mixdog 1.1.2 → 1.2.1

package/README.md

@@ -6,8 +6,12 @@
 
 <p align="center">
   The open-source coding agent that uses 63% fewer tokens.
+  <br>
+  <a href="#installation"><b>Download ↓</b></a>
 </p>
 
+https://github.com/user-attachments/assets/af515dc9-b0fb-4b6a-aaa1-1d82db792ed0
+
 <p align="center">
   <a href="https://github.com/tribgames/mixdog/releases/latest/download/mixdog-desktop-win-x64.exe"><img src="https://raw.githubusercontent.com/tribgames/mixdog/main/docs/assets/download-windows.svg" alt="Download for Windows" height="48"></a>
   &nbsp;
@@ -30,8 +34,6 @@
   <a href="#documentation"><b>Docs</b></a>
 </p>
 
-https://github.com/user-attachments/assets/af515dc9-b0fb-4b6a-aaa1-1d82db792ed0
-
 ## Highlights
 
 - [63% fewer tokens](#benchmarks) than Codex CLI on the same model, at the
@@ -116,9 +118,9 @@ shell networking — headless exec is not an offline sandbox. Run
 
 <p align="center">
   <picture>
-    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/tribgames/mixdog/main/benchmarks/terminal-bench-2.1/tb21-sol-vs-codex-dark.svg">
-    <source media="(prefers-color-scheme: light)" srcset="https://raw.githubusercontent.com/tribgames/mixdog/main/benchmarks/terminal-bench-2.1/tb21-sol-vs-codex-light.svg">
-    <img alt="Bar chart: Mixdog vs Codex CLI on Terminal-Bench 2.1" src="https://raw.githubusercontent.com/tribgames/mixdog/main/benchmarks/terminal-bench-2.1/tb21-sol-vs-codex-light.svg">
+    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/tribgames/mixdog/main/design/promo/tb21-sol-vs-codex-dark.svg">
+    <source media="(prefers-color-scheme: light)" srcset="https://raw.githubusercontent.com/tribgames/mixdog/main/design/promo/tb21-sol-vs-codex-light.svg">
+    <img alt="Bar chart: Mixdog vs Codex CLI on Terminal-Bench 2.1" src="https://raw.githubusercontent.com/tribgames/mixdog/main/design/promo/tb21-sol-vs-codex-light.svg">
   </picture>
 </p>
 
@@ -139,6 +141,7 @@ savings come from.
 - [Benchmarks](benchmarks/terminal-bench-2.1/)
 - [Code Tidy](docs/code-tidy.md)
 - [Git & GitHub](docs/git-github-integration.md)
+- [SSH servers](docs/ssh-remote.md) (Korean, with an English summary)
 - [Language servers](docs/language-servers.md)
 - [Office runtime](src/runtime/office/README.md)
 - [Development, configuration, and testing](docs/development.md)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mixdog",
-  "version": "1.1.2",
+  "version": "1.2.1",
   "private": false,
   "type": "module",
   "description": "Open-source coding agent that matches Codex on Terminal-Bench 2.1 with 63% fewer tokens.",
@@ -156,6 +156,7 @@
     "@formulajs/formulajs": "^4.6.1",
     "@google/generative-ai": "^0.24.1",
     "@huggingface/transformers": "^4.2.0",
+    "@img/sharp-wasm32": "0.35.3",
     "@modelcontextprotocol/sdk": "^1.29.0",
     "@mozilla/readability": "^0.6.0",
     "@napi-rs/canvas": "^1.0.8",

package/scripts/lib/run-node-tests.mjs

@@ -132,6 +132,15 @@ export function testIsolationEnv(scratchDir) {
 export const DEFAULT_TEST_TIMEOUT_MS = 120_000;
 export const LONG_TEST_TIMEOUT_MS = 600_000;
 
+/** Per-file timeout default for explicitly named files: a `.slow`/`.live`/
+ *  `.electron` test file keeps its lane's long bound (scripts/test.mjs laneOf)
+ *  even when run outside the lane runner. */
+export function explicitFilesTimeoutMs(fileArgs) {
+  return fileArgs.some((arg) => /\.(slow|live|electron)\.test\.mjs$/.test(String(arg)))
+    ? LONG_TEST_TIMEOUT_MS
+    : DEFAULT_TEST_TIMEOUT_MS;
+}
+
 // A batch that has not exited by then is killed with its process tree and
 // fails the run; MIXDOG_TEST_BATCH_TIMEOUT_MS overrides the default.
 const DEFAULT_BATCH_TIMEOUT_MS = 60 * 60 * 1000;

package/scripts/prune-desktop-runtime.mjs

@@ -190,6 +190,9 @@ async function prunePdfRuntime(nodeModules) {
 }
 
 async function pruneSharpWasmFallback(nodeModules, target) {
+  // The Linux daemon runs on Electron, whose system GLib breaks the native
+  // addon; src/runtime/shared/sharp-electron-linux.mjs routes it to wasm.
+  if (target.platform === 'linux') return false;
   const imgRoot = join(nodeModules, '@img');
   let packages;
   try {

package/scripts/test-direct.mjs

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
 // Explicit Node test options and file paths, without discovery or lane filters.
-import { runNodeTests } from './lib/run-node-tests.mjs';
+import { explicitFilesTimeoutMs, runNodeTests } from './lib/run-node-tests.mjs';
 
 const argv = process.argv.slice(2);
 // The same defaults as scripts/test.mjs: `mock.module` is still flag-gated on
@@ -17,5 +17,7 @@ await runNodeTests(
     '--test',
     ...unlessGiven('--test-force-exit'),
   ],
-  argv
+  argv,
+  // A named slow-lane file keeps the slow lane's per-file bound.
+  { testTimeoutMs: explicitFilesTimeoutMs(argv) }
 );

package/src/defaults/shipped-history.json

@@ -285,9 +285,11 @@
       "07802a61d913c3fb7c431774e7f4ebb7",
       "0a34c5b6011578bebdb6cca63edfcec4",
       "0ddf9f9845e1fcfc76d81592d6a71e70",
+      "0e044c22eb0786162ebb33a30f3bd5ae",
       "44171df8144101dc74bf21f72cb5b7c0",
       "48247c1645fd94aad7711082d4512c49",
       "4f94016363acccd788a724a2d414744e",
+      "56e474f78faafae39aebe921ccff112d",
       "6358ad25dd07a2f92ad65c249a4f360b",
       "7fb12aeb07b79ae113c0deb0a90c6da7",
       "804afd71ebc3f68614a4dc419af4b98e",
@@ -319,11 +321,14 @@
       "0d5c8c6e2349addf71d2ed2b4c1ccb1f",
       "0d6926aaca2890da9347f15a826546e3",
       "0f14ba0a3d8b3896b717dc144bbe49a4",
+      "15d86146cfbd58d85adf0c25cf1d683f",
       "1c317db86e1ff80a5c4102e09e36f5fb",
       "242f11864abdef90fa0cc9cb4599a579",
+      "29a347c4d3a2f773fa92c627260301ba",
       "336ced846a9e4f774cbb109417856c3b",
       "3f72247651ad9d777b6754bfd46732aa",
       "50bdf35d06913236855080fdea6b50dc",
+      "5b4586a44104b26ae7dd45da101079bf",
       "6101786127a1e68778eafcd2bf2c9ce0",
       "6711dbc9d84023b8eabfa928b8d33d18",
       "6f9e9338f560fa9acd839a67b5e2d1cc",
@@ -339,6 +344,7 @@
       "e04e6abe9a54e7d07b37f8fc32b5ad30",
       "e38d606df69865a9cf1cc48386535e8b",
       "e5ac8ea16905815e36e88d883195a617",
+      "e9b89f9193bd9d9c80ba6cf4419e58ff",
       "feabe973080344c4d52787a96043ab68",
       "ff18ac1a9f5961925aec27ffb870bae0"
     ],
@@ -360,6 +366,10 @@
       "f209e75b6ea0360929516a713474ad71",
       "fe0512acbb2fc8f80c5961b9f44aa5e2"
     ],
+    "visualize": [
+      "aa3531bfc4fd6ec3820493d0c190cc75",
+      "f65924f0691284ad31000469a9855b62"
+    ],
     "xlsx": [
       "0deebbdd009b0a0941ca4b0833dbe7f6",
       "10778ab8e1e2f7929895c034fef9ac41",

package/src/defaults/skills/pptx/SKILL.md

@@ -23,9 +23,9 @@ The model designs; the runtime converts and measures. Write HTML/CSS for the act
 |---|---|---|
 | **New deck**, or a redesign the user wants rebuilt | one `author` call per deck: HTML slides in `script` by default (`references/html.md`), or a pptxgenjs script (§2-§4) when the kit's devices are wanted or no local browser exists; a new deck is never assembled from `batch` operations | `author` returns `audit.status: 'pass'`; `render` on that session returns the pages and a `reviewToken` |
 | **Existing deck** — template fill, partial rewrite, reorder, merging decks | `open` → `snapshot` → `batch` slide and shape operations (§6); a new page is a duplicated source page, never a runtime-composed one | every `batch` returns its own `audit` for the edited slides |
-| **Read or extract** — outline, text, or figures from a deck, or the material a deck is built from | `open` + a paged `snapshot` (`pages:[…]` or `cursor:`), or `node "${MIXDOG_SKILL_DIR}/scripts/source-extract.mjs" <file>` for blocks with citable locators | blocks prefixed `[slide 4]`, `[p3]`, `[Sheet1!B4]`, `[line 42]` |
+| **Read or extract** — outline, text, or figures from a deck, or the material a deck is built from | `open` + a paged `snapshot` (`pages:[…]` or `cursor:`), or `${MIXDOG_SKILL_NODE} "${MIXDOG_SKILL_DIR}/scripts/source-extract.mjs" <file>` for blocks with citable locators | blocks prefixed `[slide 4]`, `[p3]`, `[Sheet1!B4]`, `[line 42]` |
 
-Prerequisites: the Office capability (`metadata.requires`) — its portable backend needs no installed PowerPoint, and Office on Windows is used in the background when present; fonts that render reliably for the recipient; a signed-in `media` lane only for generated pictures; an explicit `--provider` and `--model` for the reviewer and page-QC scripts, which reuse the existing sign-in and never ask for credentials.
+Prerequisites: the Office capability (`metadata.requires`) — its portable backend needs no installed PowerPoint, and Office on Windows is used in the background when present; fonts that render reliably for the recipient; a signed-in `media` lane only for generated pictures; an explicit `--provider` and `--model` for the reviewer and page-QC scripts, which reuse the existing sign-in and never ask for credentials. Run bundled scripts with `${MIXDOG_SKILL_NODE}` (on PATH in any shell), never plain `node`.
 
 ## 1. Load routing
 | File | Owns | Read when |
@@ -59,10 +59,10 @@ Prerequisites: the Office capability (`metadata.requires`) — its portable back
    **Bounded refinement**: fix factual and material visual defects before taste edits. Keep each pass focused on the highest-impact changes. After two unsuccessful corrections of the same defect, preserve the last valid artifact and report the limitation rather than cycling for a higher score. This limit is not permission to call an inadequate design complete. If an asset fails, repair or replace it; if it is removed, recompose the space it owned.
 8. **Page QC (new decks; skip for a one-page edit)**: after the measured audit, `qa render:false`, and the first render, check overflow, clipping, collision, canvas edge, contrast, and distortion; never redesign, reword, add, or delete. In Solo, Lead reviews directly. Only with workflow-approved delegation, use the runner below: each fresh fixer sees one page's image, inventory, and measured defects, uses only office, and edits a working copy. The runner adopts it only if defects did not grow and no other page changed. For this delegated route, `close` the session before running:
    ```text
-   node "${MIXDOG_SKILL_DIR}/scripts/qc-pages.mjs" <deck.pptx> --provider <provider> --model <model> --output qc-report.json [--pages 1-12] [--no-vision]
+   ${MIXDOG_SKILL_NODE} "${MIXDOG_SKILL_DIR}/scripts/qc-pages.mjs" <deck.pptx> --provider <provider> --model <model> --output qc-report.json [--pages 1-12] [--no-vision]
    ```
    Provider and model are explicit, as for the reviewer; `--no-vision` runs the geometry read only (a text-only model, or no renderer); at most 20 pages per run. Read the report (`pages[].reason`: improved, polished as unchanged, clean, scope, regression, audit_failed, execution_failed). A page whose run never replied is not a page that passed: `execution_failed` counts into `failed`, turns `ok` false, and exits the command, so rerun those pages before delivery rather than reading the report as clean. Then `open` the deck again and `render` it: the QC edits invalidate the earlier `reviewToken`, and a reopened session no longer carries the brief, which is why the brief read (`qa render:false`) comes before this step.
-9. Reader and art director pass — read the contact sheet for sequence and full-size pages for detail. For substantial new work, use the source questions in `references/review.md`; its separate reviewer procedure (`node "${MIXDOG_SKILL_DIR}/scripts/review-deck.mjs" packet.json --provider <provider> --model <model> --output review.json`) requires workflow-approved delegation. Otherwise, explicitly label the pass self-review. Check the reader's answers against the original source key, then judge message visibility, focal hierarchy, and coherent variation before consulting `deck.rhythm`.
+9. Reader and art director pass — read the contact sheet for sequence and full-size pages for detail. For substantial new work, use the source questions in `references/review.md`; its separate reviewer procedure (`${MIXDOG_SKILL_NODE} "${MIXDOG_SKILL_DIR}/scripts/review-deck.mjs" packet.json --provider <provider> --model <model> --output review.json`) requires workflow-approved delegation. Otherwise, explicitly label the pass self-review. Check the reader's answers against the original source key, then judge message visibility, focal hierarchy, and coherent variation before consulting `deck.rhythm`.
    **Output**: a slide-specific keep/fix decision with the visible reason. Resolve phrase-breaking wraps, weak evidence, and accidental empty regions; do not move a title merely to vary the sequence or change a successful page to reach a quota. Select at most two high-impact design revisions per pass; zero is valid.
    **Default — reuse the review**: small text, shape or data edits use `batch`, not re-author; every `batch` returns its own `audit` for the edited slides, read before the next edit. Portable renders report `changedPages` and `reusedPages` for up to 12 requested pages: inspect changed pages and carry forward the unchanged pages' critique. Shared resources invalidate the cache conservatively; PDF export remains whole-document. COM and long-deck contact sheets refresh normally. Always use the new complete render token, never the previous approval.
 10. `office action:'finalize' session:<id> design:{ reviewed:true, reviewToken, critique:[...] }` — reuses a complete current preview, checks the document, saves, validates, and closes. Use the token from the final complete render; after detail-page inspection, restore a complete render before finalizing. An edit invalidates the previous approval. `fixes` lists work still owed; applied fixes belong in `note`. A render/export failure leaves the deliverable unfinished: preserve the session and artifacts, report the failure, and never substitute a manual save for successful finalize. Finalize an authored deck in its authoring session: a deck reopened in `mode:'background'` is re-saved by PowerPoint, which rewrites the theme, layout, and embedded-workbook parts, and the validation baseline then reports `changed_protected_part` / `lost_protected_part` against the file it opened. Done when finalize reports the saved, validated file and the session is closed.
@@ -84,7 +84,7 @@ Prerequisites: the Office capability (`metadata.requires`) — its portable back
 ```
 **Hard rule — the plan is the page's contract**: every slide's plan line names its job, relationship, move, composition move, carriers, texture, and rhythm before any geometry exists, so the design is decided before it is drawn. **The runtime gates on it** from two slides up: `author` refuses a deck whose brief has no `slide plan:` line (`plan_missing`), whose plan leaves a slide unwritten (`plan_slide_unplanned`), whose line names no job or carriers (`plan_incomplete`), or whose page does not carry what its own line named (`plan_promise_missing`) — `reason: plan_gate` with the slides listed, the previous deck and its session untouched, and the answer is the plan line or the page, not a retry. A single-slide specimen is not gated, and a stale slide count stays information (`plan_count_mismatch`). A thin request gets a defensible brief invented from the subject, never blank fields. The `facts:`, `directions:`, and `slide plan:` lines may wrap across comment lines; each continuation line is read as more ` · ` items — so ` · ` is the separator and never appears inside a value or a source (write "관측치와 측정", not "관측치·측정"; a fact cut at an inner ` · ` silently disappears and every figure it carried becomes `number_without_fact`). → runtime `plan_gate` at author (`plan_missing`, `plan_slide_unplanned`, `plan_incomplete`, `plan_promise_missing`); `plan_count_mismatch` in qa
 **Hard rule — every figure has a fact**: `facts:` lists each number the deck will show, with its source (a page, a cell, a URL, or "user brief"). Dates and slide numbers are exempt. A chart series is one fact listing its values in order (`F3 운영비 1Q–6Q 10.5 10.8 9.1 7.9 7.0 6.5 — 재무팀 시트 B4:G4`); the values in the script's `chart()` call are copied from that line. With source material, the fact sheet is written first; without it, the facts come from the request and say so. **The runtime gates on it**: `author` reads the staged deck and refuses to land one that shows a figure no fact covers (`reason: facts_gate`, the figures listed per slide) — the previous deck and its session stay untouched, and the answer is the facts line, not a retry. A sketch or template with no source at all declares `facts: sample — <why>`: every figure is illustrative, the deck lands with `factsMode: 'sample'`, and the delivery tells the user the numbers are examples. → runtime `facts_gate` at author; `number_without_fact`, `facts_missing`, `facts_illustrative` in qa
-**Hard rule — supplied material is cited**: when the deck is built from files or pages the user gave, `sources:` names them and every fact's source points where a reader can open it — a page (`p.12`), a cell (`Sheet1!B4`), a section (`§3`), or a URL. `node "${MIXDOG_SKILL_DIR}/scripts/source-extract.mjs" <file>` prints the material as blocks already prefixed with those locators, which is what the fact sheet quotes. A figure that came from somewhere else (a conversation, a live system) says so in its own source. → runtime `fact_without_locator` (advisory)
+**Hard rule — supplied material is cited**: when the deck is built from files or pages the user gave, `sources:` names them and every fact's source points where a reader can open it — a page (`p.12`), a cell (`Sheet1!B4`), a section (`§3`), or a URL. `${MIXDOG_SKILL_NODE} "${MIXDOG_SKILL_DIR}/scripts/source-extract.mjs" <file>` prints the material as blocks already prefixed with those locators, which is what the fact sheet quotes. A figure that came from somewhere else (a conversation, a live system) says so in its own source. → runtime `fact_without_locator` (advisory)
 
 ## 4. Script contract and library boundaries
 - The kit prelude creates `pres` (`pptxgenjs`, `LAYOUT_WIDE` 13.33 × 7.5 in, coordinates in inches) and every helper; a script that declares its own `pres` runs without the prelude. Available modules: `pptxgenjs`, `sharp`, `node:fs`, `node:path`, `node:buffer`; no network and no photo library. Icons come from the injected `ICON` (256 offline Lucide stroke icons; `ICON.names` lists them, the kit's `icon()` / `iconRow()` place them) — a feature marker, a list prefix, or a stage glyph is an icon by name, never a Unicode emoji or a hand-drawn shape.
@@ -113,7 +113,7 @@ The runtime absorbs what it can (one `a:pPr` per paragraph; package validation a
 | A secondary-axis combo needs `valAxes` and `catAxes` with two entries each | PowerPoint drops the chart → runtime `chart_axis_undeclared` |
 
 ## 5. QA and finalize
-**Content before pixels**: read the deck's text once against the fact sheet — `node "${MIXDOG_SKILL_DIR}/scripts/source-extract.mjs" <deck.pptx>` on the saved file, or the snapshot's `text` — for missing material, typos, and order; a wrong order or a dropped fact is cheaper to catch in text than on a contact sheet. → manual
+**Content before pixels**: read the deck's text once against the fact sheet — `${MIXDOG_SKILL_NODE} "${MIXDOG_SKILL_DIR}/scripts/source-extract.mjs" <deck.pptx>` on the saved file, or the snapshot's `text` — for missing material, typos, and order; a wrong order or a dropped fact is cheaper to catch in text than on a contact sheet. → manual
 Then inspect every rendered slide in this order. Read actual phrase endings, compare peer baselines and gaps, trace each connector to its intended boundary, and check that labels belong to the right object. Record a visible keep/fix reason rather than inferring a pass from a clean QA report. Hard hits need correction; the runtime's actual severity and source determine what blocks finalize. "Manual" means only the rendered image shows it.
 
 **Preview fidelity**: the portable preview is LibreOffice, and it spaces Hangul and CJK *chart category labels* by character — `1분기` is drawn as `1 분기` while the chart part holds the unspaced string, and PowerPoint draws it unspaced. Check the string (`source-extract.mjs` or the chart's `c:v` values) before treating it as a defect; slide text is unaffected.

package/src/defaults/skills/pptx/references/review.md

@@ -52,12 +52,14 @@ policy. Provider and model must be explicit; it uses the existing sign-in, never
 asks to paste credentials. This is context separation, not an OS filesystem
 sandbox or proof that the model's judgement is unbiased.
 
-Run from the Project; `<pptx-skill-directory>` is the `<base-dir>` the runtime
-printed when the skill loaded (`SKILL.md` §2 step 9 carries the command with
-that directory already resolved):
+Run it from the Project with the command `SKILL.md` §2 step 9 carries, which
+the runtime resolved when the skill loaded. It starts with the launcher command
+named there (on the shell tool's PATH), not plain `node`, which cannot load the
+script in every installation. Its shape, with `<pptx-skill-directory>` the
+`<base-dir>` the runtime printed:
 
 ```text
-node "<pptx-skill-directory>/scripts/review-deck.mjs" packet.json --provider <provider> --model <model> --output review.json
+<launcher command from SKILL.md> "<pptx-skill-directory>/scripts/review-deck.mjs" packet.json --provider <provider> --model <model> --output review.json
 ```
 
 Packet format (all paths relative to the packet file unless absolute):

package/src/defaults/skills/pptx/scripts/source-extract.mjs

@@ -6,7 +6,7 @@
 // portable session that is closed without saving; anything else is read as
 // UTF-8 text. Output is bounded by --chars (default 4000) and every block is
 // cut at BLOCK_CHARS, so a large source never floods the context.
-//   node "${MIXDOG_SKILL_DIR}/scripts/source-extract.mjs" <file> [--chars 4000]
+//   ${MIXDOG_SKILL_NODE} "${MIXDOG_SKILL_DIR}/scripts/source-extract.mjs" <file> [--chars 4000]
 import { access, readFile } from 'node:fs/promises';
 import { extname, resolve } from 'node:path';
 import { pathToFileURL } from 'node:url';

package/src/defaults/skills/setup/references/actions.md

@@ -6,7 +6,7 @@ Read this when choosing a `setup` mutation.
 
 - `status` reads one domain: summary, model, agents, workflow, websearch,
   output-style, profile, autoclear, compaction, memory, features, local-provider, shell,
-  providers, mcp, skills, plugins, update, onboarding, capabilities, developer,
+  providers, mcp, skills, plugins, update, onboarding, worktree, capabilities, developer,
   desktop, appearance, projects, connection, schedules, webhooks, hooks, or
   media.
 - `open` navigates to a supported UI surface. Read `surfaces.md` for targets and
@@ -42,6 +42,7 @@ Read this when choosing a `setup` mutation.
 | Permanently delete confirmed, unchanged, unused model files | `delete_local_model` |
 | TUI system shell | `set_system_shell` |
 | Automatic updates | `set_auto_update` |
+| Worktree defaults: new tasks on the current branch or a new worktree, branch prefix, large-repository warning size, automatic cleanup | `set_worktree_settings` |
 | Remove stored provider authentication (one OAuth account at a time) | `forget_provider_auth` |
 | Select, rename, reorder, or auto-switch OAuth accounts, or allow paid credits | `set_provider_account` |
 | Developer options such as risk-gated OAuth providers | `set_developer_option` |
@@ -56,7 +57,7 @@ Read this when choosing a `setup` mutation.
 | Refresh a plugin checkout | `update_plugin` |
 | Enable or disable a plugin | `set_plugin_enabled` |
 | Remove a plugin | `remove_plugin` |
-| Install or reconfigure the MCP server a plugin ships | `enable_plugin_mcp` |
+| Turn back on and reconnect every MCP server a plugin declares | `enable_plugin_mcp` |
 | Available hosted models and their options; optionally one provider, native-search capable only, or a refreshed catalog | `list_models` |
 | Agent orchestration, independent of workflow instructions | `set_orchestration_mode` |
 | Safe MCP edit metadata, excluding raw connection/credential values | `get_mcp_server` |
@@ -85,7 +86,9 @@ Read this when choosing a `setup` mutation.
 Inspect `status model`, `status agents`, `status websearch`, or `status
 workflow` before changing the corresponding domain.
 
-- Main and agent routes also accept `modelParameters` and `contextPercent`; agent routes accept `disabled`.
+- Main and agent routes also accept `modelParameters`, `contextPercent`, and
+  `contextTokens` (an exact window in tokens that overrides `contextPercent`);
+  agent routes accept `disabled`.
   Use the model catalog for actual available options, not remembered model ids.
 - A Web Search route must support native web search; tool exposure is changed
   separately.
@@ -201,8 +204,18 @@ markers are read with `status desktop`.
 - Environment/header maps are partial patches. Omitted keys are preserved;
   `null` explicitly removes one key, including a stored credential.
 - Plugin enablement moves its contributed skills and MCP integrations together.
-- `enable_plugin_mcp` writes the MCP server a plugin ships and reconnects; it
-  fails for a plugin without an MCP script.
+- A registered plugin's MCP servers (its `.mcp.json` and manifest
+  `mcpServers`) appear automatically with `source: "plugin"` and `plugin` /
+  `pluginServer` in `status mcp`; nothing is copied into the MCP list. A plugin
+  server runs only while both the plugin and the server are enabled;
+  `set_mcp_enabled` toggles the server, `set_plugin_enabled` the plugin, and
+  `set_extension_scope` limits it (the plugin's scope still applies).
+  `add_mcp_server`, `save_mcp_server`, and `remove_mcp_server` refuse plugin
+  servers. Declaration problems and naming conflicts are reported as
+  `declarationErrors` and `warnings`.
+- `enable_plugin_mcp` turns every server the plugin declares back on and
+  reconnects them; a disabled plugin stays disabled, and it fails for a plugin
+  that declares no MCP servers.
 
 ## Desktop, Projects, and automation
 

package/src/defaults/skills/visualize/SKILL.md

@@ -0,0 +1,160 @@
+---
+name: visualize
+description: Show mockups, design options, diagrams, and finished artifacts in the side panel beside the chat.
+when_to_use: 'Mockups, previews, design options, diagrams, or showing any finished artifact; not code alone.'
+metadata:
+  requires: visualize
+dependencies:
+  tools:
+    - type: tool
+      value: visualize
+---
+
+# Visualize (visualize tool)
+
+The side panel is where the user sees the work instead of reading about it.
+This file owns when to draw, how a mockup is built and iterated, and how results
+are shown. The schema owns action fields. `references/craft.md` owns the design
+bar every mockup is held to; read it before the first mockup of a session.
+
+## 1. Which job
+**Hard rule — seeing beats reading, or it is not this skill**: draw when the
+answer is a layout, a state, a flow, a spatial relation, or a look. Scope,
+tradeoffs, and wording decisions stay in chat even when the topic is UI.
+- *Show* — any artifact you produced or changed (HTML, deck, document, sheet,
+  PDF, image, video, a local dev page): `open` it once when it is ready. Edits
+  reload the tab by themselves; never reopen to refresh.
+- *Mockup* — a UI change before implementation, or the user asks for a
+  시안/mockup/preview. §2.
+- *Options* — the direction is genuinely open (the user is unsure, or two
+  readings of the request lead to different layouts). §3.
+- *Explain* — a flow, state machine, architecture, or data path the user needs
+  to grasp: one HTML diagram page, no option machinery.
+**Hard rule — not this tool**: authoring Office files, images, or video belongs
+to their own skills; this skill only shows their output. Clicking, typing, and
+testing a page belong to `browser`.
+
+## 2. Mockup loop
+1. **Ground in the real product.** Read the files that define the touched
+   surface — tokens, theme, the component and its neighbours — before drawing.
+   For an existing app screen, prefer the real screen (a capture of the running
+   app) with only the changed parts drawn over it; invented chrome hides what
+   actually changes. Done when you can name the exact component and values the
+   change touches.
+2. **Split the change into states.** Every state the user will meet gets its
+   own labelled frame: A default, B the interaction, C the result, D failure or
+   empty where they exist. One frame per state, captioned with what changed and
+   why in one sentence.
+3. **Write** `design/mockups/<topic>/index.html` (self-contained: inline CSS
+   and JS, no network). Sample data is realistic and marked as sample once in
+   the lead line; never fake metrics or filler sections.
+4. **Show**: `open` the page, then `annotate` each changed element with a short
+   label. Marks live on the panel, not in the file.
+5. **Look before you report.** Run `visualize check` on the tab and fix every
+   reported page problem (uncaught errors, CSP violations, console errors,
+   missing assets, blocked network requests: the page must be self-contained)
+   in the file, then `check` again until it reports none; a
+   "Page problems" section on any `open`, `annotate`, or `capture` result counts
+   the same. Then run `visualize inspect` on the tab: once at the panel's
+   width and, when the design is responsive, again with `viewport`
+   `{width: 375, height: 812}`; fix every finding it lists (overflow, clipped
+   text, tiny or overlapping targets, low contrast, missing alt or labels) in
+   the file before reporting. Then `capture` the tab (or read it through
+   `browser`) and check it against `references/craft.md` §5; fix what fails
+   before the user sees the report.
+6. **Report** in chat: path, the states shown, what to confirm, and the open
+   questions. Then stop — implementation waits for approval.
+7. **Iterate in place.** Feedback edits the same file; the tab reloads and the
+   panel keeps each saved state as a version (v1…vN) the user can step back
+   to, compare side by side, or restore, so never copy the file to keep an
+   old version. The user can also point at the page: their message then carries a
+   `<visualize_feedback>` block of numbered pins (`pin N` = the pin numbered
+   N they saw), each with its label, note, tab, project-relative file and
+   `kind` (pane kind + element or region). An HTML element pin adds its
+   selector, `inside` path, trimmed HTML, key styles and box; a region pin
+   (an HTML drag, a document, PDF, image, diagram, or video frame) adds its
+   area, plus `elements` for an HTML drag and `time` for a video. The crop
+   image "visualize pin N" comes when there is one (`crop: none` otherwise).
+   Answer pin by pin by number and act on what the selector or area names;
+   secret values arrive as `[redacted]`, so never restore them from the crop
+   or guess them.
+8. **After approval**: implement from the approved mockup; its values are the
+   spec. When the work needs a record, `export` the approved tab to PNG or PDF
+   beside the page (`export` with `format` and a project-relative `output`,
+   e.g. `design/mockups/login/login-approved.png`; `version` for a kept
+   version). A single HTML page renders whole in a hidden window, so the tab need
+   not be on screen; the result lists the written path, size, and any page
+   problems. Remove any temporary data or records you created to stage
+   the mockup, and say so.
+
+## 3. Options
+**Default — one strong mockup.** Options only when the direction is open.
+**Hard rule — options differ in stance, not in pixels**: pick one axis that
+matters for this decision (density, layout model, emphasis, interaction model)
+and pull the 2–3 options apart on it. Two options that differ by accent colour
+or radius are one option.
+- Name each by its stance ("1 · one menu", "2 · inline chips") and give its
+  cost in one line.
+- `open` 2–3 targets in one call to show them side by side.
+- Make the options pickable: put `data-choice="<id>"` on each option's root
+  (and `data-choice-group="<decision>"` when one page asks more than one
+  question; ungrouped choices form one group). A click selects one per group
+  and adds `data-choice-selected` to it; the panel's default outline uses
+  `:where()`, so style `[data-choice][data-choice-selected]` to match the
+  mockup. `data-choice-label` names it when its text does not. Picks reach
+  you with the user's next message under `choices:` in the
+  `<visualize_feedback>` block as tab, file, group, choice id and label.
+- When the user picks, consolidate into one mockup and close the others; a
+  pile of options is not a result.
+
+## 4. Showing results
+- Open a result once, when it is ready to look at, not at every intermediate
+  write. Long multi-step work shows the first reviewable state, then lets
+  reloads carry the rest.
+- Decks and documents: the panel shows the converted pages; still run the
+  owning skill's own render review — the panel is for the user, not your QC.
+- Close tabs that no longer match the conversation (a resolved choice, a
+  replaced draft) so the user is never looking at stale work.
+- A web URL is opened alone, never mixed with files; marking, capturing, or
+  closing a web page belongs to `browser`.
+- `annotate` and `capture` need the tab on screen (`export` of an HTML page does not). When they report the panel
+  is hidden, tell the user to open it rather than retrying.
+- When `visualize` is unavailable, give the file path; never start your own
+  server or pretend it was shown.
+
+### Inline content or a file
+- **Inline** (`open` with `input.content` — `mermaid`, `svg`, `chart` (an
+  ECharts option as JSON), `html`, `markdown`): diagrams, charts, quick
+  sketches and notes you will not edit again. No file is written. `mermaid`,
+  `svg` and `markdown` are static, sanitized output (no scripts). `chart` is
+  interactive (tooltips, zoom) and `html` runs its scripts, as in-memory pages
+  held to the same rules as an HTML file: outside network is blocked, so an
+  `html` sketch must be self-contained, and its page problems come back in the
+  result. Re-opening with the same `title` adds a version.
+- **File** (`targets`): a real mockup that needs iteration, scripts, assets,
+  versioning on disk, annotate/Comment feedback, or a deliverable the user keeps.
+- A render error comes back as a page problem (Mermaid parse error, bad chart
+  JSON): fix the source and open again; never report it as shown.
+
+### Tweaks
+Offer tweaks when the open decision is a value, not a layout: colour, density,
+spacing, type scale, radius. The page declares them in
+`<script type="application/json" id="mx-tweaks">` (HTML files and inline
+`html`/`chart`): `[{"id":"accent","label":"Accent","type":"color","var":"--accent","default":"#3b82f6"}]`.
+Types: `color` (hex/rgb), `range` (`min`,`max`,`step`,`unit`),
+`select` (`options`), `toggle`; each sets a `--css-var` or a
+`data-*` attribute on the root, which your CSS reads. Keep it to 6 or fewer
+meaningful controls (16 is the hard cap) and give every one a real default.
+The panel adds a Tweaks button; when the user sends their values you get them
+under `tweaks:` in `<visualize_feedback>` (`accent (--accent): #2f6fed (was
+#3b82f6)`): bake them into the file as the new defaults — both the CSS value
+and the control's `default`. A malformed block is a `tweaks: <reason>` page
+problem.
+
+## 5. Reporting
+Short: what is on screen, the path, the decision you need. Do not narrate the
+HTML. Uncertain or unverified rendering is stated, not glossed.
+The panel and the chat card already show the content: never repeat its source
+(Mermaid text, chart JSON, SVG, HTML) or re-render its data as a chat table or
+text chart. Name the tab and give one or two lines on what it shows or what to
+look at.

package/src/defaults/skills/visualize/references/craft.md

@@ -0,0 +1,88 @@
+# Craft bar for mockups and visual pages
+
+The bar is "a designer on this product drew it", not "a page was generated".
+Most generated-looking design is wrong composition, not wrong colour; fix the
+cause in the order below.
+
+## 1. Source before taste
+- An existing product wins over every rule here. Lift its real tokens,
+  spacing, radii, type, copy tone, and density from source files; do not
+  approximate from memory or a screenshot when the code is readable.
+- A new surface with no system: define a small one first (4–6 named colours,
+  one or two type families with roles, a spacing step, one radius scale) and
+  write it as CSS variables at the top of the page.
+- A brief's explicit direction wins, including when it asks for a look this
+  file warns against.
+
+## 2. Name the surface, then compose
+Before any colour or component, decide in one line what the user does here;
+the composition follows from it:
+
+| Surface | User is… | Composition carries |
+|---|---|---|
+| Watch | following changing state | density, glanceable hierarchy, no marketing framing |
+| Act | working through items | selection, row actions, clear current item |
+| Compare | weighing options | aligned columns, equal structure, one differentiator |
+| Set up | configuring | grouped fields, save/validation states, little decoration |
+| Persuade / teach | being convinced or taught | one idea per section; the only place a hero belongs |
+| Browse | exploring a space | filters, results, peek/zoom |
+| Drive / inspect | commanding or drilling into one object | keyboard paths, focus, detail panes |
+
+A screen spanning two picks a primary one; never average them.
+
+## 3. Hierarchy with type and space first
+- Scale, weight, and spacing carry hierarchy before boxes, borders, icons, or
+  colour. Add a container only when it groups something the user acts on.
+- Line length under ~80 characters; numbers in tables tabular and aligned.
+- Structural marks (numbers, dividers, labels) only when they encode real
+  information — numbered steps only for a real sequence.
+- One memorable element per page; everything around it stays quiet.
+
+## 4. Words are interface
+- Name things by what the user understands, not how the system is built.
+- Buttons say what happens ("Save changes"); the same action keeps the same
+  name through the flow ("Publish" → "Published").
+- Errors say what happened and what to do, without apology. Empty states
+  invite the next action.
+- Sentence case, plain verbs, no filler or invented claims. Draft copy is
+  marked as draft.
+
+## 5. Self-review before showing
+Look at the capture, list which of these fire, then repair by cause:
+
+Composition (re-layout; recolouring will not fix these):
+1. Wrong surface — e.g. a hero on a dashboard.
+2. Everything centred because no layout was chosen.
+3. Equal-weight tile grids (icon + heading + sentence ×3) where one thing
+   should lead.
+
+Decoration (remove; replace with scale, weight, or spacing):
+4. Coloured side rails on cards.
+5. Blur or glass with no elevation system behind it.
+6. Oversized stat numbers filling space meant for content.
+7. A rounded icon above every heading; icons that do not aid scanning.
+8. Identical cards, one radius and one soft shadow on everything.
+
+Defaults (recolour / re-typeset only if the brief left them free):
+9. Indigo/violet accent or glossy gradient chosen by habit.
+10. Default font used without a reason; ALL-CAPS eyebrow labels over every
+    heading; one word accented in a headline; "→" on every link.
+
+Re-check after repairing. Do not show while 1–3 still fire.
+
+## 6. Quality floor (always, without announcing it)
+`visualize inspect` reports most floor failures (overflow, clipped text,
+small or overlapping targets, low contrast, missing alt or labels); a page is
+not shown while it lists any.
+- Visible keyboard focus, real hover and disabled states where shown.
+- Readable contrast for text and controls; hit targets ≥ 44px on touch frames.
+- Motion only when it shows what changed; respect `prefers-reduced-motion`.
+- Fits the panel width without horizontal scroll; frames wider than the panel
+  scale down rather than clip.
+- Self-contained: no CDN, web font, API or other network request. The panel
+  blocks outside network unless the user allows it (a blocked request is a
+  `blocked-network` page problem), so inline or bundle every asset in the
+  project. No fake data passed off as real, or stock imagery.
+- A value the user may want to tune (accent, density, spacing) is a CSS variable
+  on the root plus a control in `<script type="application/json" id="mx-tweaks">
+  [{"id":"accent","type":"color","var":"--accent","default":"#3b82f6"}]`.