chore(release): 0.9.0

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

.debrand

@@ -1,4 +1,4 @@
-# Upstream (michaelshimeles/paper-rs) carries no ArtCraft/Discord branding; keep it that way. Run: perl tools/craft-skills/skills/debrand/debrand.pl --check .
+# The tree carries no ArtCraft/Discord branding; keep it that way. Run: perl tools/craft-skills/skills/debrand/debrand.pl --check .
 keep ^(LICENSE-(MIT|APACHE)|\.debrand)$
 forbid (?i)artcraft
 forbid (?i)discord

.github/CODEOWNERS

@@ -0,0 +1 @@
+* @nxtime

.github/ISSUE_TEMPLATE/bug_report.md

@@ -0,0 +1,19 @@
+---
+name: Bug report
+about: Something does not work as expected
+labels: bug
+---
+
+**What happened**
+
+**What you expected**
+
+**Steps to reproduce**
+1.
+2.
+
+**Version and platform**
+PaperCraft version (the release you installed), OS, and the GPU if it is a rendering problem.
+
+**Screenshot, recording or project**
+Attach one if you can. Remove anything private first.

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,5 @@
+blank_issues_enabled: true
+contact_links:
+  - name: Security vulnerability
+    url: https://github.com/nxtime/papercraft/security/advisories/new
+    about: Report security problems privately, not as an issue.

.github/ISSUE_TEMPLATE/feature_request.md

@@ -0,0 +1,14 @@
+---
+name: Feature request
+about: Suggest an idea or a missing capability
+labels: enhancement
+---
+
+**The problem**
+What are you trying to do, and what gets in the way?
+
+**What you would like**
+
+**Alternatives you considered**
+
+**Is it already in ROADMAP.md?**

.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,24 @@
+## What and why
+
+<!-- What changed, and the problem it solves. Link the issue: Fixes #123 -->
+
+## How I tested it
+
+<!-- Tests added, commands run, platforms tried. -->
+
+## AI assistance
+
+<!-- Which tool helped and how much, or "none". See AI_POLICY.md. You must understand every line. -->
+
+## Screenshots
+
+<!-- For UI changes: before and after, at scale 2, in each theme. -->
+
+## Checklist
+
+- [ ] Tests first; new code is covered, error paths included
+- [ ] `cargo fmt --all`, `cargo clippy --all-targets` and the affected `cargo test --test` targets pass
+- [ ] Follows [RULES.md](../RULES.md) (no overflow, data safety, one undo step per tool call)
+- [ ] Docs, tool descriptions and `PROGRESS.md` updated where needed
+- [ ] Conventional commit messages
+- [ ] I read and understand every line of this change (AI-assisted or not)

.github/workflows/release.yml

@@ -0,0 +1,95 @@
+# Builds Windows, macOS and Linux on GitHub's runners when a v* tag is pushed, then publishes a
+# GitHub release with the files once the full test suite (test.yml) passes. Gitea reads .github/workflows too, so every job is gated on the
+# server URL: this pipeline only ever runs on github.com.
+name: release
+
+on:
+  push:
+    tags: ["v*"]
+  workflow_dispatch:
+    inputs:
+      tag:
+        description: Existing tag to build and release (e.g. v0.8.4)
+        required: true
+
+permissions:
+  contents: write
+
+env:
+  CARGO_TERM_COLOR: always
+  TAG: ${{ inputs.tag || github.ref_name }}
+
+jobs:
+  test:
+    uses: ./.github/workflows/test.yml
+    with: { ref: "${{ inputs.tag || github.ref }}" }
+
+  windows:
+    if: github.server_url == 'https://github.com'
+    runs-on: windows-latest
+    steps:
+      - uses: actions/checkout@v4
+        with: { ref: "${{ inputs.tag || github.ref }}" }
+      - uses: dtolnay/rust-toolchain@stable
+      - uses: Swatinem/rust-cache@v2
+      - name: Build and zip
+        shell: pwsh
+        run: |
+          cargo build --release --locked --bins
+          $v = $env:TAG.TrimStart('v')
+          New-Item -ItemType Directory -Force dist | Out-Null
+          $exes = 'papercraft','papercraft-cli','papercraft-mcp','papercraft-relay' | ForEach-Object { "target/release/$_.exe" }
+          Compress-Archive -Path $exes -DestinationPath "dist/papercraft-$v-windows-x86_64.zip"
+      - uses: actions/upload-artifact@v4
+        with: { name: windows, path: dist/* }
+
+  macos:
+    if: github.server_url == 'https://github.com'
+    runs-on: macos-latest
+    steps:
+      - uses: actions/checkout@v4
+        with: { ref: "${{ inputs.tag || github.ref }}" }
+      - uses: dtolnay/rust-toolchain@stable
+        with: { targets: aarch64-apple-darwin }
+      - uses: Swatinem/rust-cache@v2
+      - name: Package (Apple silicon, ad-hoc signed, not notarized)
+        run: packaging/macos/package.sh --arch aarch64
+      - uses: actions/upload-artifact@v4
+        with: { name: macos, path: dist/release/* }
+
+  linux:
+    if: github.server_url == 'https://github.com'
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with: { ref: "${{ inputs.tag || github.ref }}" }
+      - uses: dtolnay/rust-toolchain@stable
+      - uses: Swatinem/rust-cache@v2
+      - name: Build dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y pkg-config libssl-dev libxkbcommon-x11-dev libxcb-xkb-dev libx11-xcb-dev libwayland-dev libvulkan-dev libdbus-1-dev
+      - name: Build and tar
+        run: |
+          cargo build --release --locked --bins
+          v="${TAG#v}"
+          mkdir -p dist stage
+          cp target/release/papercraft target/release/papercraft-cli target/release/papercraft-mcp target/release/papercraft-relay stage/
+          tar -C stage -czf "dist/papercraft-$v-linux-x86_64.tar.gz" .
+          tar -C stage -czf "dist/papercraft-cli-$v-linux-x86_64.tar.gz" papercraft-cli papercraft-mcp
+      - uses: actions/upload-artifact@v4
+        with: { name: linux, path: dist/* }
+
+  publish:
+    if: github.server_url == 'https://github.com'
+    needs: [test, windows, macos, linux]
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/download-artifact@v4
+        with: { path: files, merge-multiple: true }
+      - name: Create the release
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          gh release create "$TAG" files/* --repo "$GITHUB_REPOSITORY" --title "PaperCraft ${TAG#v}" \
+            --generate-notes --verify-tag || gh release upload "$TAG" files/* --repo "$GITHUB_REPOSITORY" --clobber

.github/workflows/test.yml

@@ -0,0 +1,34 @@
+# The full test suite. Local runs cover only the targets a change touches (RULES.md, "The machine
+# you build on"); this is where everything runs, before a release can publish. Run it alone with
+# `gh workflow run test.yml -f ref=<branch or tag>`. Gitea reads .github/workflows too, so the job
+# is gated on the server URL.
+name: test
+
+on:
+  workflow_call:
+    inputs:
+      ref:
+        type: string
+        required: false
+  workflow_dispatch:
+    inputs:
+      ref:
+        description: Branch or tag to test (default the one you run from)
+        required: false
+
+env:
+  CARGO_TERM_COLOR: always
+
+jobs:
+  test:
+    if: github.server_url == 'https://github.com'
+    # Windows: the platform the app is developed on, and its runners render with a software GPU.
+    runs-on: windows-latest
+    steps:
+      - uses: actions/checkout@v4
+        with: { ref: "${{ inputs.ref || github.ref }}" }
+      - uses: dtolnay/rust-toolchain@stable
+        with: { targets: wasm32-unknown-unknown }
+      - uses: Swatinem/rust-cache@v2
+      - run: cargo test --locked --no-fail-fast
+      - run: cargo check --locked --lib --target wasm32-unknown-unknown

.playwright-mcp/console-2026-10-25T20-31-26-027Z.log

@@ -0,0 +1,2 @@
+[    3482ms] [INFO] %cDownload the React DevTools for a better development experience: https://react.dev/link/react-devtools font-weight:bold @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[    3531ms] [LOG] [HMR] connected @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719

.playwright-mcp/console-2026-10-25T20-31-41-217Z.log

@@ -0,0 +1,9 @@
+[     210ms] [LOG] [Fast Refresh] rebuilding @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[    1039ms] [LOG] [Fast Refresh] done in 930ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[    1076ms] [INFO] %cDownload the React DevTools for a better development experience: https://react.dev/link/react-devtools font-weight:bold @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[    1086ms] [LOG] [HMR] connected @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[   37298ms] [ERROR] Pattern attribute value [0-9() -]{10,16} is not a valid regular expression: Uncaught SyntaxError: Invalid regular expression: /[0-9() -]{10,16}/v: Invalid character in character class @ http://localhost:3000/motorista/entrar:0
+[   60872ms] [LOG] [Fast Refresh] rebuilding @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[   60976ms] [LOG] [Fast Refresh] done in 99ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[   61021ms] [LOG] [Fast Refresh] done in 99ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[   66490ms] [LOG] [Fast Refresh] done in 99ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719

.playwright-mcp/console-2026-10-25T20-32-58-325Z.log

@@ -0,0 +1,13 @@
+[      93ms] [LOG] [Fast Refresh] done in 99ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[     953ms] [INFO] %cDownload the React DevTools for a better development experience: https://react.dev/link/react-devtools font-weight:bold @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[     963ms] [LOG] [HMR] connected @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[    4345ms] [LOG] [Fast Refresh] rebuilding @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[    4433ms] [LOG] [Fast Refresh] done in 189ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[    4548ms] [VERBOSE] [DOM] Password forms should have (optionally hidden) username fields for accessibility: (More info: https://goo.gl/9p2vKq) %o @ http://localhost:3000/entrar/senha:0
+[    9731ms] [LOG] [Fast Refresh] done in 5486ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[    9777ms] [LOG] [Fast Refresh] done in 5486ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[   11070ms] [LOG] [Fast Refresh] done in 5486ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[   30758ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=fBVPvwqt_6CU4TU3WWYmu' failed:  @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[   31764ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=fBVPvwqt_6CU4TU3WWYmu' failed: Error in connection establishment: net::ERR_CONNECTION_REFUSED @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[   32766ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=fBVPvwqt_6CU4TU3WWYmu' failed: Error in connection establishment: net::ERR_CONNECTION_REFUSED @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[   33776ms] [LOG] [HMR] connected @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719

.playwright-mcp/console-2026-10-25T20-33-32-647Z.log

@@ -0,0 +1,79 @@
+[     805ms] [INFO] %cDownload the React DevTools for a better development experience: https://react.dev/link/react-devtools font-weight:bold @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[     813ms] [LOG] [HMR] connected @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[    1121ms] [VERBOSE] [DOM] Password forms should have (optionally hidden) username fields for accessibility: (More info: https://goo.gl/9p2vKq) %o @ http://localhost:3000/entrar/senha:0
+[    2267ms] [LOG] [Fast Refresh] rebuilding @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[    2310ms] [LOG] [Fast Refresh] done in 144ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[    9621ms] [LOG] [Fast Refresh] rebuilding @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[    9763ms] [LOG] [Fast Refresh] done in 243ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[   48693ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[   48749ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[   48826ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[   51025ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[   51539ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[   51907ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  107213ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  118297ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  123464ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  125669ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  125771ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  126799ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  126941ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  127929ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  127981ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  154930ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  156620ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  157845ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  158106ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  158454ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  158504ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  158748ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  158848ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  158955ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  159307ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  159399ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  159488ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  159901ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  159993ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  187894ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  188000ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  205603ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  207208ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  207296ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  207951ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  208126ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  208244ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  208353ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  208415ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  208590ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  208695ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  208794ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  208878ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  208986ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  209132ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  209234ms] [LOG] [Fast Refresh] done in 39171ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  228278ms] [LOG] [Fast Refresh] rebuilding @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  228308ms] [LOG] [Fast Refresh] done in 3ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  238991ms] [LOG] [Fast Refresh] done in 10713ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  239037ms] [LOG] [Fast Refresh] done in 10713ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  239175ms] [LOG] [Fast Refresh] rebuilding @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  239206ms] [LOG] [Fast Refresh] done in 3ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  239602ms] [LOG] [Fast Refresh] done in 3ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  239697ms] [LOG] [Fast Refresh] done in 3ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  264111ms] [LOG] [Fast Refresh] done in 3ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  265916ms] [LOG] [Fast Refresh] done in 3ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  321743ms] [LOG] [Fast Refresh] done in 3ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  333352ms] [LOG] [Fast Refresh] done in 3ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  369410ms] [LOG] [Fast Refresh] done in 3ms @ http://localhost:3000/_next/static/chunks/1bba_next_dist_041c6vio7na_y._.js:719
+[  413886ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=ruCdDO_vey1tElDs8WYR0' failed:  @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[  414902ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=ruCdDO_vey1tElDs8WYR0' failed: Error in connection establishment: net::ERR_CONNECTION_REFUSED @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[  415904ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=ruCdDO_vey1tElDs8WYR0' failed: Error in connection establishment: net::ERR_CONNECTION_REFUSED @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[  416906ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=ruCdDO_vey1tElDs8WYR0' failed: Error in connection establishment: net::ERR_CONNECTION_REFUSED @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[  417908ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=ruCdDO_vey1tElDs8WYR0' failed: Error in connection establishment: net::ERR_CONNECTION_REFUSED @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[  418911ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=ruCdDO_vey1tElDs8WYR0' failed: Error in connection establishment: net::ERR_CONNECTION_REFUSED @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[  423918ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=ruCdDO_vey1tElDs8WYR0' failed: Error in connection establishment: net::ERR_CONNECTION_REFUSED @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[  428928ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=ruCdDO_vey1tElDs8WYR0' failed: Error in connection establishment: net::ERR_CONNECTION_REFUSED @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[  433950ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=ruCdDO_vey1tElDs8WYR0' failed: Error in connection establishment: net::ERR_CONNECTION_REFUSED @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[  438984ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=ruCdDO_vey1tElDs8WYR0' failed: Error in connection establishment: net::ERR_CONNECTION_REFUSED @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[  444022ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=ruCdDO_vey1tElDs8WYR0' failed: Error in connection establishment: net::ERR_CONNECTION_REFUSED @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[  449184ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=ruCdDO_vey1tElDs8WYR0' failed: Error in connection establishment: net::ERR_CONNECTION_REFUSED @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225
+[  454479ms] [ERROR] WebSocket connection to 'ws://localhost:3000/_next/hmr?id=ruCdDO_vey1tElDs8WYR0' failed: Error in connection establishment: net::ERR_CONNECTION_REFUSED @ http://localhost:3000/_next/static/chunks/1bba_next_dist_client_0ru2jkh0js1s8._.js:16225

.playwright-mcp/page-2026-10-25T20-31-29-604Z.yml

@@ -0,0 +1,28 @@
+- generic [ref=e2]:
+  - complementary [ref=e3]:
+    - generic [ref=e8]:
+      - generic [ref=e9]: Rota Fresca
+      - generic [ref=e10]: Distribuição de alimentos
+    - generic [ref=e11]:
+      - heading "Do depósito à prateleira, na hora certa." [level=2] [ref=e12]
+      - paragraph [ref=e13]: Pedidos, entregas e notas da sua rede no mesmo lugar.
+  - main [ref=e14]:
+    - link "Precisa de ajuda?" [ref=e16] [cursor=pointer]:
+      - /url: mailto:suporte@rotafresca.com.br
+    - generic [ref=e17]:
+      - generic [ref=e18]:
+        - heading "Entrar" [level=1] [ref=e19]
+        - paragraph [ref=e20]: Use o e-mail da sua conta. Conforme o tipo de conta, pedimos a senha ou mandamos um link de acesso.
+      - generic [ref=e21]:
+        - generic [ref=e22]:
+          - generic [ref=e23]: E-mail
+          - textbox "E-mail" [ref=e24]:
+            - /placeholder: nome@empresa.com.br
+        - button "Continuar" [ref=e25]
+      - generic [ref=e26]: ou
+      - generic [ref=e33]:
+        - paragraph [ref=e34]: É motorista?
+        - paragraph [ref=e35]: Entre pelo app Rota Fresca no celular, com o seu número. Não precisa de e-mail.
+    - generic [ref=e36]:
+      - paragraph [ref=e37]: Recebeu um convite? Abra o link do e-mail para criar seu acesso.
+      - paragraph [ref=e38]: Termos de uso · Privacidade

.playwright-mcp/page-2026-10-25T20-31-42-327Z.yml

@@ -0,0 +1,20 @@
+- generic [active] [ref=f1e1]:
+  - generic [ref=f1e2]:
+    - generic [ref=f1e3]:
+      - generic [ref=f1e8]:
+        - generic [ref=f1e9]: Rota Fresca
+        - generic [ref=f1e10]: App do motorista
+      - banner [ref=f1e11]:
+        - heading "Entrar" [level=1] [ref=f1e12]
+        - paragraph [ref=f1e13]: Use o número de celular que a central cadastrou. Mandamos um código, sem senha.
+      - generic [ref=f1e14]:
+        - generic [ref=f1e15]: Celular
+        - generic [ref=f1e16]:
+          - generic [ref=f1e17]: "+55"
+          - textbox "(21) 98877-4410" [ref=f1e18]
+    - generic [ref=f1e19]:
+      - button "Receber código no WhatsApp" [ref=f1e20]
+      - button "Receber por SMS" [ref=f1e21]
+      - paragraph [ref=f1e22]: Trocou de número? Fale com a central.
+  - button "Open Next.js Dev Tools" [ref=f1e28] [cursor=pointer]
+  - alert [ref=f1e32]

.playwright-mcp/page-2026-10-25T20-32-59-335Z.yml

@@ -0,0 +1,32 @@
+- generic [active] [ref=f2e1]:
+  - generic [ref=f2e2]:
+    - complementary [ref=f2e3]:
+      - generic [ref=f2e9]:
+        - generic [ref=f2e10]: Rota Fresca
+        - generic [ref=f2e11]: Distribuição de alimentos
+      - generic [ref=f2e12]:
+        - heading "Do depósito à prateleira, na hora certa." [level=2] [ref=f2e13]
+        - paragraph [ref=f2e14]: Pedidos, entregas e notas da sua rede no mesmo lugar.
+    - main [ref=f2e15]:
+      - link "Precisa de ajuda?" [ref=f2e17] [cursor=pointer]:
+        - /url: mailto:suporte@rotafresca.com.br
+      - generic [ref=f2e18]:
+        - generic [ref=f2e19]:
+          - heading "Entrar" [level=1] [ref=f2e20]
+          - paragraph [ref=f2e21]: Use o e-mail da sua conta. Conforme o tipo de conta, pedimos a senha ou mandamos um link de acesso.
+        - generic [ref=f2e22]:
+          - generic [ref=f2e23]:
+            - generic [ref=f2e24]: E-mail
+            - textbox "E-mail" [ref=f2e25]:
+              - /placeholder: nome@empresa.com.br
+          - button "Continuar" [ref=f2e26]
+        - generic [ref=f2e27]: ou
+        - button "Entrar com chave de acesso (passkey)" [ref=f2e31]
+        - generic [ref=f2e35]:
+          - paragraph [ref=f2e36]: É motorista?
+          - paragraph [ref=f2e37]: Entre pelo app Rota Fresca no celular, com o seu número. Não precisa de e-mail.
+      - generic [ref=f2e38]:
+        - paragraph [ref=f2e39]: Recebeu um convite? Abra o link do e-mail para criar seu acesso.
+        - paragraph [ref=f2e40]: Termos de uso · Privacidade
+  - button "Open Next.js Dev Tools" [ref=f2e46] [cursor=pointer]
+  - alert [ref=f2e50]

.playwright-mcp/page-2026-10-25T20-33-03-457Z.yml

@@ -0,0 +1,38 @@
+- generic [active] [ref=f2e1]:
+  - button "Open Next.js Dev Tools" [ref=f2e46] [cursor=pointer]
+  - alert [ref=f2e50]: Digite sua senha
+  - generic [ref=f2e51]:
+    - complementary [ref=f2e52]:
+      - generic [ref=f2e58]:
+        - generic [ref=f2e59]: Rota Fresca
+        - generic [ref=f2e60]: Distribuição de alimentos
+      - generic [ref=f2e61]:
+        - heading "Do depósito à prateleira, na hora certa." [level=2] [ref=f2e62]
+        - paragraph [ref=f2e63]: Pedidos, entregas e notas da sua rede no mesmo lugar.
+    - main [ref=f2e64]:
+      - link "Precisa de ajuda?" [ref=f2e66] [cursor=pointer]:
+        - /url: mailto:suporte@rotafresca.com.br
+      - generic [ref=f2e67]:
+        - generic [ref=f2e68]:
+          - generic [ref=f2e69]: ana.souza@rotafresca.com.br
+          - button "Trocar" [ref=f2e74]
+        - generic [ref=f2e75]:
+          - heading "Digite sua senha" [level=1] [ref=f2e76]
+          - paragraph [ref=f2e77]: Conta de consultor. Depois da senha pedimos o código do autenticador.
+        - generic [ref=f2e78]:
+          - generic [ref=f2e79]:
+            - generic [ref=f2e80]:
+              - generic [ref=f2e81]: Senha
+              - textbox "Senha" [ref=f2e82]
+            - generic [ref=f2e83]:
+              - generic [ref=f2e84]:
+                - checkbox "Mostrar senha" [ref=f2e85]
+                - checkbox [aria-hidden]
+                - text: Mostrar senha
+              - link "Esqueci minha senha" [ref=f2e86] [cursor=pointer]:
+                - /url: /entrar/redefinir
+          - button "Entrar" [ref=f2e87]
+        - paragraph [ref=f2e88]: Pode colar a senha ou usar o gerenciador de senhas do navegador.
+      - generic [ref=f2e89]:
+        - paragraph [ref=f2e90]: Recebeu um convite? Abra o link do e-mail para criar seu acesso.
+        - paragraph [ref=f2e91]: Termos de uso · Privacidade

.playwright-mcp/page-2026-10-25T20-33-15-378Z.yml

@@ -0,0 +1,40 @@
+- generic [ref=f2e1]:
+  - button "Open Next.js Dev Tools" [ref=f2e46] [cursor=pointer]
+  - alert [ref=f2e50]: Rota Fresca
+  - generic [ref=f2e51]:
+    - complementary [ref=f2e52]:
+      - generic [ref=f2e58]:
+        - generic [ref=f2e59]: Rota Fresca
+        - generic [ref=f2e60]: Distribuição de alimentos
+      - generic [ref=f2e61]:
+        - heading "Do depósito à prateleira, na hora certa." [level=2] [ref=f2e62]
+        - paragraph [ref=f2e63]: Pedidos, entregas e notas da sua rede no mesmo lugar.
+    - main [ref=f2e64]:
+      - link "Precisa de ajuda?" [ref=f2e66] [cursor=pointer]:
+        - /url: mailto:suporte@rotafresca.com.br
+      - generic [ref=f2e67]:
+        - generic [ref=f2e68]:
+          - generic [ref=f2e69]: ana.souza@rotafresca.com.br
+          - button "Trocar" [ref=f2e74]
+        - generic [ref=f2e75]:
+          - heading "Digite sua senha" [level=1] [ref=f2e76]
+          - paragraph [ref=f2e77]: Conta de consultor. Depois da senha pedimos o código do autenticador.
+        - generic [ref=f2e78]:
+          - generic [ref=f2e79]:
+            - generic [ref=f2e80]:
+              - generic [ref=f2e81]: Senha
+              - textbox "Senha" [active] [invalid] [ref=f2e82]
+            - generic [ref=f2e83]:
+              - generic [ref=f2e84]:
+                - checkbox "Mostrar senha" [ref=f2e85]
+                - checkbox [aria-hidden]
+                - text: Mostrar senha
+              - link "Esqueci minha senha" [ref=f2e86] [cursor=pointer]:
+                - /url: /entrar/redefinir
+          - alert [ref=f2e92]:
+            - generic [ref=f2e93]: Senha incorreta. Confira e tente de novo.
+          - button "Entrar" [ref=f2e87]
+        - paragraph [ref=f2e88]: Pode colar a senha ou usar o gerenciador de senhas do navegador.
+      - generic [ref=f2e89]:
+        - paragraph [ref=f2e90]: Recebeu um convite? Abra o link do e-mail para criar seu acesso.
+        - paragraph [ref=f2e91]: Termos de uso · Privacidade

.playwright-mcp/page-2026-10-25T20-33-33-487Z.yml

@@ -0,0 +1,32 @@
+- generic [active] [ref=f3e1]:
+  - generic [ref=f3e2]:
+    - complementary [ref=f3e3]:
+      - generic [ref=f3e9]:
+        - generic [ref=f3e10]: Rota Fresca
+        - generic [ref=f3e11]: Distribuição de alimentos
+      - generic [ref=f3e12]:
+        - heading "Do depósito à prateleira, na hora certa." [level=2] [ref=f3e13]
+        - paragraph [ref=f3e14]: Pedidos, entregas e notas da sua rede no mesmo lugar.
+    - main [ref=f3e15]:
+      - link "Precisa de ajuda?" [ref=f3e17] [cursor=pointer]:
+        - /url: mailto:suporte@rotafresca.com.br
+      - generic [ref=f3e18]:
+        - generic [ref=f3e19]:
+          - heading "Entrar" [level=1] [ref=f3e20]
+          - paragraph [ref=f3e21]: Use o e-mail da sua conta. Conforme o tipo de conta, pedimos a senha ou mandamos um link de acesso.
+        - generic [ref=f3e22]:
+          - generic [ref=f3e23]:
+            - generic [ref=f3e24]: E-mail
+            - textbox "E-mail" [ref=f3e25]:
+              - /placeholder: nome@empresa.com.br
+          - button "Continuar" [ref=f3e26]
+        - generic [ref=f3e27]: ou
+        - button "Entrar com chave de acesso (passkey)" [ref=f3e31]
+        - generic [ref=f3e35]:
+          - paragraph [ref=f3e36]: É motorista?
+          - paragraph [ref=f3e37]: Entre pelo app Rota Fresca no celular, com o seu número. Não precisa de e-mail.
+      - generic [ref=f3e38]:
+        - paragraph [ref=f3e39]: Recebeu um convite? Abra o link do e-mail para criar seu acesso.
+        - paragraph [ref=f3e40]: Termos de uso · Privacidade
+  - button "Open Next.js Dev Tools" [ref=f3e46] [cursor=pointer]
+  - alert [ref=f3e50]

.playwright-mcp/page-2026-10-25T20-33-34-597Z.yml

@@ -0,0 +1,38 @@
+- generic [active] [ref=f3e1]:
+  - button "Open Next.js Dev Tools" [ref=f3e46] [cursor=pointer]
+  - alert [ref=f3e50]
+  - generic [ref=f3e51]:
+    - complementary [ref=f3e52]:
+      - generic [ref=f3e58]:
+        - generic [ref=f3e59]: Rota Fresca
+        - generic [ref=f3e60]: Distribuição de alimentos
+      - generic [ref=f3e61]:
+        - heading "Do depósito à prateleira, na hora certa." [level=2] [ref=f3e62]
+        - paragraph [ref=f3e63]: Pedidos, entregas e notas da sua rede no mesmo lugar.
+    - main [ref=f3e64]:
+      - link "Precisa de ajuda?" [ref=f3e66] [cursor=pointer]:
+        - /url: mailto:suporte@rotafresca.com.br
+      - generic [ref=f3e67]:
+        - generic [ref=f3e68]:
+          - generic [ref=f3e69]: ana.souza@rotafresca.com.br
+          - button "Trocar" [ref=f3e74]
+        - generic [ref=f3e75]:
+          - heading "Digite sua senha" [level=1] [ref=f3e76]
+          - paragraph [ref=f3e77]: Conta de consultor. Depois da senha pedimos o código do autenticador.
+        - generic [ref=f3e78]:
+          - generic [ref=f3e79]:
+            - generic [ref=f3e80]:
+              - generic [ref=f3e81]: Senha
+              - textbox "Senha" [ref=f3e82]
+            - generic [ref=f3e83]:
+              - generic [ref=f3e84]:
+                - checkbox "Mostrar senha" [ref=f3e85]
+                - checkbox [aria-hidden]
+                - text: Mostrar senha
+              - link "Esqueci minha senha" [ref=f3e86] [cursor=pointer]:
+                - /url: /entrar/redefinir
+          - button "Entrar" [ref=f3e87]
+        - paragraph [ref=f3e88]: Pode colar a senha ou usar o gerenciador de senhas do navegador.
+      - generic [ref=f3e89]:
+        - paragraph [ref=f3e90]: Recebeu um convite? Abra o link do e-mail para criar seu acesso.
+        - paragraph [ref=f3e91]: Termos de uso · Privacidade

.playwright-mcp/page-2026-10-25T20-33-35-637Z.yml

@@ -0,0 +1,33 @@
+- generic [active] [ref=f3e1]:
+  - button "Open Next.js Dev Tools" [ref=f3e46] [cursor=pointer]
+  - alert [ref=f3e50]
+  - generic [ref=f3e92]:
+    - complementary [ref=f3e93]:
+      - generic [ref=f3e99]:
+        - generic [ref=f3e100]: Rota Fresca
+        - generic [ref=f3e101]: Distribuição de alimentos
+      - generic [ref=f3e102]:
+        - heading "Do depósito à prateleira, na hora certa." [level=2] [ref=f3e103]
+        - paragraph [ref=f3e104]: Pedidos, entregas e notas da sua rede no mesmo lugar.
+    - main [ref=f3e105]:
+      - link "Precisa de ajuda?" [ref=f3e107] [cursor=pointer]:
+        - /url: mailto:suporte@rotafresca.com.br
+      - generic [ref=f3e108]:
+        - heading "Confirme que é você" [level=1] [ref=f3e113]
+        - generic [ref=f3e114]:
+          - generic [ref=f3e115]:
+            - generic [ref=f3e116]:
+              - generic [ref=f3e117]: Código
+              - textbox "Código" [ref=f3e118]
+            - generic [ref=f3e119]:
+              - checkbox "Não pedir de novo neste computador por 30 dias" [checked] [ref=f3e120]
+              - checkbox [checked] [aria-hidden]
+              - text: Não pedir de novo neste computador por 30 dias
+          - button "Confirmar" [ref=f3e121]
+        - generic [ref=f3e122]:
+          - link "Usar um código de recuperação" [ref=f3e123] [cursor=pointer]:
+            - /url: /entrar/codigo?recuperacao=1
+          - paragraph [ref=f3e124]: Perdeu o celular? Peça ao administrador da conta para refazer o autenticador.
+      - generic [ref=f3e125]:
+        - paragraph [ref=f3e126]: Recebeu um convite? Abra o link do e-mail para criar seu acesso.
+        - paragraph [ref=f3e127]: Termos de uso · Privacidade

.playwright-mcp/page-2026-10-25T20-33-42-991Z.yml

@@ -0,0 +1,47 @@
+- generic [active] [ref=f3e1]:
+  - button "Open Next.js Dev Tools" [ref=f3e46] [cursor=pointer]
+  - alert [ref=f3e50]: Bom dia, Ana
+  - generic [ref=f3e128]:
+    - complementary [ref=f3e129]:
+      - generic [ref=f3e130]:
+        - generic [ref=f3e136]:
+          - generic [ref=f3e137]: Rota Fresca
+          - generic [ref=f3e138]: Distribuição de alimentos
+        - navigation [ref=f3e139]:
+          - link "Painel" [ref=f3e140] [cursor=pointer]:
+            - /url: /painel
+          - link "Roteiro do dia" [ref=f3e143] [cursor=pointer]:
+            - /url: /roteiro
+          - link "Lojas (PDVs)" [ref=f3e146] [cursor=pointer]:
+            - /url: /lojas
+          - link "Pedidos" [ref=f3e149] [cursor=pointer]:
+            - /url: /pedidos
+          - link "Entregas" [ref=f3e152] [cursor=pointer]:
+            - /url: /entregas
+          - link "Validade e lotes" [ref=f3e155] [cursor=pointer]:
+            - /url: /validade
+          - link "Relatórios" [ref=f3e158] [cursor=pointer]:
+            - /url: /relatorios
+          - link "Integrações e notas" [ref=f3e161] [cursor=pointer]:
+            - /url: /integracoes
+          - link "Leitura de documentos" [ref=f3e164] [cursor=pointer]:
+            - /url: /documentos
+          - link "Contatos e mensagens" [ref=f3e167] [cursor=pointer]:
+            - /url: /contatos
+          - link "Configurações" [ref=f3e170] [cursor=pointer]:
+            - /url: /configuracoes
+      - generic [ref=f3e173]:
+        - generic [aria-hidden] [ref=f3e174]: A
+        - generic [ref=f3e175]:
+          - generic [ref=f3e176]: Ana Souza
+          - generic [ref=f3e177]: Consultora · Zona Oeste
+    - main [ref=f3e178]:
+      - generic [ref=f3e179]:
+        - generic [ref=f3e180]:
+          - heading "Bom dia, Ana" [level=1] [ref=f3e181]
+          - paragraph [ref=f3e182]: domingo, 25 de outubro
+        - generic [ref=f3e183]:
+          - textbox "Buscar loja, pedido ou produto" [ref=f3e187]
+          - button "Notificações" [ref=f3e188]
+          - button "Importar / Exportar" [ref=f3e189]
+          - button "Novo pedido" [ref=f3e190]

AGENTS.md

@@ -0,0 +1,77 @@
+# Guidance for coding agents
+
+PaperCraft is a native Rust design and whiteboard canvas (egui, wgpu, winit) that AI agents drive
+over MCP and a CLI. Read [CONTRIBUTING.md](CONTRIBUTING.md) and [RULES.md](RULES.md) first; this
+file is the short version for agents, and [AI_POLICY.md](AI_POLICY.md) says what the person using
+you is accountable for.
+
+## Commands
+
+```sh
+cargo build --release
+cargo fmt --all
+cargo clippy --all-targets
+cargo test --test incremental                   # only the targets your change touches
+cargo test --lib editor::changes                # unit tests of one module
+cargo check --lib --target wasm32-unknown-unknown
+```
+
+Do not run the full suite locally: it builds dozens of test binaries. It runs on GitHub before
+each release (`.github/workflows/test.yml`). `tests/update.rs` cannot run on Windows outside CI
+(it needs elevation); that failure is known.
+
+Builds run on the person's own machine, so mind its resources (RULES.md, "The machine you build
+on"): check free memory and running builds first, cap jobs at no more than half the cores
+(`cargo test -j 8 -- --test-threads=8` on 16 cores), run at low priority, run one heavy job at a
+time, and stop at once if they say it is slow.
+
+## Running the app safely
+
+Never run a test copy against the person's real projects or MCP port, and never close or restart
+their running PaperCraft. Give yours its own folder and port, and keep its PID so you can stop only
+your own process:
+
+```sh
+PAPERCRAFT_HOME=/tmp/pc PAPERCRAFT_MCP_PORT=29998 ./target/release/papercraft
+```
+
+Do not stop processes by name.
+
+## Layout
+
+`src/model` (document), `src/editor` (tools and edits), `src/render` (wgpu), `src/mcp` (tools,
+HTML/CSS mapping, lint), `src/ui` (egui panels), `src/io` (import and export), `src/project`
+(store and formats), `tests/` (scenario tests), `specs/` (feature specs), `skills/` (agent
+skills). The README's project layout lists the rest.
+
+## Conventions
+
+* Tests first; match the style of the tests next to yours, and check the behaviour is not already
+  covered.
+* Changes to the document go through `Editor::edit`: one transaction, one undo step.
+* Handle errors; do not `unwrap` user data and do not discard errors with `let _ =`.
+* Comments explain *why*, never summarise the code. No dead code, debug prints or
+  commented-out blocks.
+* Files stay under about 800 lines; split by feature.
+* Reuse the shared UI kit and theme instead of one-off values. Nothing may overflow its container.
+* Writes to projects are atomic and format changes ship a migration (see RULES.md, "Data safety").
+* Adding or removing an MCP tool means a test in `tests/mcp.rs`, a precise description, and the
+  tool count updated in the README and ROADMAP.
+* Commits are conventional (`feat(scope): summary`) and focused. Do not commit, push or tag unless
+  the person asked you to.
+
+## Designing with PaperCraft (skills, examples, screenshots)
+
+* Themes are modes, never copies: one artboard per screen, a value per mode on each variable, and
+  `set_variable_mode` to switch. Never duplicate artboards for light and dark.
+* Every colour on a themed screen is a variable.
+* Check each artboard for overflow, programmatically and by screenshot in each mode, and run
+  `lint_design`.
+* Release design locks when you finish.
+
+## Avoid
+
+* Creative additions nobody asked for; stay inside the task.
+* Trivial tests, wide or optional types added to appease the compiler, and swallowed errors.
+* Editing generated files, lockfiles or release artifacts by hand, or touching files unrelated to
+  the task. Leave other people's uncommitted work alone.

AI_POLICY.md

@@ -0,0 +1,40 @@
+# AI usage policy
+
+PaperCraft is built to be driven by AI agents, and many of its contributors use them too. AI tools
+are welcome here. What we ask is that a person stays accountable for everything submitted.
+
+## For contributors
+
+* **You own what you submit.** Read every line of your pull request before asking for review. If
+  you cannot explain what a change does and how it fits the rest of the system without asking an
+  AI, do not submit it yet.
+* **Disclose meaningful AI help.** Say which tool you used and roughly how much it wrote in the
+  pull request description (for example "drafted with Claude Code, reviewed and tested by me").
+  Autocomplete and spell-checking do not need a note.
+* **Write issues yourself.** An agent can help you research, but the report should be in your own
+  words, trimmed to what matters, with reproduction steps you ran. Long generated text that nobody
+  checked puts the burden on maintainers.
+* **Think before you generate.** Agents jump to code and make their own job easy: trivial tests,
+  wide or optional types, swallowed errors, redundant comments, copied local patterns that do not
+  fit. Push back on all of it. [RULES.md](RULES.md) lists the standards to hold the output to.
+* **Verify.** Run the checks in [CONTRIBUTING.md](CONTRIBUTING.md), use the app, and attach
+  screenshots for UI changes. "The agent said it works" is not a test result.
+* **Answer automated review.** If a bot reviews your pull request, fix its findings or reply with
+  a one-line reason a finding does not apply.
+* **No bulk or drive-by pull requests.** Many small generated changes across the tree, or changes
+  to something nobody asked for, will be closed. Open an issue first for anything beyond a small
+  fix.
+
+Pull requests that clearly skip these steps may be closed without a detailed review. Repeated
+low-effort submissions can lead to a block, under the [Code of Conduct](CODE_OF_CONDUCT.md).
+
+## For coding agents
+
+Read [AGENTS.md](AGENTS.md). It holds the commands, conventions and guardrails an agent working in
+this repository should follow.
+
+## Why
+
+Every issue and pull request is read by a person. Review time is the scarce resource, and a change
+nobody understood costs more than it saves. AI that is reviewed, tested and understood by the
+person submitting it is a good way to build PaperCraft; AI used to skip that work is not.

CLAUDE.md

@@ -0,0 +1 @@
+@AGENTS.md

CODE_OF_CONDUCT.md

@@ -0,0 +1,31 @@
+# Code of Conduct
+
+PaperCraft is built by people with different backgrounds and experience. We want everyone to be
+able to contribute without harassment.
+
+## Our standard
+
+Be respectful and constructive. Assume good intent, give specific and kind feedback, and accept
+it the same way. Disagree about the work, not the person.
+
+Not acceptable:
+
+* harassment, insults, slurs, or personal or political attacks;
+* sexualised language or imagery, or unwelcome attention;
+* threats, doxxing, or publishing someone's private information;
+* sustained disruption of discussions, or pressure to merge or answer;
+* anything else a reasonable person would consider inappropriate in a professional setting.
+
+## Scope
+
+This applies in issues, pull requests, discussions, commit messages and any space where someone
+represents the project.
+
+## Enforcement
+
+Maintainers may edit or remove comments, close or lock threads, and block people who break these
+rules. To report a problem, use GitHub's "Report content" on the comment, or contact a maintainer
+through a private channel on their GitHub profile. Reports are handled privately and in good
+faith, and retaliation against a reporter is itself a violation.
+
+The spirit of this text follows the [Contributor Covenant](https://www.contributor-covenant.org).

CONTRIBUTING.md

@@ -0,0 +1,129 @@
+# Contributing to PaperCraft
+
+Thanks for wanting to help. PaperCraft is an alpha, so there is plenty to do and the bar for a
+first contribution is low: a typo fix, a clearer error message, a missing test or a one-line bug
+fix are all welcome.
+
+Please read [RULES.md](RULES.md) before writing code: it holds the standards a change is judged by.
+By taking part you agree to the [Code of Conduct](CODE_OF_CONDUCT.md). Using an AI tool is fine;
+the [AI usage policy](AI_POLICY.md) says what you stay responsible for, and coding agents should read
+[AGENTS.md](AGENTS.md).
+
+## Ways to help
+
+* **Report a bug** or **suggest a feature** with an [issue](https://github.com/nxtime/papercraft/issues/new/choose).
+  A screenshot or a short recording helps; so does the project folder (or a page of it) that
+  triggers the bug.
+* **Pick something from [ROADMAP.md](ROADMAP.md).** Rows marked `todo` or `partial` are open. Say
+  which one you are taking in an issue first, so two people do not build the same thing.
+* **Improve the docs, skills or screenshots.** [docs/](docs), [skills/](skills) and the README are
+  as important as code.
+* **Review a pull request.** A second pair of eyes is always useful.
+
+Security problems do not go in public issues: see [SECURITY.md](SECURITY.md).
+
+## Before you start
+
+For anything bigger than a small fix, open an issue (or comment on one) describing the problem and
+how you plan to solve it. Large features start as a short spec in [specs/](specs): a `spec.md`
+(why, what, what is out of scope) and a `plan.md`. [specs/component-props](specs/component-props)
+shows the shape. A spec agreed up front saves a rewrite later.
+
+## Set up
+
+You need Rust 1.95 or newer (`rust-version` in `Cargo.toml`) and the platform libraries listed in
+the README's [Build from source](README.md#build-from-source).
+
+```sh
+git clone https://github.com/nxtime/papercraft
+cd papercraft
+cargo build --release
+```
+
+Run a test copy next to your own PaperCraft so it cannot touch your real projects or collide with
+the running app's MCP port:
+
+```sh
+PAPERCRAFT_HOME=/tmp/pc PAPERCRAFT_MCP_PORT=29998 ./target/release/papercraft
+```
+
+`PAPERCRAFT_HOME` moves the project root and the config folder; `PAPERCRAFT_MCP_PORT` moves the
+MCP server. Always set both when developing.
+
+The code depends on [`craft-ui`](https://git.marcospaulo.dev.br/masi/craft-libs), which Cargo
+fetches from a public git repository at the revision pinned in `Cargo.toml`.
+
+## Make the change
+
+1. Fork the repository and branch from `main` (`fix/short-name`, `feat/short-name`).
+2. **Write the test first.** Fix a bug by adding a test that fails, then make it pass. See
+   [Tests](#tests).
+3. Keep the change focused. One concern per pull request; unrelated cleanups go in their own.
+4. Run the checks below and fix what they report.
+5. Update the docs the change touches: the README, `docs/`, the tool description for an MCP tool,
+   the skills if agents should now behave differently.
+6. Add a short entry to [PROGRESS.md](PROGRESS.md) for user-visible changes.
+
+### Checks
+
+```sh
+cargo fmt --all
+cargo clippy --all-targets
+cargo test --test <target>                            # the targets your change touches
+cargo check --lib --target wasm32-unknown-unknown     # the web build
+```
+
+Each file in `tests/` is its own binary linking the whole app, so the full suite is slow and heavy
+on a desktop. It runs on GitHub instead: before every release (`.github/workflows/test.yml`), or
+on demand with `gh workflow run test.yml -f ref=<branch>`.
+
+Keep the tree free of the third-party branding patterns listed in `.debrand`.
+
+### Tests
+
+* Unit tests sit next to the code in `#[cfg(test)]` modules.
+* Scenario tests live in [tests/](tests). They drive the editor with synthetic pointer and
+  keyboard events; `tests/mcp.rs` covers every MCP tool on a headless session.
+* Aim for **80% coverage** on code you add or change, including the error paths.
+* GPU-dependent tests skip when no adapter exists. Set `PAPERCRAFT_REQUIRE_GPU=1` to make them
+  fail instead.
+* `tests/update.rs` cannot run on Windows (it needs elevation); that failure is known and not
+  yours.
+* For UI work, run the app, look at it, and attach a screenshot. Headless rendering needs no
+  display: `papercraft --screenshot ui.png --demo --size 1440x900 --scale 2`.
+
+## Commits
+
+[Conventional commits](https://www.conventionalcommits.org): `type(scope): summary`, with the
+types `feat`, `fix`, `refactor`, `docs`, `test`, `chore`, `perf` and `ci`.
+
+```
+fix(ui): keep the review row's revert button off the hover edge
+feat(mcp): add set_screens
+docs(readme): add the demo video
+```
+
+Write the summary in the imperative, under about 72 characters, and use the body to say *why*.
+Reference the issue (`Fixes #123`).
+
+## Pull requests
+
+* Fill in the template: what changed, why, how you tested it, screenshots for UI changes.
+* Keep it reviewable. If it grows past a few hundred lines of non-generated code, split it.
+* The branch must be up to date with `main`; rebase rather than merge.
+* Expect review comments; they are about the code, not you. Push fixes as new commits and we
+  squash on merge.
+* A maintainer merges once a review is approved.
+
+## Releases
+
+Maintainers cut releases. A `v*` tag pushed to GitHub builds Windows, macOS and Linux and
+publishes them (`.github/workflows/release.yml`). Contributors do not need to bump versions.
+
+## Licensing
+
+PaperCraft is dual-licensed under [MIT](LICENSE-MIT) or [Apache-2.0](LICENSE-APACHE). Unless you
+state otherwise, anything you submit is licensed under the same terms, with no extra conditions.
+Only submit work you wrote or have the right to submit. Code ported from another project needs a
+compatible license and an entry in [THIRD_PARTY.md](THIRD_PARTY.md); bundled assets go in
+[ATTRIBUTION.md](ATTRIBUTION.md).

Cargo.lock

@@ -784,7 +784,7 @@ dependencies = [
 [[package]]
 name = "craft-ui"
 version = "0.1.0"
-source = "git+https://git.marcospaulo.dev.br/masi/craft-libs?rev=22544b6d619d4f16779b8bdfa3f0b642a86fc065#22544b6d619d4f16779b8bdfa3f0b642a86fc065"
+source = "git+https://git.marcospaulo.dev.br/masi/craft-libs?rev=9c470aa935c4ce48f2ba4f0a17f9410c0c8660fb#9c470aa935c4ce48f2ba4f0a17f9410c0c8660fb"
 dependencies = [
  "egui",
 ]
@@ -3128,7 +3128,7 @@ dependencies = [
 
 [[package]]
 name = "papercraft"
-version = "0.7.3"
+version = "0.9.0"
 dependencies = [
  "arboard",
  "axum",
@@ -3161,6 +3161,7 @@ dependencies = [
  "resvg",
  "rfd",
  "rmcp",
+ "ruzstd",
  "semver",
  "serde",
  "serde_json",
@@ -3174,8 +3175,10 @@ dependencies = [
  "web-sys",
  "web-time",
  "wgpu",
+ "windows-sys 0.61.2",
  "winit",
  "winresource",
+ "zstd",
 ]
 
 [[package]]
@@ -3853,6 +3856,15 @@ version = "1.0.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "cf54715a573b99ac80df0bc206da022bcd442c974952c7b9720069370852e21f"
 
+[[package]]
+name = "ruzstd"
+version = "0.8.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a7c1c839d570d835527c9a5e4db7cb2198683a988cb9d7293fc8674e6bd58fc8"
+dependencies = [
+ "twox-hash",
+]
+
 [[package]]
 name = "ryu"
 version = "1.0.23"
@@ -4754,6 +4766,12 @@ dependencies = [
  "thiserror 2.0.21",
 ]
 
+[[package]]
+name = "twox-hash"
+version = "2.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "86a801b3cea342a06d468c8710662aa29e5e05e4f5c0d62f00bbb7f2ad7941c2"
+
 [[package]]
 name = "type-map"
 version = "0.5.1"
@@ -4777,7 +4795,7 @@ checksum = "f2f6fb2847f6742cd76af783a2a2c49e9375d0a111c7bef6f71cd9e738c72d6e"
 dependencies = [
  "memoffset",
  "tempfile",
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -6018,6 +6036,34 @@ version = "1.0.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "29666d0abbfad1e3dc4dcf6144730dd3a3ab225bbbdac83319345b1b44ccfc1b"
 
+[[package]]
+name = "zstd"
+version = "0.13.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e91ee311a569c327171651566e07972200e76fcfe2242a4fa446149a3881c08a"
+dependencies = [
+ "zstd-safe",
+]
+
+[[package]]
+name = "zstd-safe"
+version = "7.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "64d80649ab6db9d9f6f9c80a40becd948eda4714a0a5ac8c4d157a32231c7882"
+dependencies = [
+ "zstd-sys",
+]
+
+[[package]]
+name = "zstd-sys"
+version = "2.1.2+zstd.1.5.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "481ada9c543b2323cbb7bbdbb5249177fbf0188ed959ed2c15e4b22421c84ed3"
+dependencies = [
+ "cc",
+ "pkg-config",
+]
+
 [[package]]
 name = "zune-core"
 version = "0.5.3"

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "papercraft"
-version = "0.7.3"
+version = "0.9.0"
 edition = "2024"
 rust-version = "1.95"
 description = "A native infinite-canvas app for planning, working and sketching"
@@ -44,7 +44,7 @@ env_logger = "0.11"
 rfd = "0.17"
 fastrand = "2"
 dirs = "6"
-craft-ui = { version = "0.1.0", git = "https://git.marcospaulo.dev.br/masi/craft-libs", rev = "22544b6d619d4f16779b8bdfa3f0b642a86fc065" }
+craft-ui = { version = "0.1.0", git = "https://git.marcospaulo.dev.br/masi/craft-libs", rev = "9c470aa935c4ce48f2ba4f0a17f9410c0c8660fb" }
 i_overlay = "9.0.1"
 miniz_oxide = "0.9.1"
 usvg = "0.48"
@@ -97,8 +97,14 @@ reqwest = { version = "0.13", default-features = false, features = ["native-tls"
 # libc: SIGINT to the spawned chat CLIs.
 keyring = { version = "3", default-features = false, features = ["apple-native", "windows-native", "sync-secret-service", "crypto-rust"] }
 libc = "0.2"
+# Pages and versions on disk are zstd-compressed JSON (`project::packed`).
+zstd = "0.13"
 
 # Browser build (`src/web.rs`): canvas + WebGPU/WebGL surface, DOM events instead of winit.
+# Process RAM and CPU readings for the performance overlay (`ui::perf`).
+[target.'cfg(windows)'.dependencies]
+windows-sys = { version = "0.61", features = ["Win32_Foundation", "Win32_System_ProcessStatus", "Win32_System_Threading"] }
+
 [target.'cfg(target_arch = "wasm32")'.dependencies]
 wgpu = { version = "30", features = ["webgl"] }
 wasm-bindgen = "0.2"
@@ -106,4 +112,6 @@ wasm-bindgen-futures = "0.4"
 js-sys = "0.3"
 console_error_panic_hook = "0.1"
 console_log = "1"
+# zstd without C, for the browser: reads the packed files native builds write.
+ruzstd = "0.8"
 web-sys = { version = "0.3", features = ["Window", "Document", "Element", "HtmlElement", "HtmlCanvasElement", "KeyboardEvent", "PointerEvent", "WheelEvent", "ClipboardEvent", "DataTransfer", "MediaQueryList", "Location", "Performance", "Navigator", "Clipboard", "CssStyleDeclaration", "console", "WebSocket", "MessageEvent", "CloseEvent", "ErrorEvent", "Storage", "Crypto", "DragEvent", "FileList", "File", "Blob", "Request", "RequestInit", "Response"] }

PROGRESS.md

@@ -76,6 +76,118 @@ 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).
 
+### 2026-10-28: overflow lint rule and `get_bounds`
+
+* `lint_design` has an `overflow` rule (`src/mcp/lint_overflow.rs`): layers past their artboard or clipping parent, and text taller than its box. Scrolling frames and what is pinned in them are exempt; it is never auto-fixed. It is a pure function over a node set, so a scope filter can pass any subset.
+* New tool `get_bounds {nodeIds?, pageId?, depth?}` (`src/mcp/bounds.rs`): every layer's rect in one read (about 2000 rows per call, then `continueWith`). Tool count is now 87 (one more than before; if another branch also adds a tool, recount). Tests: `tests/mcp_bounds.rs`.
+
+### 2026-10-28: agent reads no longer look like edits
+
+* Read-only tool calls (`get_node_info`, `get_children`, ...) have honest verbs ("Looking at the design") and light the "Working" pill for 1.5 s instead of 6 s, so it goes out right after a sweep ends.
+* A sweep of reads draws at most one layer outline (a read outlines only when no other read ran in the last 2 s) and no longer pans the follow-agent camera; edits behave as before. "Is this a read" comes from `mcp::tools::edits_doc`.
+
+### 2026-10-28: 0.9.0 (presentations, critique, several agents, lighter chats)
+
+* Export the screens as a presentation PDF: cover, contents and a slide per screen (File menu, palette, `--format presentation`, `export_presentation`).
+* `critique_design` reviews finished work through personas (vision, UX, flow, QA, visual, accessibility, copy, end-user); `design_lessons` keeps the findings the person approved for every later design.
+* Several agents on one project: named CLI agents (`--as`), conflict reports and `ifRev` on writes, `claim_nodes` / `release_claims` / `list_claims`, scoped delegate builds and per-agent undo. 94 tools.
+* Screens keep their device height and scroll what does not fit; sections group screens on the canvas; pages can carry icons.
+* UI kits for shadcn, MUI, Mantine, Bootstrap and Chakra; `write_html` follows the browser more closely.
+* Memory: chat screenshots move to files (kept 14 days), thumbnails stay small, canvas images decode only for the GPU. A long chat dropped from about 1 GB to about 400 MB.
+* Fix: long answer options in a question card wrap instead of widening the whole chat.
+
+### 2026-10-28: 0.8.4 (demo video, fresh screenshots)
+
+* README opens with a time-lapse of designing a habit-tracker app (`docs/demo.gif`, full video `docs/demo.mp4`), and every screenshot is retaken from that design.
+* Fix: the Performance overlay reads its settings from the project config folder, so `PAPERCRAFT_HOME` isolates it too.
+
+### 2026-10-28: 0.8.3 (menus, panel grips, narrow windows)
+
+* Menus open 6 px below their button, are at least 180 px wide and show an icon beside each item. Below 1100 px the hamburger opens every menu in one popup.
+* Side panels: a slim rounded grip in the gap resizes them (double click resets) and `/` hides or shows both. Panels size to their content and step aside when the window is too narrow for the canvas.
+* Sharper window buttons (pixel-snapped glyphs); the hamburger morphs into a back arrow while the File / Edit strip slides open and shut.
+* Fixes: lint Fix buttons no longer cover the issue text; review rows keep their badge and revert button off the hover edge; the Export button stays on one line.
+
+### 2026-10-28: 0.8.2 (refresh-rate cap, accurate RAM/CPU, toolbar)
+
+* Settings > Performance > Refresh rate is now a real frame cap applied live (30 to 360 Hz, 60 by default); **Auto** draws as fast as the display allows and is the default on macOS.
+* Performance Monitor: RAM is the private working set and CPU is divided by logical cores on Windows, so both match Task Manager.
+* Windows and Linux toolbar: a hamburger button shows or hides the File / Edit strip (remembered); the close button uses the theme's danger colour on the acrylic surface, with hover animation and a pointing-hand cursor.
+
+### 2026-10-28: 0.8.1 (Performance Monitor on Windows, Design tab speed)
+
+* Performance Monitor: RAM (working set) and CPU now show on Windows; before they were `n/a` there. The overlay and slow-frame badge judge frames against a refresh rate you pick in Settings > Performance: 30 to 360 Hz (144, 165 and 240 included), 60 by default.
+* Design tab: the lint and the open-feedback scan are kept until the document changes instead of walking every node each frame; a 12.6k-node page no longer drops to a few FPS while the tab is open.
+* Windows and Linux: window buttons in the toolbar; controls and acrylic come from craft-ui.
+
+### 2026-10-27: 0.8.0 (screens, packed pages, canvas performance)
+
+* Screens (`specs/screens`): a frame can be declared a screen or an overlay (drawer, modal, sheet) with `set_screens`, inspector toggles, or Declare N screens on the page. Agent edits to layers hidden behind an overlay or opaque layer in their screen are refused unless `allowHidden`; partly covered ones are reported with their visible share. `get_basic_info` lists screens, `get_node_info` tells a layer's visibility, `lint_design` skips hidden layers, On click lists screens, nested screens get canvas labels.
+* Pages and saved versions are stored as zstd-compressed compact JSON (`.json.zst`, project format 3): the 68 MB Interview project is 1.3 MB on disk and loads faster. A legacy `.json` still reads and is replaced on the next save; older builds refuse format 3.
+* Canvas performance: the editor logs which layers each change touched, so the scene cache re-tessellates only those and keeps no copy of the document (443 → 346 MB on Interview). Zoom holds its geometry until it settles, tiny layers are skipped, vertices are 24 bytes, redraws follow vsync, the idle pointer no longer redraws, freed memory goes back to the system when idle, off-screen frame labels are skipped, and the layers panel keeps its row heights between frames (panning a large page: ~9 → ~0.7 ms a frame).
+* Frame labels: none on frames too narrow on screen, overlapping labels are dropped (selected and larger frames first), long names end in an ellipsis.
+* Dropdowns size to their longest option (capped), keep the name of `A / B / Name` options, scroll when long and show the full name on hover.
+* Settings: a Performance page (FPS, frame time, RAM, CPU, GPU, optional overlay and slow-frame badge), agent logos in the agent rows; the Chat page no longer collapses.
+
+### 2026-10-25: 0.8.0 (properties modal, row merges)
+
+* Properties modal (inspector, on a main: Edit properties…): a tab per property (name, kind, layer, default, delete); an options property as a Table (a row per option, a column per layer · style; colours picked from the project's variables with a swatch, visibility as a toggle, icon, image and text fields; add or remove options and columns) or as Source (each option's changes as JSON with syntax highlighting); templates from the + Property menu; Save replaces them all. The inline list is name, kind and a remove button. The `status` template writes colours the project has as variables as those variables.
+* Canvas overlays (agent glow, agent outlines, review outlines, the agent-elsewhere chip) sit below modals instead of showing through them. Two segmented controls in one row no longer share widget ids (the second did not respond).
+* `merge_variants` keeps what each moved instance shows (texts that came from its old variant), carries nested instances' property values (a badge's Tone) in the options, and takes `textLayers` for state texts only.
+
+### 2026-10-25: 0.8.0 (prototype links tool, merge icon colours)
+
+* `set_prototype_links`: agents can link a layer to a screen for Present mode (back arrows, rows, tabs); before, only the inspector could.
+* `merge_variants` compares an icon's colour (keeping its variable), now that icons are compared as one unit.
+
+### 2026-10-25: 0.8.0 (data safety fixes)
+
+* Sync no longer throws away an instance written over a nested one the way older versions stored it (its own root, no swap): it becomes a swap of the main's child at that place. Projects built before #11 lost such layers (a checked checkbox, a status badge) on the next edit of the outer main.
+* Icons: the HTML parser keeps attribute names in lowercase, so the icon and plugin code never removed the SVG's original `viewBox`; since #14 that made `<x-paper-icon>` (and icon swaps) scale twice and come out tiny. Unreleased builds only; 0.7.3 is not affected.
+* `merge_variants` refuses merges that would leave one option and changes nothing then.
+
+### 2026-10-25: 0.8.0 (merge variants, pinned bars)
+
+* `merge_variants`: per-state components (`Tab bar / Hoje`, `Status / Sucesso`...) become one component with an enum property. Each variant's differences from the base, matched by layer path (fills, colours, borders, weight, radius, visibility, icons; texts on request), become its option; every instance on every page moves to the base with its value; the base takes the family name; the emptied variants are deleted. A variant with layers the base lacks is left out and reported.
+* Absolute children of a scrolling screen stay pinned while its content scrolls (a tab bar, a floating button), and do not add to the scroll range.
+
+### 2026-10-25: 0.8.0 (fix #22)
+
+* #22: a hug column whose width comes from its parent (stretched or filled) re-measures its height once its text wraps at that width, and the parent measures again, instead of keeping a height measured while the text was narrow.
+
+### 2026-10-25: 0.8.0 (component properties)
+
+* Main components carry properties (`Node.props`): `text`, `bool`, `icon` and `enum`, whose options are changes in the overrides format (a layer name, `$root` or a path `Item > Label`, to a text or styles that may hold `text`, `icon` and `visibility`), so they restyle, retext, show, hide or re-icon any layer. Instances keep their values (`prop_values`) as overrides, so main edits keep them; changing the definitions re-applies every instance on every page.
+* Templates: `text` (a property per text layer), `active` (which item of a tab bar, nav or segmented row is active) and `status` (options with a colour tint the root and colour the label and icon).
+* `set_component_props`, `set_instance_props`, `props='{...}'` on `<x-paper-instance>`, properties in `get_components`, `{Prop}` in code link templates, and an inspector section (a control per property on an instance; Texts and Active item templates on a main). Icon swaps keep the layer's name and link (a `children` override); recoloured icon shapes inside instances now keep their colour after a sync.
+* Spec: `specs/component-props/spec.md`.
+
+### 2026-10-25: 0.8.0 (scrolling screens)
+
+* Screens keep their size and scroll what overflows: a frame's `scroll` flag (CSS `overflow: auto|scroll`, `overflow-y-auto` in the Tailwind export) makes it fixed height and clipped, and the wheel over it scrolls its content on the canvas and in Present, then pans once the content ends. The offset is view state, not saved; hit-testing and overlays follow it.
+* Toggles: a scroll button beside each screen's name (shown when it scrolls, overflows, or is hovered), Scroll overflow on a selected frame, and Screens scroll on the page for every screen at once. A fit-content screen that grew goes back to its min height.
+* Agents are told to keep screens at their size and use overflow:auto for long content.
+
+### 2026-10-25: 0.8.0 (fixes #19, #20, #21)
+
+* #19: a clipping frame draws its border after its children, so filled rows no longer hide it. Clipping stays rectangular: children's square corners still show past a rounded border.
+* #20: a border written with a colour variable (`1px solid var(--border)`) binds the stroke, a literal one unbinds it, and reads show `var(--name)` for bound strokes. `find_nodes` says when it stopped at 500 matches.
+* #21: hug (fit-content) frames grow around growing children (CSS flex-basis auto) instead of staying at their min height; with short content the min height stays and the growing child fills it, so a pinned footer works.
+
+### 2026-10-25: 0.8.0 (follow agents, skeletons)
+
+* Follow agents (Settings, Chat; on by default): in Design and Chat mode alike, the canvas pans, and switches page, to the layer any agent changes (the chat, Claude Code over MCP, a helper). Moving the canvas yourself pauses it for a moment; off, a chip offers the agent's page.
+* Skeletons: the moment the model starts writing a tool call (`ToolDraft` and `ToolInput` events from Claude Code's stream and the API), the canvas shows a shimmering skeleton over its target, or where a new artboard goes, labelled with what it is doing, until the call has run. Fresh agent content comes in through a skeleton that fades, so an outline never frames empty space.
+* Top bar: the title and the agent pill leave room measured from the buttons on the right, so nothing overlaps at narrow widths.
+
+### 2026-10-25: 0.8.0 (fixes #14 to #18)
+
+* #14: `write_html` fits an `<svg>` viewBox into its width and height like a browser (centred, aspect kept).
+* #15: a note on something placed straight in its artboard treats the artboard as the screen: tidy keeps cards off it instead of in its free space.
+* #16: `not-an-instance` no longer matches look-alikes with a different icon; its fix names the `<x-paper-instance>` to write (swap_instance refuses plain frames).
+* #17: instances with their own text re-measure it when the main's text style changes, so hug buttons keep fitting their labels (also covers the size reset after a swap).
+* #18: `lint_design` reports borders whose hex equals a color variable, and `fix` binds them (`stroke`), so a mode switch reaches them.
+
 ### 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.
@@ -139,6 +251,17 @@ are credited in `THIRD_PARTY.md`.
 * Agent view lists runs and supports review, whole-run revert, per-change undo and accept all. Review adds canvas outlines and a finish toast; palette commands review or revert the last run.
 * Id-paired page diffs power review; History version compare UI remains todo. Completed with full restore (including variables, modes and tokens), tool edit classification, and wasm support.
 
+### 2026-10-28: Incremental lint
+
+Agents re-linted whole pages after every edit. A bounded change log (`src/editor/changes.rs`, the
+last 10000 layer changes, each with the client that made it) now records which layers every
+commit, undo and redo touched; variable edits count as touching the layers bound to them. Write
+tools return `rev`, `get_basic_info` reports it, `get_changes {since, page?, exclude?, by?}` lists
+the changes, and `lint_design {since | changedOnly}` checks only those layers plus their
+ancestors, descendants and siblings (`src/mcp/incremental.rs`), with `covered` saying what ran.
+Tests: `tests/incremental.rs`, unit tests next to the code, the attribution check in
+`tests/mcp.rs::parallel_sessions_get_their_own_runs`.
+
 ### 2026-10-21: Design libraries and offline CLI
 
 * Variables are now collections with modes, slash groups, four types and aliases, stored in `design.json` (project version 2; 0.4.0 refuses migrated projects). Each page keeps a derived cache; every edit is one undo step per page, and the page-only removals are gone. Per-page active mode per collection.

README.md

@@ -12,9 +12,17 @@
 </p>
 
 <p align="center">
-  <img src="docs/screenshots/readme/hero.png" alt="PaperCraft with the Redesign page of its own design file open: layers on the left, the canvas in the middle, the inspector on the right" width="100%">
+  <a href="docs/demo.mp4">
+    <img src="docs/demo.gif" alt="PaperCraft's editor breaking apart into its panels on a 3D stage, then chat mode with an agent, annotations pinned to a sign-in screen, the variables and libraries modals, and the home screen" width="100%">
+  </a>
   <br>
-  <sub>PaperCraft editing its own design file (the Redesign page).</sub>
+  <sub>A 40-second tour: the editor, chat mode, annotations and design systems. <a href="docs/demo.mp4">Watch it in full quality</a> (MP4, 1080p60). Rendered from the app's own design with <a href="tools/showcase/README.md">tools/showcase</a>, not screen-recorded.</sub>
+</p>
+
+<p align="center">
+  <img src="docs/screenshots/readme/hero.png" alt="PaperCraft with the Pulse habit-tracker design open: layers on the left, three annotated screens on the canvas, the inspector on the right" width="100%">
+  <br>
+  <sub>The finished design: three screens, a component library and annotations.</sub>
 </p>
 
 <p align="center">
@@ -25,6 +33,7 @@
   <a href="#keyboard-shortcuts">Shortcuts</a> ·
   <a href="#status-and-known-limits">Status</a> ·
   <a href="#project-layout">Layout</a> ·
+  <a href="#contributing">Contributing</a> ·
   <a href="#license-and-credits">License</a>
 </p>
 
@@ -54,11 +63,11 @@ It is an alpha. It is good for building screens, design systems and sketches, an
 <table>
   <tr>
     <td width="50%" valign="top">
-      <img src="docs/screenshots/readme/variables.png" alt="The Variables modal listing the Colors collection with four Catppuccin modes" width="100%">
-      <br><sub>Variables modal: collections, modes, groups and types.</sub>
+      <img src="docs/screenshots/readme/variables.png" alt="The Variables modal listing the Colors collection with Light and Dark values for brand, foreground and background" width="100%">
+      <br><sub>Variables modal: collections, modes (Light and Dark here), groups and types.</sub>
     </td>
     <td width="50%" valign="top">
-      <img src="docs/screenshots/readme/assets.png" alt="The Assets side view with components grouped into Banner, Button, Checkbox and Chip" width="100%">
+      <img src="docs/screenshots/readme/assets.png" alt="The Assets side view with the Habit row component and the shadcn Button kit" width="100%">
       <br><sub>Assets view: components grouped by name, from this file and enabled libraries.</sub>
     </td>
   </tr>
@@ -68,12 +77,32 @@ It is an alpha. It is good for building screens, design systems and sketches, an
       <br><sub>Libraries modal: publish this file, enable libraries from the store.</sub>
     </td>
     <td width="50%" valign="top">
-      <img src="docs/screenshots/readme/agent-review.png" alt="Agent run review: a list of changes on the left and green, yellow and red outlines on the canvas" width="100%">
-      <br><sub>Review of an agent run, with the changes outlined on the canvas.</sub>
+      <img src="docs/screenshots/readme/agent-review.png" alt="The Agent side view: a finished run with Review and Revert, above the list of earlier runs" width="100%">
+      <br><sub>The Agent view: runs, with Review and Revert for the latest.</sub>
     </td>
   </tr>
 </table>
 
+<table>
+  <tr>
+    <td width="50%" valign="top">
+      <img src="docs/screenshots/readme/toolbar-panels.png" alt="The toolbar with the menu button folded into a back arrow, the File to Theme strip, and the floating left panel with a slim grip beside it" width="100%">
+      <br><sub>Toolbar and panels: the menu button shows or hides the File / Edit strip, and the slim grip between a panel and the canvas resizes it (double click resets). <kbd>/</kbd> hides both side panels.</sub>
+    </td>
+    <td width="50%" valign="top">
+      <img src="docs/screenshots/readme/menus.png" alt="The Tools menu open below its button, each tool with an icon and its shortcut" width="100%">
+      <br><sub>Menus open a little below their button, are never narrower than 180 px, and carry an icon next to each item.</sub>
+    </td>
+  </tr>
+  <tr>
+    <td width="50%" valign="top">
+      <img src="docs/screenshots/readme/annotations.png" alt="The Annotations side view listing five notes grouped by screen, with their category chips" width="100%">
+      <br><sub>Annotations: rule, behaviour, content and data notes, kept with the layers they describe.</sub>
+    </td>
+    <td width="50%" valign="top"></td>
+  </tr>
+</table>
+
 **Prototyping and review**
 * Prototype links between frames and a **Present** mode that follows them with a camera glide.
 * **Version history:** named and automatic saved versions per page; rename, restore, delete.
@@ -90,7 +119,7 @@ Projects live in `~/Documents/papercraft/<name>/` (`project.json`, one JSON file
 
 ## Agents
 
-PaperCraft exposes **70 tools** (`src/mcp/tools.rs`) three ways. They run the same code inside the app.
+PaperCraft exposes **94 tools** (`src/mcp/tools.rs`) three ways. They run the same code inside the app.
 
 | Way | Notes |
 |---|---|
@@ -141,7 +170,7 @@ EOF
 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%">
+  <img src="docs/screenshots/readme/chat.png" alt="Chat mode: the chat column on the left with its starter prompts and input, and the annotated habit-tracker screens on the canvas" width="100%">
   <br>
   <sub>Chat mode with a finished agent run: review, revert or accept from the thread.</sub>
 </p>
@@ -162,12 +191,13 @@ Releases are at [git.marcospaulo.dev.br/masi/papercraft/releases](https://git.ma
 
 | Build | File |
 |---|---|
+| Windows (x86_64) | `papercraft-<version>-windows-x86_64.zip`: `papercraft.exe`, `papercraft-cli.exe`, `papercraft-mcp.exe` and `papercraft-relay.exe` |
 | macOS app (Apple silicon) | `papercraft-<version>-macos-aarch64.dmg`, containing PaperCraft.app |
 | Linux (x86_64) | `papercraft-<version>-linux-x86_64.tar.gz` |
 | CLI tools, macOS | `papercraft-cli-<version>-macos-aarch64.zip` (`papercraft-cli` and `papercraft-mcp`) |
 | CLI tools, Linux | `papercraft-cli-<version>-linux-x86_64.tar.gz` |
 
-macOS builds are ad-hoc signed and not notarized: on another Mac, open the app with right-click, Open the first time. Other platforms: build from source.
+The Windows zip is unsigned: SmartScreen may ask you to confirm the first run. macOS builds are ad-hoc signed and not notarized: on another Mac, open the app with right-click, Open the first time. Other platforms: build from source.
 
 ### Build from source
 
@@ -179,9 +209,10 @@ cargo build --release
 ./target/release/papercraft ~/Documents/papercraft/Plan    # open a project folder
 ./target/release/papercraft --demo                         # sample board, not saved
 packaging/macos/package.sh                                 # PaperCraft.app on a DMG, plus the CLI zip
+pwsh packaging/windows/package.ps1 -Arch x64               # Windows MSI, portable zip and CLI zip
 ```
 
-**macOS:** `xcode-select --install` for the linker. It renders with Metal through wgpu. **Linux:** install the runtime libraries (`sudo apt install libxkbcommon-x11-0 libxcb-xkb1 libvulkan1 mesa-vulkan-drivers` on Debian or Ubuntu); it runs on X11 and Wayland through Vulkan, with GL as a fallback, and file dialogs use the XDG desktop portal.
+**Windows:** the MSVC build tools (Visual Studio Build Tools with the C++ workload); it renders through wgpu on Direct3D 12 or Vulkan. **macOS:** `xcode-select --install` for the linker. It renders with Metal through wgpu. **Linux:** install the runtime libraries (`sudo apt install libxkbcommon-x11-0 libxcb-xkb1 libvulkan1 mesa-vulkan-drivers` on Debian or Ubuntu); it runs on X11 and Wayland through Vulkan, with GL as a fallback, and file dialogs use the XDG desktop portal.
 
 To run a test copy next to your own PaperCraft, give it its own folder and port: `PAPERCRAFT_HOME=/tmp/pc PAPERCRAFT_MCP_PORT=29998 ./target/release/papercraft`.
 
@@ -228,24 +259,32 @@ The scenario tests drive the editor with synthetic pointer and keyboard events;
 | 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 |
+| / | Show / hide the side panels | | |
 
 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.
 
 ## Status and known limits
 
 > [!IMPORTANT]
-> PaperCraft is an alpha. [ROADMAP.md](ROADMAP.md) lists what is missing, in priority order; [PROGRESS.md](PROGRESS.md) is the milestone history.
+> PaperCraft is an alpha (0.8.x). It is used daily on Windows, macOS and Linux, and its project format migrates forward, but expect rough edges. [ROADMAP.md](ROADMAP.md) lists what is missing, in priority order; [PROGRESS.md](PROGRESS.md) is the milestone history.
+
+**What works today:** local projects with pages and an infinite canvas; vector, text, auto layout, constraints, components with properties and variants, variables with modes, shared libraries and screens; annotations, lint and Present mode with prototype links; PNG, SVG and PDF export, Figma, Excalidraw and SVG import; the MCP server, CLI (online and offline) and in-app Chat mode; versions, agent runs with review and revert; optional relay sync.
+
+**Known limits:**
 
 * **No rotation.** Nodes are axis-aligned boxes; rotated imports are baked into paths.
 * **Text** is Inter only, with no shaping (no ligatures, complex scripts or RTL) and no font picker. IME preedit is not displayed.
 * **Rendering approximations.** Frame clipping is a rectangle, so rounded frames do not clip their corners. Overlay, dodge, burn, light, difference, exclusion and the HSL blend modes draw as Normal on the canvas (they are exact in SVG and PDF export). There is no layer or background blur yet.
-* **Variables:** gradients keep fixed colors; there are no text styles and no component variants or properties.
+* **Variables:** gradients keep fixed colors, and there are no text styles (named type styles that update every text layer).
+* **Components:** properties (text, bool, icon, enum) and `merge_variants` cover variants, but there are no variant sets with several axes, and no responsive per-breakpoint variants or breakpoint preview.
+* **Prototyping:** links and flows work in Present mode; smart transitions and hover states do not exist yet.
 * **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, 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.
+* **Scale:** cached frames take about 3 ms at 10k visible nodes; it is not tuned for 100k.
 * **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.
+* **Releases:** the Gitea releases ship Windows x86_64. macOS (Apple silicon) and Linux x86_64 builds come from the packaging scripts or from source; the Linux build is not part of the current releases.
 * **PDF** text is outlined, not selectable.
 
 ## Project layout
@@ -280,12 +319,14 @@ analysis/            behaviour studies and design notes
 docs/                guides and screenshots
 ```
 
+## Contributing
+
+Bug reports, ideas, docs and code are welcome. Start with [CONTRIBUTING.md](CONTRIBUTING.md) (setup, checks, commits, pull requests) and [RULES.md](RULES.md) (the standards a change is reviewed against). AI tools are welcome under the [AI usage policy](AI_POLICY.md), and coding agents should read [AGENTS.md](AGENTS.md). Security problems go through [SECURITY.md](SECURITY.md), and everyone follows the [Code of Conduct](CODE_OF_CONDUCT.md).
+
 ## License and credits
 
 Dual-licensed under [MIT](LICENSE-MIT) or [Apache-2.0](LICENSE-APACHE), at your option.
 
 Bundled fonts, icons and other assets keep their own licenses, listed in [ATTRIBUTION.md](ATTRIBUTION.md); Inter is © The Inter Project Authors under the SIL Open Font License 1.1. Some algorithms are ported from MIT-licensed projects (rough.js, hachure-fill, Excalidraw), with their notices in [THIRD_PARTY.md](THIRD_PARTY.md).
 
-PaperCraft started as a fork of [storytold/paper-rs](https://git.marcospaulo.dev.br/storytold/paper-rs); thanks to its contributors.
-
 PaperCraft is not affiliated with Paper or Figma. Only the behavior of other tools was studied (see `analysis/`); no code, shaders, icons, fonts or assets from them are used.

ROADMAP.md

@@ -22,7 +22,7 @@ Agents can already build a design system through the tools; people cannot yet ma
 | 7 | Number variables (spacing, radius, font size) | **done** | Typed number, string and boolean variables in collections with groups and aliases; fields bind (size, gap, padding, radius, font size, line height, opacity, visible) in the inspector and `bind_variable`. |
 | 8 | Stroke and gradient variables | partial | Strokes and text color bind to variables. Gradients keep fixed colors. |
 | 9 | Text styles | todo | Named type styles (Heading / L, Body / M) that update every text layer using them. |
-| 10 | Component variants and properties | todo | Variants are separate components named `Button / Primary`. No variant sets, boolean or text properties. |
+| 10 | Component variants and properties | partial | Main components carry text, bool, icon and enum properties; `merge_variants` folds per-state components into one. Multi-axis variant sets todo. |
 | 11 | Swap instance, detach by tool | **done** | `swap_instance` and `detach_instance` MCP tools + UI controls in Design tab and Inspector. |
 | 12 | Cross-page and shared libraries | **done** | Instances of a main on any page; publish a project as a library, enable, update and disable it in other projects (Libraries modal, `*_library` tools). |
 | 13 | Export design tokens | **done** | `get_variables` with `format: "css"`, `"json"` (W3C DTCG + modes), or `"tailwind"` (`@theme`), plus Copy CSS / Copy JSON in Design tab. |
@@ -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** | 67 tools, one undo step per call, `papercraft-cli list / schema / call`. |
+| 15 | MCP and CLI tool surface | **done** | 94 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). |
@@ -63,7 +63,7 @@ Agents can already build a design system through the tools; people cannot yet ma
 | 32 | Version history | **done** | Named and automatic saved versions per page: rename, restore, delete (`src/versions.rs`). Restore now includes variables and tokens; compare todo. |
 | 33 | Figma import | **done** | `.fig` binary (Kiwi wire decoder + zip container) and Figma REST API JSON import in `src/io/fig.rs`. |
 | 34 | Multiplayer and cloud sync | partial | Optional shared projects through the relay, with merge of concurrent edits to one page. Presence (cursors, selections) todo. Stays optional. |
-| 35 | Release builds for Windows and Linux | todo | The code targets them (wgpu, winit); packaging today is the macOS DMG and the `papercraft-cli` zip. |
+| 35 | Release builds for Windows and Linux | partial | Windows x86_64 zips are released and `packaging/windows` builds an MSI; Linux x86_64 is not in the current releases. |
 
 ## 5. Quality
 

RULES.md

@@ -0,0 +1,124 @@
+# Project rules
+
+The standards a change to PaperCraft is reviewed against. [CONTRIBUTING.md](CONTRIBUTING.md) says
+*how* to contribute; this file says what "good" means once you do.
+
+## Code
+
+* **Immutability by default.** Return new values rather than mutating shared ones. The document
+  changes only through `Editor::edit` transactions, so every change is one undo step and is
+  logged for the renderer's cache.
+* **Small, cohesive files.** 200 to 400 lines is typical, 800 is the ceiling. Split a module
+  when it grows past that. Organise by feature, not by type.
+* **Small functions** (under about 50 lines) and **shallow nesting** (4 levels at most); prefer
+  early returns.
+* **Handle errors explicitly.** No `unwrap` or `expect` on user data, file contents or network
+  responses; return an error with context. User-facing messages are plain language. Never swallow
+  an error silently.
+* **Validate at the boundaries**: MCP and CLI arguments, imported files (Figma, Excalidraw, SVG),
+  project folders and anything from the relay. Treat all of it as untrusted.
+* **No magic numbers.** Name thresholds, sizes and delays as constants; UI measures come from
+  the theme, not literals.
+* **No dead code, debug prints or commented-out blocks.**
+* **Match the surrounding code**: naming, comment density and idiom. Comments say *why*, not
+  *what*.
+* **Cross-platform.** Code must build on Windows, macOS and Linux and for the web build. Gate
+  platform code with `cfg`, never assume a path separator, and keep
+  `cargo check --lib --target wasm32-unknown-unknown` passing.
+* **No new heavy dependencies** without discussion. Licenses must be compatible with
+  MIT OR Apache-2.0.
+
+## Data safety
+
+PaperCraft stores people's work. Treat that as the top priority.
+
+* Writes to projects are **atomic** (`src/io/atomic.rs`); never write a page in place.
+* A format change bumps the project format version and ships a **migration**. Older data must
+  keep loading; if a newer build writes something an older build cannot read, the older build
+  must refuse it rather than corrupt it.
+* Agent and sync paths must never silently drop a layer. If a merge cannot keep something, it
+  says so.
+* Destructive operations are undoable or confirmed.
+
+## The machine you build on
+
+A full build links dozens of test binaries and can take every core and most of the memory of a
+desktop. The person is still using that machine, often with PaperCraft open.
+
+* **Look before you start a heavy job.** Check free memory, CPU load and what is already running
+  (the person's app, another build, other agents) before a build, a test run or clippy. If memory
+  is short or another build is running, wait or ask.
+* **Leave room.** Cap Cargo at no more than half the logical cores (`-j 8` on a 16-core machine,
+  or `CARGO_BUILD_JOBS`) and the test harness too (`-- --test-threads=8`); go lower when memory is
+  tight, since linking is the peak. Run heavy jobs at low priority (`nice` on Unix, Below Normal
+  on Windows).
+* **One heavy job at a time.** Never run two builds side by side; each target folder is a full
+  build's worth of compiling and linking. Parallel agents share one build slot.
+* **Run only what your change touches**: `cargo check` before `cargo build`, and the test
+  targets that cover the change (`cargo test --test incremental`). The full suite runs on GitHub
+  before a release (`.github/workflows/test.yml`), not on a desktop.
+* **Stop at once when the person says the machine is slow**, then check that no compiler, linker
+  or test process of yours is left over.
+* **Clean up.** Delete scratch target folders and worktrees when the work is merged; they grow to
+  tens of gigabytes.
+
+## UI
+
+* **Use the shared kit.** Controls, spacing, radii and colours come from `craft-ui`,
+  `src/ui/kit.rs` and the theme, not one-off values.
+* **Themes are modes.** Never branch on a theme to pick a colour; read it from the palette.
+* **Nothing overflows.** Text truncates with an ellipsis, rows keep their controls inside, and
+  panels step aside when the window is narrow. Test a narrow window and a long name.
+* **Targets are at least 44 px** where a person taps or clicks something primary, and everything
+  stays reachable from the keyboard.
+* **Animation is quick**: eased and under about 200 ms.
+* **Look at it.** Run the app or take a headless screenshot at scale 2 in each theme before
+  asking for review.
+
+## MCP tools and the CLI
+
+* Every tool has a precise one-paragraph description, a JSON schema, and a test in
+  `tests/mcp.rs`. Descriptions are what agents read; write them for that reader.
+* One tool call is **one undo step**.
+* Names follow the existing style (`verb_noun`); reuse Paper-compatible names and argument shapes
+  where a tool has an equivalent.
+* Reads are cheap and bounded. Apply limits, say when you truncated, and never dump the whole
+  document by default.
+* The server listens on `127.0.0.1` only and rejects browser origins. Do not weaken that.
+* Update the tool count in the README and ROADMAP when you add or remove tools.
+
+## Designs made with PaperCraft
+
+These apply to example projects, templates, screenshots and skill content, and are what the
+skills teach agents.
+
+* **Themes are modes, never copies.** One artboard per screen. Give each variable a value per
+  mode and switch with `set_variable_mode`. Do not duplicate artboards as light and dark
+  versions, and do not rebuild a screen with other hex values.
+* **Every colour on a themed screen is a variable** (`var(--name)` on fills and text,
+  `bind_variable` for strokes, opacity and visibility).
+* **No overflow.** Nothing extends past its artboard or clipping parent. Scrolling content is
+  clipped inside its viewport and ends above floating bars with a fade and enough bottom padding.
+  Check each artboard programmatically *and* by screenshot in each mode.
+* Run `lint_design` and fix what it finds: contrast, the 4 px grid, type scale and 44 px targets.
+* Release design locks when you are done so the next person can inspect.
+
+## Skills and docs
+
+* Skills in [skills/](skills) are instructions for agents: keep them short, concrete and tested
+  against a real agent run.
+* Docs describe what exists. Do not document a feature that is not built; put plans in `specs/`
+  and `ROADMAP.md`.
+* Screenshots come from the real app, in a generic project (no personal data, local paths or
+  real names), at scale 2.
+
+## Branding
+
+The tree carries no third-party product branding; `.debrand` lists the forbidden patterns and
+the check runs in review. Describe PaperCraft on its own terms.
+
+## Review
+
+A pull request is approved when it has no unresolved **critical** (security, data loss) or
+**high** (bug, significant quality) issues. Medium and low comments are suggestions the author
+may take or decline with a reason.

SECURITY.md

@@ -0,0 +1,29 @@
+# Security policy
+
+## Supported versions
+
+PaperCraft is an alpha. Security fixes go into the latest release and `main`.
+
+## Reporting a vulnerability
+
+**Do not open a public issue.** Use GitHub's private reporting: the repository's **Security** tab,
+then **Report a vulnerability**. Include what you found, how to reproduce it, the version and
+platform, and the impact you see.
+
+A maintainer will acknowledge the report as soon as they can and keep you updated until it is
+fixed. Please give us reasonable time to ship a fix before you disclose it.
+
+## What counts
+
+Examples of in-scope problems:
+
+* the MCP server (`127.0.0.1:29980`) or the relay accepting requests it should reject (browser
+  origins, DNS rebinding, another user);
+* path traversal or code execution from a project folder, a design-system or library package, or
+  an imported Figma, Excalidraw or SVG file;
+* the updater or the agent integrations running something the person did not approve;
+* secrets (an Anthropic key, relay tokens) written to disk unprotected or sent anywhere
+  unexpected.
+
+PaperCraft is local-first: it opens no network port beyond the loopback MCP server unless you
+share a project through a relay you run.

THIRD_PARTY.md

@@ -62,3 +62,15 @@ SOFTWARE.
 
 Inter 4.1 by Rasmus Andersson — SIL Open Font License 1.1 (`assets/fonts/OFL.txt`). Regular,
 Medium, SemiBold, Bold, Italic and Bold Italic are bundled unmodified.
+
+Interface fonts, SIL Open Font License 1.1, Latin subsets as served by Google Fonts, unmodified
+(Regular, Medium, SemiBold): Geist (Vercel), Figtree (Erik Kennedy), DM Sans (Colophon Foundry,
+Jonny Pinhorn), Nunito (Vernon Adams, Jacques Le Bailly), Quicksand (Andrew Paglinawan),
+Fredoka (Milena Brandão, Hafontia), and Varela Round (Joe Prince; Regular only).
+
+## Agent logos
+
+`assets/agents/*.svg` are from LobeHub Icons (`@lobehub/icons-static-svg` 1.80.0,
+https://github.com/lobehub/lobe-icons), MIT License, Copyright (c) LobeHub. They are used
+unmodified to identify each coding agent in the agent picker. The marks are trademarks of
+their owners.

assets/agents/claude.svg

@@ -0,0 +1 @@
+<svg fill="currentColor" fill-rule="evenodd" height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Claude</title><path d="M4.709 15.955l4.72-2.647.08-.23-.08-.128H9.2l-.79-.048-2.698-.073-2.339-.097-2.266-.122-.571-.121L0 11.784l.055-.352.48-.321.686.06 1.52.103 2.278.158 1.652.097 2.449.255h.389l.055-.157-.134-.098-.103-.097-2.358-1.596-2.552-1.688-1.336-.972-.724-.491-.364-.462-.158-1.008.656-.722.881.06.225.061.893.686 1.908 1.476 2.491 1.833.365.304.145-.103.019-.073-.164-.274-1.355-2.446-1.446-2.49-.644-1.032-.17-.619a2.97 2.97 0 01-.104-.729L6.283.134 6.696 0l.996.134.42.364.62 1.414 1.002 2.229 1.555 3.03.456.898.243.832.091.255h.158V9.01l.128-1.706.237-2.095.23-2.695.08-.76.376-.91.747-.492.584.28.48.685-.067.444-.286 1.851-.559 2.903-.364 1.942h.212l.243-.242.985-1.306 1.652-2.064.73-.82.85-.904.547-.431h1.033l.76 1.129-.34 1.166-1.064 1.347-.881 1.142-1.264 1.7-.79 1.36.073.11.188-.02 2.856-.606 1.543-.28 1.841-.315.833.388.091.395-.328.807-1.969.486-2.309.462-3.439.813-.042.03.049.061 1.549.146.662.036h1.622l3.02.225.79.522.474.638-.079.485-1.215.62-1.64-.389-3.829-.91-1.312-.329h-.182v.11l1.093 1.068 2.006 1.81 2.509 2.33.127.578-.322.455-.34-.049-2.205-1.657-.851-.747-1.926-1.62h-.128v.17l.444.649 2.345 3.521.122 1.08-.17.353-.608.213-.668-.122-1.374-1.925-1.415-2.167-1.143-1.943-.14.08-.674 7.254-.316.37-.729.28-.607-.461-.322-.747.322-1.476.389-1.924.315-1.53.286-1.9.17-.632-.012-.042-.14.018-1.434 1.967-2.18 2.945-1.726 1.845-.414.164-.717-.37.067-.662.401-.589 2.388-3.036 1.44-1.882.93-1.086-.006-.158h-.055L4.132 18.56l-1.13.146-.487-.456.061-.746.231-.243 1.908-1.312-.006.006z"></path></svg>

assets/agents/cline.svg

@@ -0,0 +1 @@
+<svg fill="currentColor" fill-rule="evenodd" height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Cline</title><path d="M17.035 3.991c2.75 0 4.98 2.24 4.98 5.003v1.667l1.45 2.896a1.01 1.01 0 01-.002.909l-1.448 2.864v1.668c0 2.762-2.23 5.002-4.98 5.002H7.074c-2.751 0-4.98-2.24-4.98-5.002V17.33l-1.48-2.855a1.01 1.01 0 01-.003-.927l1.482-2.887V8.994c0-2.763 2.23-5.003 4.98-5.003h9.962zM8.265 9.6a2.274 2.274 0 00-2.274 2.274v4.042a2.274 2.274 0 004.547 0v-4.042A2.274 2.274 0 008.265 9.6zm7.326 0a2.274 2.274 0 00-2.274 2.274v4.042a2.274 2.274 0 104.548 0v-4.042A2.274 2.274 0 0015.59 9.6z"></path><path d="M12.054 5.558a2.779 2.779 0 100-5.558 2.779 2.779 0 000 5.558z"></path></svg>

assets/agents/cursor.svg

@@ -0,0 +1 @@
+<svg fill="currentColor" fill-rule="evenodd" height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Cursor</title><path d="M22.106 5.68L12.5.135a.998.998 0 00-.998 0L1.893 5.68a.84.84 0 00-.419.726v11.186c0 .3.16.577.42.727l9.607 5.547a.999.999 0 00.998 0l9.608-5.547a.84.84 0 00.42-.727V6.407a.84.84 0 00-.42-.726zm-.603 1.176L12.228 22.92c-.063.108-.228.064-.228-.061V12.34a.59.59 0 00-.295-.51l-9.11-5.26c-.107-.062-.063-.228.062-.228h18.55c.264 0 .428.286.296.514z"></path></svg>

assets/agents/gemini.svg

@@ -0,0 +1 @@
+<svg fill="currentColor" fill-rule="evenodd" height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Gemini</title><path d="M20.616 10.835a14.147 14.147 0 01-4.45-3.001 14.111 14.111 0 01-3.678-6.452.503.503 0 00-.975 0 14.134 14.134 0 01-3.679 6.452 14.155 14.155 0 01-4.45 3.001c-.65.28-1.318.505-2.002.678a.502.502 0 000 .975c.684.172 1.35.397 2.002.677a14.147 14.147 0 014.45 3.001 14.112 14.112 0 013.679 6.453.502.502 0 00.975 0c.172-.685.397-1.351.677-2.003a14.145 14.145 0 013.001-4.45 14.113 14.113 0 016.453-3.678.503.503 0 000-.975 13.245 13.245 0 01-2.003-.678z"></path></svg>

assets/agents/githubcopilot.svg

@@ -0,0 +1 @@
+<svg fill="currentColor" fill-rule="evenodd" height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>GithubCopilot</title><path d="M19.245 5.364c1.322 1.36 1.877 3.216 2.11 5.817.622 0 1.2.135 1.592.654l.73.964c.21.278.323.61.323.955v2.62c0 .339-.173.669-.453.868C20.239 19.602 16.157 21.5 12 21.5c-4.6 0-9.205-2.583-11.547-4.258-.28-.2-.452-.53-.453-.868v-2.62c0-.345.113-.679.321-.956l.73-.963c.392-.517.974-.654 1.593-.654l.029-.297c.25-2.446.81-4.213 2.082-5.52 2.461-2.54 5.71-2.851 7.146-2.864h.198c1.436.013 4.685.323 7.146 2.864zm-7.244 4.328c-.284 0-.613.016-.962.05-.123.447-.305.85-.57 1.108-1.05 1.023-2.316 1.18-2.994 1.18-.638 0-1.306-.13-1.851-.464-.516.165-1.012.403-1.044.996a65.882 65.882 0 00-.063 2.884l-.002.48c-.002.563-.005 1.126-.013 1.69.002.326.204.63.51.765 2.482 1.102 4.83 1.657 6.99 1.657 2.156 0 4.504-.555 6.985-1.657a.854.854 0 00.51-.766c.03-1.682.006-3.372-.076-5.053-.031-.596-.528-.83-1.046-.996-.546.333-1.212.464-1.85.464-.677 0-1.942-.157-2.993-1.18-.266-.258-.447-.661-.57-1.108-.32-.032-.64-.049-.96-.05zm-2.525 4.013c.539 0 .976.426.976.95v1.753c0 .525-.437.95-.976.95a.964.964 0 01-.976-.95v-1.752c0-.525.437-.951.976-.951zm5 0c.539 0 .976.426.976.95v1.753c0 .525-.437.95-.976.95a.964.964 0 01-.976-.95v-1.752c0-.525.437-.951.976-.951zM7.635 5.087c-1.05.102-1.935.438-2.385.906-.975 1.037-.765 3.668-.21 4.224.405.394 1.17.657 1.995.657h.09c.649-.013 1.785-.176 2.73-1.11.435-.41.705-1.433.675-2.47-.03-.834-.27-1.52-.63-1.813-.39-.336-1.275-.482-2.265-.394zm6.465.394c-.36.292-.6.98-.63 1.813-.03 1.037.24 2.06.675 2.47.968.957 2.136 1.104 2.776 1.11h.044c.825 0 1.59-.263 1.995-.657.555-.556.765-3.187-.21-4.224-.45-.468-1.335-.804-2.385-.906-.99-.088-1.875.058-2.265.394zM12 7.615c-.24 0-.525.015-.84.044.03.16.045.336.06.526l-.001.159a2.94 2.94 0 01-.014.25c.225-.022.425-.027.612-.028h.366c.187 0 .387.006.612.028-.015-.146-.015-.277-.015-.409.015-.19.03-.365.06-.526a9.29 9.29 0 00-.84-.044z"></path></svg>

assets/agents/goose.svg

@@ -0,0 +1 @@
+<svg fill="currentColor" fill-rule="evenodd" height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Goose</title><path d="M21.595 23.61c1.167-.254 2.405-.944 2.405-.944l-2.167-1.784a12.124 12.124 0 01-2.695-3.131 12.127 12.127 0 00-3.97-4.049l-.794-.462a1.115 1.115 0 01-.488-.815.844.844 0 01.154-.575c.413-.582 2.548-3.115 2.94-3.44.503-.416 1.065-.762 1.586-1.159.074-.056.148-.112.221-.17.003-.002.007-.004.009-.007.167-.131.325-.272.45-.438.453-.524.563-.988.59-1.193-.061-.197-.244-.639-.753-1.148.319.02.705.272 1.056.569.235-.376.481-.773.727-1.171.165-.266-.08-.465-.086-.471h-.001V3.22c-.007-.007-.206-.25-.471-.086-.567.35-1.134.702-1.639 1.021 0 0-.597-.012-1.305.599a2.464 2.464 0 00-.438.45l-.007.009c-.058.072-.114.147-.17.221-.397.521-.743 1.083-1.16 1.587-.323.391-2.857 2.526-3.44 2.94a.842.842 0 01-.574.153 1.115 1.115 0 01-.815-.488l-.462-.794a12.123 12.123 0 00-4.049-3.97 12.133 12.133 0 01-3.13-2.695L1.332 0S.643 1.238.39 2.405c.352.428 1.27 1.49 2.34 2.302C1.58 4.167.73 3.75.06 3.4c-.103.765-.063 1.92.043 2.816.726.317 1.961.806 3.219 1.066-1.006.236-2.11.278-2.961.262.15.554.358 1.119.64 1.688.119.263.25.52.39.77.452.125 2.222.383 3.164.171l-2.51.897a27.776 27.776 0 002.544 2.726c2.031-1.092 2.494-1.241 4.018-2.238-2.467 2.008-3.108 2.828-3.8 3.67l-.483.678c-.25.351-.469.725-.65 1.117-.61 1.31-1.47 4.1-1.47 4.1-.154.486.202.842.674.674 0 0 2.79-.861 4.1-1.47.392-.182.766-.4 1.118-.65l.677-.483c.227-.187.453-.37.701-.586 0 0 1.705 2.02 3.458 3.349l.896-2.511c-.211.942.046 2.712.17 3.163.252.142.509.272.772.392.569.28 1.134.49 1.688.64-.016-.853.026-1.956.261-2.962.26 1.258.75 2.493 1.067 3.219.895.106 2.051.146 2.816.043a73.87 73.87 0 01-1.308-2.67c.811 1.07 1.874 1.988 2.302 2.34h-.001z"></path></svg>

assets/agents/kimi.svg

@@ -0,0 +1 @@
+<svg fill="currentColor" fill-rule="evenodd" height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Kimi</title><path d="M21.846 0a1.923 1.923 0 110 3.846H20.15a.226.226 0 01-.227-.226V1.923C19.923.861 20.784 0 21.846 0z"></path><path d="M11.065 11.199l7.257-7.2c.137-.136.06-.41-.116-.41H14.3a.164.164 0 00-.117.051l-7.82 7.756c-.122.12-.302.013-.302-.179V3.82c0-.127-.083-.23-.185-.23H3.186c-.103 0-.186.103-.186.23V19.77c0 .128.083.23.186.23h2.69c.103 0 .186-.102.186-.23v-3.25c0-.069.025-.135.069-.178l2.424-2.406a.158.158 0 01.205-.023l6.484 4.772a7.677 7.677 0 003.453 1.283c.108.012.2-.095.2-.23v-3.06c0-.117-.07-.212-.164-.227a5.028 5.028 0 01-2.027-.807l-5.613-4.064c-.117-.078-.132-.279-.028-.381z"></path></svg>

assets/agents/mistral.svg

@@ -0,0 +1 @@
+<svg fill="currentColor" fill-rule="evenodd" height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Mistral</title><path clip-rule="evenodd" d="M3.428 3.4h3.429v3.428h3.429v3.429h-.002 3.431V6.828h3.427V3.4h3.43v13.714H24v3.429H13.714v-3.428h-3.428v-3.429h-3.43v3.428h3.43v3.429H0v-3.429h3.428V3.4zm10.286 13.715h3.428v-3.429h-3.427v3.429z"></path></svg>

assets/agents/openai.svg

@@ -0,0 +1 @@
+<svg fill="currentColor" fill-rule="evenodd" height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>OpenAI</title><path d="M9.205 8.658v-2.26c0-.19.072-.333.238-.428l4.543-2.616c.619-.357 1.356-.523 2.117-.523 2.854 0 4.662 2.212 4.662 4.566 0 .167 0 .357-.024.547l-4.71-2.759a.797.797 0 00-.856 0l-5.97 3.473zm10.609 8.8V12.06c0-.333-.143-.57-.429-.737l-5.97-3.473 1.95-1.118a.433.433 0 01.476 0l4.543 2.617c1.309.76 2.189 2.378 2.189 3.948 0 1.808-1.07 3.473-2.76 4.163zM7.802 12.703l-1.95-1.142c-.167-.095-.239-.238-.239-.428V5.899c0-2.545 1.95-4.472 4.591-4.472 1 0 1.927.333 2.712.928L8.23 5.067c-.285.166-.428.404-.428.737v6.898zM12 15.128l-2.795-1.57v-3.33L12 8.658l2.795 1.57v3.33L12 15.128zm1.796 7.23c-1 0-1.927-.332-2.712-.927l4.686-2.712c.285-.166.428-.404.428-.737v-6.898l1.974 1.142c.167.095.238.238.238.428v5.233c0 2.545-1.974 4.472-4.614 4.472zm-5.637-5.303l-4.544-2.617c-1.308-.761-2.188-2.378-2.188-3.948A4.482 4.482 0 014.21 6.327v5.423c0 .333.143.571.428.738l5.947 3.449-1.95 1.118a.432.432 0 01-.476 0zm-.262 3.9c-2.688 0-4.662-2.021-4.662-4.519 0-.19.024-.38.047-.57l4.686 2.71c.286.167.571.167.856 0l5.97-3.448v2.26c0 .19-.07.333-.237.428l-4.543 2.616c-.619.357-1.356.523-2.117.523zm5.899 2.83a5.947 5.947 0 005.827-4.756C22.287 18.339 24 15.84 24 13.296c0-1.665-.713-3.282-1.998-4.448.119-.5.19-.999.19-1.498 0-3.401-2.759-5.947-5.946-5.947-.642 0-1.26.095-1.88.31A5.962 5.962 0 0010.205 0a5.947 5.947 0 00-5.827 4.757C1.713 5.447 0 7.945 0 10.49c0 1.666.713 3.283 1.998 4.448-.119.5-.19 1-.19 1.499 0 3.401 2.759 5.946 5.946 5.946.642 0 1.26-.095 1.88-.309a5.96 5.96 0 004.162 1.713z"></path></svg>

assets/agents/qwen.svg

@@ -0,0 +1 @@
+<svg fill="currentColor" fill-rule="evenodd" height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Qwen</title><path d="M12.604 1.34c.393.69.784 1.382 1.174 2.075a.18.18 0 00.157.091h5.552c.174 0 .322.11.446.327l1.454 2.57c.19.337.24.478.024.837-.26.43-.513.864-.76 1.3l-.367.658c-.106.196-.223.28-.04.512l2.652 4.637c.172.301.111.494-.043.77-.437.785-.882 1.564-1.335 2.34-.159.272-.352.375-.68.37-.777-.016-1.552-.01-2.327.016a.099.099 0 00-.081.05 575.097 575.097 0 01-2.705 4.74c-.169.293-.38.363-.725.364-.997.003-2.002.004-3.017.002a.537.537 0 01-.465-.271l-1.335-2.323a.09.09 0 00-.083-.049H4.982c-.285.03-.553-.001-.805-.092l-1.603-2.77a.543.543 0 01-.002-.54l1.207-2.12a.198.198 0 000-.197 550.951 550.951 0 01-1.875-3.272l-.79-1.395c-.16-.31-.173-.496.095-.965.465-.813.927-1.625 1.387-2.436.132-.234.304-.334.584-.335a338.3 338.3 0 012.589-.001.124.124 0 00.107-.063l2.806-4.895a.488.488 0 01.422-.246c.524-.001 1.053 0 1.583-.006L11.704 1c.341-.003.724.032.9.34zm-3.432.403a.06.06 0 00-.052.03L6.254 6.788a.157.157 0 01-.135.078H3.253c-.056 0-.07.025-.041.074l5.81 10.156c.025.042.013.062-.034.063l-2.795.015a.218.218 0 00-.2.116l-1.32 2.31c-.044.078-.021.118.068.118l5.716.008c.046 0 .08.02.104.061l1.403 2.454c.046.081.092.082.139 0l5.006-8.76.783-1.382a.055.055 0 01.096 0l1.424 2.53a.122.122 0 00.107.062l2.763-.02a.04.04 0 00.035-.02.041.041 0 000-.04l-2.9-5.086a.108.108 0 010-.113l.293-.507 1.12-1.977c.024-.041.012-.062-.035-.062H9.2c-.059 0-.073-.026-.043-.077l1.434-2.505a.107.107 0 000-.114L9.225 1.774a.06.06 0 00-.053-.031zm6.29 8.02c.046 0 .058.02.034.06l-.832 1.465-2.613 4.585a.056.056 0 01-.05.029.058.058 0 01-.05-.029L8.498 9.841c-.02-.034-.01-.052.028-.054l.216-.012 6.722-.012z"></path></svg>

assets/agents/tencent.svg

@@ -0,0 +1 @@
+<svg fill="currentColor" fill-rule="evenodd" height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Tencent</title><path d="M9.976 1L24 9.8l-10.587.015L10.723 23H5.489L8.18 9.8H3.244L1 5.4h8.077L9.976 1z"></path></svg>

docs/agents.md

@@ -1,6 +1,6 @@
 # Driving PaperCraft with agents
 
-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.
+PaperCraft exposes one set of 88 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 |
 |---|---|---|
@@ -108,6 +108,18 @@ write_html {
 EOF
 ```
 
+## Checking only what changed
+
+On a big page a full `lint_design` after every edit repeats the same output, and a `get_node_info` call per layer is slow (about 53 s for 397 layers). Check the change, then check everything once.
+
+* Every write tool result includes `rev`, the document revision after the call. `get_basic_info` reports the current `rev`. Reads never change it.
+* `get_changes {since, page?}` returns `{rev, since, truncated, changes: [{page, ids}]}`: what the edits since `since` touched. `truncated: true` means the log no longer reaches that far back, so do a full pass.
+* `lint_design {changedOnly: true}` lints since this client's last lint; `{since: rev}` lints since a revision you name. A scoped lint checks the touched layers plus their ancestors, descendants, siblings and every layer bound to a changed variable, and reports its coverage in `covered`, for example `{checked: 12, total: 397, since: 40, partial: true}`. Plain `lint_design` is a full pass (`partial: false`). The `overflow` rule flags a layer that extends past its parent or clipping artboard (scroll frames are excluded).
+* `get_bounds {nodeIds?, pageId?, depth?}` returns `columns` `[id, parent, x, y, w, h, absX, absY, clip]` and one row per layer for a subtree in one call (at most 2000 rows; then `truncated` and `continueWith`, the nodeIds for the next call). Use it instead of looping `get_node_info`.
+* Always finish with one full `lint_design` (no arguments) and a screenshot per artboard in each theme mode before reporting. A partial lint speeds up the loop; it does not replace that pass.
+
+In the app, read-only calls no longer draw per-layer outlines or keep the "Working" pill lit for long. The agent workflow, with an example, is in [incremental-checks.md](../skills/papercraft/references/incremental-checks.md).
+
 ## Design-system tools
 
 * Variables: `get_variables` and `set_variables` take collections, groups (`color/brand/primary`), types (`color`, `number`, `string`, `boolean`), `mode`, `alias`, `description` and `hidden`; the old `{"name":"#hex"}` call still works and lands in the first collection. `create_collection` makes a collection with modes, `delete_variables` removes by name (refuses aliased ones unless `force`), `bind_variable` binds or detaches a field of layers (`stroke`, `text-color`, `width`, `height`, `gap`, `padding`, `radius`, `font-size`, `line-height`, `opacity`, `visible`).

docs/design-system.md

@@ -93,6 +93,8 @@ Fill-based sets (Phosphor, Material) follow variables; Lucide is stroke-based an
 | `layer-name` | warning | default names (Frame, Rectangle, Group...) on layers with children |
 | `hard-coded-color` | warning, info | a literal color that equals a variable (warning), or a literal used repeatedly when variables exist (info) |
 
+| `overflow` | error | a layer extending past its artboard or clipping parent (by how many px, on which side), or text taller than its box. Scrolling frames and layers pinned in them are exempt. Never auto-fixed by `fix: true`: resize, move or let the parent grow. `get_bounds` returns every rect in one call |
+
 Instances are skipped (their main component is linted). The rules are conventions, not law: pass `scale` to match your own type scale and ignore `info` items you disagree with. Lint does not judge taste; always look at `get_screenshot` too.
 
 ## Libraries

skills/papercraft-critique/SKILL.md

@@ -1,180 +1,110 @@
 ---
 name: papercraft-critique
-description: Pre-emit critique gate for work built in PaperCraft. Five panels (designer, critic, brand, accessibility, copy) score a screen 0-10 against PaperCraft's own objective evidence (lint_design, verify_code, get_screenshot, get_design_spec, get_variables), the weighted composite must reach 8.0, and a failed gate ships the best round instead of the last one. Use when an agent is about to hand over a PaperCraft screen and wants a tool-backed review instead of a self-praise.
+description: Critique finished PaperCraft work the way a design team does - against the person's objective, through independent reviewers (the person's vision, UX heuristics, flow across screens, QA states, visual, accessibility, copy, an end user from the audience), each finding rated 0-4 and backed by tool evidence, then fixed, reported and, when the person approves, kept as a lesson. Use when finish_working_on_nodes offers a critique, when asked to review or critique a design, or before handing over a screen or flow.
 ---
 
 # papercraft-critique
 
-A gate, not a compliment. Before a screen leaves the app it goes through five panels. Each panel is a rubric pass over the same artifact, and each panel may only cite what a PaperCraft tool actually returned. The panels produce five scores, the scores produce one composite, and the composite decides whether the work ships. Mechanics live in the `papercraft` skill, taste in `papercraft-design`; this skill is the gate.
+A critique judges a design against what it is for, not against taste. The tool `critique_design`
+hands you the evidence and the reviewers; this skill is how to run them. Mechanics live in the
+`papercraft` skill, taste in `papercraft-design`. Per-reviewer checklists:
+[personas.md](references/personas.md).
 
-Reference: [panels.md](references/panels.md) (what each panel inspects, which tool backs it, its 0–10 bands).
+## 1. When
 
-## 1. When the gate runs
+* `finish_working_on_nodes` returns `next`: ask the person (`ask_user`) whether to critique now.
+* The person asks "review this", "is it ready?", "critique it".
+* Before handing over a whole screen or flow. Not for one component, an early draft or a copy tweak.
 
-Run it when the resolver (section 2) says on, or when the request is a review of finished work:
+## 2. Objective first
 
-* a whole screen is about to be handed over: landing page, dashboard, sign-in, settings, an onboarding flow, a whiteboard plan;
-* a redesign of an existing screen, after the edits and a clean `lint_design`;
-* the user asked "does this look right?", "review this", "critique it", "is it ready to ship?".
+Critique without an objective is opinion. Before judging, have:
 
-Do not run it for: a single component or icon, an exploratory draft before a direction is chosen (the `papercraft-design` section 0 step), or a one-line copy tweak. The gate costs a lint and a screenshot per round; spend it on work that is finished.
+* **Objective**: what the work must achieve ("a returning customer reorders in two taps").
+* **Audience**: who uses it, in what context (phone, in a hurry, first visit, low vision).
+* **Top tasks**: the two to five things people come to do.
 
-If the gate is on for a run, it must produce a status. Never skip it silently: report `Degraded` with the reason.
+Take them from the brief, the spec, the person's notes (`get_feedback`) and references. Missing?
+Ask with `ask_user` and pass them to `critique_design {objective, audience}`.
 
-## 2. Enable resolver
+## 3. The reviewers
 
-Four levels, highest priority first. Take the first level that gives an explicit answer; "unset" falls through to the next one. A higher level saying *off* is final — a lower level cannot turn the gate back on.
-
-| Order | Level | PaperCraft equivalent | How to read it |
-|---|---|---|---|
-| 1 | Per-task policy | the instruction the agent was given for this run | "run the critique gate before you finish", "review this screen" |
-| 2 | Per-project setting | a toggle stored in the project's own metadata | **not implemented today** — PaperCraft has no project-level flag. Treat this level as unset and fall through to the environment flag or the per-task instruction |
-| 3 | Environment flag | `PAPERCRAFT_CRITIQUE_ENABLED` | `1` / `true` / `always` = on, `0` / `false` / `never` = off, unset = fall through |
-| 4 | Rollout phase | the release phase of the skill itself | off by default: the gate ships disabled until a run or a project turns it on |
-
-So today the gate turns on in one of two ways: the user asks for it in the run's instructions, or `PAPERCRAFT_CRITIQUE_ENABLED=1` is in the environment. A future project setting — stored with the project, not with the user — would sit between those two and win over the environment, so a project could opt in permanently without an env var.
-
-## 3. Order of operations
-
-Objective checks first, subjective panels second. A panel never scores above evidence that contradicts it.
-
-One round is: **evidence → repair → judge → score**.
-
-1. **Snapshot.** Note the `pageId` and the artboard `nodeId` under review and keep them for the whole gate. A page switch between rounds invalidates the evidence.
-2. **Evidence.** `lint_design` (`nodeId`, `pageId`, no `fix` yet) for the raw findings — the payload counts `errors`, `warnings` and `info`, groups them in `byRule`, and lists each issue with its `rule`, `severity`, `node`, `message` and `fix`; `get_screenshot` of the same node (`scale: 2` to read small text); `get_variables` for the token set and the active mode; `get_design_spec` (`nodeId`, `target`) when the design is headed for code. If there is code to check, `verify_code` (`nodeId`, `target`, `code`).
-3. **Repair the hard evidence.** Fix every `lint_design` error and warning and re-run the lint until it is clean; fix every `verify_code` violation until `passed` is true. A panel cannot be scored on a design that still carries a lint error.
-4. **Judge the five panels** on the evidence from steps 2–3 plus the screenshot. Every finding cites the tool result or the screenshot region behind it.
-5. **Score.** Five scores 0–10, then the composite (section 5). Composite ≥ 8.0 → status `Shipped`, stop, `finish_working_on_nodes`.
-6. **Otherwise fix and go again.** Take the findings in composite-impact order, fix them, run round N+1. Maximum 3 rounds.
-7. **Ship the best round** if round 3 is still below the threshold (section 6).
-
-The panels are rubric passes by the agent that holds the tools: one agent, five passes, five different rubrics. A runtime that can spawn subagents may run the panels in parallel, but the evidence vetoes, the composite and the round counter belong to the coordinator.
-
-## 4. Hard evidence vetoes a panel
-
-`lint_design` and `verify_code` are not suggestions. Each finding belongs to one panel, and that panel cannot score above the band its evidence forces.
-
-| Evidence | Rule | Owner panel | Cap |
-|---|---|---|---|
-| `lint_design` error | `contrast` | Accessibility (the Critic reads it too) | 3.9 |
-| `lint_design` warning | `text-size` (under 12px) | Accessibility | 6.9 |
-| `lint_design` warning | `spacing-grid` (off the 4px grid) | Designer | 6.9 |
-| `lint_design` warning | `touch-target` (under 44px) | Accessibility | 6.9 |
-| `lint_design` warning | `layer-name` (default names) | Designer | 6.9 |
-| `lint_design` warning | `hard-coded-color` | Brand | 6.9 |
-| `lint_design` info | `type-scale` | Designer | none — must be listed |
-| `verify_code` error | `missing-text` | Copy | 3.9 |
-| `verify_code` error | `hardcoded-color` | Brand | 3.9 |
-| `verify_code` error | `forbidden-absolute` | Designer | 3.9 |
-| `verify_code` error | `missing-layout` | Designer | 3.9 |
-| `verify_code` warning | `missing-component` | Brand | 6.9 |
-| `verify_code` warning | `spacing-drift` | Designer | 6.9 |
-
-An error caps the owning panel at 3.9 (Blocker band), a warning at 6.9 (Weak band), an `info` item caps nothing but must appear in the panel's notes. `hard-coded-color` arrives in both flavours: a warning when the color equals a variable that exists, an info when it is a literal used several times — the warning caps Brand, the info does not. The cap lifts as soon as the finding is gone; the next round re-scores the panel freely.
-
-A finding is evidence for its owner and a hint for the others: `contrast` also informs the Critic, `spacing-grid` also informs the Brand (was it the token or the layout?), but only the owner is capped.
-
-## 5. The panels and the composite
-
-| Panel | Judges | Weight |
+| Persona | Asks | Method |
 |---|---|---|
-| Designer | layout, composition, hierarchy, spacing rhythm, type scale in use | **0 in v1** — it still scores and its findings still drive the next round |
-| Critic | does it meet the brief; contrast, weight, readability | 0.4 |
-| Brand | token compliance, on-brand color, voice | 0.2 |
-| Accessibility | WCAG, focus rings, semantic structure, alt text | 0.2 |
-| Copy | voice, tone, terseness, error-message quality | 0.2 |
+| `vision` | Does it achieve the objective, as the person imagined it? | Objective, element, effect, why (Connor and Irizarry) |
+| `ux` | Can people use it without help? | Nielsen's 10 heuristics |
+| `flow` | Do the screens make one coherent journey per task? | Task walkthrough across screens, 4 questions per step |
+| `qa` | What breaks with real data and states? | State and edge-case matrix, pass/fail/n.a. |
+| `visual` | Does the eye land right, does it hold together? | Hierarchy, balance, contrast, gestalt, rhythm, tokens |
+| `accessibility` | Can everyone use it? | W3C POUR applied to a design |
+| `copy` | Is every word clear and specific? | Plain language, labels, errors, empty states |
+| `end-user` | Where would this user get stuck or give up? | Persona walkthrough of the top task |
 
-```
-composite = critic x 0.4 + brand x 0.2 + accessibility x 0.2 + copy x 0.2
-```
+`vision` always runs: the person's intent is the yardstick for the others. Pick the rest with
+`personas` (default: all). The person's own words outrank any reviewer.
 
-The weights sum to 1.0, so the composite is also 0–10. The Designer carries weight 0 in v1 on purpose: it is the panel most likely to argue with itself, so its score is reported and its findings are acted on, but it does not move the composite.
+## 4. Independence
 
-Each panel scores 0–10 in half-point steps. The composite is the weighted sum, rounded to two decimals.
+Several reviewers find far more than one, but only when they work apart (Nielsen: one
+evaluator finds about a third of the problems, three to five most of them). Reviewers who
+discuss or debate converge and amplify each other's bias.
 
-| Composite | Verdict |
+* **With subagents**: one subagent per persona, given only its persona brief, the objective and
+  the evidence. No reviewer sees another's findings before rating.
+* **Alone**: one pass per persona, in order, writing that persona's findings before reading the
+  next brief. Do not revise earlier personas' findings during later passes.
+* **You coordinate**: you merge and judge; you never let reviewers argue a finding out.
+
+## 5. Evidence
+
+From `critique_design`: lint (fix every error and warning first), the `flow` graph (arrows between
+screens, `deadEnds`, `unreached`), the lessons. Then `get_screenshot` each artboard at scale 1
+(2 for small text) in every variable mode, `get_bounds` for overflow, `get_feedback` for the
+person's notes. Every finding cites one of these: a tool result or a screenshot region. No
+evidence, no finding.
+
+## 6. A finding
+
+`persona`, `heuristic or check`, `element` (layer name), `seen` (the observation), `effect` (on
+the objective), `evidence`, `severity`, `suggestion`. Keep the observation and the suggestion
+apart: critique is analysis, not brainstorming. Each persona also names at least one thing that
+works and should stay.
+
+Severity (Nielsen), from frequency, impact and persistence:
+
+| | |
 |---|---|
-| ≥ 8.0 | Ship |
-| < 8.0 | Not shippable as is: fix and re-run, or ship the best round after round 3 |
+| 0 | not a problem |
+| 1 | cosmetic: fix if time allows |
+| 2 | minor: low priority |
+| 3 | major: fix before handing over |
+| 4 | catastrophe: blocks the task, fix now |
 
-The band legend is the same for every panel:
+Every broken lesson (`lessons`) is a finding of at least 3.
 
-| Band | Range | Means |
-|---|---|---|
-| Blocker | 0.0–2.9 | Broken: the panel's own evidence is violated. Cannot ship |
-| Poor | 3.0–4.9 | Works against the brief or the brand; ships visible damage |
-| Weak | 5.0–6.9 | Passable but noticeably off; expect a second round |
-| Acceptable | 7.0–7.9 | No blocker, but a reviewer would ask for one change |
-| Good | 8.0–9.4 | Ships as is |
-| Excellent | 9.5–10.0 | Reference quality for this project |
+## 7. Merge, fix, report
 
-What each panel inspects, which tool backs it and where its scores land: [panels.md](references/panels.md).
+1. **Merge** into one list: join duplicates, count the personas that found each (two or more is
+   confirmed), severity is the mean of their independent ratings.
+2. **Fix** every 3 and 4 with the smallest edit (`update_styles`, `set_text_content`), then check
+   that finding again with the same evidence. The work is ready when no 3 or 4 is open and lint is
+   clean. Three fix rounds at most; then hand over with the open findings stated.
+3. **Report** to the person: what works, then findings by severity, each with what was seen, the
+   evidence, the suggestion and whether it was fixed. Phrase open ones as questions about the
+   objective ("How does a first-time user find Reorder from here?"), not verdicts.
 
-## 6. Rounds and the fallback
+## 8. Lessons
 
-* At most **3 rounds** per gate. The counter is per gate, not per panel.
-* A round ends when the composite is computed.
-* Rounds 2 and 3 exist to fix what the panels found. Fix in composite-impact order: the panel with the largest `weight x (10 - score)` first.
-* **`fallbackPolicy: ship_best`.** If the third round is still below 8.0, the status is `Below threshold` and what ships is the best round, not the last one. Best = highest composite; a tie goes to the earlier round (fewer changes). A later round edited the canvas, so put the canvas back into the best round's state before handing over — undo back to it, or re-apply that round's fix list in order.
-* Ship the best round's artifact and report the failing score. `Below threshold` is a status, not a licence to hide the number.
+Ask the person with `ask_user` which findings to remember, one question per finding: for every
+project, for this project, or not. Save only what they approve, with
+`design_lessons {add: {rule, why, scope}}`: the rule is one line written as an instruction
+("Screens keep their device height; the main column scrolls"), `why` is what went wrong. Never
+save a lesson the person did not approve, and never save praise. Lessons come back in
+`get_basic_info` and in the next `critique_design`; `design_lessons {remove: "L3"}` drops one.
 
-## 7. Status vocabulary
+## Sources
 
-Every gate ends with exactly one status.
-
-| Status | When | What ships |
-|---|---|---|
-| `Shipped` | composite ≥ 8.0 | the reviewed artifact, unchanged |
-| `Below threshold` | still < 8.0 after round 3 | the best-scoring round's artifact (`ship_best`), with the score reported |
-| `Timed out` | the gate ran out of its budget (rounds or minutes) before a verdict | the best round so far, with the timeout stated |
-| `Interrupted` | the user or the session stopped the gate mid-run | the last complete round's artifact |
-| `Degraded` | a panel could not be judged properly (section 8) | the artifact, but the verdict is provisional and the reason is named |
-
-`Degraded` outranks `Shipped`: a degraded panel is capped at 6.9, so a degraded run reports `Degraded`, not `Shipped`, even when the composite clears 8.0.
-
-## 8. Degraded reasons
-
-A panel that cannot be judged on real evidence is not scored as if it were fine. It is capped at 6.9, the composite is computed with the cap, and the run reports `Degraded` with the reason and the panel it hit.
-
-| Reason | Cause | Handling |
-|---|---|---|
-| `malformed_block` | a tool returned something the gate could not parse (lint payload in an unexpected shape, a truncated screenshot answer) | retry the tool once; if it still fails, judge the panel on the evidence you do have and name the missing part |
-| `oversize_block` | the evidence was too big to read whole (a screenshot scaled down past legibility, a tree deeper than the context, a lint payload whose `truncated` flag is set) | narrow the scope: screenshot or lint a subtree, split the screen into regions and judge region by region |
-| `adapter_unsupported` | the requested target or format is not supported here (a `verify_code` target the project does not build) | drop that check; the panel falls back to the screenshot plus the design rules |
-| `protocol_version_mismatch` | the MCP server exposes a different tool set than the gate expects (a tool renamed, an argument changed) | stop using the broken tool, finish what you can, report `Degraded`; do not guess the schema |
-| `missing_artifact` | the thing under review is gone (artboard deleted, page switched, stale node id) | re-snapshot with `get_basic_info` and restart the gate; if the artifact is truly gone, report `Degraded` |
-
-Degradation is a report, not an excuse: name the reason, the panel and the evidence you lost.
-
-## 9. What the gate reports
-
-End the gate with the status, the numbers and the findings, in that order:
-
-```
-Critique gate - Sign in (page p1, artboard 3fa9c1:7)
-Round 2 of 3 - composite 8.44 / threshold 8.00 - status Shipped
-Designer 7.5 (weight 0) | Critic 8.5 | Brand 8.5 | Accessibility 8.5 | Copy 8.2
-Composite = 8.5x0.4 + 8.5x0.2 + 8.5x0.2 + 8.2x0.2 = 8.44
-```
-
-Then, per finding: the panel, what was seen, the evidence behind it (tool and finding, or the region of the screenshot), the fix, and whether it was fixed in this round or is still open.
-
-Machine-readable form, when the caller wants JSON:
-
-```json
-{
-  "status": "Shipped",
-  "threshold": 8.0,
-  "composite": 8.44,
-  "round": 2,
-  "rounds": [{ "round": 1, "composite": 7.35 }, { "round": 2, "composite": 8.44 }],
-  "scores": { "designer": 7.5, "critic": 8.5, "brand": 8.5, "accessibility": 8.5, "copy": 8.2 },
-  "weights": { "designer": 0.0, "critic": 0.4, "brand": 0.2, "accessibility": 0.2, "copy": 0.2 },
-  "degraded": [],
-  "findings": [{ "panel": "critic", "evidence": "get_screenshot: headline at 16px over a photo", "severity": "warning", "state": "fixed" }]
-}
-```
-
-## 10. Attribution
-
-The five-panel critique contract — the roles, the weights, the composite, the 8.0 threshold, the three rounds and the `ship_best` fallback — is inspired by OpenDesign's Critique Theater (https://github.com/nexu-io/open-design, Apache-2.0). The contract is re-implemented here in PaperCraft's own words and against PaperCraft's own tools; no text was copied.
+Adam Connor and Aaron Irizarry, *Discussing Design* (objective, element, effect, rationale);
+Nielsen Norman Group: 10 usability heuristics, severity ratings, cognitive walkthroughs, design
+critiques, visual design principles; W3C WAI accessibility principles (POUR); work on panels of
+independent LLM judges (debate amplifies bias, a coordinating judge resists it).

skills/papercraft-critique/references/panels.md

@@ -1,182 +0,0 @@
-# Panel rubrics
-
-Five panels, one artifact, one set of evidence. A panel is a pass with a fixed rubric: what it inspects, which tool output backs it, what it does **not** judge, and the band each finding lands in. The panels share the band legend in [../SKILL.md](../SKILL.md) (Blocker 0–2.9, Poor 3–4.9, Weak 5–6.9, Acceptable 7–7.9, Good 8–9.4, Excellent 9.5–10).
-
-## How to run one panel
-
-1. Read the evidence you were handed: the lint payload, the screenshot, the design spec, the variables, the code result.
-2. Write findings. Every finding cites its evidence — the lint rule and node, the region of the screenshot, the spec's MUST rule, the `verify_code` violation.
-3. Score the panel 0–10. The score is the panel's verdict on the whole screen, not an average of its findings: two serious defects on one region cost more than two trivial ones on two regions, but a screen with three medium defects is not "Good" either.
-4. Apply the caps. Hard evidence caps the owning panel (SKILL.md section 4); degraded evidence caps it at 6.9 (SKILL.md section 8).
-
-Evidence a panel may not use: taste, guesses about a user's preferences, or a claim about something no tool showed. When a panel wants to judge what no tool exposes — whether an image is meaningful, whether a flow makes sense — it raises the question as an open item instead of scoring it as a defect.
-
-Tool arguments used below are the current names. Note that `lint_design`'s `scale` is the list of allowed font sizes, not a zoom; the zoom argument of `get_screenshot` is `scale: 1 | 2`.
-
----
-
-## Designer — weight 0 in v1
-
-**Judges:** layout and composition, alignment, spacing rhythm, hierarchy, the type scale in use, how things are grouped, screen states, overflow, density.
-
-**Evidence it uses**
-
-| Tool | How |
-|---|---|
-| `get_screenshot` (`nodeId`, `scale: 2`) | the primary evidence: alignment, edges, rhythm, hierarchy, clipping, empty areas |
-| `lint_design` (`nodeId`, `pageId`) | `spacing-grid`, `type-scale`, `layer-name` findings |
-| `get_design_spec` (`nodeId`, `target`) | flow and sizing (`fixed`, `hug`, `fill`), grouping, where the spacing rules were broken |
-| `get_tree_summary` (`nodeId`, `depth`) | structure depth, unnamed frames, accidental nesting |
-
-**Does not judge:** contrast and WCAG (Accessibility), token and palette choices (Brand), wording (Copy).
-
-**Bands**
-
-| Band | What the Designer saw |
-|---|---|
-| 0.0–2.9 | Broken layout: frames overlap, content sits outside the artboard, text is clipped, the screen reads as a pile |
-| 3.0–4.9 | Poor composition: everything on the default grid, equal weight everywhere, no grouping, ragged edges |
-| 5.0–6.9 | Weak: spacing off the 4px grid in places, sizes off the type scale, mixed grouping (space *and* containers everywhere), a state missing |
-| 7.0–7.9 | Acceptable: coherent, but one region is crowded or one rhythm is broken |
-| 8.0–9.4 | Good: clear hierarchy, aligned to a few strong edges, consistent rhythm, states present |
-| 9.5–10.0 | Excellent: deliberate rhythm, one memorable idea, whitespace used as a material |
-
-**Fixes that move the score:** snap spacing to the 4/8 scale (`lint_design` with `fix: true`), pull one size step per level, give each group either a container or space — not both, align to one left edge, add the missing empty/loading/error state.
-
----
-
-## Critic — weight 0.4
-
-**Judges:** whether the screen meets the brief (the request this run is executing), whether visual weight matches the priority, readability as a human experiences it, and whether the real cases are covered.
-
-**Evidence it uses**
-
-| Tool | How |
-|---|---|
-| the brief | the user's request and the `papercraft-design` section 0 direction, read against the screenshot |
-| `get_screenshot` (`nodeId`, `scale: 2`) | one primary action, reading order, what is above the fold, legibility by eye |
-| `lint_design` (`nodeId`, `pageId`) | `contrast`, `text-size`, `type-scale` findings as the objective part of readability |
-| `get_design_spec` (`nodeId`, `target`) | what the screen actually contains, against what was asked for |
-| `verify_code` (`nodeId`, `target`, `code`) | when the design is code-bound: whether the implementation kept the intent |
-
-**Does not judge:** which palette (Brand), line-level wording (Copy), implementation details beyond fidelity to the design.
-
-**Bands**
-
-| Band | What the Critic saw |
-|---|---|
-| 0.0–2.9 | Misses the brief: wrong screen or wrong content, the primary action is absent or ambiguous |
-| 3.0–4.9 | Fights the brief: right elements, wrong order or emphasis, or hard to read (contrast errors, tiny type) |
-| 5.0–6.9 | Weak: it does the job, but the weight is flat and something important sits below the fold or is buried |
-| 7.0–7.9 | Acceptable: meets the brief; one area stays unclear or the priority is not obvious |
-| 8.0–9.4 | Good: the brief is met, the primary action is obvious, readability is solid |
-| 9.5–10.0 | Excellent: unmistakable hierarchy, the screen reads in one pass, the edge cases are covered |
-
-**Fixes that move the score:** make one thing clearly primary and quiet the rest, move the important content up, cut a competing element, re-read the brief line by line and check each requirement against the screenshot.
-
----
-
-## Brand — weight 0.2
-
-**Judges:** token compliance, on-brand color, semantic color use, mode coverage, product voice consistency, component reuse.
-
-**Evidence it uses**
-
-| Tool | How |
-|---|---|
-| `get_variables` (`pageId`, `format`) | the variables, tokens, active `mode` and available `modes` — the reference palette |
-| `lint_design` (`nodeId`, `pageId`) | `hard-coded-color` findings |
-| `get_computed_styles` (`nodeIds`) / `find_nodes` (`filters`) | fills and text colors that are not following a variable |
-| `get_design_spec` (`nodeId`, `target`) | which tokens the design already binds |
-| `get_components` / `get_tree_summary` | components built and instanced versus copies |
-| `verify_code` (`nodeId`, `target`, `code`) | `hardcoded-color`, `missing-component` in the implementation |
-
-**Does not judge:** composition (Designer), wording line by line (Copy), whether the layout serves the brief (Critic).
-
-**Bands**
-
-| Band | What the Brand saw |
-|---|---|
-| 0.0–2.9 | Off-system: hard-coded colors where tokens belong, a foreign palette, no modes |
-| 3.0–4.9 | Off-brand: technically on-token but the accent is used everywhere, or the screens read as different products |
-| 5.0–6.9 | Weak: one or two hard-coded colors, the accent covers much of the screen, components copied instead of instanced |
-| 7.0–7.9 | Acceptable: on-system, with one token or component inconsistency |
-| 8.0–9.4 | Good: every color is a token, the accent marks the primary action only, modes defined and checked, components reused |
-| 9.5–10.0 | Excellent: coherent across screens and modes, the palette carries meaning |
-
-**Fixes that move the score:** `lint_design` with `fix: true` relinks hard-coded colors that match a variable, add the missing variable before using the color, cut accent back to the primary action, build the repeated part once and place instances, check the second mode with `set_variable_mode`.
-
----
-
-## Accessibility — weight 0.2
-
-**Judges:** WCAG contrast, the text-size floor, touch-target size, focus and hover states, semantic structure, alt text, and signalling that does not rely on color alone.
-
-**Evidence it uses**
-
-| Tool | How |
-|---|---|
-| `lint_design` (`nodeId`, `pageId`) | `contrast`, `text-size`, `touch-target` findings |
-| `get_screenshot` (`nodeId`, `scale: 2`) | whether a visible focus or hover state was designed at all, and how the error state reads |
-| `get_design_spec` (`nodeId`, `target`) | which text is a title, how the screen is grouped, what each control is |
-| `find_nodes` (`textValue`, `filters`) | image and icon nodes, unlabeled icon buttons |
-| `get_tree_summary` (`nodeId`, `depth`) | semantic structure: heading levels, groups, control order |
-
-**Does not judge:** visual taste (Designer), palette taste (Brand), voice (Copy).
-
-**Bands**
-
-| Band | What Accessibility saw |
-|---|---|
-| 0.0–2.9 | Fails WCAG: contrast errors, text under 12px, targets under 44px, meaning carried by color alone |
-| 3.0–4.9 | Poor: borderline contrast, unlabeled controls, an error state that is red text and nothing else |
-| 5.0–6.9 | Weak: passes on the desktop artboard but fails on the mobile one, one image without alt text |
-| 7.0–7.9 | Acceptable: compliant baseline, one gap (no visible focus state, one icon-only button) |
-| 8.0–9.4 | Good: WCAG clean, targets pass, semantics and labels present, focus and hover states designed |
-| 9.5–10.0 | Excellent: every state specified, sensible keyboard order, no meaning by color alone |
-
-**Fixes that move the score:** fix the color, not the check, until `lint_design` reports no contrast error, name interactive layers `Button…` / `Tab /…` so the target rule can see them and grow them to 44px, pair every error color with an icon or a word, design the focus and hover variants as components, give every image an alt text.
-
-PaperCraft has no focus-ring or alt-text checker: judge those from the design itself (a `Button / Focused` variant, alt text in the spec) and report what is missing as a gap, not as a tool finding.
-
----
-
-## Copy — weight 0.2
-
-**Judges:** voice and tone against the product's voice, terseness, labels and microcopy, error-message quality, empty states, placeholders versus labels, the absence of filler.
-
-**Evidence it uses**
-
-| Tool | How |
-|---|---|
-| `get_design_spec` (`nodeId`, `target`) | the authoritative copy: every text node's content, in order |
-| `get_screenshot` (`nodeId`, `scale: 2`) | how the copy reads in context: line breaks, truncation, all-caps abuse |
-| `find_nodes` (`textValue`) | repeated strings, leftover placeholders, "Lorem ipsum", inconsistent labels for the same action |
-| `verify_code` (`nodeId`, `target`, `code`) | `missing-text` — copy that was lost between the design and the code |
-
-**Does not judge:** typography and layout (Designer), color (Brand), whether the screen meets the brief (Critic).
-
-**Bands**
-
-| Band | What Copy saw |
-|---|---|
-| 0.0–2.9 | Wrong: placeholder text, lorem ipsum, or copy that contradicts the brief |
-| 3.0–4.9 | Poor: filler adjectives, two voices on one screen, errors that blame the user |
-| 5.0–6.9 | Weak: wordy, mixed formality, vague labels ("Submit" where "Save changes" is meant) |
-| 7.0–7.9 | Acceptable: clear but plain; one error message says only "Error" |
-| 8.0–9.4 | Good: terse, one voice, every action labeled with a verb, errors say what happened and what to do next |
-| 9.5–10.0 | Excellent: reads like a product that shipped, not like a mockup |
-
-**Fixes that move the score:** cut every word that does not inform, give each action a verb label ("Save changes", not "OK"), write errors as *what happened + what to do next*, fill empty states with an explanation and the next action, replace placeholder copy with real content, keep one formality level per screen.
-
-PaperCraft has no spell-checker or copy linter: the copy panel reads the spec and the screenshot, and it flags what a human reviewer would flag — it does not invent grammar rules.
-
----
-
-## Scoring discipline
-
-* Half-point steps only: 7.5, 8.0, 8.5. Never 8.37.
-* The score is a verdict on the screen, not an average of findings.
-* A panel that finds nothing wrong and cites no evidence scores 7.0–7.9, not 10. A 10 means the screen is a reference for the next one.
-* Caps bind: hard evidence (SKILL.md section 4) and degraded evidence (section 8) are applied after the verdict, never negotiated with.
-* No panel may raise another panel's score. If the Critic likes the hierarchy, that stays a Designer finding.

skills/papercraft-critique/references/personas.md

@@ -0,0 +1,110 @@
+# Reviewer checklists
+
+Each reviewer works from its own list, alone, and cites evidence for every finding. Mark each
+item pass, fail or n/a; a state or screen that was never designed is an open question, not a pass.
+
+## vision: the person's vision
+
+For each key element on each screen, four questions (Connor and Irizarry):
+
+1. **Objective**: which part of the objective is this element for?
+2. **Element**: which elements serve it?
+3. **Effect**: do they achieve it, for this audience?
+4. **Why** or why not?
+
+Also: everything the brief, references and feedback notes asked for is there; nothing contradicts
+an open or resolved note (`get_feedback`); the tone matches the references. Unknown product facts
+are questions for the person, never guesses.
+
+## ux: usability heuristics (Nielsen)
+
+1. Visibility of system status: progress, saving, selection shown at once.
+2. Match with the real world: the users' words and order, not the system's.
+3. User control and freedom: a marked exit, undo, cancel.
+4. Consistency and standards: one word and one look per thing, platform conventions.
+5. Error prevention: risky actions confirmed, impossible choices removed.
+6. Recognition rather than recall: options visible, no remembering across screens.
+7. Flexibility and efficiency: shortcuts for experts that novices need not see.
+8. Aesthetic and minimalist design: nothing that competes with what matters.
+9. Recognize, diagnose, recover from errors: plain language, the problem, the fix.
+10. Help and documentation: needed rarely, and then where the task is.
+
+## flow: coherent journeys
+
+Walk each top task across screens (the `flow` graph when arrows connect screens, else the screen
+order and names; ask the person when the intended order is unclear). At each step, the four
+walkthrough questions: will they try to do this, will they notice the control, will they connect
+it to their goal, will they see progress afterwards?
+
+* Every task has an entry point (including first run and empty state) and a clear end
+  (success screen or state).
+* Every screen has a way forward and a way back or out; no dead ends (`deadEnds`), no screen
+  nothing leads to (`unreached`) unless it is an entry.
+* Each action that can fail has its error state; each destructive one a confirmation and undo.
+* Navigation, back and primary actions sit in the same place with the same labels on every screen.
+* What the user entered or chose shows on the following screens (the cart count, the name typed).
+* Steps a user must remember between screens are shown instead (recognition over recall).
+* The number of steps fits the objective ("two taps" means two).
+
+## qa: states and edge cases
+
+Per screen and component, pass, fail or n/a:
+
+* States: default, hover, focus, pressed, disabled, selected, loading, empty, error, success.
+* Content: the longest realistic text and names, minimum content, zero, one and many items,
+  missing images and avatars, large numbers, special characters.
+* Modes and sizes: every variable mode (screenshot each); the narrow and wide screen sizes when the
+  design has them.
+* Layout: nothing overflows or clips (`get_bounds`, lint `overflow`), screens keep their device
+  height and scroll inside, fixed bars stay fixed.
+* Consistency: repeated pieces are instances of the same component (lint `not-an-instance`).
+
+## visual
+
+* One focal point per screen; the primary action is the most prominent control.
+* Scale and hierarchy: size, weight and colour follow importance.
+* Balance and alignment: shared edges, a steady grid.
+* Gestalt proximity: related items closer together than unrelated ones.
+* Rhythm: spacing from the 4px scale, the same gaps for the same relationships.
+* Type: sizes on the scale, at most three levels per screen.
+* Tokens: every colour a variable, consistent radii, components reused.
+
+## accessibility (W3C POUR, for a design)
+
+**Perceivable**
+* Text contrast 4.5:1, 3:1 for large text (18px, or 14px bold) and for control borders and icons.
+* Colour is never the only signal (errors also have an icon or text; links are underlined or bold).
+* Text is real text, not pictures of text; icons that act have a label or tooltip.
+* The layout holds at 320px wide and with text at 200% (a narrow artboard shows it).
+
+**Operable**
+* Touch targets 44px (24px minimum with spacing).
+* A visible focus state is designed for every interactive component.
+* Headings, a page title and a visible current location (active nav item).
+* Destructive actions confirmed or undoable; every gesture has a button alternative.
+* Nothing flashes; motion can be reduced.
+
+**Understandable**
+* Fields keep a visible label (a placeholder alone disappears).
+* Errors say what went wrong and how to fix it, next to the field.
+* Navigation and labels are the same across screens.
+* Plain language; abbreviations explained.
+
+**Robust**
+* Controls are named for their role (layer names, code links to real components), so the code
+  hand-off keeps them buttons and inputs, not divs.
+
+## copy
+
+* The users' words, not internal jargon; verbs on buttons ("Save changes", not "OK").
+* No placeholder or lorem text, no "Item 1".
+* Errors: what happened and what to do. Empty states: what this is and the first action.
+* Neutral wording about people whose gender is unknown (their name, or they).
+* Consistent terms: one word per concept across all screens.
+
+## end-user: a walkthrough as the audience
+
+Take the `audience` (else ask). Write as that person, in the first person, for each top task:
+what I see first, what I think it does, what I try, where I hesitate, and the first point where I
+would give up. One persona per run; launch another critique with a different `audience` for a
+second user type.

skills/papercraft-design/SKILL.md

@@ -33,7 +33,7 @@ Allowed spacing values: **4, 8, 12, 16, 24, 32, 48, 64** (and 96 for page sectio
 ## 2. Tokens first (variables)
 
 * All colors are variables (`var(--name)`), named by role: `bg`, `surface`, `border`, `text`, `text-muted`, `accent`, `on-accent`, `success`, `warning`, `danger`. Two to three surface levels at most (`bg` < `surface` < `surface-2`).
-* Define at least two modes if the product may be dark or light (`set_variables` with `mode`), and check both: lint contrast in each mode after `set_variable_mode`.
+* Define at least two modes if the product may be dark or light (`set_variables` with `mode`), and check both: lint contrast in each mode after `set_variable_mode`, or look with `get_screenshot {mode}` without switching. Themes are modes, never copies: one artboard per screen, never a light and a dark duplicate (`lint_design` flags them as `theme-copy`).
 * Semantic colors carry meaning only: red means error or destructive, green means success. Do not use them decoratively.
 * If a color you need is not a variable, add the variable, then use it. `lint_design` reports hard-coded colors that equal a variable.
 * Spacing, radius and type sizes are number variables (`space/md`, `radius/sm`, `font/body`): define them once, in a collection of their own, and bind fields with `bind_variable`. Keep the values on the lists in this skill. Group names with `/` and alias semantic names to base ones (`color/button/bg` to `color/brand/primary`).
@@ -105,7 +105,7 @@ Run it before `finish_working_on_nodes`:
 1. Direction written down; the screen has one memorable idea and one primary action.
 2. Every color is a variable; both modes checked if there are two.
 3. Spacing only 4, 8, 12, 16, 24, 32, 48, 64, 96; type only from the scale; radii from one family.
-4. `lint_design` has zero errors and zero warnings, and the `info` items were considered.
+4. A full `lint_design` (no arguments) has zero errors and zero warnings, and the `info` items were considered. Between edits, lint only what changed: see `papercraft/references/incremental-checks.md`.
 5. Screenshot at scale 2 looked at: hierarchy, alignment, wrapping, clipping, empty areas, contrast by eye.
 6. Interactive things at least 44px on touch and named `Button…` / `Tab /…`.
 7. Repeated parts are components; states exist for every interactive component and for empty, loading and error screens.

skills/papercraft/SKILL.md

@@ -17,7 +17,7 @@ papercraft-cli call get_basic_info      # fails with exit 1 if the app is not ru
 
 * **Check the project is the right one.** `get_basic_info` returns `fileName` and `path`. Build only in a project that belongs to the work (a NotesCraft design file is for NotesCraft screens, not for redesigning PaperCraft). For a different product: `create_file {name}`, then `open_file {fileId}`, and say so. `open_file` swaps the project the user is looking at, so tell them. Pass `fileId` on later calls when more than one project is in play.
 * Not running: ask the user to start PaperCraft and open a project (or `create_file` then `open_file`). `PAPERCRAFT_MCP_PORT` must match if the port was changed. Do not try to launch the GUI yourself unless asked.
-* Read `get_basic_info` (pages, artboards, sizes, fonts), then `get_selection` (what the user is looking at) and `get_variables` and `get_components`. Reuse what exists before creating anything.
+* Read `get_basic_info` (pages, artboards, sizes, fonts, and `lessons`: the person's rules from past critiques; follow every one), then `get_selection` (what the user is looking at) and `get_variables` and `get_components`. Reuse what exists before creating anything.
 * `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 -`.
@@ -33,8 +33,10 @@ papercraft-cli call get_basic_info      # fails with exit 1 if the app is not ru
 3. **Build in small steps.** `create_artboard`, then one `write_html` per visual group (header, card, row) so the user sees progress. Name every layer with `layer-name`.
 4. **Look.** `get_screenshot` of the artboard after each group or two. Scale 2 to read small text. Read it as a designer: hierarchy, alignment, spacing rhythm, clipping, wrapping.
 5. **Lint.** `lint_design` on the artboard. Fix every error and warning; consider the `info` items. Run it again until errors and warnings are zero.
+   Between edits, lint only what changed: keep the `rev` each write returns and run `lint_design {changedOnly: true}` (or `{since: rev}`). `get_changes {since}` shows what an edit touched; `get_bounds` measures a whole subtree in one call instead of looping `get_node_info`.
+   Always finish with one full `lint_design` (no `since` or `changedOnly`: `covered.partial` is false) and a screenshot per artboard in each theme mode before reporting. Details: [incremental-checks.md](references/incremental-checks.md).
 6. **Fix with the smallest edit.** `update_styles` and `set_text_content` on the existing nodes, not a rewrite. Rewrite a group with `write_html` mode `replace` only when its structure is wrong.
-7. **Finish.** `finish_working_on_nodes` (saves). Tell the user what you made in a sentence or two, with artboard names, not ids.
+7. **Finish.** `finish_working_on_nodes` (saves). Tell the user what you made in a sentence or two, with artboard names, not ids. When its result has `next`, offer a critique with `ask_user`; on yes run `critique_design` and follow its steps (the `papercraft-critique` skill has the full gate). Keep a finding as a lesson (`design_lessons {add}`) only when the person approved it.
 
 Lint passes are necessary, not sufficient: contrast and grids say nothing about composition. The screenshot is the real test.
 
@@ -44,8 +46,12 @@ Lessons from redesigning a 3-page, 130-screen project with up to six agents at o
 
 * **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.
+* **See what the others changed** before building on their work: `get_changes {since: <your last rev>, exclude: "self"}` lists layers by the client that changed them. `truncated: true` means re-read everything.
+* **Name yourself and fence your area.** With the CLI pass `--as <name>` (or `PAPERCRAFT_AGENT`), so writes, conflicts and claims carry your name. `claim_nodes {ids, note}` the sections you will rewrite; others' writes there are refused until you `finish_working_on_nodes` or `release_claims`. A denial names the holder: work elsewhere, never wait in a loop.
+* **Writes report conflicts.** A result with a `swarm` block lists layers another agent changed since you last looked: read them with `get_changes` before going on. Pass `ifRev: <your last rev>` to have the write refused instead (`"error": "conflict"`).
+* **Undo takes back only your own steps** (`undo`, scope `mine`); it is refused with `undo_blocked` when someone changed the same layers since. Fix forward then; `scope: "page"` undoes whoever was last.
 * **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).
+* **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` (a different icon is a different component); fix with `write_html` mode replace and the `<x-paper-instance>` it names.
 * **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.
 
@@ -89,7 +95,7 @@ For a screen with dozens of parts, drive the CLI from a short Python helper (`ca
 ```
 
 Artboards: mobile 390×844, tablet 768×1024, desktop 1440×900, large desktop 2560×1440 (a 1440p display; ask which the user uses and design that first). A narrow window (1280×800) is its own artboard: collapse the sidebar to an icon rail, make secondary actions icon-only with a tooltip, keep one labelled primary action.
-Dot-grid canvas (no pattern fill exists): one `<svg>` with a single path of tiny round-capped dashes, absolute at 0,0 inside the canvas frame: `<svg style="position:absolute;left:0px;top:0px" width="W" height="H" viewBox="0 0 W H"><path d="M16 16h0.01M48 16h0.01..." stroke="var(--surface-2)" stroke-width="2.5" stroke-linecap="round" fill="none"/></svg>`. One vector layer, about 3,000 dots at a 32px step. Its colour is frozen at write time. Create with `height:"fit-content"` when the content decides the height. Replace margin with `gap` and `padding`.
+Dot-grid canvas (no pattern fill exists): one `<svg>` with a single path of tiny round-capped dashes, absolute at 0,0 inside the canvas frame: `<svg style="position:absolute;left:0px;top:0px" width="W" height="H" viewBox="0 0 W H"><path d="M16 16h0.01M48 16h0.01..." stroke="var(--surface-2)" stroke-width="2.5" stroke-linecap="round" fill="none"/></svg>`. One vector layer, about 3,000 dots at a 32px step. Its colour is frozen at write time. A screen keeps its device's size; at a device width (390, 1440...) `height:"fit-content"` becomes the device height. Content taller than the screen scrolls by itself: the main column under a fixed sidebar and header, else the screen (write_html says which). Pass `overflow:"visible"` to let a board grow (a landing page drawn in full). Replace margin with `gap` and `padding`.
 
 Full flows: [recipes.md](references/recipes.md).
 
@@ -100,9 +106,11 @@ 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` |
+| Sizes and positions of a subtree | `get_bounds` (one call, `columns` + `rows`; `continueWith` when `truncated`) |
+| What changed since a revision | `get_changes {since}` (`exclude: "self"` for other agents' edits) |
 | 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) |
+| See it | `get_screenshot` (`nodeId`, `scale` 1 or 2, `annotations: true` to include notes, `mode` to preview a theme without switching) |
 | 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) |
@@ -112,18 +120,22 @@ Full flows: [recipes.md](references/recipes.md).
 | 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` (also a nested badge inside a row instance: a swap that survives main edits), `get_components`, `<x-paper-instance>` |
+| Component properties | One component with properties instead of one per state: `set_component_props` (`text`, `bool`, `icon`, `enum` options in the overrides format; templates `text`, `active` for tab bars and navs, `status` with [{value, color}]), then `set_instance_props {nodeIds, values}` or `<x-paper-instance component="Badge" props='{"Status":"Rejeitada"}'/>`; `get_components` lists them |
+| Merge per-state components | `merge_variants {componentIds, name}` turns `Tab bar / Hoje`, `Tab bar / Lojas`... into one `Tab bar` with an enum, moves every instance and deletes the emptied variants |
+| Prototype | `set_prototype_links {links:[{nodeId, to}]}`: back arrows, rows and tabs go to their screens in Present mode |
+| Screens | Fixed size (390x844, 1440x900...); content longer than the screen scrolls by itself (the main column under a fixed sidebar and header, else the screen), or set `overflow: auto` on the frame that should scroll; it keeps its size and the overflow scrolls on the canvas and in Present (inspector: Scroll overflow; page: Screens scroll) |
 | 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 |
+| Design in a library's look, hand off its code | `apply_ui_kit` (shadcn, mui, mantine, bootstrap, chakra: ~48 components and 5 screen layouts each, re-colour through their variables; 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` |
+| Check quality | `lint_design` (`changedOnly: true` or `since` between edits, a full pass at the end; `overflow` flags layers past their artboard or clipping parent) |
+| Export | `export` (PNG, SVG, PDF to files), `export_combined_pdf`, `export_presentation` (a client deck: cover, contents, a slide per screen) |
 | Point at things for the user | `select_nodes` (`zoom: true`) |
 | Take back | `undo`, `redo` |
 | Save | `finish_working_on_nodes` |
@@ -156,7 +168,7 @@ Full flows: [recipes.md](references/recipes.md).
 | `papercraft-cli` exits 1, "isn't running" | The app is closed, no project is open, or the port differs. Start it and open a project; check `PAPERCRAFT_MCP_PORT`. |
 | Call fails with an unknown node id | The id is from another page or was deleted. Re-read with `get_children`; pass `pageId`. |
 | `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. |
+| Content is cut off | A frame inside the screen clips it (`overflow: hidden`, or a fixed height under a hugging parent). Give the column that should scroll `overflow: auto`; for a board that should grow, `overflow: "visible"` and `height: "fit-content"`. |
 | 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 | 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. |

skills/papercraft/references/compact-workflow.md

@@ -37,6 +37,6 @@ How to keep designs small, keep design and code in step, and spend few tokens do
 
 ## Sharing and several agents
 
-* Several agents may work on one open project through one MCP server: calls run one at a time, each is one undo step, so they do not corrupt each other. Give each agent its own page or artboard prefix and say so in the brief, so they do not restyle each other's nodes.
+* Several agents may work on one open project through one MCP server: calls run one at a time, each is one undo step, so they do not corrupt each other. Give each agent its own page or artboard prefix and say so in the brief, so they do not restyle each other's nodes. Each agent names itself (`papercraft-cli --as <name>`), claims what it rewrites (`claim_nodes`), reads the conflicts its writes report, and undoes only its own steps.
 * Several people on a shared project: pages sync as files and are merged when both sides changed the same page (nodes by id, child lists as sets). Edits to different nodes both survive; the same property changed twice keeps the local value. Ask people to keep a page per person when restyling the same frame.
 * While an agent works the app shows an animated "working" pill and a ring around the canvas; call `finish_working_on_nodes` when done so it stops.

skills/papercraft/references/html-subset.md

@@ -28,14 +28,16 @@
 
 Supported: `display:flex` (and `inline-flex`), `flex-direction`, `flex-wrap`, `gap`, `row-gap`, `column-gap`, `padding` (shorthand, per side, `padding-inline`, `padding-block`), `align-items`, `justify-content` (including `space-between`), `align-self` (`start`, `center`, `end`, `stretch`), `flex-grow` / `flex: 1`, `width`, `height` (px, %, `fit-content`), `min-width`, `max-width`, `min-height`, `max-height`, `overflow:hidden` (clips), `position:absolute` with `left`/`top`/`right`/`bottom`, `display:none`, `visibility`.
 
-Not supported (check `ignoredStyles` to be sure): `margin` (use `gap` and `padding`), `display:grid`, tables, `float`, `calc()`, viewport units, `transform` and rotation.
+Rewritten before layers are made, each reported in `warnings`: `margin` becomes the parent's `gap` (when uniform) or spacer layers, `margin-*:auto` a growing spacer and `margin:0 auto` centring; `display:grid` becomes rows of cells (`fr` tracks grow, px tracks keep their width, `repeat(auto-fill, minmax())` counts columns from the width); inline boxes (chips, badges) become boxes in a wrapping row; `<input>` and `<textarea>` become a box showing their placeholder or value. Prefer `gap` and `padding` anyway: the result is the same and shorter.
+
+Not supported (check `ignoredStyles` to be sure): tables, `float`, `calc()`, viewport units, `transform` and rotation.
 
 Tips:
 
 * Artboards are `display:flex; flex-direction:column`. Children of a column stretch to the full width unless they have a width or `align-items` / `align-self` says otherwise.
 * To center something, put it in a flex parent with `align-items:center; justify-content:center`.
 * To make a row fill the remaining space, give the child `flex-grow:1`.
-* Use `height:"fit-content"` on artboards that should grow with content; fixed heights clip.
+* A screen keeps its device's size; at a device width (390, 1440...) `height:"fit-content"` becomes the device height. Content taller than the screen scrolls by itself: the main column under a fixed sidebar and header, else the screen (write_html says which). Pass `overflow:"visible"` to let a board grow (a landing page drawn in full).
 
 ## Text
 

skills/papercraft/references/incremental-checks.md

@@ -0,0 +1,50 @@
+# Incremental checks
+
+Check what you just changed between edits, and check everything once at the end. This keeps the loop fast and the context small without skipping the final full pass.
+
+## Why
+
+* A full `lint_design` on a big page walks every layer and returns every finding. Run after each edit, it repeats the same output and burns tokens.
+* Looping `get_node_info` over nodes is slower still: one measured loop over 397 nodes took about 53 s. One `get_bounds` call returns the same geometry for a whole subtree in milliseconds, in one result.
+* Reads never change the document, so they never change `rev`. Only writes do.
+
+## The flow
+
+1. **Note `rev`.** Every write tool result includes `rev`, the document revision after the call. `get_basic_info` reports the current one. Keep the last value.
+2. **Edit.** Make the change (`write_html`, `update_styles`, `set_text_content`, ...). Note the new `rev` from the result.
+3. **Lint what changed.** `lint_design {"changedOnly": true}` checks since this client's last lint. `lint_design {"since": <rev>}` checks since a revision you name. A scoped lint covers the touched layers plus their ancestors, descendants and siblings, and every layer bound to a changed variable. The result says how much it covered, for example `{"checked": 12, "total": 397, "since": 40, "partial": true}`.
+4. **See what an edit really touched.** `get_changes {"since": <rev>}` returns `{rev, since, truncated, changes: [{page, ids}]}`. Add `page` to limit it to one page. If `truncated` is `true`, the log no longer reaches that far back: do a full pass instead of trusting the list.
+5. **Measure in one call.** `get_bounds {"nodeIds": [...]}` (or `pageId`, with `depth`) returns `columns` `[id, parent, x, y, w, h, absX, absY, clip]` and `rows`, one array per layer, parents first. If `truncated` is `true`, call it again with `continueWith` as `nodeIds` (same `depth`). Never loop `get_node_info` to read sizes or positions.
+6. **Finish with the full check.** Before you report, always:
+   * run `lint_design` with no arguments (a full pass, `partial: false`) and fix every error and warning, including `overflow`;
+   * take a `get_screenshot` of every artboard you touched in each theme mode (`mode`, scale 2).
+
+A partial lint is a speed-up, not a verdict. Only the final full pass can say the design is clean.
+
+## Example
+
+```json
+get_basic_info {}
+// {"rev": 40, ...}
+
+write_html {"targetNodeId": "a1:1", "mode": "insert-children", "html": "<div class='Card'>...</div>"}
+// {"rev": 41, ...}
+
+lint_design {"changedOnly": true}
+// {"errors": 0, "warnings": 1, "covered": {"checked": 12, "total": 397, "since": 40, "partial": true}, ...}
+
+update_styles {"nodeIds": ["a1:7"], "styles": {"width": "320px"}}
+// {"rev": 42, ...}
+
+get_changes {"since": 41}
+// {"rev": 42, "since": 41, "truncated": false, "changes": [{"page": "p1", "ids": ["a1:7"]}]}
+
+get_bounds {"nodeIds": ["a1:1"], "depth": 3}
+// {"columns": ["id", "parent", "x", "y", "w", "h", "absX", "absY", "clip"],
+//  "rows": [["a1:1", null, 0, 0, 390, 844, 0, 0, true], ...]}
+
+lint_design {}
+// full pass, "partial": false, zero errors and warnings
+get_screenshot {"nodeId": "a1:1", "scale": 2, "mode": "light"}
+get_screenshot {"nodeId": "a1:1", "scale": 2, "mode": "dark"}
+```

skills/papercraft/references/recipes.md

@@ -57,7 +57,7 @@ For the next screens, create each artboard at 390x844, reuse the components by `
 
 ## 2. Landing page (1440 wide, height grows with content)
 
-Create the artboard with `height:"fit-content"` so it never clips, then add one section per call.
+Create the artboard with `height:"fit-content", overflow:"visible"` so it grows with its sections instead of keeping the 900px screen height, then add one section per call.
 
 ```sh
 AB=$(papercraft-cli call create_artboard '{"name":"Landing","styles":{"width":"1440px","height":"fit-content","backgroundColor":"var(--bg)","display":"flex","flexDirection":"column","paddingBottom":"96px"}}' | jq -r .id)

skills/papercraft/references/tools.md

@@ -79,6 +79,17 @@ Details of one node: type, name, size, visibility, lock state, parent, children
 | `nodeId` | string | yes | The node to inspect. |
 
 
+### `get_bounds` (read-only)
+
+Rects of a whole subtree in one read, instead of one get_node_info per layer. Returns `columns` [id, parent, x, y, w, h, absX, absY, clip] and `rows` (one array per layer, parents before children): x/y/w/h are relative to the parent, absX/absY are page coordinates, clip is true when the layer clips its content (clipping or scrolling frame), parent is null for a top-level layer. At most 2000 rows per call: then `truncated` is true and `continueWith` lists the node ids to pass as nodeIds. Use it to verify nothing leaves its artboard, and `lint_design`'s `overflow` rule to list the offenders.
+
+| Argument | Type | Required | Notes |
+|---|---|---|---|
+| `depth` | number |  | Levels returned, counting each given node as 1 (default: all). |
+| `nodeIds` | string[] |  | Subtrees to measure; omit for all artboards on the page. |
+| `pageId` | string |  | Page to use (default: the active page). |
+
+
 ### `get_children` (read-only)
 
 Direct children of a node (or of the page root when nodeId is "root"), with ids, names, types, child counts and positions.
@@ -207,6 +218,7 @@ 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. |
+| `mode` | string |  | Render in this variable mode (e.g. "Light") without switching the document: preview every theme of one artboard while other agents keep working in the current mode. |
 | `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. |
@@ -232,11 +244,24 @@ Export several nodes into one PDF, one page per node, ordered top-to-bottom then
 | `path` | string |  | Output file (default: Downloads/<project>.pdf). |
 
 
+### `export_presentation`
+
+Export pages as a presentation PDF, the way a designer sends work to a client: a cover, contents, a divider per page, one 1920×1080 slide per screen (its name and size; narrow side panels share a slide) and a closing slide, numbered. Screens are the page's declared screens, else its top-level frames of 300×300 or more; a section holding no screen (a design-system board) gets a slide of its own. They stay vector. Returns the file path.
+
+| Argument | Type | Required | Notes |
+|---|---|---|---|
+| `pageIds` | array of string |  | Pages to include, in this order (default: every page). |
+| `path` | string |  | Output file (default: Downloads/<title>.pdf). |
+| `title` | string |  | Deck title (default: the project name). |
+| `subtitle` | string |  | Cover subtitle (default: screen and page counts). |
+| `author` | string |  | Shown on the cover after the date. |
+
+
 ## Write
 
 ### `create_artboard`
 
-Creates an artboard (top-level frame) and returns its id; then add content with write_html({mode: "insert-children"}). `styles` are camelCase CSS; width and height are required (e.g. "1440px"). Artboards default to display:flex, flexDirection:column and a white background, and are placed in an empty spot unless left/top are given. Default sizes: desktop 1440×900, tablet 768×1024, mobile 390×844.
+Creates an artboard (top-level frame) and returns its id; then add content with write_html({mode: "insert-children"}). `styles` are camelCase CSS; width and height are required (e.g. "1440px"). Artboards default to display:flex, flexDirection:column and a white background, and are placed in an empty spot unless left/top are given. Default sizes: desktop 1440×900, tablet 768×1024, mobile 390×844. At a device width, height "fit-content" becomes the device height; content taller than a screen scrolls in its main column (else the screen). overflow "visible" lets the board grow instead.
 
 | Argument | Type | Required | Notes |
 |---|---|---|---|
@@ -248,7 +273,7 @@ Creates an artboard (top-level frame) and returns its id; then add content with
 
 Parse HTML with inline styles into layers. Write incrementally: each call should add one visual group (a header, one list row, a card) so the user sees progress.
 - mode "insert-children" appends the HTML inside targetNodeId ("root" = the page); "replace" replaces targetNodeId with it.
-- Inline styles only. Flexbox (display:flex, flex-direction, gap, padding, align-items, justify-content, align-self, flex-grow, flex-wrap) is the layout model; block elements stack vertically. position:absolute with left/top is supported. No margin, grid or tables.
+- Inline styles only. Flexbox (display:flex, flex-direction, gap, padding, align-items, justify-content, align-self, flex-grow, flex-wrap) is the layout model; block elements stack vertically. position:absolute with left/top is supported. Margins become gaps or spacers, display:grid becomes rows of cells, text inputs show their placeholder (each reported in warnings). No tables.
 - Text: font-size, font-weight (≥600 = bold), font-style, color, line-height, letter-spacing, text-align, text-decoration (including wavy and text-decoration-color); <b>/<i>/<u>/<s>/<span style> make rich text. Font family is always Inter.
 - Boxes: background(-color/-image: colors, linear/radial gradients, url()), border (and border-top/right/bottom/left), border-radius, opacity, overflow:hidden, mix-blend-mode, width/height (px, %, fit-content), min-width/max-width/min-height/max-height.
 - <img src> takes paper-asset:///abs/path, file:// URLs, absolute paths or data: URIs. <svg> markup is imported as vector layers.
@@ -502,9 +527,31 @@ Search installed open icon sets (Lucide, Tabler, Phosphor, Heroicons, Material).
 | `set` | string |  |  |
 
 
+### `critique_design`
+
+Critique finished work the way a design team does (offer it with ask_user when finish_working_on_nodes says the design changed). Judges against the person's objective, through independent reviewers: vision (the person's brief and notes), ux (Nielsen's heuristics), flow (coherent journeys across screens), qa (states and edge cases), visual, accessibility (W3C POUR), copy and end-user (a walkthrough as the audience). Returns the reviewers' briefs, the finding format and 0-4 severity scale, lint, the arrows between screens (dead ends, unreached screens), the person's lessons and the steps: review each persona independently, merge, fix severity 3-4, report, then ask which findings to keep as lessons.
+
+| Argument | Type | Required | Notes |
+|---|---|---|---|
+| `nodeId` | string |  | Artboard or subtree to critique; omit for the whole page. |
+| `objective` | string |  | What the design must achieve, for whom, and its top tasks (from the brief or the person). |
+| `audience` | string |  | Who the end-user reviewer plays ("first-time shopper on a phone"). |
+| `personas` | array of string |  | Reviewers to run (default: all): vision, ux, flow, qa, visual, accessibility, copy, end-user. |
+
+
+### `design_lessons`
+
+The person's design lessons: rules from past critiques that every design must follow (get_basic_info lists them too; read them before designing). Without arguments, lists those that apply in the open project. `add` stores one ONLY after the person approved it in this conversation (ask_user); never on your own. `remove` deletes one by id.
+
+| Argument | Type | Required | Notes |
+|---|---|---|---|
+| `add` | object |  | `{rule, why, scope}`: rule is one line, an instruction; why is what went wrong; scope `all` (default) or `project`. |
+| `remove` | string |  | Id of a lesson to delete ("L3"). |
+
+
 ### `lint_design`
 
-Check a page or subtree for common design mistakes and get a fix for each: text contrast (WCAG 4.5:1, 3:1 for large text), text under 12px, font sizes off the type scale, gap/padding off the 4px grid, buttons and tabs under 44px tall, default layer names (Frame, Rectangle), and hard-coded colors that equal a variable. Run it before finish_working_on_nodes and fix errors and warnings; `info` items are suggestions. Pass `fix: true` to automatically snap spacing/padding to 4px, relink matching color variables, snap text sizes, and enlarge small touch targets.
+Check a page or subtree for common design mistakes and get a fix for each: text contrast (WCAG 4.5:1, 3:1 for large text), text under 12px, font sizes off the type scale, gap/padding off the 4px grid, buttons and tabs under 44px tall, default layer names (Frame, Rectangle), layers or text overflowing their artboard or clipping parent (rule `overflow`, never auto-fixed; scrolling frames are exempt), text too narrow for its words (text-collapsed), and hard-coded colors that equal a variable. Run it before finish_working_on_nodes and fix errors and warnings; `info` items are suggestions. Pass `fix: true` to automatically snap spacing/padding to 4px, relink matching color variables, snap text sizes, and enlarge small touch targets.
 
 | Argument | Type | Required | Notes |
 |---|---|---|---|
@@ -673,12 +720,48 @@ Move the user's view: switch to pageId (as clicking it in the page list does) an
 
 ### `undo`
 
-Undo the last change on a page (the user's or an agent's).
+Undo a change. An agent's undo (scope `mine`, its default) takes back its own newest step and leaves later steps by others in place; it is refused (`undo_blocked`, with the steps in the way) when one of them touched the same layers. Scope `page` undoes the page's last step, whoever made it.
+
+| Argument | Type | Required | Notes |
+|---|---|---|---|
+| `scope` | string |  | `mine` or `page`. |
 
 
 ### `redo`
 
-Redo the last undone change on a page.
+Redo the last undone change on a page. With scope `mine` (an agent's default) only when that step was the caller's own.
+
+| Argument | Type | Required | Notes |
+|---|---|---|---|
+| `scope` | string |  | `mine` or `page`. |
+
+
+### `claim_nodes`
+
+Several agents on one project: claim the layers you will rewrite (each with everything under it) or a shared resource, so other agents' writes there are refused while you work. All or nothing; a denial names the holder and when it expires (never wait in a loop: do other work). The claim lasts ttlSeconds after your last call and goes with finish_working_on_nodes or release_claims. The person is never blocked. Needs the running app.
+
+| Argument | Type | Required | Notes |
+|---|---|---|---|
+| `ids` | array of string |  | Layer ids. |
+| `resources` | array of string |  | `variables`, `components`, `pages`. |
+| `ttlSeconds` | integer |  | Default 300 (10 to 1800); renewed by each of your calls. |
+| `note` | string |  | What you are doing, shown to others. |
+
+
+### `release_claims`
+
+Release your claims: one by claimId, those on ids or resources, or all of them when nothing is given.
+
+| Argument | Type | Required | Notes |
+|---|---|---|---|
+| `claimId` | string |  |  |
+| `ids` | array of string |  |  |
+| `resources` | array of string |  |  |
+
+
+### `list_claims` (read-only)
+
+Who is working where: every live claim with its holder, layers, resources, note and seconds left.
 
 
 ### `finish_working_on_nodes`

specs/component-props/spec.md

@@ -0,0 +1,46 @@
+# Component properties
+
+## Why
+
+A status badge, a tab bar or a list row is drawn once as a component, but every use differs:
+the status is Autorizada, Rejeitada or Na SEFAZ; a different tab is active; the label and the
+icon change. Today each case is its own component (`Badge / Green`, `Tabs / Lojas`, `Tabs /
+Pedidos`) or hand edits on layers, so a design needs a component per state and an agent edits
+layer by layer.
+
+## What
+
+1. **Properties live on the main component** (`Node.props`), each with a name and a type:
+   * `text`: the text of a layer (`layer`).
+   * `bool`: shows or hides a layer.
+   * `icon`: the icon of an icon layer (`Icon / set:name`), any `set:name`.
+   * `enum`: a list of options; each option is a set of changes in the `overrides` format
+     (`{"Label": "Rejeitada", "$root": {"background": "var(--danger-bg)"}, "Dot": {"icon": "lucide:x"}}`),
+     so an option can restyle, retext, show, hide or re-icon any layer. Fully custom.
+2. **Instances keep their values** (`Node.prop_values` on the instance root) and apply them as
+   overrides, so edits to the main keep them; editing the definitions re-applies every
+   instance's values on every page.
+3. **Templates** make the common cases one call: `text` (a text property per text layer),
+   `active` (an enum choosing which item of a row, tab bar or nav is active, with active and
+   inactive styles) and `status` (an enum of statuses, each a colour that tints the root and
+   colours its text and icon).
+4. **Where to use it**: `set_component_props` (define, or from a template),
+   `set_instance_props` (many instances in one call), `get_components` (lists properties),
+   `<x-paper-instance component="Badge" props='{"Status":"Rejeitada"}'/>` in `write_html`,
+   the inspector (a control per property on an instance; templates on a main), and code
+   links (`{Status}` in a usage template is the instance's value).
+
+## Not now
+
+Variant sets made of several mains, nested property exposure (a row exposing its badge's
+properties), numeric properties.
+
+## Acceptance
+
+* A badge with an enum Status (Autorizada, Rejeitada) shows the option's colours and label on
+  each instance; switching the value updates only that instance.
+* Editing the main (gap) keeps every instance's values; editing an option's colour updates
+  every instance using it.
+* The `active` template on a five-item tab bar makes one property whose options are the item
+  names; each instance shows its active item highlighted.
+* An icon property swaps the icon and survives a sync of the main.

specs/design-critique/spec.md

@@ -0,0 +1,42 @@
+# Design critique and lessons
+
+## Problem
+When an agent finishes a design it says it is done; nobody reviews the work unless the person asks,
+and the `papercraft-critique` skill only reaches agents that load skills (not the in-app chat). The
+same mistakes come back in the next design (a page that grows instead of scrolling, gendered copy,
+clipped labels) because nothing the person corrected is remembered.
+
+## Goal
+After a design is finished the person is offered a critique; the corrections they approve are
+remembered and shown to every agent before its next design, so the same mistake is not repeated.
+
+## Non-Goals
+- Scoring by a model inside the app: the agent does the judging, the tool gathers the evidence.
+- Saving anything the person did not approve.
+- A settings screen for lessons (the tool lists and removes them).
+- Syncing lessons between machines.
+
+## User Stories
+- As a designer, I want to be offered a critique when an agent finishes, so that I don't have to remember to ask.
+- As a designer, I want to choose which critique findings become lessons, so that only what I agree with sticks.
+- As a designer, I want agents to follow my lessons in later designs, so that I don't correct the same thing twice.
+- As a designer, I want lessons for one product to stay in that product, so that a brand rule does not leak elsewhere.
+- As a designer, I want to remove a lesson, so that a rule I no longer want stops applying.
+
+## Acceptance Criteria
+- [ ] Given edits since the last critique, when an agent calls `finish_working_on_nodes`, then the result tells it to offer the person a critique (`ask_user`), and it does not after a critique with no edits since.
+- [ ] Given a page or artboard, when `critique_design` runs, then it returns the objective (or that it is missing and must be asked), the chosen reviewers (vision, ux, flow, qa, visual, accessibility, copy, end-user) with their methods, the finding format and 0-4 severity scale, the lint, the flow between screens and the lessons that apply.
+- [ ] Given screens joined by arrows, when `critique_design` runs, then the flow lists the links, the dead ends (reached, leading nowhere) and the screens nothing leads to or from.
+- [ ] Given several reviewers, the steps have them review independently (one subagent each, or one pass each) and a coordinator merge, never debate.
+- [ ] Given the person approved a finding, when `design_lessons` adds it, then it is stored in the project store and listed by `get_basic_info` and `critique_design` from then on, in every session.
+- [ ] Given a lesson saved with scope `project`, when another project is open, then it is not listed.
+- [ ] Given a lesson id, when `design_lessons` removes it, then it is no longer listed.
+- [ ] The tool descriptions say a lesson is saved only after the person approved it.
+
+## Constraints
+- Stack: Rust, MCP tools in `src/mcp`; lessons in a JSON file at the project store root, written atomically.
+- Compat: missing or unreadable lessons file means no lessons, never an error for other tools.
+- Security: lessons are plain text the person approved; no secrets.
+
+## Open Questions
+- None.

specs/presentation-export/plan.md

@@ -0,0 +1,20 @@
+# Presentation export v2: plan
+
+All in `src/io/deck.rs` unless noted; the PDF writer and tools stay as they are.
+
+## Tasks (each with a test in `deck.rs`)
+1. **Notes per screen.** `notes_of(doc, screen)`: attached notes whose target is the screen or inside it, in reading order of their targets. Category colour from the project's categories, else the built-ins.
+2. **Annotated screen slide.** `Slide::Screen` with notes: the screen fitted in the left 64% of the stage, numbered accent markers at each noted layer's top-left, the note column on the right (number, title bold, first body line muted, wrapped). `NOTES_PER_SLIDE` = 6; more continue on a repeat slide titled `… (notes 2/2)`.
+3. **States to an appendix.** Main flow: the first state, its caption `+ N states`; the rest collected per page into Appendix slides after the last page (a divider `Appendix`, rows of two titled `Page · Screen`).
+4. **Map slide.** After each divider: up to 12 key screens (one per state group, boards excluded) as a grid of thumbnails with names; more pages of the map when needed.
+5. **Open questions.** Notes containing `?`, listed with their screen's name, before the end slide; split at 8 per slide.
+6. **Counts.** Contents and dividers say `N screens · M boards`.
+
+## Risks
+- Notes on tiny layers: the marker sits on the layer's top-left, clamped inside the screen.
+- Long note bodies: one line in the column, cut with an ellipsis (text wrap at fixed width, max rows 2).
+- Projects without notes: nothing changes for their slides.
+
+## Data (needs the owner's go-ahead, not code)
+- Interview: about 40 screens are named without ` / ` (`Challenges List No Match`), so they don't group as states; rename to `Screen / Challenges / No match`.
+- Interview: 76 detached notes (their targets were deleted in the v1 cleanup).

specs/presentation-export/spec.md

@@ -0,0 +1,44 @@
+# Presentation export
+
+## Problem
+Exporting a project as PDF gives one bare page per frame: no cover, no titles, no order beyond the canvas, and boards whose content sits on top of them as sibling layers came out empty. Sending designs to a client or team means rebuilding the deck by hand in another tool.
+
+## Goal
+One export produces a client-ready PDF deck of a project's screens, the way a freelance designer presents work.
+
+## Non-Goals
+- Authored narrative (problem statements, research, rationale paragraphs): the deck uses what the project knows (names, pages, sizes).
+- Editable slide documents (Keynote, PowerPoint) or slide editing in the app.
+- Theming the deck from the project's variables.
+
+## User Stories
+- As a designer, I want to export a page or the whole project as a presentation, so that I can send it to a client without rebuilding slides.
+- As an agent, I want a tool and a CLI flag for the same export, so that I can deliver designs as a deck.
+- As a reviewer, I want each screen on its own slide with its name, so that I can refer to it in feedback.
+
+## Acceptance Criteria
+- [ ] Given the Export section, when the format is Presentation and the scope is All pages, then one PDF is written with a cover, a contents slide, a divider per page that has screens, one slide per screen, and a closing slide.
+- [ ] Given a page with declared screens, then the deck shows those screens; otherwise its top-level frames of at least 300 × 300.
+- [ ] Given narrow, tall screens (side panels), then up to five share a slide, each with its name.
+- [ ] Given a board frame with layers laid over it as siblings, then its slide shows those layers (as the canvas does).
+- [ ] Every slide but the cover has the deck title and a page number; screen slides show the page name, the screen name and its size.
+- [ ] Screens stay vector in the PDF (scaled, not rasterised).
+- [ ] `papercraft --export deck.pdf --format presentation --scope all <project>` and the MCP tool `export_presentation {pageIds?, path?, title?, subtitle?, author?}` produce the same deck.
+
+## v2: presentable like a designer's deck
+Research (client-presentation guides, the pptx skill's slide rules): show the big picture before details, annotate key decisions, keep filler (states, look-alike screens) out of the main flow, close on open questions, keep one visual system.
+
+- [x] Given a page, after its divider, then a map slide shows its key screens as a captioned thumbnail grid in reading order.
+- [x] Given a screen with notes on it (or on its layers), then its slide shows the screen with numbered markers on the noted layers and a column of the numbered notes (title, first line of the body, the category's colour); more notes than fit continue on a second slide of the same screen.
+- [x] Given states of one screen (`Sessions / List`, `Sessions / Empty`), then the main flow shows the first, captioned with how many states it has, and the others are in an Appendix section at the end, two per slide.
+- [x] Given notes that ask a question (a `?`), then an Open questions slide before the end lists them with their screen.
+- [x] Given a page with boards, then the contents and its divider count screens and boards apart.
+- [x] Detached notes (their layer was deleted) are left out.
+
+## Constraints
+- Stack: Rust, the existing PDF writer (`io/pdf.rs`); no new dependencies, no browser.
+- Performance: a 120-screen project exports in seconds.
+- Compat: PDF 1.7, readable in Preview, Acrobat and browsers.
+
+## Open Questions
+None: defaults chosen (dark slides, 1920 × 1080, deck title from the file name in the app and CLI).

specs/screens/plan.md

@@ -0,0 +1,29 @@
+# Screens — plan
+
+Built against `spec.md` (decisions resolved 2026-10-27).
+
+## Model
+- `Node::screen` and `Node::overlay` (bools, omitted when false). Shipped in project format v3, which is not released yet, so no extra version bump is needed: older builds already refuse v3.
+
+## Editor (`src/editor/screens.rs`)
+- `Role`, `refusal`, `declared`, `screen_of`, `candidates`, `device_of`, and `Editor::set_role`, which changes roles as one undo step.
+- `coverage`: inside the layer's screen, it finds layers painted after it that are visible and either overlays or opaque (fill at 95% opacity or more), clips them to the screen, and returns the union area as a visible share. A layer under 1% visible counts as hidden; under 99% visible, as partly covered.
+- `Editor::screens()` returns the declared screens, or the top-level frames when none are declared. "Screens scroll" therefore stops toggling sections.
+- Frame labels: declared screens nested in sections get labels. A frame's own ancestors don't count as covering its label.
+
+## MCP
+- `set_screens {nodeIds, role}`.
+- The guard runs on every tool in `GUARDED`. It reads `nodeId`, `nodeIds`, `parentId` and `targetNodeId` at any depth in the arguments.
+  - It refuses hidden targets unless `allowHidden` is set.
+  - It appends a `partlyCovered` or `editedWhileHidden` note to the result.
+- `get_basic_info.screens`, `get_node_info.screen`, `lint_design.skippedHidden` (hidden layers are dropped from the findings).
+
+## UI
+- The frame inspector has Screen and Overlay toggles, shown only where the role is allowed.
+- The page inspector has a "Declare N screens" button listing the candidates, as one undo step.
+- "On click" lists `ed.screens()`.
+
+## Known ceilings
+- Covers are clipped to the screen only, not to frames in between.
+- Lint autofix still edits hidden layers.
+- Export and design-to-code don't use screens yet (planned for later).

specs/screens/spec.md

@@ -0,0 +1,75 @@
+# Screens
+
+## Problem
+PaperCraft has no notion of a "screen": a screen, a section that groups screens, a slot that labels one, and a component main are all just frames. Big projects like Interview (42k layers) nest screens three or four frames deep (Section > Screens > Row > Slot > Screen), so every frame-based feature guesses wrong:
+- The canvas shows labels only for top-level frames, so the nested screens get none.
+- The "On click" link list offers sections and labels instead of screens.
+- Scrolling picks whichever scroll-enabled frame is deepest.
+- Agents cannot tell which frame is the device viewport.
+
+The owner reports a second, related problem. When a drawer or modal is open over a screen, agents still edit, lint and "fix" the content hidden underneath it. The owner can't see that work, it can't be checked in a screenshot, and it often breaks the open-state design. Evidence: owner reports in this session (2026-10-27) about agents working behind open drawers and screen scroll caught by the section frame.
+
+## Goal
+A frame can be declared a Screen. Agents and the editor then treat it as one device viewport, and agent edits to layers fully hidden inside it are refused with a clear reason.
+
+## Non-Goals
+- Prototyping or interactive preview (flows, transitions, overlays opening on click).
+- Responsive variants or auto-generating breakpoints from one screen.
+- Occlusion between separate screens on the canvas, or between sections. Only layers inside one screen count.
+- Pixel-exact occlusion through transparency, blur or blend modes. Translucent covers (scrims) are judged by a rule, not by rendered pixels.
+- Changing how frames that are not screens behave.
+- Migrating existing projects automatically without the owner choosing to.
+
+## User Stories
+- As a designer, I want to mark a frame as a Screen (with a device size), so that PaperCraft and agents know it is a page of the product, not a container.
+- As a designer, I want screens to show their name and device on the canvas even when nested in sections, so that I can read a large page at a glance.
+- As a designer, I want "On click" and other screen pickers to list only screens, so that I can find the target among hundreds of frames.
+- As a designer working with an agent, I want the agent stopped from editing layers that are fully hidden behind an open drawer or modal, so that it can't change what I can't see.
+- As an agent, I want each layer's visibility in its screen (visible, partly covered, hidden) and what covers it, so that I can choose the right target or open the right state first.
+- As a designer, I want scrolling over a screen to scroll that screen's scroll area, not its section, so that I can review long screens in place.
+- As a designer, I want to convert existing screen-like frames into Screens in one step, so that big projects get the benefit without hand-marking each one.
+
+## Acceptance Criteria
+Declaring screens:
+- [ ] Given a frame, when the owner or an agent declares it a Screen, then it is saved as a Screen with a device size, survives save and reload, and the change is one undo step.
+- [ ] Given a frame that is a component main, an instance, or inside another Screen, when someone tries to declare it a Screen, then the request is refused with a message that names the reason.
+- [ ] Given a Screen, when it is undeclared, then it goes back to an ordinary frame and its layers and layout do not change.
+- [ ] Given a project file written before this feature, when it opens, then it loads with no Screens and nothing else changes.
+- [ ] Given a project that has Screens, when an older PaperCraft build opens it, then the behaviour matches the project's existing version-compatibility rules (refused or loaded without screen data), never silent data loss.
+
+Agent tools:
+- [ ] Given a page with Screens, when an agent asks for the page summary, then Screens are listed as screens (name, device, size, section they sit in), separate from other frames.
+- [ ] Given a Screen where an opaque drawer covers a layer entirely, when an agent edits that layer (styles, text, delete, move, write content into it), then the edit is refused. The reason names the covering layer and is returned as an error, not a silent no-op.
+- [ ] Given the same layer only partly covered, when an agent edits it, then the edit succeeds and the result says the layer is partly covered, by which layer, and the visible fraction.
+- [ ] Given the owner, not an agent, edits a fully hidden layer in the app, then the edit is allowed (the guard applies to agent tools only).
+- [ ] Given an agent that passes an explicit override on a refused edit, then the edit succeeds and the result records that it was hidden. (Pending Open Question 3.)
+- [ ] Given a layer that is hidden only because its parent clips it or it scrolled out of view, then it is not treated as covered. Scroll and clipping are reported separately from covering.
+- [ ] Given a Screen with a hidden layer, when an agent lints that Screen, then hidden layers are skipped (or reported once as "hidden") instead of producing fix-me findings.
+- [ ] Given an agent reads a layer's info, then the response includes its visibility state in its Screen: visible, partly covered, hidden, or not in a screen.
+
+Editor:
+- [ ] Given Screens nested in sections, when the canvas is zoomed out, then each Screen shows a name label (subject to the existing label overlap rules), and sections keep theirs.
+- [ ] Given the "On click" picker, when a project has Screens, then it lists Screens (grouped by section) and no other frames.
+- [ ] Given the pointer over a Screen whose own scroll area can still move, when the owner scrolls, then that area scrolls and the section around the Screen does not.
+- [ ] Given a page with screen-like frames (same size as a known device, or named "Screen / ..."), when the owner runs "Convert to Screens", then the owner sees the candidates and confirms, and the declaration is one undo step.
+
+Performance:
+- [ ] Given the Interview Admin page (19k layers), when an agent edits a layer inside a Screen, then the hidden check adds no more than 5 ms to the call.
+
+## Constraints
+- Stack: Rust, egui 0.36, wgpu; existing MCP tools and `papercraft-cli` (online and offline) must behave the same.
+- Performance: no extra cost per canvas frame; the hidden check runs only when an agent tool targets layers.
+- Security: none new. The refusal is a safety rail for agents, not a permission system.
+- Compat: the project format version bump follows the existing format migration rules. Web (wasm) build keeps parity.
+
+## Open Questions
+None open. Resolved 2026-10-27 (owner: "go with your recommendations"):
+1. Cover: a layer later in paint order inside the same Screen that is opaque (solid or image fill at 95%+ effective opacity), or any layer marked as an overlay (drawer, modal, sheet), even if translucent.
+2. Partly covered: an edit passes when at least 1% of the layer's area is visible; the result reports the visible share.
+3. Override: yes, agents can pass an explicit flag; the result records that the layer was hidden.
+4. Overlays: a direct child of a Screen can be marked as an overlay. To work underneath it, agents hide it (an ordinary undo step); no separate toggle.
+5. Device: presets (mobile 390×844, tablet 768×1024, desktop 1440×900, large 2560×1440) or free size. The Screen's height is the viewport; taller content scrolls.
+6. Labels: top-level frames keep labels, and nested Screens gain them. Other nested frames get none.
+7. Convert to Screens: candidates are frames named "Screen / …", frames that are the screen inside a "Slot / …" frame, or top-level frames at a preset size. The owner confirms in the app; agents declare by id only.
+8. Never inside a component main or another Screen.
+9. First release: On click, page summary, lint scope, layer info. Export and design-to-code follow later.

specs/swarm-safety/plan.md

@@ -0,0 +1,85 @@
+# Plan
+
+Read [spec.md](spec.md) first. Phase 0 is the change log and `rev` that another change is adding; nothing here starts before it merges (the harness skeleton and identity can start against its interface).
+
+## Architecture in one paragraph
+
+A new module tree `src/mcp/swarm/` holds the target-agnostic logic (compiles for wasm): `mod.rs` (the `Swarm` struct: claims, `seen` map, settings, and `Swarm::check(...) -> Gate` / `Swarm::finish(...)` / `swarm::attach(client, result)`), `targets.rs` (tool -> targets, `reach`, resources), `conflicts.rs` (the rule over the change log), `claims.rs` (state machine, TTL, overlap), `gate.rs` (the ordered checks, so claims and conflicts are separate files). `App::run_mcp_calls` and `cli::Offline::call` wrap their existing `Tracker::call` with `Swarm::check` before and `swarm::attach` after; nothing else in `exec.rs` changes, so the tools stay oblivious. Per-client undo lives in `history.rs`/`editor` with a thin change in `exec::undo_redo`. Offline and save-time work is in `cli.rs`, `session.rs`, `project/`, `merge.rs`.
+
+## Phases
+
+Size: S under 150 changed lines, M 150-500, L over 500, tests included. "Needs" lists real dependencies; phases with no overlap in files can run in parallel (wave column).
+
+| # | Phase | Needs | Wave | Size | Files touched |
+|---|---|---|---|---|---|
+| 0 | Change log, `rev`, `get_changes`, scoped lint (other agent, external) | | - | | `src/mcp/changes*`, `exec.rs`, `tools.rs` |
+| 1a | Identity (protocol) | 0 | 1 | M | `src/mcp/mod.rs` (`Client.parent`), `src/mcp/server.rs` (`agent_of`), `src/bin/papercraft-cli.rs` (`--as`, `--parent`, env, URL query), `src/cli.rs` (`Offline::new(.., client)`), `src/mcp/delegate.rs` (env), `src/runs.rs` (label for named sessions, run id in log entries) |
+| 1b | Harness skeleton + gate skeleton | 0 | 1 | M | `tests/common/swarm.rs`, `tests/swarm.rs`, `src/mcp/swarm/{mod,gate}.rs` (empty `Gate` chain, `attach`), call sites in `src/app.rs` and `src/cli.rs` |
+| 2 | Conflicts: `ifRev`, rule, warn/strict, settings | 1a, 1b | 2 | L | `src/mcp/swarm/{conflicts,targets}.rs`, `src/mcp/tools.rs` (`ifRev`/`onConflict` on write tools, `reach` test), `src/ui/settings.rs` + `src/settings`-side struct, `src/ui/chat_msgs.rs` (warning chip) |
+| 3 | Claims: tools, state machine, auto-release, delegate scope | 1a, 1b | 2 | L | `src/mcp/swarm/claims.rs`, `src/mcp/tools.rs` (3 tools, `delegate.scope`), `src/mcp/exec.rs` (3 tool arms call `Swarm`), `src/app.rs` (claim before delegate spawn, turn end, project switch), `src/mcp/delegate.rs` (release on exit) |
+| 4 | UI: agent colours, presence pin, claim outlines, Agent view, Break | 1a, 3 (claims part) | 3 | L | `src/theme.rs` (`agent_1..8`), `src/ui/mod.rs` (`mark_agent`, `AgentMark`), `src/ui/panels.rs` (Agent view), canvas painter, `src/ui/palette.rs` (Break all claims) |
+| 5 | Per-client undo and run revert under concurrency | 0, 1a | 2 | M | `src/history.rs` (`author`, `undo_by`, root set-diff), `src/editor/mod.rs`, `src/mcp/exec.rs::undo_redo`, `src/mcp/tools.rs` (`scope`), `src/runs.rs` (`blocker`, `revert`, `foreign_changed`), `src/ui/panels.rs` (Undo this step, kept-layers list) |
+| 6 | Offline: per-call lease, page refresh, app takes the lock, holder info, `Tracker` in `Offline`, `activity.json`, offline claims file | 0, 1a | 2 | L | `src/cli.rs`, `src/session.rs` (`open`), `src/project/mod.rs` (stamp), `src/bin/papercraft-cli.rs`, changes persistence file |
+| 7 | Save-time stamp check, three-way merge, merge report | none (6 helps) | 1 | M | `src/session.rs::save`, `src/merge.rs` (`Merged`/notes), `src/project/packed.rs` (keep bytes), `src/editor/pages.rs` (`restore_snapshot` reuse) |
+| 8 | Skill, docs, tool counts | 2, 3 | 4 | S | `skills/papercraft/SKILL.md`, `docs/agents.md`, `README.md`, `ROADMAP.md`, `src/mcp/tools.rs::INSTRUCTIONS`, chat system prompt (`src/ui/chat.rs`) |
+| 9 | Swarm scenarios and coverage gate | per phase | with each | M | `tests/swarm.rs` grows with each phase |
+
+**Order and parallelism.**
+
+* Wave 1: 1a, 1b, 7 in parallel (disjoint files: 7 touches only session/merge/packed).
+* Wave 2 (after 1a and 1b merge): 2, 3, 5, 6 in parallel. They share only `src/mcp/tools.rs` (different tool entries: add each in its own `fn` group to avoid conflicts) and `src/mcp/exec.rs` (3 and 5 add arms in different places). 2 and 3 plug into the `Gate` chain from 1b as separate files, so they do not edit each other's code. 5 and 6 are independent of 2 and 3.
+* Wave 3: 4 (UI), once 3 exposes the claims API; the colour/pin half of 4 only needs 1a and can start in wave 2 if split (`4a` colours and pin, `4b` claims UI).
+* Wave 4: 8.
+* Merge conflict hot spots to coordinate: `tools.rs` (tool list and the `edits_doc` classification: new tools `claim_nodes`/`release_claims` edit nothing; `list_claims` read-only), `app.rs::run_mcp_calls` (1b owns the call-site edit; 3 adds the delegate claim in its own function), `panels.rs` (4 and 5 add different sections).
+
+## Test plan
+
+**Harness (Phase 1b, grows per phase).** `tests/common/swarm.rs`:
+
+* `Sim { host: SessionHost, tracker: Tracker, swarm: Swarm, clock: FakeClock }` over a temp store (`temp_store`, as `tests/cli_offline.rs`), headless (no GPU unless a step needs a screenshot), single-threaded: determinism comes from scripted order, not threads.
+* `SimClient { client: Client }`; `Step { at_ms, client, tool, args, expect }`; `Sim::run(script) -> Vec<Outcome>`. Every step goes through the **same entry the app uses** (`swarm::call(host, tracker, swarm, client, name, args, now)`, the function `App::run_mcp_calls` and `Offline::call` both call), so the harness cannot drift from production.
+* `Sim::human(edit)` applies a canvas edit through `Editor::edit` with the human author, to test `You`.
+* Scenario builders: `disjoint(n)` (N clients, own artboards), `overlap(n)` (shared frame), `shared_resources()`, `chain()` (delegate parent/child), `churn(seed)` (random ops from `fastrand::Rng::with_seed`).
+* **Oracle.** A deliberately naive reference (`Oracle`) re-implements the section 2 table over the list of executed ops (per-node last-writer map, ancestor walk, per-tool `reach`) in O(n^2) and is compared with the product's `conflicts` and with the final document: every accepted write's value is present unless a later accepted write to the same property exists (no lost update). `churn` runs 200 seeds in CI, 5000 under `--ignored`.
+* Offline: `OfflineSim` runs 3 `Offline` instances on threads against one folder with real locks (different `File` handles conflict on every platform), each doing N edits to its own artboard plus M to a shared one; asserts 3N + M edits in a fresh `Project::load_all`, no torn file, no panic, and that a held idle lock (shell between lines) does not block.
+
+**Per-phase unit tests**
+
+* 1a: `agent_of` parsing and sanitizing, reserved names, `Client` equality with `parent`, env -> flags precedence, `delegate::run` passes env (existing unix test extended with `env | grep`), stable session for `--as`, run label for named sessions.
+* 2: each relation row positive and negative (self, ancestor own-fields vs `children`, descendant only for `subtree`, `children` reach not conflicting on sibling append, variable set vs bind, component, page); implicit reference per `(client, artboard)`; `ifRev` > current -> `history_reset`; strict leaves `rev`, doc, history, runs untouched (assert equality of all four); mode precedence table; human edit reported; conflict cap 20; every tool classified.
+* 3: grant, deny with each `relation`, atomic denial (nothing claimed), own-claim merge, TTL renewal by reads, expiry with fake clock, each auto-release row, limits, `Block` vs `Warn`, parent-chain exemption, human exempt, resource names and `Claim required`, delegate scope required for `build`, claim exists before spawn and is gone after helper exit (stub command).
+* 4: `agent_color` stable and in range; screenshot tests at scale 2 in both themes (`tests/agent_view.rs` style) for outlines, claims tags, Agent view, long names; Break removes the claim and the holder gets the notice.
+* 5: `undo_by` independence matrix (disjoint ok; same node blocked; same parent children blocked; root set-diff ok; design/variables/page_modes/annotations; main component blocked); redo mine; default scope; human Ctrl+Z unchanged; run revert partial with a foreign edit, all-foreign, "Revert anyway".
+* 6: lease around each call; interleaved shells; page refresh keeps other pages' history; app `Session::open` waits and reports the holder; holder info in the lock file; `activity.json` roundtrip, bounded, absent file; offline run/version created; `claims.json` roundtrip and expiry.
+* 7: unchanged stamp writes; changed stamp merges and reports notes; undecodable or newer-format file refuses; merged result is one undo step; no layer dropped (property test over `merge::merge`: node count of merged >= union of kept ids).
+
+**Coverage.** `cargo llvm-cov` (or `cargo tarpaulin` where the maintainer already uses it) on `src/mcp/swarm/`, `history.rs` changes and `cli.rs`; gate at 80% lines for the new modules. UI painting is covered by screenshots, not counted.
+
+## Risks
+
+| Risk | Mitigation |
+|---|---|
+| Conflict noise: agents ignore `conflicts` that fire on every call | `children` is a separate field so sibling appends do not fire; the implicit reference is per artboard; `reach` limits `descendant` to subtree tools; conflicts cap at 20; measure on the harness `churn` scenario (target: no conflict on pure disjoint work, by construction) |
+| Under-reporting from the implicit reference (coarse, artboard-level) | `ifRev` is exact; the skill tells swarm agents to use it |
+| Default `undo` scope change breaks single-agent habits | `scope: "page"` keeps the old behaviour; the tool description says so; only differs when someone else edited in between |
+| Claims fence off the canvas (a stalled agent) | TTL with renewal only by the holder's calls, per-client and per-project limits, human Break, parent exemption |
+| MCP session close not observable with rmcp's `LocalSessionManager` | TTL fallback; Phase 3 checks whether the manager can notify, else adds a `Drop`-guard on the per-session service handle |
+| Per-call offline lease and page refresh cost on large projects | refresh is by stamp, only changed pages; `verify_large.rs`/`perf.rs` get a case |
+| `PageBase` memory (compressed bytes of every saved page) | bytes already exist at save time; keep only for pages saved in this session; evict when the page is unloaded by `lazy_load` |
+| Advisory locks on network folders | unsupported, documented; holder info and stamp check (Phase 7) still catch lost updates after the fact |
+| Merge hides a semantic clash (both changed the same scalar, local wins) | the merge report surfaces each such scalar; the human decides |
+| Change log assumptions differ from what Phase 0 ships | all dependence is in `conflicts.rs` and the entry struct; reconcile before wave 2 |
+| `tools.rs` / `app.rs` merge conflicts between parallel agents | file ownership table above; each phase adds code in new files and one-line call sites |
+
+## Open questions for the maintainer
+
+1. **Default for claims:** `Block` (proposed) or `Warn`? Block makes claims meaningful but can stop an agent in a way the user sees as "stuck"; the denial message names the holder and time.
+2. **`undo` default scope `mine`:** accept the behaviour change for single-agent use, or keep `page` as the default and make agents opt in through the skill? (Proposed: `mine`.)
+3. **Session close hook:** is it acceptable to patch or wrap rmcp's session manager to learn of a closed session, or is TTL expiry enough?
+4. **Human `locked` flag vs agents:** should agent writes honour `Node.locked` (reject with a clear error)? It is a one-line check in `swarm::targets` and would make "lock this for me" work with agents; it also changes today's behaviour where agents edit locked layers. Out of scope here unless you say otherwise.
+5. **`delegate` `scope` required for `build`:** is a hard error acceptable, or should it default to the caller's current claims?
+6. **Offline runs:** is it right that offline edits now create runs and auto-versions (disk cost per CLI invocation), or should they be opt-in (`--run`)?
+7. **Merge on save (Phase 7):** worth its memory and complexity now, or defer until a sync-less second-writer case actually occurs? It is independent and can be dropped without affecting the rest.
+8. **Notice/conflict placement:** trailing `swarm` content block (proposed) vs adding keys to the tool's JSON when it is an object. The block is uniform for text-only and image results; keys are easier for some clients to parse.
+9. **Naming collision:** `rev` (log) vs `Editor::revision` (redraw counter). Rename the editor field (touches many files) or keep both and document?
+10. **Where does the human's mid-edit text field count?** Proposed: nowhere (committed on blur). Confirm.

specs/swarm-safety/spec.md

@@ -0,0 +1,354 @@
+# Swarm safety: several agents on one project
+
+## Why
+
+One project can now have many writers at once: the in-app Chat agent, one or more external MCP clients (Claude Code, Codex), `papercraft-cli` calls online and offline, `delegate`d helpers, the person at the canvas, and (through the relay) other machines. Today the *transport* is safe: every online call is serialized on the UI thread and is one undo step, so nothing tears. What is missing is the *semantics*: an agent that read a frame, thinks for 40 s and then overwrites it does not learn that another agent rebuilt it meanwhile; two agents can both edit `variables`; `undo` undoes whoever was last; an offline CLI and a second process can race on the folder. This spec makes those cases visible, mostly non-blocking, and recoverable.
+
+Design stance: **cooperative and fail-fast**. No tool call ever waits on another agent (no deadlock is possible). Conflicts are reported by default and rejected only on request. The human always wins and is never blocked.
+
+## Builds on (do not redesign)
+
+A change log with a document revision is being added by another change. This spec assumes exactly this interface and names it `rev`:
+
+* `rev: u64`, bumped once per committed write. Every write result returns `rev`. It is **not** `Editor::revision` (`src/editor/mod.rs:253`), which is a redraw counter that `App::run_mcp_calls` bumps after every call.
+* A bounded in-memory log of entries `{rev, client, tool, page, nodes[], fields[], variables?, at}`; `client` is the label of the caller; the human is `You`.
+* `get_changes {since, exclude:"self"}` and `lint_design {since | changedOnly}`.
+
+Two small additions are needed from that log and are tracked as Phase 0 follow-ups below: entries must say **which fields** of a node changed and keep `children` as its own field name (so "another agent added a child" is distinguishable from "another agent restyled the frame"), and the log must survive restart (see 6.4). If the other change lands differently, only `src/mcp/swarm/conflicts.rs` adapts.
+
+## What exists today (checked 2026-10-28)
+
+* **One chokepoint, online.** `server.rs::call_tool` builds a `Call {name, args, client, reply}`, `Dispatch` sends it down an mpsc channel and wakes the loop with `UserEvent::Mcp`; `App::run_mcp_calls` (`src/app.rs:1562`) runs each call in order via `Tracker::call` (`src/runs.rs:522`), then `UiState::mark_agent`, `log_agent`, and `steer::attach`. The chat's API backend uses the same channel (`App::tool_dispatch`), and its CLI backends reach the same HTTP server with `?chat=<id>` (`server.rs::client_of`, `chat_of`). `refs`, `imagegen` and `delegate` are the exception: they run on a spawned thread (`app.rs:1573`), because the helper calls back into the app through `papercraft-cli`.
+* **Identity.** `Client {name, session}` (`src/mcp/mod.rs`). `name` is the MCP `clientInfo` title or name, or `CHAT_CLIENT`; `session` is the `mcp-session-id` header or the chat id. `Tracker` keeps one open run per `Client` (`Tracker.open`, `open_of`) and labels the second session of a name "Claude Code · 2" (`Tracker::label`). `papercraft-cli` connects with `().serve(...)` (`bin/papercraft-cli.rs:309`): default client info, a **new MCP session per process**, so each `papercraft-cli call` is a different `Client` and opens its own run. A delegated helper is just another such CLI process; nothing links it to its parent.
+* **Touched layers.** `Editor.touch_log` is filled by `log_touched` on every commit while `Tracker::call` has it open; `take_touched` hands (page id, node) pairs to `UiState::mark_agent`, which keeps the topmost ones as `AgentMark {page_id, node, verb, client, at}` (cap `AGENT_MARKS_MAX`, fade `AGENT_MARK_SECS`). `RunPage.nodes` persists the node set per run and page; `runs::blocker` / `keep_own` / `revert` already use it to refuse or narrow a revert when a *later run* touched the same nodes.
+* **Layer `lock`** is a human flag (`Node.locked`, `is_effectively_locked`): it stops canvas hit-testing and canvas editing. No tool in `src/mcp/exec.rs` checks it, so agents write through locked layers today. This spec does not change that (open question 4).
+* **Undo** is per page (`ed.with_page(page, ..)`, `History.undo/redo` stacks of `Change`), global across authors. The `undo` / `redo` tools (`exec.rs::undo_redo`) pop the top `Change` of the page whoever made it. A `Change` has no author. A design-system edit puts the same `design` before/after in a step on **every** page (`history.rs`, `Change.design`).
+* **Offline.** `cli::Offline` (`src/cli.rs`) wraps `SessionHost` and calls `exec::call` directly: no `Tracker`, no runs, no versions, identity-less. It already has two locks: `APP_LOCK` (`.papercraft-app.lock` in the projects root, held by the app for its whole life from `app.rs:1714`) and `PROJECT_LOCK` (`.papercraft-cli.lock` in the project folder, taken by `lock_for_write` before the first editing call, with a 5 s wait, then a **reload** of the project from disk). `refuse_if_open` refuses `--offline` on a project the app reports as `open` in `list_resources`. All are OS advisory locks (`File::try_lock`), so a crash releases them. Gaps are listed in section 6.
+* **Saving.** `Session::save` (`src/session.rs:217`) writes every dirty page, `design.json` and the manifest atomically (`io::atomic::write`, temp + rename) but **never looks at what is on disk**: it is last-writer-wins per file. Only shared projects have a three-way page merge (`sync.rs` + `merge::merge`, base copy in `.sync-base/`).
+* **Skills.** `skills/papercraft/SKILL.md` teaches single-agent use; nothing about sharing a project.
+
+## Scope
+
+### 1. Identity
+
+Every call is attributed to an `AgentId`:
+
+```rust
+pub struct Client {            // src/mcp/mod.rs, extended; stays Eq + Hash (Tracker key)
+    pub name: String,          // "Claude Code", "Codex", "Chat", "You", "Actions", or --as value
+    pub session: String,       // MCP session id, chat id, or stable agent session (below)
+    pub parent: Option<String>,// session of the agent that delegated this one
+}
+```
+
+`run` is not a field of `Client`: the run is derived (`Tracker.open_of(client)`), and the change-log entry stores `run_id` when one is open so Review can group them.
+
+**Sources of identity**
+
+| Caller | name | session | parent |
+|---|---|---|---|
+| MCP client | `clientInfo` title or name (as today) | `mcp-session-id` | none |
+| Chat turn (API or CLI backend) | `Chat` | chat id | none |
+| `papercraft-cli` online | `--as <name>` > `$PAPERCRAFT_AGENT` > `cli` | `agent:<name>` (stable across invocations) | `--parent` / `$PAPERCRAFT_PARENT` |
+| `papercraft-cli` offline / `shell` | same as online | same | same |
+| `delegate` helper | `<parent label> / <role>` e.g. `Chat / build` | `d<n>-<chat id>` | the caller's session |
+| Actions menu | `Actions` | `""` | none |
+| The person at the canvas | `You` (reserved, cannot be claimed by `--as`) | `""` | none |
+
+* **Why a query string, not `clientInfo`.** `papercraft-cli` has no per-call client info and a fresh MCP session per invocation. The chat already proves the pattern: `?chat=<id>`. The CLI appends `?agent=<name>&parent=<session>` to the MCP URL; `server.rs::client_of` reads it (`agent_of(query)` next to `chat_of`) and builds the stable session `agent:<name>`. Names are sanitized (`[A-Za-z0-9 ._/-]{1,40}`); `You`, `Chat` and `Actions` are rejected for `--as`. An unnamed CLI is `cli` and prints one stderr hint: "set --as <name> so other agents can tell you apart".
+* **Delegation.** `delegate::run` sets `PAPERCRAFT_AGENT` and `PAPERCRAFT_PARENT` on the helper process (it already sets `PAPERCRAFT_PROMPT` and `PATH`), and the preamble tells the helper to pass nothing else: `papercraft-cli` reads the env. Grandchildren inherit the env, so chains work (`parent` is the immediate delegator).
+* **Offline.** `Offline::new(store, project, client)` takes the same `Client`; the CLI binary builds it from the flags and env above. Offline gets a `Tracker` (Phase 6) so offline work also produces runs, versions and attributed log entries, not an anonymous blob.
+* **Run labels.** `Tracker::label` keeps its numbering for MCP clients with no stable name. A named CLI agent is not numbered (its session is the name).
+
+**Where it shows**
+
+* *Canvas.* `mark_agent` already carries `client` (the label). The outline and its tag are drawn in the agent's colour: `agent_color(label) -> theme slot`, a stable hash of the label into 8 theme tokens `agent_1..agent_8` (added to `src/theme.rs` next to the other palette entries; both themes; no hex in UI code). The tag reads `Codex . Styling` (label, then verb); a delegate reads `Chat / build`. A *presence pin* (a small coloured dot plus the label, no text editing cursor exists for agents) sits on the top-left corner of the last touched layer for `AGENT_MARK_SECS`. The human's own edits get no mark.
+* *Agent view.* One group per `Client` with its colour dot, label, state (working per `AGENT_WORKING_SECS`, idle, finished), runs, and its claims (section 3). Delegates are nested under their parent. Rows are one line, truncate with an ellipsis, and have 44 px targets for actions.
+* *Change log and `get_changes`.* Entries carry `{client, session, parent, run}`; `exclude:"self"` matches on `(name, session)`.
+
+### 2. Optimistic concurrency
+
+**Reference rev.** The point in history the caller's picture of a layer is from.
+
+* Explicit: any write tool accepts `ifRev` (integer, the `rev` the caller last synced to).
+* Implicit (default, no `ifRev`): the server remembers `seen[(client, page, artboard)] = rev at the end of the client's last call that read or wrote anything under that artboard` (any tool whose resolved node ids sit under it; `get_screenshot`, `get_tree_summary`, `get_jsx`, writes, ...). A client with no history on an artboard has reference = the rev when its session first called. Tradeoff: coarse (artboard, not node) so a read of the header does not hide a stale body, in the direction of *over*-reporting; it costs one small map per client, dropped when the client has been silent for an hour.
+
+**The rule.** A write call has a target set `T` (resolved before the call from `nodeId`, `nodeIds`, `targetNodeId`, parent, page and variable arguments by `swarm::targets::of(name, args, ed)`, the same resolution `runs::target_pages` already does for pages). It conflicts when the log has an entry with `rev > ref` by a **different client** (including `You`) whose node set matches one of:
+
+| relation | an entry conflicts when it ... | applies to |
+|---|---|---|
+| `self` | changed a node in `T` (any field) | all writes |
+| `ancestor` | changed the **own fields** (not `children`) of an ancestor of a node in `T`, or deleted or moved one | all writes (size, layout, clip, visibility of a parent change what a child edit means) |
+| `descendant` | changed any node under a node in `T` | writes with `reach: subtree` (see below) |
+| `variable` | changed (value, rename, delete) a variable the call **sets**, or deleted/renamed one the call **binds** with `var(--x)` or `bind_variable` | variable tools; node writes that bind variables |
+| `component` | changed the main component of an instance in `T`, or an instance's main when `T` is a main | writes to instances and mains |
+| `page` | created, deleted, renamed or reordered pages | page tools |
+
+Per-tool `reach` (table in `src/mcp/swarm/targets.rs`, one test asserts every tool in `tools::definitions()` is classified, like the `edits_doc` test):
+
+* `node`: `update_styles`, `set_text_content`, `rename_nodes`, `set_code_link`, `bind_variable`, ... (own fields only; descendants are not affected).
+* `subtree`: `write_html` in replace mode, `delete_nodes`, `move_nodes` (also its source and destination parents' `children`), `detach_instance`, `apply_ui_kit`, anything that rebuilds children.
+* `children`: `write_html` `insert-children`, `create_shape` under a parent, `duplicate_nodes`: adds below the target; conflicts only through `self` / `ancestor` (target deleted or moved), so two agents appending to the same frame do not conflict.
+* `none`: `create_artboard`, `create_page` (no existing target; page relation only).
+
+**Reading the result.** Conflicts are computed *before* the call runs, against the log, on the UI thread, so there is no check-then-act gap.
+
+```json
+// warn (default): the call is applied; trailing content block of the result
+{"swarm": {"rev": 58, "refRev": 51, "applied": true,
+  "conflicts": [
+    {"node": "1:12", "name": "Header", "relation": "self",
+     "by": {"client": "Codex", "session": "agent:Codex", "label": "Codex"},
+     "rev": 55, "tool": "update_styles", "at": "2026-10-28T10:02:11Z",
+     "changed": ["fill", "paddingTop"]}
+  ],
+  "hint": "Codex changed 1 layer you just wrote. get_changes {since: 51, exclude: \"self\"}, re-read those layers, adjust."}}
+```
+
+```json
+// strict: nothing applied; isError result, text is this JSON
+{"error": "conflict", "applied": false, "rev": 57, "refRev": 51,
+ "conflicts": [ ... same objects ... ],
+ "hint": "Re-read the layers above (get_changes {since: 51, exclude: \"self\"}), then retry with ifRev 57."}
+```
+
+The result stays a `ToolResult = Result<Vec<Out>, String>`: strict errors are the JSON string above (stable `error` codes: `conflict`, `claimed`, `needs_claim`, `undo_blocked`); warnings ride in a trailing `Out::Text` block, exactly how `steer::attach` appends its note, so tools' own JSON is untouched and image results keep working. `conflicts` is capped at 20 with `"more": n`.
+
+**Modes**
+
+| Setting | How set | Behaviour |
+|---|---|---|
+| `warn` (default) | nothing | apply; report conflicts against the implicit or explicit reference |
+| `strict` per call | `ifRev` present (and no `onConflict: "warn"`) | reject if any conflict against `ifRev`; nothing applied |
+| `strict` global | Settings, Agents, "Reject conflicting writes"; env `PAPERCRAFT_CONFLICTS=strict` | like strict per call, using the implicit reference when `ifRev` is absent |
+| `warn` per call | `onConflict: "warn"` | override the above for this call |
+
+Precedence: `onConflict` > presence of `ifRev` > env > setting > default.
+
+`ifRev` newer than the current `rev` (the log was reset, for example the app restarted without a persisted log) is reported as one conflict `{"relation": "history_reset"}`: reported in warn, rejected in strict.
+
+**Recovery protocol** (what the tool hint and the skill say): on `conflict` or a non-empty `conflicts`, call `get_changes {since: <refRev>, exclude: "self"}`, `get_tree_summary` / `get_screenshot` on only the listed nodes, decide whether the other change is compatible, redo your edit if not, and continue with `ifRev` = the `rev` you got back. Never retry the same call blindly.
+
+**The human.** Canvas edits are recorded in the log as `You` through the editor commit path (a `HUMAN` author when no tool call is running), never checked, never blocked, and reported to agents by the same `conflicts` structure (`by.client == "You"`). An agent write is not blocked by a selection or an open text edit; the in-progress text edit is committed on Esc/blur as a normal human write and conflicts back in the other direction.
+
+### 3. Soft claims
+
+A claim is a short lease that says "I am working here". It is advisory by default and enforced for other agents (not the human) when the setting says so.
+
+```json
+// claim_nodes
+{"name": "claim_nodes", "inputSchema": {"type": "object", "properties": {
+  "ids":        {"type": "array", "items": {"type": "string"}, "description": "Layer ids (\"<page>:<n>\"). Each claims the layer and everything under it."},
+  "resources":  {"type": "array", "items": {"enum": ["variables", "components", "pages", "settings"]}, "description": "Shared resources that cannot be split by layer."},
+  "ttlSeconds": {"type": "integer", "minimum": 10, "maximum": 1800, "default": 300},
+  "note":       {"type": "string", "maxLength": 120, "description": "Shown to the person and to other agents."}},
+  "anyOf": [{"required": ["ids"]}, {"required": ["resources"]}]}}
+// -> {"claimId": "c7", "claimed": ["1:12"], "resources": [], "expiresAt": "...", "ttlSeconds": 300, "rev": 58}
+// release_claims   {"claimId"?: "c7", "ids"?: [...], "resources"?: [...]}   (nothing = all of mine)
+// -> {"released": ["c7"]}
+// list_claims      {}  (read-only)  -> {"claims": [{claimId, holder, ids, resources, note, expiresInSeconds}]}
+```
+
+Resource names are the "named shared resource" form of a claim: `variables` (collections, variables, modes, tokens, the design system), `components` (creating or editing main components), `pages` (create, delete, rename, reorder, group), `settings` (project-level settings, libraries).
+
+**Semantics**
+
+* **Subtree.** A claim on a layer covers it and all descendants. It does not cover ancestors' own fields.
+* **Overlap.** Another agent's claim `X` overlaps a request `Y` when `X` is the same layer, an ancestor, or a descendant of `Y`. A request overlapping *another client's* live claim is denied. Overlap with the caller's own claims merges (the existing claim is extended and renewed).
+* **Atomic.** `claim_nodes` takes all of `ids` and `resources` or none. A denial is a normal result, not a thrown error:
+
+```json
+{"claimed": [], "denied": [
+  {"id": "1:12", "relation": "ancestor", "heldBy": "Codex", "claimId": "c3", "ids": ["1:3"],
+   "note": "rebuilding Settings screen", "expiresInSeconds": 212, "retryAfterSeconds": 212}]}
+```
+
+* **Limits.** At most 16 live claims per client and 64 per project (`MAX_CLAIMS_PER_CLIENT`, `MAX_CLAIMS`), so a runaway agent cannot fence off the canvas.
+* **Lease, not lock.** The expiry is `now + ttl` and is renewed by any call from the holder (read or write) while the claim lives, so an active agent never loses a claim and a stalled one loses it `ttl` after its last call. Expiry is checked lazily on every claim check and by `Tracker::tick`; there is no timer thread.
+* **What a claim does for others** (Settings, Agents, "Claimed layers": `Block` (default) or `Warn`; per call `onClaim` is not offered, the holder decides by claiming):
+  * `Block`: a write by another agent whose resolved targets (with `reach`) overlap a claim is rejected before it runs: `{"error": "claimed", "applied": false, "claims": [{claimId, heldBy, ids, note, expiresInSeconds}], "hint": "Work elsewhere or retry after expiresInSeconds; ask the holder through your orchestrator."}`.
+  * `Warn`: applied, with `claimWarnings` in the trailing `swarm` block.
+  * Reads and screenshots are never blocked. The parent chain of a delegate (delegator, its delegator, ...) is never blocked by the delegate's claim, only warned, so an orchestrator can always step in.
+  * The human is never blocked. A human edit inside a claim is logged as `You`, so the holder sees it as a conflict on its next call, and the canvas shows the claim as "edited".
+* **Resources.** Writes that touch a shared resource map to a resource name by `swarm::targets::resources_of(name, args)`: `set_variable*`, `create_variable*`, `set_variable_mode`, token tools -> `variables`; component tools and writes whose target is a main component or inside one -> `components`; `create_page`, `delete_page`, `rename_page`, reorder and group tools -> `pages`; project settings and library tools -> `settings`. A resource claim held by someone else blocks (or warns) like a layer claim. Setting "Shared resources" (Settings, Agents): `Open` (default: no claim needed) or `Claim required` (a write to a shared resource without holding its claim is rejected with `needs_claim`; this is how an orchestrator reserves them for itself: it claims and holds them, and helpers cannot).
+* **Auto-release**
+
+| Event | Effect |
+|---|---|
+| `release_claims` | released |
+| `finish_working_on_nodes {nodeIds}` | the caller's claims on those layers (or all of its claims when no ids) are released, then the run ends as today |
+| Chat turn ends (`Tracker::finish`, `app.rs:772`) | all claims of that chat |
+| `delegate` helper process ends, times out or is stopped | all claims of the helper |
+| Run ends by idle (`RUN_IDLE_SECS` = 30) | **kept**: an agent thinking for 45 s is not finished; the TTL covers a dead one |
+| MCP session closes (transport DELETE) | all claims of that session, if the session manager reports it (open question 3); else TTL |
+| TTL passes with no call from the holder | expired |
+| Project switch (`Tracker::sync_project` sees another dir) | all claims |
+| App restart | gone (claims are memory-only online; MCP session ids do not survive a restart) |
+| Human "Break claim" | removed; the holder's next result carries `notices: [{"type": "claim_broken", "by": "You", "claimId": "c7"}]` |
+
+* **State machine of a claim**
+
+| State | On | Next |
+|---|---|---|
+| (none) | `claim_nodes`, no overlap with others | Active |
+| (none) | `claim_nodes`, overlap with another client | denied, no state created |
+| Active | holder calls any tool | Active (expiry renewed) |
+| Active | holder `claim_nodes` overlapping its own | Active (merged, renewed, one id) |
+| Active | `release_claims`, `finish_working_on_nodes`, turn/helper end, project switch | Released |
+| Active | no call from holder for `ttl` | Expired |
+| Active | human Break | Broken |
+| Expired / Released / Broken | any | removed at once; the change log keeps `claim_*` entries |
+
+* **Deadlock freedom by construction.** No tool waits. A claim attempt returns immediately (granted or denied with holder and expiry). A claim is all-or-nothing, so no agent holds a partial set while asking for more; claims always expire, so no chain of holders can be unbounded. The remaining failure is *starvation*, not deadlock; the denial tells the caller `retryAfterSeconds` and the protocol (section 4) tells it to do other work first.
+* **Delegation.** The `delegate` tool gains `scope` (`{"ids": [...], "resources": [...]}`), required for role `build`, ignored for `qa` and `read` (they do not edit). `run_mcp_calls` takes the claim **on the UI thread before spawning the helper** (the spawn happens at `app.rs:1573`), with holder = the helper's `Client` (name `<parent> / build`, `parent` = caller session) and a TTL of 15 min (the helper's timeout is `TIMEOUT` = 900 s), released when the process ends. If the scope is denied, the delegate call returns the `denied` list and no process starts. The parent can read and write inside the child's claim (warned, not blocked).
+
+**Seeing and breaking claims (human)**
+
+* *Canvas:* a claimed layer gets a dashed outline in the holder's colour and a small tag `Codex . 4:12 left` (a painted lock-and-clock icon, truncated note in the tooltip). Right-click on a claimed layer: "Claimed by Codex (4:12 left, rebuilding Settings)", "Break claim".
+* *Agent view:* a `Claims` section per agent: target names, note, remaining time, a "Break" button (44 px). Resources show as chips (`variables`).
+* *Editing under a claim:* allowed; the first human edit inside a claim shows a one-line status ("Codex has claimed this; it will be told") and no modal.
+
+### 4. Swarm partitioning guidance
+
+For the orchestrator (a person's prompt, or a lead agent). Prefer to split by things that do not share state, in this order:
+
+| Split by | Why it is safe | Notes |
+|---|---|---|
+| **Page** | pages have their own `Document`, history and file; the only shared state is the design system and component sources | best default for large work |
+| **Artboard** | one artboard = one subtree; claims and conflicts are cheap | the usual unit inside a page |
+| **Component** | a main component and its instances | claim `components` unless every agent owns a different main |
+| **Variables** | one agent owns the design system for the swarm | do this *first*, then fan out |
+| **Annotations** | stored per page | partition by page like layers |
+
+Shared resources that cannot be partitioned (each of these is a single value in the project: `design.json`, the manifest page list, component sources shared across pages via `flush_cross_pages`, project settings):
+
+| Resource | Rule |
+|---|---|
+| Variables, modes, tokens | Orchestrator defines them up front. Later changes: `claim_nodes {resources:["variables"]}`, change, release, in a short window. Never switch the document mode (`set_variable_mode`) while others take screenshots: it changes what every agent sees. Do it last or have the orchestrator do it |
+| Components | One owner per main component. Anyone may place instances; only the owner edits the main. Use `components` claim for new components |
+| Page list and order | Orchestrator creates pages first (`pages` claim for later changes) |
+| Project settings, libraries | Orchestrator only (`settings` claim held for the swarm, setting "Claim required") |
+
+**Recommended agent protocol** (the block for `skills/papercraft/SKILL.md`, about 20 lines):
+
+```
+Sharing a project with other agents
+1. Identify yourself: papercraft-cli --as <short name> (or set PAPERCRAFT_AGENT). MCP clients are named by the client.
+2. Read before you write: note `rev` (get_basic_info). Work only in the layers you were assigned.
+3. Claim what you will rewrite: claim_nodes {ids:[<your artboard>], ttlSeconds:600, note:"<what you do>"}.
+   If denied, do other work or report back; never wait in a loop.
+4. Edit with ifRev = the rev your last result returned.
+5. If a result has `conflicts` or error `conflict`: get_changes {since: <refRev>, exclude:"self"}, re-read only those layers, redo what is no longer right. Do not blindly retry.
+6. After a batch: lint_design {changedOnly:true}, fix findings in your layers only.
+7. Shared things (variables, components, pages, settings): claim the resource, change, release in the same minute.
+8. Finish: finish_working_on_nodes {nodeIds:[...]} (releases your claims) and report which layers you changed, which conflicts you saw and how you resolved them.
+9. Never undo with scope "page"; use undo {scope:"mine"}.
+Orchestrators: give each helper a disjoint `scope` in `delegate`; create pages and variables before fanning out.
+```
+
+### 5. Undo and revert under concurrency
+
+**Per-client undo.** `undo` and `redo` get `scope`:
+
+```json
+{"name": "undo", "inputSchema": {"type": "object", "properties": {
+  "pageId": {"type": "string"},
+  "scope": {"enum": ["mine", "page"], "default": "mine",
+    "description": "mine: undo only your latest step, if no later step by someone else touched the same layers. page: undo the page's latest step whoever made it."}}}}
+// -> {"done": true, "what": "write_html", "step": "s41", "rev": 61}
+// -> refused: {"done": false, "reason": "undo_blocked", "step": "s41", "blockedBy": [{"client": "Codex", "rev": 59, "tool": "update_styles", "nodes": ["1:12"]}]}
+```
+
+* **Authorship.** `History` stores per `Change` an `author: Arc<str>` (the client label, `You` for canvas edits), its `rev` and a step id. Coalesced canvas steps (`coalesce_key`) keep the human author.
+* **Algorithm** (`History::undo_by(doc, author) -> Result<Option<Undone>, Blocked>`): find the newest `Change` `m` by the caller on the page. Let `later` be the changes above it. `m` is *independent* when, for every `l` in `later` by anyone: entry ids of `l` and `m` are disjoint, `m.design`/`m.variables`/`m.page_modes`/`m.annotations` are `None` or the same field is `None` in `l`, and `m` touched no main component (a `touched_mains` step may rewrite instances on other pages). If independent, apply `m` backwards, splice it out of the undo stack, and push it to the redo stack. If not, return `Blocked` with the offending `l`s. Conservative on purpose: if both changed `children` of the same parent they both have an entry for the parent, so it is refused. The root list is handled by set difference (remove ids `m` added, re-insert ids `m` removed after their old predecessor), so two agents each creating an artboard do not block each other's undo.
+* **Redo** `scope: "mine"` re-applies the caller's last undone step if the same independence holds in reverse; as always, any new commit on the page clears the redo stack, so an agent cannot redo across someone else's later edit (documented, not worked around).
+* **Default.** `scope` defaults to `mine` for every client except `You`, because "undo" from an agent should never remove a person's or another agent's work. The old behaviour is `scope: "page"`. This is a behaviour change for single-agent use when the human edits between agent steps; it is deliberate (open question 2).
+* **The human's Ctrl+Z / Ctrl+Shift+Z** is unchanged: it undoes the newest step on the page whatever its author, because that is what the canvas shows and what the person expects. The status line says whose step it was ("Undid Codex: write_html"). The Agent view run rows get "Undo this step" using the same independence check, with the refusal explained in a toast.
+* **Cross-page steps.** A design-system edit adds a step on every page; its independence check is against later `design` changes only, so undoing one's variable edit works while others edit layers.
+
+**Run revert** (`runs::revert`, `revert_change`, "Restore to here") under concurrency:
+
+* The existing `blocker(runs, run)` refuses when a *later run* touched the same nodes. Extend it with the change log: a node of `RunPage.nodes` that has a log entry by another client (including `You` and clients with no run, like offline CLI before Phase 6) with `rev` after the run's first call is **foreign-changed**.
+* Outcomes of "Revert run":
+  1. *No foreign-changed nodes:* as today (full page restore when the run is the only writer, node-level when `shared`).
+  2. *Some:* revert only the clean nodes (the existing node-level path used for `shared` pages, now chosen whenever *any* foreign entry exists on the page since the run started, not only when another run exists), skip the foreign-changed ones, and say so: "Reverted 14 layers. Kept 3 that Codex or you changed afterwards." with a list and a per-row "Revert anyway" (restores that node to the snapshot, overwriting the later change; one undo step, shows what will be lost).
+  3. *Everything is foreign-changed:* nothing is reverted; the state stays `Review`.
+* An agent-driven revert (a future tool) is never forced; only the person can use "Revert anyway".
+* The auto-version `Before agent run N` is a whole-page snapshot taken when *that client's* run starts, so it contains other agents' in-flight work. `RunPage.nodes` / `keep_own` already restrict review and revert to the run's own nodes; the change log makes that precise for clients that never had a run.
+
+### 6. Offline CLI and multi-process
+
+**6.1 What races today**
+
+| Race | Today | Verdict |
+|---|---|---|
+| Offline A vs offline B, same project | `PROJECT_LOCK` taken at first edit, held until the process exits; B waits 5 s then fails; A reloads at lock time | Safe but unfair: a long `shell` starves everyone (G1) |
+| Offline vs running app, project open in the app | `--offline` refused via `list_resources.open` for `open_file` / `--project` | Safe |
+| Offline vs running app, project **not** open in the app, app opens it later | the app's `Session::open` never looks at `PROJECT_LOCK`; it loads mid-edit state, then saves over the CLI's later edits | Race (G2) |
+| Save over a changed file (Dropbox, a second machine's copy, a manual copy) | blind overwrite | Last writer wins (G3) |
+| Offline writes carry no identity, runs, versions or log entries | `Offline::call` bypasses `Tracker` | Gap (G4) |
+
+**6.2 Per-call lease (G1).** `Offline::call` takes the project lock **around each editing call**, not for the process lifetime: `lock -> refresh -> exec -> save -> unlock`. Waits stay `PROJECT_LOCK_WAIT` (5 s), but now a held lock lasts one call (milliseconds to a second), so shells and scripts interleave. `refresh` compares the on-disk stamp (`manifest.modified` plus `len + mtime` of each page file and `design.json`) with the stamp recorded at the last load/save; for each page whose stamp changed it reloads **that page only** (replacing its `Document` and resetting only its history); unchanged pages keep their history, so `:undo` in a shell works across interleaving as long as that page was not changed meanwhile. Because every offline editing call saves before unlocking, the process has no dirty pages at refresh time, so the refresh never needs a merge.
+
+**6.3 The app takes the lock too (G2).** `Session::open` (native) takes `PROJECT_LOCK` with a 2 s wait before loading and keeps it while the project is open (released on close or switch, and by the OS on crash). An offline process that finds it held by the app reports "PaperCraft has this project open" (same text as `refuse_if_open`) instead of "another papercraft-cli is writing". The lock file's content records the holder (`{"pid", "client", "since"}`) so errors can say "held by papercraft-cli --as Codex (pid 4120, 12 s)". The windows between app open and offline lease are one call long, thanks to 6.2.
+
+**6.4 Identity and the log offline (G4).** `Offline` gets a `Client` and a `Tracker`; the run closes when the process ends or after `RUN_IDLE_SECS`, and the auto-version rules apply, so Review/Revert cover offline work. The change log (Phase 0) is persisted as `activity.json` in the project folder (bounded: `ACTIVITY_KEEP` = 500 entries, plus `epoch` and `rev`), written atomically by the process holding the lock after each editing call, and loaded by the app on open and by offline processes at lease time. This is what makes `ifRev` and `get_changes` meaningful across offline processes and across an app restart. A project without the file starts at `rev = 0` with a new epoch (`history_reset` for older `ifRev`). Format: additive, optional, ignored by older builds.
+
+**6.5 Claims offline.** With no app running, claims live in `claims.json` in the project folder, read and written under the lease (expired ones are dropped on load); the holder is the stable `agent:<name>` session, so a claim survives across that agent's CLI invocations. With the app running, offline is only used on projects the app does not hold, and the app discards `claims.json` when it opens the project (agents using the app are on MCP sessions that the file cannot represent). Every offline process reads the file at lease time, so claims are enforced between offline processes too.
+
+**6.6 Save-time check, merge or refuse (G3).** `Session::save` keeps per page the packed bytes it last read or wrote (`PageBase`, zstd, about the size of the file) and the file stamp. Before writing a dirty page it compares the stamp on disk. Unchanged (the normal case): write. Changed: three-way merge with `merge::merge(base, local, disk)` (decoded JSON), write the result atomically, and install it in the editor as one undo step "Merged changes from disk" (via `Editor::restore_snapshot`, which already exists for versions). `merge` is extended to return a report (`Merged { value, notes }`, notes for each scalar where both sides changed and local won, and each renumbered id) because RULES.md says a merge that cannot keep something must say so; the notes go to the status line and the Agent view. Refuse (save skipped, status "Disk changed and cannot be merged: ...") when either side fails to decode or the page file has a newer format version. `design.json` and the manifest are small objects; they use the same merge with `base` = last written. The shared-project sync engine keeps its own `.sync-base/` merge; this check runs first and does not replace it.
+
+**6.7 App running, offline requested.** Unchanged: plain `call` goes through the app; `--offline` on a project the app has open is refused; `--offline` on another project works under the lease. A CLI that finds `APP_LOCK` held but no server yet keeps waiting (`app_starting`).
+
+### 7. Chat vs MCP
+
+Both reach `App::run_mcp_calls`, so identity, conflicts, claims, rev, runs and the log are identical. Differences, all unchanged by this spec:
+
+| Aspect | Chat | External MCP / CLI |
+|---|---|---|
+| Identity | `Chat` + chat id, from `?chat=` or the API backend's dispatch | client info / `--as` |
+| Steering | `steer::attach` appends the user's mid-turn messages to the next tool result | none |
+| Run end | `Tracker::finish` when the turn ends; also releases claims | idle 30 s or `finish_working_on_nodes` |
+| Auto mode | answers `ask_user` with the first option; unaffected | not applicable |
+| Sub-agents | `delegate` (off the UI thread) | an MCP client's own sub-agents are not visible; they appear as the same or new sessions |
+
+The swarm block must surface in chat the same way: `swarm::attach(client, r)` runs for every client right before `steer::attach`, so the chat model sees `conflicts`, `claimWarnings` and `notices` in its tool results. In the chat UI, `chat_msgs` renders a result with `swarm.conflicts` or a `claimed` / `conflict` error as a warning chip on the tool card ("Codex changed Header at rev 55") with a "Show" action that selects the node; the chat system prompt gets the protocol block of section 4 (the chat agent is a swarm member, and its `delegate` calls carry `scope`).
+
+### 8. Relay and other machines
+
+Out of scope: real-time cross-machine claims, and cross-machine `rev` (each machine has its own counter). In scope by reference only:
+
+* A relay sync is **page granularity** (`sync.rs`): when both sides changed a page it merges it with `merge::merge` against `.sync-base/` and the folder wins ties; the open project reloads (`Session::sync_poll`) unless it has unsaved edits (those win and go up on the next save).
+* A synced merge lands in the editor as a reload, so it is recorded in the log as one entry by `client: "Sync"`, `nodes` = the pages' changed nodes if cheap to compute, else `page: <id>, whole: true`. Agents then see it as a conflict (`by: Sync`) on their next write to that page. `Tracker` already flags "Includes changes synced from others" in Review; this reuses that signal.
+* Claims are local to the machine that holds the app; a remote editor is like the human: never blocked, always reported (as `Sync`).
+* Remote agents on the relay (`mcp/relay.rs`) arrive through the same MCP path on the host machine and are ordinary clients there.
+
+## Not in scope
+
+* Real-time collaborative cursors for humans.
+* Automatic semantic merge of conflicting edits on the same property (the agent decides).
+* Enforcing the human `Node.locked` flag against agents (open question 4).
+* Blocking a read while another agent is writing (reads see the committed state).
+* Per-property claims; layer granularity only.
+* Cross-machine claims or `rev`.
+* Cross-page atomic undo.
+* Locking on network filesystems where advisory locks are unreliable: documented as unsupported, not detected.
+
+## Acceptance
+
+* Two simulated clients editing disjoint artboards produce no `conflicts`; overlapping edits produce exactly the conflicts the brute-force oracle computes (section 2 table), in `warn` and `strict`, for every relation.
+* A strict rejection leaves the document, the log (`rev`), the undo stack and runs unchanged.
+* `claim_nodes` on a layer makes another client's `update_styles` on it or a descendant fail with `claimed` (`Block`) or succeed with `claimWarnings` (`Warn`); the holder is unaffected; the human is never blocked. A claim expires after `ttl` of holder silence (fake clock), is released by `finish_working_on_nodes`, by turn end, by helper end, by project switch, and by Break; overlapping and ancestor/descendant requests are denied atomically.
+* `delegate` of role `build` without `scope` is an error; with `scope` the helper's claim exists before the process starts and is gone after it ends.
+* `undo {scope:"mine"}` undoes the caller's newest step when later steps by others touch other layers and returns `undo_blocked` with `blockedBy` when they touch the same ones; Ctrl+Z still undoes the newest step.
+* Run revert with a later foreign edit to one of its nodes reverts the others and reports the kept one; revert never changes a layer the run did not touch.
+* Two offline `Offline` instances on one folder (threads, real locks) each doing N edits end with 2N edits on disk and in a fresh load; the shell with an idle prompt does not block the other.
+* `Session::open` on a project whose `PROJECT_LOCK` is held waits, then reports the holder. Saving a page whose file changed on disk merges and reports notes, or refuses; a layer is never silently dropped.
+* A project without `activity.json`, `claims.json` or the new run fields loads unchanged; an older build ignores them.
+* Tool schemas and descriptions exist for `claim_nodes`, `release_claims`, `list_claims`, the new `ifRev` / `onConflict` / `scope` arguments, each tested in `tests/mcp.rs`; README and ROADMAP tool counts are updated.
+* Screenshot inspection at scale 2 in both themes of: agent colours on outlines, claim outline and tag, Agent view with two agents and a delegate, Claims section, a long agent name and a long note (no overflow, 44 px targets).
+* `cargo test --no-fail-fast`, `cargo check --lib --target wasm32-unknown-unknown` (swarm logic is target-agnostic; file locks are native only) and the Windows check pass. 80% line coverage on the new `src/mcp/swarm/` modules and the changed lines elsewhere.
+
+## Open questions for the maintainer
+
+See `plan.md`.