page or nodeId for a subtree. `textValue` matches text case-insensitively with * wildcards anchored to the whole value. `filters` are {styleName, styleValue} pairs (AND), values may use * wildcards; colors match by value.
+Find nodes by text content and/or computed style. Searches the whole project by default; pass pageId for one page or nodeId for a subtree. `textValue` finds text containing it, ignoring case and line breaks; with * wildcards it must match the whole value ("Get *"). `filters` are {styleName, styleValue} pairs (AND), values may use * wildcards; colors match by value.
 
 | Argument | Type | Required | Notes |
 |---|---|---|---|
@@ -195,6 +206,7 @@ Render a node (or the whole page when nodeId is omitted or "root") to a PNG and
 
 | Argument | Type | Required | Notes |
 |---|---|---|---|
+| `annotations` | boolean |  | Also draw the note cards, badges and leader lines of notes on the captured layers; the image grows to include the cards. |
 | `nodeId` | string |  | The node to capture; omit for the whole page. |
 | `out` | string |  | Optional file path to save the PNG to (also used by papercraft-cli). |
 | `scale` | number |  | Render scale: 1 (default) or 2. |
@@ -426,7 +438,7 @@ Main components of every page (or one with `pageId`): id, name, `componentKey` (
 
 ### `import_components`
 
-Copy the main components anywhere on `fromPageId` (including inside artboards, typically on a `Design System` page) onto this one, so you can place instances here with create_instance or <x-paper-instance component="Name"/> / component-id="ID". Components with the same name already here are skipped. Variables and tokens are already shared by every page; components are copies, not live links.
+Copies mains onto this page (detached from the source); only for moving a design system, not for using components: create_instance and swap_instance take a main on any page or an enabled library directly, with no import. Copies the main components anywhere on `fromPageId` (including inside artboards); components with the same name already here are skipped. Variables and tokens are already shared by every page.
 
 | Argument | Type | Required | Notes |
 |---|---|---|---|
@@ -446,7 +458,7 @@ Turn a frame or artboard into a main component. Build reusable parts (button, ro
 
 ### `create_instance`
 
-Place an instance of a component. Edit text and styles on the instance's descendants with set_text_content / update_styles (they become overrides); edits to the main component flow to every instance. Returns the instance id and its text layers.
+Place an instance of a component. componentKey may be a component on any page or an enabled library; no import needed (the instance stays linked to that main). Edit text and styles on the instance's descendants with set_text_content / update_styles (they become overrides); edits to the main component flow to every instance. Returns the instance id and its text layers.
 
 | Argument | Type | Required | Notes |
 |---|---|---|---|
@@ -461,7 +473,7 @@ Place an instance of a component. Edit text and styles on the instance's descend
 
 ### `swap_instance`
 
-Swap an instance root to point at a different main component (any page), preserving overrides on descendants matched by name and layer type.
+Swap an instance root to point at a different main component, preserving overrides on descendants matched by name and layer type. componentKey may be a component on any page or an enabled library; no import needed.
 
 | Argument | Type | Required | Notes |
 |---|---|---|---|
@@ -544,6 +556,62 @@ Stop using a library: bindings to its variables are detached (values kept), its
 | `libraryId` | string | yes |  |
 
 
+## Annotations
+
+### `annotate`
+
+Add notes to node ids or exact layer names on a page in one undo step. First text line is label; remaining lines are body. Type is an allowed category key (default rule); put fields as "Name: value" lines in text. Placement is automatic (tidy re-lays-out). Returns [number, node, type, label] rows; tidy alone returns {tidied: moved}.
+
+| Argument | Type | Required | Notes |
+|---|---|---|---|
+| `frame` | string |  | With tidy: only the notes on this frame. |
+| `notes` | array of object |  | `{node, text, type?}`. |
+| `pageId` | string |  |  |
+| `tidy` | boolean |  | Re-layout all notes (of frame, else the page) as part of this undo step; notes may then be omitted. |
+
+Check the placement with `get_screenshot {"nodeId": "<frame>", "annotations": true}`.
+
+### `get_annotation_template` (read-only)
+
+Read the page's annotation template: [type key, guidance, field names] for each allowed type. Call once before annotate.
+
+| Argument | Type | Required | Notes |
+|---|---|---|---|
+| `pageId` | string |  |  |
+
+
+### `set_annotation_template`
+
+Switch the annotation template by id or name, for one page or the whole project. Annotate only accepts the template's types.
+
+| Argument | Type | Required | Notes |
+|---|---|---|---|
+| `pageId` | string |  |  |
+| `scope` | string |  | `page` or `project`. |
+| `template` | string | yes |  |
+
+
+### `upsert_annotation_template`
+
+Create or replace an annotation template, optionally defining new categories (key, name, #hex color, icon name) first. Reuse existing keys freely. One undo step for the edit; setAs page adds a second. Returns {id, name, types}.
+
+| Argument | Type | Required | Notes |
+|---|---|---|---|
+| `categories` | array of object |  | Category styles `{key, name, color, icon}`; an existing key is updated, a new key needs all four. |
+| `pageId` | string |  |  |
+| `setAs` | string |  | `project` or `page`. |
+| `template` | object | yes | `{id?, name, display?: full or badges, categories: [{key, enabled?, guidance?, fields?}]}`. |
+
+
+### `delete_annotation_template`
+
+Delete an annotation template by id or name. The last template cannot be deleted; a project default pointing at it reverts to Product spec.
+
+| Argument | Type | Required | Notes |
+|---|---|---|---|
+| `template` | string | yes |  |
+
+
 ## Session
 
 ### `get_feedback` (read-only)
@@ -581,7 +649,7 @@ Mark notes on a layer resolved, with a short reply saying what you changed. With
 
 ### `select_nodes`
 
-Select nodes on the page the user is viewing so they can see what you mean; zoom: true also frames them in the viewport. An empty list clears the selection.
+Select nodes on the page the user is viewing so they can see what you mean; zoom: true also frames them in the viewport (at most 100%). An empty list clears the selection. For another page or a closer zoom use set_viewport.
 
 | Argument | Type | Required | Notes |
 |---|---|---|---|
@@ -589,6 +657,20 @@ Select nodes on the page the user is viewing so they can see what you mean; zoom
 | `zoom` | boolean |  |  |
 
 
+### `set_viewport`
+
+Move the user's view: switch to pageId (as clicking it in the page list does) and fit nodeIds or a world rect, optionally at an explicit zoom (0.01 to 256). With only zoom it zooms about the view's centre. Refused while the user is moving the canvas (within 10 s of their last pan or zoom) unless force: true. Returns the new viewport ({pageId, rect: [x, y, w, h] in world units, zoom}); get_basic_info and get_selection report the current one.
+
+| Argument | Type | Required | Notes |
+|---|---|---|---|
+| `force` | boolean |  | Move the view even if the user is panning or zooming. |
+| `nodeIds` | array of string |  | Nodes to fit in the view. |
+| `padding` | number |  | Screen margin when fitting, in px (default 64). |
+| `pageId` | string |  | Page to show; defaults to the page of nodeIds, else the current one. |
+| `rect` | array of number |  | World rect [x, y, width, height] to fit, instead of nodeIds. |
+| `zoom` | number |  | Explicit zoom level (1 = 100%); centres the nodes or rect at this zoom instead of fitting. |
+
+
 ### `undo`
 
 Undo the last change on a page (the user's or an agent's).

specs/annotations/spec.md

@@ -0,0 +1,71 @@
+# Annotations: notes with leader lines that explain a design
+
+## Why (user request, 2026-10-21)
+
+Teams annotate designs in Figma with callouts and arrows that explain what each part does: business rules, behaviour, states, data sources, edge cases. They are toggled on for review and handoff and off for presenting. PaperCraft has comments (threads, open/resolved) but nothing that documents a design in place.
+
+## Model
+
+* An **annotation** belongs to a page: `{ id, target: NodeId (or a point), anchor: side or offset on the target, label: short title, body: markdown text, category, number, position: where the note sits on the canvas, author: user | agent, created }`. Stored in the page file under `annotations` (serde default; old pages load unchanged).
+* **Categories** are project-level, editable: default `Business rule`, `Behaviour`, `Content`, `Data`, `Accessibility`, `Dev note`, each with a colour (theme palette) and icon.
+* Annotations survive moves and resizes of the target (anchor is relative); deleting the target keeps the annotation detached with a warning badge.
+
+## Classification drives the look (user, 2026-10-21)
+
+* The **category is the only styling input**. Icon, colour, badge shape and default label prefix come from the category; an annotation stores no icon or colour of its own. Changing a category's colour or icon restyles every note in it.
+* Categories carry a short **type key** used everywhere (`rule`, `behaviour`, `content`, `data`, `a11y`, `dev`); custom categories get a key too.
+* **Hiding information**, three levels, all undo-free view state (not saved in the design):
+  * per note: collapse to its number badge only (click the badge to expand);
+  * per category: show/hide from the category filter;
+  * global (⌘⇧A cycles; ⇧A is Add auto layout): full notes → badges only → hidden.
+  * Exports and Present mode use "hidden" unless chosen otherwise.
+
+## Templates (user, 2026-10-21)
+
+People annotate differently per project and even per page (a product spec, a dev handoff, an accessibility review). A **template** says which kinds of notes are wanted and what each should contain.
+
+* `AnnotationTemplate { id, name, categories: [ { key, enabled, prompt: short guidance for what this note should say, fields: [name] (optional structured lines, e.g. "Applies when", "Owner") } ], display: full | badges }`.
+* Stored in `design.json` (project-level list); the **project default** is one template, and a **page can override** it (`page.annotation_template`, serde default). New categories added in a template become available project-wide.
+* **Built-in templates** (editable copies, not locked): `Product spec` (rule, behaviour, content, data), `Dev handoff` (behaviour, data, dev, a11y), `Accessibility review` (a11y with fields "Issue", "WCAG", "Fix"), `Content review` (content with "Tone", "Max length"), `Minimal` (rule only).
+* **UI**: the Annotations side view header has a template picker (project default or this page) and a "Customize…" modal: toggle categories, edit name/colour/icon/key, edit each category's guidance and fields, duplicate/reset a template. The Annotate popover only offers the template's enabled categories, and shows the fields as labelled lines.
+* **Ask AI**: an "Annotate with AI" button in the side view (and a chat suggestion "Annotate this page") sends a prompt to the chat: annotate the selected frame or the page using the active template. The agent reads the template with one cheap call.
+* MCP: `get_annotation_template {pageId?}` returns compact `[[key, guidance, fields]]` for the enabled categories (a few hundred tokens); `annotate` rejects types the page's template disables (error lists the allowed keys) and accepts `fields` as `"Name: value"` lines inside `text`. `set_annotation_template {pageId?, template}` switches it.
+
+## Cheap for agents
+
+* One tool call can add many notes: `annotate {notes: [{node, text, type}]}` where `node` is a node id or a layer name on the active page, `text` is the note (first line becomes the label, the rest the body), `type` is the category key (default `rule`). No positions, icons or colours: placement is automatic (beside the target, outside its frame, stacked without overlap).
+* `get_annotations` returns compact rows `[n, node, type, text]`, not full objects; `{full: true}` for everything.
+* `update_annotation` / `delete_annotations` take ids or numbers ("frame Sign in #3").
+* The tool description stays under ~60 words.
+
+## Canvas
+
+* A note is a small card (number badge, label, first lines of the body) connected to its target by a **leader line**: elbow or straight, ending in a dot or arrow on the target's anchor, plus a thin outline around the target while hovered.
+* Notes are draggable; the line re-routes (reuse the connector router, `src/connector.rs`).
+* **Toggle**: View → Annotations (⌘⇧A) shows or hides all; a filter per category; never drawn in Present mode or exports (notes live only in the editor overlay; an export option to include them is a later item).
+* Numbering is per frame, in reading order, so a frame reads "1, 2, 3…".
+
+## UI
+
+* **Annotate tool** (N; A is the Frame tool): click a layer, type the note; or select layers → right-click → Annotate.
+* **Annotations side view**: list per frame with number, category chip, label; click to select and zoom; filter by category; edit inline.
+* **Inspector**: an Annotations section for the selected layer.
+
+## Agents and handoff
+
+* MCP: `annotate` (batch, see "Cheap for agents"), `get_annotations`, `update_annotation`, `delete_annotations`; classified in `edits_doc`.
+* `get_design_spec` and `get_jsx` include annotations as comments (`// 1 Business rule: …`), so code generated from a frame carries the rules.
+* Chat: the `@ Selection` context includes the selected layers' annotations; the agent can be asked to "annotate this screen".
+
+## Mock-ups
+
+`PaperCraft` design file, page `Annotations`: `Canvas / Annotated frame` (several categories, leader lines, numbering), `Canvas / Annotations hidden`, `Panel / Annotations` (344x1440), `Inspector / Annotation` (344x1440), `Popover / New annotation`.
+
+## Acceptance
+
+* Add, edit, move, hide/show, filter, delete annotations; undo for each.
+* Moving or resizing a target keeps the leader line attached; deleting it detaches the note with a badge.
+* Old projects load unchanged; page files without `annotations` stay byte-identical on save.
+* Exports and Present mode never include annotations.
+* MCP tools round trip; `get_design_spec` includes them.
+* Screenshot check against the mock-ups; tests for model, anchoring, numbering, serialization, MCP.

specs/annotations/tasks.md

@@ -0,0 +1,24 @@
+# Tasks
+
+Stop on failure. UI tasks go to Sonnet, never Haiku. Agents clean up after themselves.
+
+## Wave 1 (with chat-agents wave 1)
+
+- [x] 1. **Mock-ups** on page `Annotations`. Sonnet.
+
+## Wave 2 (parallel with chat-agents wave 2)
+
+- [x] 2. **Model and storage**: `Annotation`, categories in `design.json`, page field (serde default, byte-identical save for old pages), anchoring relative to target, numbering per frame in reading order, undoable edits through `Editor`. Tests. Sonnet.
+- [x] 3. **MCP tools** `add_annotation`, `get_annotations`, `update_annotation`, `delete_annotations`; `get_design_spec` / `get_jsx` include them as comments; classified in `edits_doc`. Codex Sol (Sonnet if unavailable).
+
+- [x] 3b. **Templates**: model in `design.json` + page override, built-ins, `get_annotation_template` / `set_annotation_template`, `annotate` validates against the page's template. Sonnet (after 3).
+
+## Wave 3 (after wave 2 and the mock-ups)
+
+- [x] 4. **Canvas**: note cards, leader lines (connector router), drag, toggle ⇧A, category filter, hidden in Present and exports unless asked. Sonnet.
+- [x] 5. **Annotate tool, side view (template picker, Customize modal, Annotate with AI), inspector section, popover**. Sonnet.
+
+## Wave 4
+
+- [x] 6. **Review** (Codex Sol, read-only) and fixes (Sonnet); chat `@ Selection` context includes annotations.
+- [x] 7. **Docs** (Haiku): `docs/annotations.md`, README, `docs/agents.md`, ROADMAP, PROGRESS.

specs/chat-agents/spec.md

@@ -0,0 +1,38 @@
+# Chat agents: any agent, found reliably, and a picker that scales
+
+## Why (user report, 2026-10-21)
+
+* Claude Code and Codex show "Not found on PATH" although both are installed in `~/.local/bin`. Root cause: an app opened from the Dock or Finder gets the minimal macOS PATH (`/usr/bin:/bin:/usr/sbin:/sbin`); `launchctl getenv PATH` is empty. The user's login shell also finds `pi`, `opencode`, `gemini`, `goose` (`/opt/homebrew/bin`) and `cursor-agent` (`~/.local/bin`).
+* The setup state offers `npm install -g …` as the way to install agents. Claude Code and Codex have their own installers; PaperCraft should not tell people how to install agents at all.
+* Only two agents are supported; the user wants every popular agent, generically.
+* The agent picker popup has no shadow and overlaps the setup card; the Settings → Chat page has broken spacing, padding and height (fields not aligned, the page bottom cut off).
+* Sending a message did nothing: no `chats/` folder was created, no error was shown.
+* The composer allows one `@ Selection` chip; the user wants several.
+
+## Scope
+
+1. **Find agents like the shell does.** At startup (native, off the UI thread), resolve the login shell PATH once: `$SHELL -l -c 'printf %s "$PATH"'` with a 3 s timeout; merge with the process PATH and common folders (`~/.local/bin`, `/opt/homebrew/bin`, `/usr/local/bin`, `~/.cargo/bin`, `~/.bun/bin`, `~/.npm-global/bin`, Volta and nvm current bins). Use it for agent detection and for spawning agents (set `PATH` on the child). A path override in Settings still wins.
+2. **ACP backend** (`src/chat/acp.rs`): a generic Agent Client Protocol client (JSON-RPC 2.0 over the agent's stdio): `initialize`, `session/new` with `cwd` = project folder and `mcpServers` = PaperCraft's server (HTTP entry when the agent advertises `mcpCapabilities.http`, else a stdio entry running the bundled `papercraft-mcp` bridge with `PAPERCRAFT_MCP_PORT`), `session/prompt` with text (and image) content blocks, streaming `session/update` notifications mapped to our `Event`s (agent message chunks → Text, thought chunks → Thinking, tool_call / tool_call_update → ToolStart/ToolEnd, plan → Plan items), `session/cancel` for Stop, `session/request_permission` answered from our approval flow (papercraft tools allowed; destructive ones and any non-papercraft tool ask the user), `session/load` to resume when supported. The PaperCraft skill text goes in as the first prompt's leading context (ACP has no system prompt field).
+3. **Agent registry, as data** (`src/chat/agents.rs`): `{ id, name, detect: binary name, launch: command + args, kind: Native | Npx, docs_url }`. Native entries are found on the resolved PATH: Gemini CLI (`gemini --acp`), Cursor (`cursor-agent acp`), OpenCode (`opencode acp`), Goose (`goose acp`), Kimi (`kimi acp`), Devin (`devin acp`), Junie (`junie --acp=true`), Mistral Vibe (`vibe-acp`), Hermes (`hermes acp`), OpenHands (`openhands acp`), Stakpak (`stakpak acp`), VT Code (`vtcode acp`), Poolside (`pool acp`), Droid (`droid exec --output-format acp-daemon`). `npx`-launched entries (Qwen Code, Kilo, Cline, Auggie, Copilot CLI, Qoder, Codebuddy, Pi via `pi-acp`, Claude/Codex via their ACP adapters) are listed as available but off until the user enables them once, because enabling runs code fetched by `npx`. Claude Code and Codex keep their direct adapters (already proven) and are the default for them. Exact flags are verified against each installed agent's `--help` where installed; the rest keep the documented command and fail with the agent's own error.
+4. **Custom agent**: name + command + args (ACP over stdio). No provider lists beyond the registry.
+5. **No install instructions.** A missing agent shows "Not found" plus "Set path…" and a link to the agent's own docs (`docs_url`), nothing else.
+6. **Agent picker becomes a panel**, not a dropdown: opened from the agent chip, sized like the chat column (overlay with shadow and a backdrop), sections Detected, Enable (npx), API endpoints, Custom, a search field, and per row: icon, name, status chip (Ready, Needs enable, Not found, Signed out if detectable), version, path. Clicking a row selects it for this chat.
+7. **Settings → Chat** redesign with the same list (one row per agent with a details disclosure: path override, args override, enable), API endpoints as compact rows with an edit disclosure, consistent 16/24 px padding, labels aligned on one grid, the page scrolls instead of being cut off.
+8. **Setup state** in the chat column: one card "Connect an agent" listing detected agents to pick from (usually already some), plus "Add an API endpoint" and "Add custom agent"; no install commands.
+9. **Send bug**: reproduce, find the root cause, fix, and add a regression test; any start failure must show an error row in the thread (never a silent no-op) and keep the typed text.
+10. **Multiple selections**: each `@` adds a chip (selection at that moment, labelled by layer names); chips stack, can be removed one by one; all are sent as context lines with node ids.
+
+## Not in scope
+
+* Installing agents.
+* ACP terminal/file-system client methods beyond what PaperCraft needs (answer `fs/*` and `terminal/*` requests with "not supported" unless the agent requires them to work; most agents use their own file tools).
+
+## Acceptance
+
+* With the app launched from the Dock: Claude Code, Codex, Pi, OpenCode, Gemini, Goose and Cursor are detected on this machine.
+* A turn with Gemini CLI, OpenCode, Goose and Cursor (each installed here) creates an artboard through PaperCraft's tools via ACP; a Pi turn works after enabling `pi-acp`.
+* Stop cancels an ACP turn within 3 s.
+* Picker panel, Settings → Chat and the setup state match new mock-ups at 1440p and 1280x800; nothing clipped; shadows on overlays.
+* Sending a message always either starts a turn or shows an error row.
+* Three `@` chips can be added and removed independently.
+* Tests: PATH merge, registry detection with a fake PATH, ACP JSON-RPC handshake and update mapping against a fake agent process (recorded fixture), permission flow, cancel.

specs/chat-agents/tasks.md

@@ -0,0 +1,23 @@
+# Tasks
+
+Stop on failure. UI tasks go to Sonnet (or Codex Sol), never Haiku. Each agent cleans its temp homes, logs and drivers when done; one shared `target/`.
+
+## Wave 1 (parallel)
+
+- [x] 1. **Mock-ups** on page `Chat agents` (agent panel, setup, Settings / Chat, multiple selections). Sonnet.
+- [x] 2. **PATH resolution, no install hints, send bug, multiple `@` chips** (spec items 1, 5, 9, 10). Sonnet.
+
+## Wave 2 (after wave 1)
+
+- [x] 3. **ACP backend** `src/chat/acp.rs`: JSON-RPC client, session/new with PaperCraft's MCP server (HTTP or the `papercraft-mcp` stdio bridge), prompt, update mapping, permission flow, cancel, load. Fake-agent fixture tests; real turns with Gemini CLI, OpenCode, Goose and Cursor (installed here). Sonnet.
+- [x] 4. **Agent registry and custom agents** `src/chat/agents.rs`: data table, detection on the resolved PATH, npx enable state, custom agent entries in ChatPrefs, BackendKind/Backend for ACP agents. Codex Sol (Sonnet if unavailable).
+
+## Wave 3 (after wave 2 and the mock-ups)
+
+- [x] 5. **Agent panel and setup state** in the chat column (spec items 6, 8). Sonnet.
+- [x] 6. **Settings → Chat redesign** (spec item 7). Sonnet.
+
+## Wave 4
+
+- [x] 7. **Review** (Codex Sol, read-only), fixes (Sonnet).
+- [x] 8. **Docs** `docs/chat.md` (agents, ACP, custom), README. Haiku.

specs/chat-canvas/gaps-found.md

@@ -0,0 +1,24 @@
+# Gaps found while driving PaperCraft on the Interview project (2026-10-21..23)
+
+Observed by the orchestrating agent and its subagents through MCP / papercraft-cli on the user's Mac.
+
+## MCP / agent side
+1. `get_screenshot` never draws annotations, so no agent can verify note placement; every check needed a macOS screen capture of the live window.
+2. No tool to switch the page the user is viewing or to set the viewport (zoom / pan / fit a rect); `select_nodes {zoom:true}` only works on the active page and caps zoom at 1.0.
+3. Agents can't see what the user sees (viewport rect, zoom, selection on screen) except `get_selection`.
+4. `~/.local/bin/papercraft-cli` was stale (schema/list missing new tools) while `call` reached the app — CLI and app versions drift with no warning.
+5. Applying a code change to the running app needs quit + reinstall + reopen (no dev reload / no "restart and reopen project" path).
+6. Long batched edits (thousands of bindings, renames) run as many separate calls; no progress shown to the user in chat mode for CLI-driven agents beyond the "Working…" pill.
+7. Cleanup agents made destructive mistakes (deleted button copies, lost icons on LG buttons) with no per-agent review/diff of what changed; undo is global, not per agent run.
+8. Tidy/placement was blind to zoom-dependent card size (fixed), to screens inside sections (fixed), and lines detoured inside screens (fixed).
+9. `get_tree_summary` depth semantics surprised scripts (rows under "Screens" needed depth 3).
+
+## App / 0.6.2 follow-ups already noted
+- Run card stale after undo; lingering "Working" pill; Stop reported as error.
+- No-GPU panic should be a readable error; Wayland clipboard without DISPLAY.
+- Email truncation (no text-overflow ellipsis).
+- 13.7 s open once after install (first run, not reproduced).
+10. `find_nodes` text search returned nothing for strings that exist on the page (Candidate redesign agent, 2026-10-23) — reproduce and fix.
+11. `get_tree_summary` truncates long text to ~40 chars with no option for full text, so copy review across many screens is impossible without per-node reads.
+12. `annotate {tidy:true}` with no notes returns `[]`, which agents read as failure (fix in A1: `{tidied: n}`).
+13. When the app stops mid-run the CLI silently falls back to offline writes; several agents then write the same project files concurrently with no lock (lost-update risk). Also a CLI call during app start-up reports "app not running" and goes offline.

specs/chat-canvas/research.md

@@ -0,0 +1,82 @@
+# Chat ↔ canvas: research
+
+How canvas + AI products let a user act through chat and point at the canvas, and what PaperCraft already has. Researched 2026-10-23. Secondary sources (forums, reviews) are marked; treat their details as approximate.
+
+## 1. What PaperCraft does today (code)
+
+* **One path to the canvas: tool calls.** Every backend (Claude Code, Codex, ACP, API endpoints) edits only through the 69 MCP tools. CLI/ACP agents reach them over HTTP MCP (`src/mcp/server.rs`); API endpoints call `crate::mcp::exec` on the UI thread (`src/chat/tools.rs`, `ToolDispatch`). The user can't trigger anything in-process from chat except sending text.
+* **Context the agent gets:** `skills/papercraft/SKILL.md` (14 KB) plus the project name and folder (`ChatState::system_prompt`, `src/ui/chat.rs:602`). There is no design-system digest, no viewport and no screenshot of what the user sees.
+* **Pointing:** the `@ Selection` chip (⌘L), `add_selection` in `src/ui/chat.rs`. It sends `[Selection] <label> (<node ids>)` plus the annotations on those layers (`selection_notes`). It has no image, no region and no point. Annotations' "Annotate with AI" (`src/ui/annotations.rs:557`) is the one canvas-to-chat hand-off: it drafts a prompt, adds a chip and switches to Chat mode, but does not send.
+* **Showing agent work:** the activity pill (`agent_pill`, `AGENT_PILL_FRAME_MS`, `agent_verb` in `src/ui/mod.rs:310-360`) shows a verb only, with no location. Canvas follow (`follow`/`pan_to`, `src/ui/chat.rs:1056`) pans to the node in the call args, on the active page only. The review overlay (`review_overlay`, `src/ui/mod.rs:1929`) draws added/changed/removed outlines, but only after the run ends and is opened for review (docs/chat.md, Limits).
+* **Runs:** `src/runs.rs` has one global `Tracker`. A run is closed by a 30 s idle, `finish_working_on_nodes` or the end of a chat turn. Each run has auto-versions per page, plus `revert`, `revert_all` (checkpoint "Restore to here") and `revert_change` (one change). The chat has a Checkpoint row and a Run card (Review/Revert/Accept). MCP calls carry no client identity (`call_tool` ignores `RequestContext`), so parallel subagents land in one run (gap 7).
+* **Screenshots:** `get_screenshot` renders nodes with `io::png::render`. Annotations are an egui overlay (`src/ui/annotations_canvas.rs`), so they never appear (gap 1). `src/model/annotation_layout.rs` (new) already gives pure card sizes, placement and leader routing in world units.
+* **Camera tools:** `select_nodes {zoom}` works only on the active page and fits with a 1.0 zoom cap (`src/mcp/exec.rs:5266`). There is no page switch and no viewport read.
+* **Design-system data that already exists:** `get_variables` (vars, tokens, modes; css/json/tailwind), `get_components` (page mains with use counts plus `ed.library_docs` mains), `get_annotation_template`, `lint_design {fix}` with `lint::autofix` (rules: text-size, type-scale, spacing-grid, touch-target, layer-name, hard-coded-color), `get_design_spec`, `diff_frames`, `swap_instance`, `bind_variable`. There are no component variants in the model.
+* **Rendering stack:** custom `egui_winit` + `egui_wgpu` (`src/app.rs`, `src/render/frame.rs`), not eframe, so `ViewportCommand::Screenshot` is not handled. `render/gpu.rs` already has `start_readback` (texture to buffer).
+
+## 2. Products
+
+### tldraw: agent starter kit, Make Real, fairies
+* Agent context: the user message, the **selection, the viewport**, a screenshot of the agent's view, a simplified shape list in view, **off-screen shapes summarised as clusters (bounds + count)**, recent user actions, and lint warnings. The user can add shapes, **a point or an area** as extra context. ([starter kit](https://tldraw.dev/starter-kits/agent), [AI docs](https://tldraw.dev/docs/ai))
+* The agent **moves its own viewport** to look at parts of the canvas. Actions stream in and apply one at a time. Each action shows in chat with an icon and a one-line description. **Modes** (working, critique) limit the tool set per task. ([v4.4 release](https://tldraw.dev/releases/v4.4.0))
+* Make Real: select shapes, click one button, and the selection's screenshot goes to the model. You iterate by drawing on top of the result. ([talk summary](https://gitnation.com/contents/make-real-tldraws-accidental-ai-play))
+* Fairies (Dec 2025 experiment): agents shown as small characters on the canvas. **Where an agent is shows what it is working on**, and several run concurrently on shared space. ([talk](https://gitnation.com/contents/agents-on-the-canvas-with-tldraw), [tldraw offline](https://tldraw.dev/blog/tldraw-offline))
+* Works: viewport and region as first-class context, a cheap off-screen summary, visible presence.
+
+### Figma: AI, Make, MCP
+* Quick actions (⌘K): Rename layers (only default-named layers), Make prototype, Make designs. These are **one-click scoped actions on a selection**, with no chat. ([quick actions](https://help.figma.com/hc/en-us/articles/23570416033943-Use-quick-actions), [Figma AI](https://www.figma.com/blog/introducing-figma-ai/))
+* First Draft can't use your own design system; it uses Figma-built libraries. ([help](https://help.figma.com/hc/en-us/articles/23955143044247))
+* **Make kits**: components and styles bundled with **usage guidelines**, so the model learns *how* to use the system, not just what is in it. Attachments (PDF, CSV/JSON, SVG) ground prompts in real data. ([blog](https://www.figma.com/blog/introducing-make-kits-and-make-attachments/))
+* Best practices for AI to read a DS: semantic slash names, larger reusable blocks (cards, headers) over atoms, auto layout with sizing intent, full variant and property sets, variables for every color, space, radius and type, and a usage-examples page. ([help](https://help.figma.com/hc/en-us/articles/38978644498199-Best-practices-to-help-Figma-AI-understand-your-design-system))
+* Failure mode reported by users: Make "only recognises a random subset of components" and "created new ones that looked like my components". (forum, 2025: [1](https://forum.figma.com/suggest-a-feature-11/figma-make-not-using-my-library-components-42235/index1.html), [2](https://forum.figma.com/share-your-feedback-26/figma-make-library-support-42423)) The lesson: putting the library "in context" is not enough. Reuse must be checked.
+* Dev Mode MCP: `get_variable_defs` returns the variables used **in the selection**, Code Connect maps nodes to code components, and `create_design_system_rules` writes a rules file for agents. Selection-based prompting works only with the desktop server. ([guide](https://github.com/figma/mcp-server-guide))
+* Works: selection-scoped context, DS rules as a compact digest, one-click actions.
+
+### Cursor Design Mode (browser)
+* Click an element to target it. **Shift+drag selects an area.** You can draw circles or boxes over a **frozen frame** of the viewport, and voice is accepted. ⌘L adds to chat, ⌥+click adds without the comment step. The agent gets element identity (xpath, component, computed styles) **plus a screenshot**. You can queue several edits while agents run. ([docs](https://cursor.com/docs/agent/design-mode))
+* Agent checkpoints: a timeline in the chat, preview at a point, restore. Per-file Keep/Undo appears in the review panel. Reject bugs that deleted whole files show the risk of coarse undo. ([forum](https://forum.cursor.com/t/per-change-keep-undo-buttons-missing-after-agent-edits-only-undo-all-available/158983), [forum](https://forum.cursor.com/t/cursor-deletes-entire-file-when-changes-are-rejected/151760))
+* Pain points (forum): a new pick overwrote the previous one, and the chip always lands at the end of the prompt. ([1](https://forum.cursor.com/t/new-visual-editor-prevents-adding-multiple-elements-in-chat-box/145962), [2](https://forum.cursor.com/t/insert-browser-tab-element-selection-at-cursor-in-composer-not-always-at-end/167268))
+* Works: identity + pixels together, area selection, multiple chips.
+
+### v0 design mode / Lovable visual edits
+* v0: hover highlights and click selects. The **selected element shows as a badge in the prompt**, and the prompt is scoped to it. Direct panel edits are free; prompts cost tokens. A later version added annotations on elements **sent as one batched prompt**. ([v0 docs](https://v0.app/docs/design-mode), [changelog summary](https://www.createwith.com/tool/v0/updates/v0-launches-design-mode-with-visual-editing-and-layer-management))
+* Lovable: select one or more elements (⌘-click for multi), then describe the change. The selection is the context. ([docs](https://docs.lovable.dev/features/preview-toolbar))
+* Works: two lanes, deterministic edits for free and prompts for intent. Comments become a batch prompt.
+
+### Google Stitch
+* **Annotate**: draw a box around an element, type in a bubble, and only that element changes. You can also circle or write notes on the screen. **Variations**: select a screen and generate N (default 3) side by side, with a creativity range. "Create variant" keeps the original. ([review](https://designproject.io/blog/google-stitch-review/), [guide](https://tutorialsdojo.com/what-google-stitch-actually-does-and-where-it-falls-short/), secondary)
+* Works: on-canvas prompt bubble, variations as siblings on the canvas, never overwriting.
+
+### Canva Magic Edit
+* Brush the area, then prompt one change per pass. Canva advises against giving size in words; the brushed area carries it. ([help](https://www.canva.com/help/using-magic-edit/))
+* Works: the region carries the spatial intent, so the prompt stays short.
+
+### Miro Sidekicks, Framer, Excalidraw, Penpot
+* Miro: select clusters, then ask a task-specialised "Sidekick" (persona/mode) to transform them, for example into sticky notes. It has an MCP server for outside agents. ([TrustRadius](https://tr-web.dev.trustradius.com/products/miro/details), [MCP](https://miro.com/pl/ai/mcp/), secondary)
+* Framer Wireframer: prompt a layout, then keep iterating in chat or jump to the canvas. Workshop builds components that match the site's style. ([update](https://www.framer.com/updates/wireframer))
+* Excalidraw: AI is in the paid tier (text-to-diagram, wireframe-to-code). Community MCP servers give agents a live canvas. ([listing](https://www.stork.ai/de/excalidraw), [mcp_excalidraw](https://docsearch.algolia.com/mcp/docs/repo/yctimlin/mcp_excalidraw))
+* Penpot: experimenting with MCP for agents that read and edit design data. ([summary](https://mcpmarket.com/es/news/9fad1b76-0f9d-4411-b557-348e8f85888c), secondary)
+
+## 3. Patterns, ranked by value for PaperCraft
+
+| Pattern | Seen in | PaperCraft today | Gap |
+|---|---|---|---|
+| Selection as scoped context, with pixels | Cursor, v0, Lovable, Figma MCP | `@ Selection` ids + notes | no image, no region |
+| Area/lasso or point as context | Cursor, tldraw, Stitch, Canva | none | missing |
+| One-click scoped actions on selection | Figma ⌘K, Lovable panel | Annotate menu only | add Optimize, Bind tokens, Rename, Fix lint |
+| Deterministic lane vs model lane | v0, Lovable | `lint_design {fix}` exists, not in UI | expose without the LLM |
+| DS digest in context + reuse check | Figma Make kits, `create_design_system_rules` | only the skill text | add a digest and a reuse lint |
+| Agent viewport + presence | tldraw agent, fairies | pan follow + a verb pill | show where; let the agent read and set the view |
+| Live change outlines during a run | tldraw (streamed), Cursor inline diff | after the run only | draw live |
+| Per-run timeline + revert one | Cursor checkpoints | yes (runs.rs) | per-client runs; fix stale card |
+| Variations side by side | Stitch | none | duplicate + keep one |
+| Comments batched into one prompt | v0 annotations | feedback + annotations exist | "Send open notes to chat" |
+| Voice | Cursor, Stitch | none | P2 |
+
+## 4. Conclusions
+
+1. The best tools send **identity + pixels + region**. PaperCraft sends identity only. Adding a screenshot crop to chips and an area pick is the biggest lever for "point at the canvas".
+2. Split actions into **two lanes**. Deterministic actions (lint autofix, bind hard-coded colors to the nearest variable, swap copies for instances, rename default layers) run in-process through `runs::Tracker::call`, so they land as a reviewable run with no model. Intent actions (Optimize layout, Variations, Make consistent) send a scoped prompt from a template.
+3. A DS in context is necessary but not enough (the Figma Make complaints). Pair the digest with a **reuse check**: lint "looks like component X but isn't an instance".
+4. Presence beats a pill. Outline the nodes the current run touched, live, and say where the agent is (page, frame).
+5. Ghost previews in egui are costly (a second document render path). The cheaper equivalent already exists: apply live, show the run outlines, Accept/Revert per run or per change. Variations give "preview" by construction: siblings on the canvas, keep one.

specs/chat-canvas/spec.md

@@ -0,0 +1,124 @@
+# Chat ↔ canvas: spec
+
+What and why for (a) the gaps in `gaps-found.md`, (b) a chat that acts and points at the canvas, and (c) chat grounded in the project's design system. Research and the current-state map are in `research.md`. Cost: **S** < 1 day, **M** 1–3 days, **L** > 3 days. Everything here uses egui primitives already in the tree (Area overlays, painter, popups) and no new crates.
+
+## Principles
+
+* **Two lanes.** Deterministic actions run in-process with no model. Intent actions send a scoped prompt. Both go through `runs::Tracker::call`, so each lands as one reviewable, revertable run.
+* **Identity + pixels + region.** What the user points at goes to the agent as node ids, a crop image, and a world rect.
+* **Reuse before draw.** The agent sees the design system every turn, and a lint flags work that should have been an instance or a token.
+* **Never steal the camera.** Agent camera moves obey the existing `SELF_FOLLOW_QUIET` rule (`src/ui/chat.rs`).
+
+## P0
+
+### P0.1 Annotations in screenshots (gap 1) — S/M
+`get_screenshot {annotations?: bool}`. When true, the image includes note cards, badges and leader lines for notes whose target is inside the captured roots.
+* How: clone the page doc, add each card as temporary frame + text nodes at `annotation_layout` placement (`card_size`, the placement and routing functions), add leader lines as path nodes, then call the existing `io::png::render`. Captured bounds grow to include the cards. No second renderer.
+* Accept: a test renders a frame with 2 notes, and the PNG bounds include both cards. With `false` or omitted, the output is byte-identical to today. `docs/annotations.md` and `skills/papercraft/references/tools.md` are updated.
+
+### P0.2 Viewport: read, set, switch page (gaps 2, 3) — S
+* `get_basic_info` and `get_selection` gain `viewport: {pageId, rect:[x,y,w,h] world, zoom}` from `ed.camera` and `ed.viewport`.
+* New tool `set_viewport {pageId?, nodeIds? | rect?, zoom?, padding?}`. It switches the active page when `pageId` differs (as the page list does) and fits the nodes or rect. `zoom` sets an explicit level within the canvas's min/max, instead of the 1.0 cap in `select_nodes`. It is refused with `"user is moving the canvas"` when the camera moved within `SELF_FOLLOW_QUIET`, unless `force:true`.
+* `select_nodes` keeps its behaviour. Its error on another page points to `set_viewport`.
+* Accept: tests for the page switch, fit, explicit zoom and the quiet refusal. The tool count in `docs/agents.md` is updated.
+
+### P0.3 Runs per agent client + per-run review (gap 7, run-card follow-ups) — M
+* `PaperServer::call_tool` reads `context.peer.peer_info()` (rmcp `clientInfo.name`/`title`) and the session id, and passes `client` on `Call`. Chat turns pass their chat id. `Tracker` keeps one open run per client key, instead of one global `open`. `Run` gains `client: String` (serde default `""`).
+* Agent view and chat: each run shows its client ("Claude Code · subagent 2"). Revert of one run is already in `runs::revert`; refuse it when a later run touched the same nodes, and offer "Restore to here" (`revert_all`).
+* Fix the run card going stale after undo (refresh `ReviewDiff` on `history.generation()` change). The "Working" pill clears on turn end. Stop shows as "Stopped", not an error.
+* Accept: two MCP sessions editing at once give two runs with the right counts. Reverting one leaves the other's nodes. The `tests/mcp.rs` and `runs.rs` tests pass.
+
+### P0.4 Design-system context in every chat turn (user request) — M
+Today `system_prompt()` (`src/ui/chat.rs:602`) is the skill text plus the project name. Add a **DS digest**, built in-process from data the tools already expose:
+* **Variables/tokens**: collection → group → names, with modes. Values only for colors, listed as `name=#hex/#hex(dark)`. Tokens (space, radius, type) as scales. Source: same as `get_variables`.
+* **Components**: page mains and enabled-library mains (`get_components`, `ed.library_docs`), as `componentKey · name · w×h · uses`. Sorted by uses, then name.
+* **Libraries**: enabled libraries with their update status (`libraries::status`).
+* **Annotation template**: the active template's enabled types (`get_annotation_template`).
+* **Rules line**: "Reuse: use `create_instance`/`swap_instance` for anything matching a component; colors must be `var(--…)`; spacing on tokens."
+
+Budget and caching:
+* Cap the digest at ~1,500 tokens (≈6 KB, a named constant). When over the cap, keep the top components by use, collapse variable groups to counts, and append "call get_components / get_variables for the rest".
+* Cache it on `ChatState`, keyed by a hash of (variables, tokens, component keys + names, enabled library versions, template id). Rebuild only on a key change.
+* API backend: the digest goes in the system prompt, in its own cache-breakpoint block after the skill (Claude host). CLI/ACP backends: they get the system prompt once per session, so when the hash changes between turns, prepend `[Design system changed] <digest>` to the next user message.
+* Settings → Chat gets a toggle, "Send design-system summary" (default on), because it costs tokens.
+
+**Reuse check** (the Figma Make failure mode):
+* New lint rule `not-an-instance`: a frame whose structure matches a component main (same child kinds/count, size within 4 px, same text roles) but that is not an instance. Its fix is "swap_instance with componentKey X". Computed with the same structure walk `diff_frames` uses.
+* `lint_design {fix:true}` does **not** auto-swap; swapping is a user action (P0.5).
+
+Accept: a unit test of the digest for a fixture project, checking it is under the cap with top-N kept. A test that the hash is unchanged when unrelated nodes move. A lint test where a hand-made copy of a button is flagged and an instance is not. The context meter shows the added tokens.
+
+### P0.5 Scoped actions on a selection: "Optimize" menu (user request) — M
+An **Actions** menu on the selection. It opens from the canvas right-click menu (extend `canvas_menu` in `src/ui/annotations.rs:438`, which today has only "Annotate"), from an "Actions" chip next to `@ Selection` in the composer, and from the command palette (`src/ui/palette.rs`). It works in Design and Chat modes.
+
+Deterministic lane (in-process, no model, one run each, instant):
+| Action | Runs | Exists |
+|---|---|---|
+| Fix lint here | `lint_design {nodeId, fix:true}` | yes |
+| Bind to tokens | for each hard-coded color, space or radius in the subtree, `bind_variable` to the nearest variable/token within a tolerance; the rest are listed in the result | bind exists; nearest-match is new (S) |
+| Use components | for each `not-an-instance` hit, `swap_instance` | needs P0.4 lint |
+| Rename layers | `rename_nodes` for default names (`Frame 12`, `Text`) from text content or role | new heuristic (S) |
+| Annotate | opens the existing popover | yes |
+
+Intent lane (sends a prompt to chat with a selection chip and crop; switches to Chat mode like "Annotate with AI" does, but **sends**, because the action is explicit):
+* **Optimize**: "Tighten layout and spacing to the token scale, fix contrast to AA, keep the content. Prefer existing components."
+* **Make consistent with DS**: "Replace hand-made parts with instances of the listed components, bind every color and space to variables."
+* **Annotate with AI** (existing), **Write states** (empty, loading, error), **Variations ×3** (P1.3).
+Prompts live in one const table in `src/ui/chat_msgs.rs`, next to `SUGGESTIONS`.
+
+Review: every action, in either lane, ends as a run, so the existing review overlay, Run card and Revert/Accept apply. A deterministic action shows a toast "Bound 14 colors · Review · Revert".
+
+Accept: each deterministic action has a test on a fixture (counts, one undo step, one run). The intent actions put the right prompt and chips into a sent message (pattern: `ai_prompt_and_chat_hand_off_do_not_send`).
+
+### P0.6 CLI/app version check (gap 4) — S
+* `papercraft-cli call`/`shell` compares its own `CARGO_PKG_VERSION` with the app's `serverInfo.version` (`src/mcp/server.rs:39`). On mismatch it prints once to stderr: `papercraft-cli 0.6.1 ≠ app 0.6.2; reinstall the CLI`.
+* `list`/`schema` use the app's `tools/list` when the app is running and fall back to the built-in list offline, so new tools show even from a stale CLI.
+* Accept: a test with a fake server version prints the warning. `list` against a running app includes a tool absent from the built-in list.
+
+## P1
+
+### P1.1 Point at the canvas from chat: pick and area — M
+* The composer gets **Pick** (crosshair; click a layer or Shift+drag an area; Esc cancels) next to `@ Selection`. It does not change the user's selection.
+* `SelChip` (`src/ui/chat.rs:201`) gains `rect: Option<[f64;4]>` (world) and `image: Option<path>`. An area chip is sent as `[Area] page P, rect x,y,w,h, contains: <ids…>` plus the crop image. Layer chips also attach a crop.
+* Crop: `get_screenshot`'s render of the intersecting roots, clipped to the rect (S; add a `clip` rect to `io::png::render`), with annotations per P0.1. Images go through the existing attachment path (image blocks for API; `.papercraft-chat-tmp/` paths for CLI/ACP).
+* ⌥-click on the canvas in Chat mode adds a layer chip directly (Cursor pattern). Chips are inserted at the composer cursor.
+* Accept: an area chip lists the contained node ids and attaches a PNG. Several picks give several chips.
+
+### P1.2 Agent presence and live outlines — M
+* While a run is open, outline the nodes from the last N calls' `follow_node` keys (`chat_msgs::follow_node`) on the canvas in the accent color, fading over 3 s. Label the newest outline with the verb (`agent_verb`) and the client name: an **agent cursor**. Draw it as one `egui::Area` like `review_overlay`, on any page; a chip "Agent on page X →" shows when the agent works on another page.
+* This works for MCP clients too (not only chat), using the per-client run from P0.3.
+* After the turn: today's review overlay. During the turn: cheap touched-outlines only (no full diff per call).
+* Accept: a UI test shows the overlay present while a run is open and gone 3 s after the last call. There is no repaint loop when idle (respect `AGENT_PILL_FRAME_MS`).
+
+### P1.3 Variations side by side — S/M
+* Action "Variations ×3" duplicates the frame 3 times to the right (`duplicate_nodes`, offset by width + 80), names them `<Frame> · Variant A/B/C`, and sends the agent one prompt per variant with distinct directions (compact, spacious, bold). One run.
+* Each variant gets an on-canvas tag with **Keep**. Keep deletes the other variants and the original, or replaces the original with the kept one when "replace original" is on. It is one undo step.
+* Accept: Keep leaves exactly one frame with the original's name and position.
+
+### P1.4 On-canvas Ask popover — S
+* ⌘K on a selection (or "Ask…" in the Actions menu) opens a small prompt box anchored above the selection, using the annotation popover's placement. Enter sends to chat with the chip and crop, Chat mode stays as it is, and the reply streams in the chat column. This gives the Stitch/v0 "type on the element" feel without leaving Design mode.
+* Accept: sending from the popover yields the same message as chip + composer.
+
+### P1.5 Progress for long batches (gap 6) — S
+* The run tracker exposes `calls` and `added/changed/removed` live. The pill shows `Styling · 212 calls · 1,340 changes` for any client, and the chat's step group shows the count while running.
+
+### P1.6 Notes to prompt — S
+* "Send open notes to chat" on a frame: open feedback + annotations of type `todo`/`question` become one batched prompt with a chip (the v0 annotations pattern). It reuses `selection_notes`.
+
+## P2
+
+* **True ghost preview** (L): apply agent edits to a shadow `Document` and draw it translucent over the canvas with Accept/Reject before commit. Costly: every tool would have to target the shadow doc and the renderer composites two docs. P1.2 + P1.3 + run revert cover most of the value; revisit only if users ask.
+* **Dev reload** (gap 5, M): `papercraft --reopen` saves, re-execs the new binary and reopens the project and page. It is a menu item in debug builds.
+* **Chat modes** (S): "Critique" (read-only tools only, returns a list of issues as annotations) vs "Build". It filters `tool_specs()` per mode, as tldraw does.
+* **Sketch to UI** (M): draw on a whiteboard page, then "Make real" sends the crop with the DS digest and the prompt to build a frame next to it.
+* **Voice** (M): push-to-talk into the composer through the OS dictation API. It is out of scope until the other items ship.
+* **Docs** (S, gap 9): document `get_tree_summary` depth with an example; consider `depth` counting from the artboard.
+
+## Out of scope
+
+* New model backends, in-app agent install, web build chat.
+* Gap 8 (fixed). The 0.6.2 platform items (GPU panic, Wayland clipboard, email ellipsis) are tracked in their own specs.
+
+## Order
+
+P0.2, P0.1, P0.6 (small, unblock agents), then P0.4 (digest + lint), then P0.5 (actions, needs the lint), then P0.3. After that P1.1 and P1.2, then the rest.

specs/chat-canvas/tasks.md

@@ -0,0 +1,19 @@
+# Tasks
+
+Spec: `spec.md`. Stop on failure. One shared `target/`; never install or quit the user's PaperCraft (design agents are using it); test builds use their own `PAPERCRAFT_HOME` / `PAPERCRAFT_MCP_PORT` and are killed by PID. Each agent runs `cargo test --release` and `cargo clippy --release --all-targets` clean before reporting, and deletes its temp homes.
+
+## Wave A (parallel, disjoint files)
+
+- [x] A1. **MCP for agents**: P0.2 viewport (`get_basic_info`/`get_selection` viewport, `set_viewport`), P0.1 `get_screenshot {annotations}`, P0.6 CLI version check + live `list`/`schema`, `annotate {tidy}` returns `{tidied: n}` when no notes are given, `get_tree_summary` depth doc (P2 docs). Files: `src/mcp/exec.rs`, `src/mcp/tools.rs`, `src/io/png.rs`, `src/bin/*cli*`, docs. Tests.
+- [x] A2. **Design-system context**: P0.4 digest (new module, cached on ChatState, cap const, system prompt block, CLI/ACP "[Design system changed]" prefix, Settings toggle, context meter) + `not-an-instance` lint rule in `src/mcp/lint.rs`. Tests.
+- [x] A3. **Runs per client**: P0.3 (`src/runs.rs`, `src/mcp/server.rs` client name, agent view/run card labels, per-run revert guard, stale card after undo, Working pill clears, Stop not an error). Tests.
+
+## Wave B (after A)
+
+- [x] B1. **Actions menu**: P0.5 both lanes + toast, prompts table in `chat_msgs.rs`.
+- [x] B2. **Pointing**: P1.1 Pick/area chips with crops, ⌥-click, P1.4 ⌘K Ask popover.
+- [x] B3. **Presence**: P1.2 agent cursor + live outlines, P1.5 live counts in pill, P1.6 notes to prompt.
+
+## Wave C
+
+- [ ] C1. Review (read-only) against spec, fixes (not done yet); docs (done) (`docs/chat.md`, `docs/agents.md`, README tool count); install once design agents are done; live check.

specs/chat-tokens/spec.md

@@ -0,0 +1,86 @@
+# Chat: fewer tokens, model and effort, image zoom
+
+## Why
+
+A Claude Code chat turn in the Interview project cost $0.56 (API-equivalent). Most of it was
+a 50k-token prompt prefix written to the cache and re-read on every tool step. Measured with
+`claude -p --model haiku` and the chat's own flags, before anything the user typed:
+
+| Claude Code flags | Prompt tokens |
+|---|---|
+| what the chat sent | 42.7k |
+| `--disable-slash-commands` (no user skills list) | 31.4k |
+| `--setting-sources project,local` (no user CLAUDE.md, hooks, plugins) | 30.6k |
+| both + `--tools Read,TodoWrite,ToolSearch` | 20.6k |
+
+`--tools` without `ToolSearch` makes it worse (43.9k): the CLI then loads every PaperCraft
+tool schema up front. Bash, Edit, Write and the web tools were never usable anyway: in `-p`
+mode only `mcp__papercraft__*` is allowed, everything that needs permission is denied.
+
+The API backend sends all 70 tool schemas (about 77 KB, ~19k tokens) on every request.
+
+## What
+
+1. **Lean Claude Code sessions** (Settings → Chat, on by default): the flags above.
+   Off keeps the user's own CLAUDE.md, hooks, plugins and skills.
+2. **API tool search**: on `api.anthropic.com`, the tools a design turn rarely needs are sent
+   with `defer_loading` behind the BM25 tool-search tool; the core set stays loaded.
+3. **Ask before building**: the chat system prompt tells the agent to ask up to three short
+   questions (with proposed defaults) when a request is too broad to build well, and to
+   avoid screenshots it does not need.
+4. **Model and effort for Claude Code**: the agent panel's Model/Effort block, today only for
+   Claude API endpoints, also shows for Claude Code. Claude Code no longer gets the default
+   API endpoint's model forced on it (a non-Claude default endpoint sent its model id).
+5. **Images: smaller to send, zoom for detail**: attachments are sent at most 1568 px on the
+   long edge (about 1.15 MP, at most ~1.6k image tokens instead of ~4.8k). When an image was
+   scaled down, the original is saved next to the sent one; the `zoom_image` tool returns a
+   region of it at full resolution. It reads only files under the project's chat folder.
+
+## Acceptance
+
+* Claude Code args contain the lean flags when the setting is on, and none of them when off.
+* API request: core tools have no `defer_loading`, the others do, and the tool-search tool is
+  present; non-Claude hosts get the plain full list.
+* A Claude Code chat with a picked model/effort passes `--model`/`--effort`; without a pick it
+  passes neither.
+* A 4000×1000 PNG is sent as 1568×392 and its original is kept; `zoom_image` on the sent path
+  returns a crop of the original; a path outside the chat folder is refused.
+
+## Follow-up: brief, design system, finish (after the binder test)
+
+The binder page (Claude Code, lean) built straight away at 1440x1024 with 44px buttons
+everywhere, no components for the 9 repeated slots, and a 350-word report. The 62.8k context
+was a 20.3k prefix (lean, as measured) plus the agent's own 19.3k tokens of HTML and the tool
+results; each part was written to the cache once. Lean mode also dropped the user's
+`"model": "opus"` from `~/.claude/settings.json`, so the turn ran on Sonnet 5.5.
+
+6. **`ask_user` tool + question card**: one to five questions, two to five options, palette
+   options with swatches; the chat draws clickable chips, "Send answers" and "Use your
+   defaults". The tool only checks the shape and tells the agent to end its turn.
+7. **Chat rules**: a brief before the first design (frame size, density, palette, type, scope,
+   references), variables and components first, sizes from the agreed density, a five-bullet
+   report and next steps through `ask_user`.
+8. **Lint**: `touch-target` asks 24px (WCAG 2.2 AA) inside screens 1024px or wider, 44px on
+   narrower screens and standalone components; the autofix follows.
+9. **Lean keeps the user's model**: with no pick, the chat passes the `model` from the user's
+   Claude Code settings.
+
+## Follow-up: long chats (the Jira chat at 230k)
+
+The Jira chat reached 230k of context; every step re-read it. Claude Code's `total_cost_usd`
+adds up over a resumed session, so turns showed the running total ($6.72 for a $0.47 turn).
+
+10. **Meter**: a 1M window (current Claude models), counts as `1.6M`; amber past 150k.
+11. **Per-turn cost**: `Message::session_cost` keeps the CLI's running total; a resumed turn shows
+    the difference (older saved turns are derived at display time).
+12. **Fresh session past 150k (now 120k)** (CLI backends): not resumed; the new session starts with a recap
+    (the last 8 requests, the last report) and reads the canvas with tools.
+13. **Compact now**: clicking the meter queues the same fresh start for the next message.
+14. **API context editing**: `clear_tool_uses_20250919` past 60k input, keep 4, clear ≥15k, tool
+    inputs too (beta `context-management-2025-06-27`).
+15. **Research sub-agent**: with Web on, Claude Code gets a `researcher` agent on Haiku
+    (`--agents`); it searches and fetches, the main context gets the kept paths only. A research
+    turn on a fresh chat: $0.26 in all, main context 25.8k (the late in-session one cost $1.71).
+    Sub-agent text stays out of the chat; its tool calls (and galleries) show.
+16. **Turn budget** (Settings → Chat, default 10M tokens read): a turn past it is stopped like
+    Stop, with a note.

specs/interview-cleanup/A-report.md

@@ -0,0 +1,44 @@
+# Package A report: Interview Design System
+
+Date 2026-10-23. Project `Interview` (fileName verified). Backup: `/Volumes/Storage/Projects/Interview.backup-20261023-0112`. Screenshots in `A/` (before-*, after-*). Mapping files: `variable-mapping.csv`, `component-mapping.csv` (old name, new name, componentKey, action; 280 rows).
+
+## 1. Variables (collection `Colors`)
+- 161 -> 137 variables (25 merged, +1 `color/brand/primary-strong`). All names are `color/<group>/<name>` (csv `new` column with `color/` prefix). SYS-03: named by role, not value.
+- Rename = create new + repoint every bound fill (`update_styles`) + delete old. The app has no rename tool, and `delete_variables` detaches fills, so repointing was required. 4293+ bound fills repointed on all 3 pages. No dangling `var(--old)` found afterwards.
+- Merges (csv "merge" rows) repointed to the canonical variable then deleted (no leftover aliases needed because nothing binds them).
+- SYS-02 (semantic alias over primitive value): `color/brand/pressed` and `color/text/link` are aliases of `color/brand/primary-strong` (#0958d9); `color/status/info` is an alias of `color/brand/primary`.
+- Contrast: `#1677ff` + white = 4.10:1 FAIL, `#0958d9` + white = 6.16:1 pass (`playbook contrast`), `#0958d9` on `#e6f4ff` 5.50:1 pass (SYS-05). 40 DS fills (Button / Primary x3, Avatar x5, sidebar active items, dots, brand marks, OK button) rebound to `primary-strong`; 16 literal `#1677ff` text nodes now use `text/link` colour. DS contrast errors 128 -> 67 on the original components.
+- Final groups (Colors): avatar 5, bg 6, border 4, brand 6 (primary, primary-strong, pressed, hover, secondary, focus-ring), button 1, camera 2, code 2, control 1, diff 2, hero 2, ide 31, status 9, sidebar 8, syntax 7, syscheck 6, tag 18 (tag/<hue>/{bg,border,text}), text 6, whiteboard 9, auth 12.
+- Colours the page still uses as hard-coded literals (text colours, borders) are not bound; that is B/C work (tools bind fills only).
+
+## 2. Tokens
+- Old `Tokens` collection (32: 26 number, 3 shadow strings, 4 z-index strings) deleted; shadows and z-index dropped (unsupported in PaperCraft, unreferenced). Empty `Tokens` collection remains (no delete-collection tool; remove in the Variables modal).
+- New collection `Spacing & Size` (34 Number variables): `space/{4,8,12,16,24,32,48,64}` (12 added), `radius/{sm4,md8,lg12,xl16,pill999}`, `font/size/{xs12,sm14,md16,lg20,xl24,2xl32,3xl40,4xl48}` (18 dropped), `font/leading/{xs16..2xl40}`, `layout/page-padding(-lg)`, `layout/page-width/{xs480,sm420,md720,lg900,xl1100}` (3 duplicate widths removed). SYS-01/SYS-02.
+- Nothing is bound to them yet (tokens are substituted when written). Use `bind_variable` in later work.
+
+## 3. Components (Design System page: 211 -> 260 mains)
+- Consolidated: Ant Button Primary/Default/Danger (9), Ant Tag (5), Ant Segmented (3), Ant Table (2) dropped after `swap_instance` of their DS cells to the kept set (19 swaps, 19 mains and 19 cells deleted). Ant Text / Text Danger kept as `Button / Text` and `Button / Danger Text` (SM/MD/LG). AntD showcase groups removed, cells moved into `Group / Buttons`.
+- 78 mains renamed to `Group / Variant [/ SM|LG]` (see csv). Button MD height 34 -> 32 (SM 28, LG 40). `Page Header` kept as `Header / Page` (no action slot; not equal to `Header / Admin`, so not merged).
+- Fixed: `Header / Candidate` height 0 -> 134 (Header Card height 100% -> fit-content). 8 `Alert / *` 456/478/522 -> 38 high (message text width 0px -> fit-content).
+- Moved 68 page-local components onto the DS page (17 Admin, 51 Candidate): each moved with `move_nodes toPageId`, instances on the source page re-pointed with `swap_instance` by componentKey, instance layer names updated. Packed at x=1560, y 8623..13765 on the DS page. `Candidate Message / Success` renamed `Banner / Success` (name clash with the Admin one, different design).
+- Skipped (not moved): `T Hdr`, `T Donut`, `T Spark`, `T Stat` (Admin template stubs, 0 uses; equivalents exist in DS: Section Title, Donut, Sparkline, Stat). Delete them in package B.
+- Not done: collapse of combinatorial variants (Segmented Active N, Tabs Bar N, 11 sidebars, content-baked Challenge Option/Timeline, colour-named Stat/Sparkline), missing states (hover/disabled etc.), Input/Select/Textarea LG state matrix (SYS-06/07 follow-ups), `Light` mode rename of `Default` (no tool).
+- Side effect: while moving, Admin's own copies of `Button / Primary Small`, `Button / Primary Large`, `Button / Default` were moved by mistake (the copy filter used renamed names) and then deleted from DS; no Admin instance pointed at them (nested ones live inside other instances). B should re-import or swap to the DS mains below.
+- Admin/Candidate still carry their old copies of the other 211 DS components under old names; B/C: use `component-mapping.csv` and `swap_instance componentKey`.
+
+### Kept component keys (packages B and C)
+`componentKey` for every kept or moved main is in `component-mapping.csv` (column 3). Frequently needed:
+Button / Primary `64e1d4aaf56fe6a4:5980`, Primary / SM `:5982`, Primary / LG `:5984`, Secondary `:5986`, Secondary / SM `:5988`, Secondary / LG `:5990`, Ghost `:5992`, Danger `:5998`, Dark `:6004`, Icon / Bell `:6010`.
+Header / Admin `64e1d4aaf56fe6a4:6480`, Header / Page `:6476`, Header / Candidate `:6489`, Header / Admin / With Action `1c5e5918`.
+Form Field / Default `:6663`, Card / Default `:6054`, Alert / Error `:6728`, Input / Default `:6575`, Sidebar / Admin `:6906`, Sidebar / Candidate `:7370`.
+(Key prefix is the DS page id `64e1d4aaf56fe6a4`.) Moved components have short keys (e.g. `bd7f34e5`), see csv.
+
+## 4. Annotations
+Template `Dev handoff` set on the DS page. 11 notes (dev, a11y, behaviour): one Primary per screen and contrast pair, tag colour by status, Data Table single set, Segmented, Input states, Alert, Tab hit area, Form Field, Header variants, Pill.
+
+## 5. Lint (DS page, byRule)
+Before: contrast 128, hard-coded 11, layer-name 7, spacing 57, touch 16, type-scale 42 (list truncated at 200; counts exact).
+After (whole page): contrast 194, hard-coded 366, layer-name 49, spacing 180, touch 30, type 135. The rise is the 68 moved local components (contrast 127, hard-coded 225, layer-name 42, spacing 123, touch 9, type 93). The original components: contrast 67 (-61), hard-coded 146 (more literals now match a single variable after merges, so lint flags them), layer-name 7, spacing 57, touch 21 (Button MD 32 < 44, desktop density accepted, documented in annotation), type-scale 42. Remaining contrast: `#52c41a`/white 2.3, `#8c8c8c` 3.4 (placeholders), `#bfbfbf` 1.8 (disabled), tag text pairs, IDE greys: design decisions left for follow-up. `lint-ds-before.json` saved in `A/`.
+
+## Leftovers for the user
+Empty `Tokens` collection; mode `Default` not renamed; 15 `Label / *` stubs at x=30000 not deleted; Colors board has no swatch for `brand/primary-strong`; `brand/hover` (#4096ff) with white text fails contrast (hover state unused in DS).

specs/interview-cleanup/A2-report.md

@@ -0,0 +1,35 @@
+# Package A2 report: Design System page cleanup
+
+Date 2026-10-23. Project `Interview` (fileName verified). Page `Design System` only. Evidence in `A2/`: `issues-before.json` (all 959 issues, linted per top-level frame to avoid the 200 cap), `issues-7.json` (final), `after-brand-swatch.png`, `after-components.png`. Backup: `/Volumes/Storage/Projects/Interview.backup-20261023-0112`.
+
+## Lint counts (DS page, byRule, exact)
+| rule | before A2 | after A2 | pre-A baseline |
+|---|---|---|---|
+| contrast (errors) | 194 | **0** | 128 (67 on original components) |
+| hard-coded-color | 366 | **7** | 11 |
+| spacing-grid | 180 | **0** | 57 |
+| touch-target | 30 | 30 (accepted, see below) | 16 |
+| type-scale (info) | 135 | 135 (not in scope) | 42 |
+| layer-name | 49 | 49 (not in scope) | 7 |
+Goal met: contrast 0 (target <= 67).
+
+## What changed
+- Hard-coded colours: 359 nodes bound by exact hex match (352 text colours, 12 fills) via `update_styles` `var(--color-...)`. White text bound to `color/text/inverse`, not `bg/surface` (same hex, text role). Note: `bind_variable text-color` was not recognised by lint, `update_styles color:var()` was; use that.
+- Contrast (SYS-05, A11Y checked with `playbook contrast` maths, all pairs >= 4.5):
+  - white on `#1677ff`/`#4096ff` -> parent fills rebound to `color/brand/primary-strong` (#0958d9, 6.16); 16 fills + 5 matching 1px strokes bound.
+  - white on `#52c41a` (Score Dots) -> `color/status/success-text` (#237804, 5.59); white on `#ff4d4f` -> `color/status/error` (5.57).
+  - `#5a5a5a` on `#1a1a1a` -> `color/ide/text-muted` (4.91); `#888`/`#666` on `#252526|#2a2a2a|#252525|#181818|#161616` -> `color/ide/text-dim` (>= 7.5); `#5a80a8` on `#1c2a40` -> `ide/text-dim`; `#13a8a8` on `#1a3540` -> `syntax/type` (6.33); sidebar `#64676c` -> `sidebar/text-muted` (4.51).
+  - `#8c8c8c`/`#bfbfbf`/`#bbb`/`#878787`/`#848484` on light -> `color/text/secondary` (6.7-7.0). Note placeholders and disabled input text now read as secondary text; disabled look is carried by the grey fill only.
+  - `#52c41a`, `#16a34a` text on light -> `status/success-text`; white on `#faad14`/`#13a8a8` (Score Dots 2, Theory Next) -> `color/text/primary`.
+  - Variable VALUES changed (global, affects Admin and Candidate too, all pairs now pass): `color/tag/gold/text` #ad6800 -> #874d00, `color/tag/green/text` #389e0d -> #237804, `color/tag/cyan/text` #08979c -> #006d75. Bound fills followed automatically.
+  - `Pipeline Row / Dark` was a transparent component meant for a dark card: gave it `color/ide/card` background, stage text `ide/text-muted`, name `text/inverse`.
+- Spacing: all 180 off-grid nodes bound to `space/*` (nearest in-scale value: 1,2,3,5->4; 6,7,9->8; 10,11->12; 14,18->16; 22->24; gap 6->8). On-grid padding was not rebound (would be thousands of bindings with no lint effect). Visual check of buttons, tags, rows, toolbars: nothing broke.
+- Radii: all uniform radii bound to `radius/*`: 289 -> sm (3,4,5), 442 -> md (6,7,8), 75 -> lg (12), 153 -> pill (99,999). Multi-value radii (per-corner) not touched; no radii above 12 other than the pill ones existed.
+- Colors board: added swatch `color-brand-primary-strong` (#0958d9) in the Brand card.
+- `Label / *` stubs: they are plain label frames, not components (`get_components` lists none), and each captions a real component at x=30000. Deleted only the 5 whose component has 0 instances (`Link Field / Copied`, `Challenge Option / Two Sum Off`, `... / URL shortener On`, `... / JavaScript closures Off`, `Timeline / Step 3`). The other 10 kept because their component has instances. Delete them too if you read the instruction as "all 15".
+
+## Skipped / accepted
+- 7 hard-coded: `#000000` text on `Stars / 0..5` (6, no exact variable; `text/primary` is #000000e0, near match not in the audit mapping) and `#bae0ff` on `Funnel Row` (1, no variable).
+- Touch targets (30): `Button *` SM 24-28, MD 32-34, LG 40, `Button / Undo|Clear|Delete` 24, bell 36. Accepted desktop density: playbook A11Y-09 (24 px AA minimum, 44 default on touch only) and PAT-06 (about 27-30 px practical floor for dense content). All >= 24 px. Annotation on the DS page already documents it. Candidate pages are desktop IDE flows.
+- Untouched (out of scope): type-scale 135, layer-name 49, `brand/hover` (#4096ff) with white text (state unused).
+- Side effect to check: `set_variables` (tag text values) made lint report hard-coded text again on untouched nodes (text bindings seemed to be flattened for some nodes on the DS page); rebinding fixed it. Admin/Candidate pages were not inspected for this; if B/C see text colours turn literal, that is the cause.

specs/interview-cleanup/A2/fills-changed.json

@@ -0,0 +1 @@
+{"status-error": ["64e1d4:6351"], "status-success-text": ["64e1d4:6390", "64e1d4:6392", "64e1d4:6394", "64e1d4:6396", "64e1d4:6403", "64e1d4:6405", "64e1d4:6407", "64e1d4:6409", "64e1d4:6411"], "brand-primary-strong": ["64e1d4:10020", "64e1d4:10057", "64e1d4:10473", "64e1d4:10484", "64e1d4:10495", "64e1d4:10506", "64e1d4:10677", "64e1d4:10696", "64e1d4:10720", "64e1d4:10801", "64e1d4:9898", "64e1d4:9925", "64e1d4:9934", "64e1d4:9936", "64e1d4:9943", "64e1d4:9955"]}

specs/interview-cleanup/A2/lib.py

@@ -0,0 +1,5 @@
+import json,subprocess
+def call(tool,**a):
+    r=subprocess.run(['papercraft-cli','call',tool,'-'],input=json.dumps(a),capture_output=True,text=True)
+    try: return json.loads(r.stdout)
+    except Exception: return {'_raw':r.stdout,'_err':r.stderr}

specs/interview-cleanup/A2/rebind.py

@@ -0,0 +1,18 @@
+from lib import *
+import json,re,collections,sys
+P='64e1d4aaf56fe6a4'
+flat=json.load(open('flat.json'));rev={n.replace('/','-'):n for n in flat}
+I=json.load(open(sys.argv[1]))
+hc=[i for i in I if i['rule']=='hard-coded-color' and 'is the value' in i['message']]
+ids=sorted({i['node'] for i in hc});st={}
+for k in range(0,len(ids),40): st.update(call('get_computed_styles',pageId=P,nodeIds=ids[k:k+40]))
+g=collections.defaultdict(list);f=collections.defaultdict(list)
+for i in hc:
+    hx,css=re.match(r'(#\w+) is the value of --(\S+)',i['message']).groups()
+    s=st[i['node']]
+    if s.get('backgroundColor','').lower()==hx and s.get('color','').lower()!=hx: f[css].append(i['node']);continue
+    if hx=='#ffffff': css='color-text-inverse'
+    if hx in('#0958d9','#1677ff'): css='color-text-link'
+    g[css].append(i['node'])
+ups=[{'nodeIds':n,'styles':{'color':f'var(--{c})'}} for c,n in g.items()]+[{'nodeIds':n,'styles':{'backgroundColor':f'var(--{c})'}} for c,n in f.items()]
+print(call('update_styles',pageId=P,updates=ups))

specs/interview-cleanup/A2/relint.py

@@ -0,0 +1,9 @@
+from lib import *
+import json,collections,sys
+ab=call('get_basic_info',pageId='64e1d4aaf56fe6a4')['artboards']
+allis=[]
+for a in ab:
+    r=call('lint_design',pageId='64e1d4aaf56fe6a4',nodeId=a['id'])
+    for i in r.get('issues',[]): i['top']=a['id'];i['topname']=a['name'];allis.append(i)
+json.dump(allis,open(sys.argv[1],'w'))
+print(len(ab),len(allis),collections.Counter(i['rule'] for i in allis))

specs/interview-cleanup/A2/vars.json

@@ -0,0 +1,752 @@
+{
+"color": {
+"auth": {
+"bg": {
+"$type": "color",
+"$value": "#0f1117"
+},
+"card": {
+"$type": "color",
+"$value": "#ffffff0d"
+},
+"card-head": {
+"$type": "color",
+"$value": "#ffffff99"
+},
+"card-line": {
+"$type": "color",
+"$value": "#ffffff14"
+},
+"desc": {
+"$type": "color",
+"$value": "#ffffff99"
+},
+"footer": {
+"$type": "color",
+"$value": "#ffffff8c"
+},
+"icon-bg": {
+"$type": "color",
+"$value": "#ffffff12"
+},
+"icon-fg": {
+"$type": "color",
+"$value": "#ffffff8c"
+},
+"live-bg": {
+"$type": "color",
+"$value": "#16a34a2e"
+},
+"live-fg": {
+"$type": "color",
+"$value": "#4ade80"
+},
+"row": {
+"$type": "color",
+"$value": "#ffffffc7"
+},
+"sub": {
+"$type": "color",
+"$value": "#ffffffa6"
+}
+},
+"avatar": {
+"indigo": {
+"$type": "color",
+"$value": "#2f54eb"
+},
+"orange": {
+"$type": "color",
+"$value": "#fa541c"
+},
+"pink": {
+"$type": "color",
+"$value": "#eb2f96"
+},
+"purple": {
+"$type": "color",
+"$value": "#722ed1"
+},
+"teal": {
+"$type": "color",
+"$value": "#13a8a8"
+}
+},
+"bg": {
+"layout": {
+"$type": "color",
+"$value": "#f5f5f5"
+},
+"overlay": {
+"$type": "color",
+"$value": "#00000073"
+},
+"selected": {
+"$type": "color",
+"$value": "#e6f4ff"
+},
+"subtle": {
+"$type": "color",
+"$value": "#fafafa"
+},
+"surface": {
+"$type": "color",
+"$value": "#ffffff"
+},
+"tooltip": {
+"$type": "color",
+"$value": "#000000d9"
+}
+},
+"border": {
+"default": {
+"$type": "color",
+"$value": "#d9d9d9"
+},
+"hover": {
+"$type": "color",
+"$value": "#bfdbff"
+},
+"strong": {
+"$type": "color",
+"$value": "#8c8c8c"
+},
+"subtle": {
+"$type": "color",
+"$value": "#f0f0f0"
+}
+},
+"brand": {
+"focus-ring": {
+"$type": "color",
+"$value": "#1677ff1a"
+},
+"hover": {
+"$type": "color",
+"$value": "#4096ff"
+},
+"pressed": {
+"$type": "color",
+"$value": "{color.brand.primary-strong}"
+},
+"primary": {
+"$type": "color",
+"$value": "#1677ff"
+},
+"primary-strong": {
+"$type": "color",
+"$value": "#0958d9"
+},
+"secondary": {
+"$type": "color",
+"$value": "#69b1ff"
+}
+},
+"button": {
+"dark": {
+"$type": "color",
+"$value": "#1f1f1f"
+}
+},
+"camera": {
+"label": {
+"$type": "color",
+"$value": "#ffffff99"
+},
+"ring": {
+"$type": "color",
+"$value": "#ffffff40"
+}
+},
+"code": {
+"bg": {
+"$type": "color",
+"$value": "#1e1e1e"
+},
+"fg": {
+"$type": "color",
+"$value": "#d4d4d4"
+}
+},
+"control": {
+"switch-off": {
+"$type": "color",
+"$value": "#00000040"
+}
+},
+"diff": {
+"added": {
+"$type": "color",
+"$value": "#3fb950"
+},
+"removed": {
+"$type": "color",
+"$value": "#f85149"
+}
+},
+"hero": {
+"chip": {
+"$type": "color",
+"$value": "#ffffff33"
+},
+"text": {
+"$type": "color",
+"$value": "#ffffffeb"
+}
+},
+"ide": {
+"blue-bg": {
+"$type": "color",
+"$value": "#1a2a3a"
+},
+"card": {
+"$type": "color",
+"$value": "#161616"
+},
+"card-border": {
+"$type": "color",
+"$value": "#252525"
+},
+"card-label": {
+"$type": "color",
+"$value": "#484848"
+},
+"card-text": {
+"$type": "color",
+"$value": "#dddddd"
+},
+"console": {
+"$type": "color",
+"$value": "#181818"
+},
+"css": {
+"$type": "color",
+"$value": "#c586c0"
+},
+"deep": {
+"$type": "color",
+"$value": "#1a1a1a"
+},
+"err-bg": {
+"$type": "color",
+"$value": "#1a0808"
+},
+"err-border": {
+"$type": "color",
+"$value": "#3a1010"
+},
+"fail-bg": {
+"$type": "color",
+"$value": "#3a1a1a"
+},
+"fail-border": {
+"$type": "color",
+"$value": "#5a2020"
+},
+"faint": {
+"$type": "color",
+"$value": "#555555"
+},
+"gutter": {
+"$type": "color",
+"$value": "#5a5a5a"
+},
+"line": {
+"$type": "color",
+"$value": "#2a2a2a"
+},
+"panel": {
+"$type": "color",
+"$value": "#252526"
+},
+"pass-bg": {
+"$type": "color",
+"$value": "#1a3a1a"
+},
+"pass-border": {
+"$type": "color",
+"$value": "#3fb95035"
+},
+"preview": {
+"$type": "color",
+"$value": "#0d1117"
+},
+"rail": {
+"$type": "color",
+"$value": "#1f1f1f"
+},
+"selected": {
+"$type": "color",
+"$value": "#2a2d2e"
+},
+"sep": {
+"$type": "color",
+"$value": "#333333"
+},
+"submit-bg": {
+"$type": "color",
+"$value": "#1c2a40"
+},
+"submit-fg": {
+"$type": "color",
+"$value": "#5a80a8"
+},
+"teal-bg": {
+"$type": "color",
+"$value": "#1a3540"
+},
+"text": {
+"$type": "color",
+"$value": "#cccccc"
+},
+"text-dim": {
+"$type": "color",
+"$value": "#bbbbbb"
+},
+"text-faint": {
+"$type": "color",
+"$value": "#666666"
+},
+"text-muted": {
+"$type": "color",
+"$value": "#888888"
+},
+"warn": {
+"$type": "color",
+"$value": "#d29922"
+},
+"warn-bg": {
+"$type": "color",
+"$value": "#2a2a14"
+}
+},
+"sidebar": {
+"badge": {
+"$type": "color",
+"$value": "#ffffff40"
+},
+"bg": {
+"$type": "color",
+"$value": "#11161d"
+},
+"border": {
+"$type": "color",
+"$value": "#ffffff12"
+},
+"hover": {
+"$type": "color",
+"$value": "#ffffff0f"
+},
+"tag-bg": {
+"$type": "color",
+"$value": "#1677ff40"
+},
+"text": {
+"$type": "color",
+"$value": "#ffffffa6"
+},
+"text-faint": {
+"$type": "color",
+"$value": "#ffffff59"
+},
+"text-muted": {
+"$type": "color",
+"$value": "#ffffff73"
+}
+},
+"status": {
+"error": {
+"$type": "color",
+"$value": "#cf1322"
+},
+"info": {
+"$type": "color",
+"$value": "{color.brand.primary}"
+},
+"score-poor": {
+"$type": "color",
+"$value": "#ff4d4f"
+},
+"score-warn": {
+"$type": "color",
+"$value": "#faad14"
+},
+"star": {
+"$type": "color",
+"$value": "#faad14"
+},
+"success": {
+"$type": "color",
+"$value": "#52c41a"
+},
+"success-text": {
+"$type": "color",
+"$value": "#237804"
+},
+"warning": {
+"$type": "color",
+"$value": "#d48806"
+},
+"warning-text": {
+"$type": "color",
+"$value": "#874d00"
+}
+},
+"syntax": {
+"comment": {
+"$type": "color",
+"$value": "#6a9955"
+},
+"fn": {
+"$type": "color",
+"$value": "#dcdcaa"
+},
+"gutter": {
+"$type": "color",
+"$value": "#5a5a5a"
+},
+"keyword": {
+"$type": "color",
+"$value": "#569cd6"
+},
+"number": {
+"$type": "color",
+"$value": "#b5cea8"
+},
+"string": {
+"$type": "color",
+"$value": "#ce9178"
+},
+"type": {
+"$type": "color",
+"$value": "#4ec9b0"
+}
+},
+"syscheck": {
+"bg": {
+"$type": "color",
+"$value": "#f0f0f0"
+},
+"camera": {
+"$type": "color",
+"$value": "#111111"
+},
+"fail": {
+"$type": "color",
+"$value": "#dc2626"
+},
+"mic": {
+"$type": "color",
+"$value": "#22c55e"
+},
+"ready": {
+"$type": "color",
+"$value": "#16a34a"
+},
+"ready-bg": {
+"$type": "color",
+"$value": "#16a34a1a"
+}
+},
+"tag": {
+"blue": {
+"bg": {
+"$type": "color",
+"$value": "#e6f4ff"
+},
+"border": {
+"$type": "color",
+"$value": "#91caff"
+},
+"text": {
+"$type": "color",
+"$value": "#0958d9"
+}
+},
+"cyan": {
+"bg": {
+"$type": "color",
+"$value": "#e6fffb"
+},
+"border": {
+"$type": "color",
+"$value": "#87e8de"
+},
+"text": {
+"$type": "color",
+"$value": "#08979c"
+}
+},
+"gold": {
+"bg": {
+"$type": "color",
+"$value": "#fffbe6"
+},
+"border": {
+"$type": "color",
+"$value": "#ffe58f"
+},
+"text": {
+"$type": "color",
+"$value": "#ad6800"
+}
+},
+"green": {
+"bg": {
+"$type": "color",
+"$value": "#f6ffed"
+},
+"border": {
+"$type": "color",
+"$value": "#b7eb8f"
+},
+"text": {
+"$type": "color",
+"$value": "#389e0d"
+}
+},
+"purple": {
+"bg": {
+"$type": "color",
+"$value": "#f9f0ff"
+},
+"border": {
+"$type": "color",
+"$value": "#d3adf7"
+},
+"text": {
+"$type": "color",
+"$value": "#531dab"
+}
+},
+"red": {
+"bg": {
+"$type": "color",
+"$value": "#fff1f0"
+},
+"border": {
+"$type": "color",
+"$value": "#ffccc7"
+},
+"text": {
+"$type": "color",
+"$value": "#cf1322"
+}
+}
+},
+"text": {
+"disabled": {
+"$type": "color",
+"$value": "#00000073"
+},
+"inverse": {
+"$type": "color",
+"$value": "#ffffff"
+},
+"link": {
+"$type": "color",
+"$value": "{color.brand.primary-strong}"
+},
+"placeholder": {
+"$type": "color",
+"$value": "#00000040"
+},
+"primary": {
+"$type": "color",
+"$value": "#000000e0"
+},
+"secondary": {
+"$type": "color",
+"$value": "#000000a6"
+}
+},
+"whiteboard": {
+"amber": {
+"$type": "color",
+"$value": "#d29922"
+},
+"bar": {
+"$type": "color",
+"$value": "#1c1c2e"
+},
+"bg": {
+"$type": "color",
+"$value": "#13131f"
+},
+"blue": {
+"$type": "color",
+"$value": "#6ab0ff"
+},
+"border": {
+"$type": "color",
+"$value": "#252540"
+},
+"orange": {
+"$type": "color",
+"$value": "#ff9f43"
+},
+"purple": {
+"$type": "color",
+"$value": "#c792ea"
+},
+"text": {
+"$type": "color",
+"$value": "#9999bb"
+},
+"white": {
+"$type": "color",
+"$value": "#e0e0e0"
+}
+}
+},
+"font": {
+"leading": {
+"2xl": {
+"$type": "dimension",
+"$value": "40px"
+},
+"lg": {
+"$type": "dimension",
+"$value": "28px"
+},
+"md": {
+"$type": "dimension",
+"$value": "24px"
+},
+"sm": {
+"$type": "dimension",
+"$value": "20px"
+},
+"xl": {
+"$type": "dimension",
+"$value": "32px"
+},
+"xs": {
+"$type": "dimension",
+"$value": "16px"
+}
+},
+"size": {
+"2xl": {
+"$type": "dimension",
+"$value": "32px"
+},
+"3xl": {
+"$type": "dimension",
+"$value": "40px"
+},
+"4xl": {
+"$type": "dimension",
+"$value": "48px"
+},
+"lg": {
+"$type": "dimension",
+"$value": "20px"
+},
+"md": {
+"$type": "dimension",
+"$value": "16px"
+},
+"sm": {
+"$type": "dimension",
+"$value": "14px"
+},
+"xl": {
+"$type": "dimension",
+"$value": "24px"
+},
+"xs": {
+"$type": "dimension",
+"$value": "12px"
+}
+}
+},
+"layout": {
+"page-padding": {
+"$type": "dimension",
+"$value": "24px"
+},
+"page-padding-lg": {
+"$type": "dimension",
+"$value": "48px"
+},
+"page-width": {
+"lg": {
+"$type": "dimension",
+"$value": "900px"
+},
+"md": {
+"$type": "dimension",
+"$value": "720px"
+},
+"sm": {
+"$type": "dimension",
+"$value": "420px"
+},
+"xl": {
+"$type": "dimension",
+"$value": "1100px"
+},
+"xs": {
+"$type": "dimension",
+"$value": "480px"
+}
+}
+},
+"radius": {
+"lg": {
+"$type": "dimension",
+"$value": "12px"
+},
+"md": {
+"$type": "dimension",
+"$value": "8px"
+},
+"pill": {
+"$type": "dimension",
+"$value": "999px"
+},
+"sm": {
+"$type": "dimension",
+"$value": "4px"
+},
+"xl": {
+"$type": "dimension",
+"$value": "16px"
+}
+},
+"space": {
+"12": {
+"$type": "dimension",
+"$value": "12px"
+},
+"16": {
+"$type": "dimension",
+"$value": "16px"
+},
+"24": {
+"$type": "dimension",
+"$value": "24px"
+},
+"32": {
+"$type": "dimension",
+"$value": "32px"
+},
+"4": {
+"$type": "dimension",
+"$value": "4px"
+},
+"48": {
+"$type": "dimension",
+"$value": "48px"
+},
+"64": {
+"$type": "dimension",
+"$value": "64px"
+},
+"8": {
+"$type": "dimension",
+"$value": "8px"
+}
+}
+}

specs/interview-cleanup/A3-report.md

@@ -0,0 +1,32 @@
+# Package A3 report: DS additions, DS defects, remaining swaps, layout fixes
+
+Date 2026-10-23. Project `Interview` (verified). Screenshots in `A3/` (before-*/after-*). Undo is per page (use `pageId`).
+
+## 1. Design System additions (DS page, lint 0 on each)
+- `Card / Container` (64e1d4:10850): empty shell, padding 24, gap 16, radius 12, border, fill `color/bg/surface`. Title slot not added (children go straight in).
+- `Button / Primary / LG / Icon` (10830), `Button / Secondary / LG / Icon` (10834): 40 px, leading plus icon (phosphor:plus, the set the old copies used; not Lucide). Instances cannot swap icons, so three extra Secondary variants for the icons the Candidate copies used: `... / Icon / Bell Ringing` (10838), `/ Retry` (10842), `/ Email` (10846). Placed right of `Button / Icon / Bell`, y=40.
+- Alerts: `Alert / Error / With Icon` (6730) already existed (A naming); reused, no new `/ Icon` component. Info/Success/Warning icon variants exist too, not needed.
+- `Pagination / Single` (10853): one active page, placed right of `Pagination / Default`.
+
+## 2. DS defects fixed
+- `Sidebar / Admin` user footer (main + the 16 Admin instances): User Meta now shrinks (width 100%, minWidth 0, clipped), name row wraps so the ADMIN tag drops under the name instead of overlapping, email sits in a new `User Email Clip` frame (16 px, overflow hidden) so it is cut, not wrapped. Clip is a hard cut (no ellipsis; text-overflow is unsupported). Instances do not follow structural main edits, so the clip frame was added to each of the 16 instances by hand.
+- `Input / Number`: Value text width 0 -> 100% grow (main) and on the 25 Admin instances. No more wrap.
+
+## 3. Swaps (instances now accept children: move_nodes into an instance works)
+- Admin: 4 Cards (15375, 15467, 15564, 15675) -> `Card / Container` instances with their form content moved inside (Card Body wrapper flattened; padding/gap copied; heights unchanged, identical screenshots). Alert 15470 -> `Alert / Error / With Icon` (text "Create failed").
+- Candidate: 21 Cards -> `Card / Container` (same flatten). 8 Join cards had fixed heights; fit-content made pairs unequal, so pairs set to the taller value (466/530). 3 icon Alerts (16358, 16465, 17207) -> `Alert / Error / With Icon` (text grows, wraps). 5 icon Secondary buttons -> the matching icon variants (width fill; 44 -> 40 px high). Pagination 17090 -> `Pagination / Single` (fill, right aligned, 16 px top padding).
+- Extra: Admin v3 `New session` LG button (24627) got its plus icon back (`Button / Primary / LG / Icon`). Other LG buttons that lost an icon in package B were not identified; not restored.
+- Components uses: Card / Container 25, Pagination / Single 1, icon buttons 5.
+
+## 4. Layout fixes
+- Candidate Session home: screen/main/content/page-center set fit-content (900 -> 936 px), sidebar stretches (100%), nav height fixed so footer sits at the bottom. Last card row no longer clipped. Other Session screens untouched.
+- Admin 7 tall edit screens (Fullstack, Sql, Theory Empty, Theory, Mcq Validation, Mcq, New Template Applied): sidebar wrapper and instance stretch to screen height, nav sized so footer is at the bottom (no more vertical centring).
+- v3 decline block (Entry Request / Waiting, 592 px variant on the Decline dialog screen): Who column shrinkable with 8 px right padding, Meta row wraps, so helper text no longer runs under Deny. Fixed in DS main and in the 592 instance (card 82 -> 106 px). Padding is a literal 8 (no bind tool in this build).
+
+## Lint after
+DS new components 0; Sidebar / Admin 2 info type-scale (13px, pre-existing). Admin: touch-target 22 (accepted by B), else 0. Candidate: hard-coded 8 (Google logo, intentional), else 0.
+
+## Skipped / notes
+- No Lucide icons (old copies used phosphor); no `/ Icon` alert rename; no title slot on Card / Container.
+- Pre-existing, seen not fixed: Candidate Access `Send code` button clipped ("Send co") and verify-error alert wider than its card; Session home `Enter session` button clipped.
+- Instances of `Entry Request / Waiting` at 656/1068 px were not changed (not overflowing).