mixdog 0.9.139 → 0.9.140

package/LICENSES/MIT.txt

@@ -8,24 +8,18 @@ component listed here. Each component keeps its own copyright notice.
   https://github.com/microsoft/vscode
   Copyright (c) 2015 - present Microsoft Corporation
 
+  Visual Studio Code Codicons (icon font; the icons are CC BY 4.0)
+  https://github.com/microsoft/vscode-codicons
+  Copyright (c) Microsoft Corporation
+
   GitHub Pull Requests and Issues (VS Code extension)
   https://github.com/microsoft/vscode-pull-request-github
   Copyright (c) Microsoft Corporation. All rights reserved.
 
-  Files
-  https://github.com/files-community/Files
-  Copyright (c) 2018 - present Files Community
-
-  OpenCode
-  https://github.com/anomalyco/opencode
-  Copyright (c) 2025 opencode
-
-  pi
-  https://github.com/earendil-works/pi
-  Copyright (c) 2025 Mario Zechner
-
-  Orca
-  Copyright (c) 2026 Lovecast Inc.
+  ripgrep (grep/ignore crates plus the packaged rg executable; offered as
+  MIT OR Unlicense — the MIT terms are taken here)
+  https://github.com/BurntSushi/ripgrep
+  Copyright (c) 2015 Andrew Gallant
 
   Ink (vendored under vendor/ink, which keeps its own license file)
   https://github.com/vadimdemedes/ink

package/NOTICE.md

@@ -1,9 +1,10 @@
 # NOTICE
 
-Mixdog itself is MIT-licensed (see `LICENSE`). Parts of it were written
-against the public source of the projects listed below. Their terms are
-preserved here in one place so the individual source files can stay free of
-scattered attribution comments.
+Mixdog itself is MIT-licensed (see `LICENSE`). The license sections below
+cover third-party code and data that Mixdog actually carries. The closing
+"Behavioral references" section records projects that only informed the
+implementation, with no code taken. Everything is kept here in one place so
+the individual source files stay free of scattered attribution comments.
 
 Full license texts live in `LICENSES/`:
 
@@ -12,6 +13,10 @@ Full license texts live in `LICENSES/`:
 - `LICENSES/Apache-2.0.txt` — the Apache License 2.0.
 - `LICENSES/codex-NOTICE.txt` — the upstream NOTICE file of the Codex CLI,
   carried forward as Apache-2.0 section 4(d) requires.
+- `LICENSES/ripgrep-NOTICE.txt` — the ripgrep notice for the search crates
+  compiled into the native binaries.
+- `LICENSES/editor-assets-NOTICE.txt` — the editor icon and language-data
+  notice referenced by the generated files.
 
 ## MIT
 
@@ -34,6 +39,14 @@ language contributions used to generate the Monaco pane languages.
 `editor-ansi.ts` the terminal ANSI palette values, and `seti-icons.ts` /
 `editor-languages.ts` are generated from the shipped extension data.
 
+### Visual Studio Code Codicons — Copyright (c) Microsoft Corporation
+
+<https://github.com/microsoft/vscode-codicons>
+
+The chrome-level UI glyph font used by the desktop and web renderer. The code
+is MIT; the icons themselves are licensed under Creative Commons Attribution
+4.0 International (CC BY 4.0).
+
 ### GitHub Pull Requests extension — Copyright (c) Microsoft Corporation
 
 <https://github.com/microsoft/vscode-pull-request-github>
@@ -46,39 +59,17 @@ The default pull-request query set and its category labels in
 Bundled glyphs and colour tables. Full license text in
 `apps/desktop/THIRD-PARTY-NOTICES.txt`.
 
-### Files — Copyright (c) Files Community
-
-<https://github.com/files-community/Files>
-
-Folder-pane grouping keys, date-span labels, size buckets and the discrete
-layout size ladder in `apps/desktop/src/renderer/FolderPane.lazy.tsx`.
-
-### OpenCode — Copyright (c) Anomaly / SST
-
-<https://github.com/anomalyco/opencode>
+### ripgrep — Copyright (c) 2015 Andrew Gallant
 
-Transcript auto-scroll gesture grammar, virtual-timeline anchoring and the
-streaming-markdown projection model.
+<https://github.com/BurntSushi/ripgrep>
 
-- `apps/desktop/src/renderer/use-transcript-follow.ts`, `TranscriptList.tsx`,
-  `transcript-measure.ts`, `transcript-rows.ts`, `transcript-virtual-cache.ts`,
-  `StreamingMarkdownBody.tsx`, `streaming-markdown.ts`, `Conversation.tsx`
+The `grep` and `ignore` library crates are compiled into the native search
+binary, and the packaged `rg` executable ships as the grep tool's fast path.
+ripgrep is offered under the MIT License OR the Unlicense; Mixdog takes the
+MIT terms. The upstream notice is preserved as `LICENSES/ripgrep-NOTICE.txt`.
 
-### pi — Copyright (c) 2025 Mario Zechner
-
-<https://github.com/earendil-works/pi>
-
-`trimPartialClosingFences()` in `src/tui/markdown/stream-fence.mjs`, ported
-from the TUI markdown component.
-
-### Orca — Copyright (c) 2026 Lovecast Inc.
-
-Editor-surface and tab-hierarchy structure, the terminal host portal, the
-keep-awake power-save blocker and the hosted-review link derivation.
-
-- `apps/desktop/src/renderer/EditorPane.lazy.tsx`, `PaneSurfaceGate.tsx`,
-  `PullRequestsPane.tsx`, `SourceControlDock.tsx`
-- `apps/desktop/src/main/agent-awake.ts`
+- `native/mixdog-graph/` (the `grep` and `ignore` crates)
+- `src/runtime/agent/orchestrator/tools/builtin/` (the `rg` fast path)
 
 ### Ink — Copyright (c) Vadim Demedes
 
@@ -107,3 +98,44 @@ upstream NOTICE file is preserved as `LICENSES/codex-NOTICE.txt`.
 The Apache-2.0 terms require this notice to travel with any redistribution of
 the derived files. Full license text: `LICENSES/Apache-2.0.txt`
 (<https://www.apache.org/licenses/LICENSE-2.0>).
+
+## Behavioral references
+
+The projects below informed Mixdog through publicly observable behavior, a
+published wire contract, or a documented algorithm. No source code from them
+is present in Mixdog, so no license obligation travels with the result; they
+are recorded here because the source files themselves carry no attribution
+comments.
+
+- **Chromium** — the workspace tab strip follows Chromium's tab-strip layout
+  semantics: two layout domains around the crossover width, the active-tab
+  floor, the inactive sliver floor, the left-to-right remainder grant and the
+  icon visibility ladder. The implementation is independent TypeScript and CSS
+  on Mixdog's own constants (`apps/desktop/src/renderer/WorkspaceTabStrip.tsx`,
+  `apps/desktop/src/renderer/desktop/04-workspace-tabs.css`).
+- **GitHub Desktop** — the Source Control dock grammar: commit-form layout,
+  changed-file status glyph semantics, single-sentence path colouring, history
+  context-menu gating, and detached-checkout / tag rules. Painted entirely on
+  Mixdog's own semantic tokens.
+- **OpenCode** — transcript auto-scroll gesture grammar, virtual-timeline
+  anchoring and the streaming-markdown projection model.
+- **Orca** — editor-surface and tab-hierarchy structure, the terminal host
+  portal, the keep-awake power-save blocker and the hosted-review link
+  derivation.
+- **Files** — the folder pane's grouping keys, date-span labels, size buckets
+  and discrete layout size ladder
+  (`apps/desktop/src/renderer/FolderPane.lazy.tsx`).
+- **Claude Code** (Anthropic) — terminal input tokenizing and keypress
+  parsing, the TUI selection word model in `vendor/ink`, and interactive turn
+  semantics such as queued-message recall, message-selector rewind and
+  background-task lifetime. The Anthropic OAuth route additionally sends the
+  client identity that Anthropic's token edge validates; those values are a
+  wire requirement, not a derivation.
+- **opencode-antigravity-auth** — MIT
+  (<https://github.com/NoeFabris/opencode-antigravity-auth>). The Google Cloud
+  Code Assist wire contract used by the Antigravity OAuth provider: OAuth
+  client parameters, endpoint fallback order, client impersonation headers,
+  request envelope and thinking-signature handling.
+- **assistant-ui, Chatbox, Cherry Studio, Jan, Zed** — the heading-size and
+  list-density survey behind the desktop markdown ladder in
+  `apps/desktop/src/renderer/desktop/22-markdown.css`.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mixdog",
-  "version": "0.9.139",
+  "version": "0.9.140",
   "private": false,
   "type": "module",
   "description": "Standalone mixdog coding-agent CLI/TUI workspace.",
@@ -108,7 +108,6 @@
     "chalk": "^5.6.2",
     "cli-highlight": "^2.1.11",
     "diff": "^9.0.0",
-    "discord.js": "^14.26.4",
     "ink": "^7.1.0",
     "jsdom": "^27.0.0",
     "kiwi-nlp": "^0.21.0",
@@ -129,12 +128,6 @@
     "zod": "^3.25.76"
   },
   "overrides": {
-    "discord.js": {
-      "undici": "6.28.0"
-    },
-    "@discordjs/rest": {
-      "undici": "6.28.0"
-    },
     "onnxruntime-node": {
       "global-agent": "^4.1.3"
     },

package/src/lib/output-style-meta.cjs

@@ -0,0 +1,115 @@
+// Output-style metadata contract: id normalization, frontmatter parsing,
+// catalog ordering, and lookup. Required by the CJS rules builder (prompt
+// injection) and imported by the ESM session runtime (pickers, config), so one
+// schema governs both instead of two parallel copies drifting apart.
+//
+// Frontmatter schema for `output-styles/<id>.md`:
+//   name                — canonical id (defaults to the file name)
+//   title | label       — display label (defaults to a title-cased id)
+//   description         — one-line picker text
+//   aliases             — comma-separated alternate ids; the only alias source
+//   partial             — true hides the file from the selectable catalog
+//   keep-shared-format  — false replaces common.md instead of extending it
+'use strict';
+
+const DEFAULT_OUTPUT_STYLE_ID = 'simple';
+// Catalog order: deepest first, then unknown/custom styles by label.
+const OUTPUT_STYLE_ORDER = ['detailed', 'simple', 'minimal', 'extreme-minimal'];
+
+/** Slug form of a style id. Aliases live in frontmatter, never in code. */
+function normalizeOutputStyleId(value) {
+  const slug = String(value ?? '').trim().toLowerCase()
+    .replace(/[_\s]+/g, '-')
+    .replace(/^-+|-+$/g, '');
+  return /^[a-z0-9.-]+$/.test(slug) ? slug : '';
+}
+
+/** Separator-insensitive key so `one-line`, `one_line`, and `oneline` match. */
+function outputStyleCompactKey(value) {
+  return normalizeOutputStyleId(value).replace(/[_.-]+/g, '');
+}
+
+function titleCaseOutputStyle(id) {
+  return String(id || '')
+    .split(/[_.-]+/)
+    .filter(Boolean)
+    .map((part) => `${part.charAt(0).toUpperCase()}${part.slice(1)}`)
+    .join(' ') || 'Default';
+}
+
+function parseOutputStyleFrontmatter(markdown) {
+  const match = String(markdown || '').match(/^---\r?\n([\s\S]*?)\r?\n---/);
+  const meta = {};
+  if (!match) return meta;
+  for (const line of match[1].split(/\r?\n/)) {
+    const kv = line.match(/^([A-Za-z0-9_-]+):\s*(.*?)\s*$/);
+    if (!kv) continue;
+    meta[kv[1]] = kv[2].replace(/^['"]|['"]$/g, '').trim();
+  }
+  return meta;
+}
+
+function outputStyleFlag(value, fallback) {
+  const raw = String(value ?? '').trim().toLowerCase();
+  if (!raw) return fallback;
+  if (/^(?:true|1|yes|on)$/.test(raw)) return true;
+  if (/^(?:false|0|no|off)$/.test(raw)) return false;
+  return fallback;
+}
+
+/**
+ * Frontmatter → catalog entry. Returns null for `partial: true` files (the
+ * shared format partial) and for files with no usable id.
+ */
+function outputStyleMetaFromMarkdown(markdown, fileName) {
+  const meta = parseOutputStyleFrontmatter(markdown);
+  if (outputStyleFlag(meta.partial, false)) return null;
+  const fileId = normalizeOutputStyleId(String(fileName || '').replace(/\.md$/i, ''));
+  const id = normalizeOutputStyleId(meta.name) || fileId;
+  if (!id) return null;
+  return {
+    id,
+    label: String(meta.title || meta.label || '').trim() || titleCaseOutputStyle(id),
+    description: String(meta.description || '').trim(),
+    aliases: String(meta.aliases || '').split(',')
+      .map((alias) => normalizeOutputStyleId(alias))
+      .filter(Boolean),
+    // Built-in and custom styles both inherit the shared format partial; a
+    // style opts out only by declaring `keep-shared-format: false`.
+    keepSharedFormat: outputStyleFlag(meta['keep-shared-format'], true),
+  };
+}
+
+function sortOutputStyles(styles) {
+  return [...(styles || [])].sort((a, b) => {
+    const ai = OUTPUT_STYLE_ORDER.indexOf(a.id);
+    const bi = OUTPUT_STYLE_ORDER.indexOf(b.id);
+    if (ai !== bi) return (ai < 0 ? 999 : ai) - (bi < 0 ? 999 : bi);
+    return a.label.localeCompare(b.label, 'en', { sensitivity: 'base' });
+  });
+}
+
+/** Resolve an id, label, or frontmatter alias, ignoring separators and case. */
+function matchOutputStyle(value, styles) {
+  const id = normalizeOutputStyleId(value);
+  const compact = outputStyleCompactKey(value);
+  if (!id && !compact) return null;
+  return (styles || []).find((style) => {
+    if (style.id === id || outputStyleCompactKey(style.id) === compact) return true;
+    if (outputStyleCompactKey(style.label) === compact) return true;
+    return (style.aliases || []).some((alias) => alias === id || outputStyleCompactKey(alias) === compact);
+  }) || null;
+}
+
+module.exports = {
+  DEFAULT_OUTPUT_STYLE_ID,
+  OUTPUT_STYLE_ORDER,
+  normalizeOutputStyleId,
+  outputStyleCompactKey,
+  titleCaseOutputStyle,
+  parseOutputStyleFrontmatter,
+  outputStyleFlag,
+  outputStyleMetaFromMarkdown,
+  sortOutputStyles,
+  matchOutputStyle,
+};

package/src/lib/rules-builder.cjs

@@ -35,6 +35,12 @@
 
 const fs = require('fs');
 const path = require('path');
+const {
+  DEFAULT_OUTPUT_STYLE_ID,
+  matchOutputStyle,
+  outputStyleMetaFromMarkdown,
+  sortOutputStyles,
+} = require('./output-style-meta.cjs');
 
 /**
  * Read a single section from mixdog-config.json (unified config).
@@ -187,6 +193,7 @@ function buildLanguageSection(dataDir) {
   const lines = [
     `- Default user-facing language${source}: ${language.prompt}. Use it for all user-facing text (preambles, progress, questions, reports, notices), overriding output style; switch only when the user does or asks.`,
     `- Keep code identifiers, paths, commands, symbols, API names, and exact errors in original form.`,
+    `- Never coin a word-for-word translation of source jargon; use the established original term or a plain functional description.`,
   ];
   return `# Language\n\n${lines.join('\n')}`;
 }
@@ -221,84 +228,10 @@ function omitToolRoutes(text, omitTools = []) {
   return out.replace(/[ \t]+\n/g, '\n').replace(/\n{3,}/g, '\n\n');
 }
 
-const DEFAULT_OUTPUT_STYLE_ID = 'simple';
-const OUTPUT_STYLE_ORDER = ['detailed', 'simple', 'minimal', 'extreme-minimal'];
-const OUTPUT_STYLE_ALIASES = new Map([
-  ['extreme', 'extreme-minimal'],
-  ['extremesimple', 'extreme-minimal'],
-  ['extreme-simple', 'extreme-minimal'],
-  ['extreme_simple', 'extreme-minimal'],
-  ['extrememinimal', 'extreme-minimal'],
-  ['extreme_minimal', 'extreme-minimal'],
-  ['mono', 'extreme-minimal'],
-  ['oneline', 'extreme-minimal'],
-  ['one-line', 'extreme-minimal'],
-  ['one_line', 'extreme-minimal'],
-]);
-
-function normalizeOutputStyleId(value) {
-  const raw = String(value || '').trim().toLowerCase();
-  if (!raw) return '';
-  const slug = raw.replace(/[_\s]+/g, '-').replace(/^-+|-+$/g, '');
-  const compact = slug.replace(/[_.-]+/g, '');
-  if (OUTPUT_STYLE_ALIASES.has(slug)) return OUTPUT_STYLE_ALIASES.get(slug);
-  if (OUTPUT_STYLE_ALIASES.has(compact)) return OUTPUT_STYLE_ALIASES.get(compact);
-  return /^[a-z0-9.-]+$/.test(slug) ? slug : '';
-}
-
-function outputStyleCompactKey(value) {
-  return normalizeOutputStyleId(value).replace(/[_.-]+/g, '');
-}
-
-function titleCaseOutputStyle(id) {
-  return String(id || '')
-    .split(/[_.-]+/)
-    .filter(Boolean)
-    .map((part) => `${part.charAt(0).toUpperCase()}${part.slice(1)}`)
-    .join(' ') || 'Default';
-}
-
-function parseOutputStyleFrontmatter(markdown) {
-  const match = String(markdown || '').match(/^---\r?\n([\s\S]*?)\r?\n---/);
-  const meta = {};
-  if (!match) return meta;
-  for (const line of match[1].split(/\r?\n/)) {
-    const kv = line.match(/^([A-Za-z0-9_-]+):\s*(.*?)\s*$/);
-    if (!kv) continue;
-    meta[kv[1]] = kv[2].replace(/^['"]|['"]$/g, '').trim();
-  }
-  return meta;
-}
-
-function outputStyleFlag(value, fallback) {
-  const raw = String(value ?? '').trim().toLowerCase();
-  if (!raw) return fallback;
-  if (/^(?:true|1|yes|on)$/.test(raw)) return true;
-  if (/^(?:false|0|no|off)$/.test(raw)) return false;
-  return fallback;
-}
-
-function readOutputStyleEntry(filePath, source) {
-  const raw = readOptional(filePath);
-  if (!raw) return null;
-  const meta = parseOutputStyleFrontmatter(raw);
-  if (/^(?:true|1|yes)$/i.test(String(meta.partial || '').trim())) return null;
-  const fileId = normalizeOutputStyleId(path.basename(filePath).replace(/\.md$/i, ''));
-  const id = normalizeOutputStyleId(meta.name) || fileId;
-  if (!id) return null;
-  return {
-    id,
-    label: String(meta.title || meta.label || '').trim() || titleCaseOutputStyle(id),
-    aliases: String(meta.aliases || '').split(',')
-      .map((alias) => normalizeOutputStyleId(alias))
-      .filter(Boolean),
-    // Built-in and custom styles both inherit the shared format partial; a
-    // style opts out only by declaring `keep-shared-format: false`.
-    keepSharedFormat: outputStyleFlag(meta['keep-shared-format'], true),
-    source,
-    raw,
-  };
-}
+// Framing line under the style header: the block owns user-facing prose only,
+// so a style (including a `keep-shared-format: false` replacement) never reads
+// as guidance for code or tool payloads.
+const OUTPUT_STYLE_ANCHOR = 'The section below governs how you write user-facing text; it outranks any other formatting habit and does not apply to code or tool calls.';
 
 function listOutputStyleEntries({ PLUGIN_ROOT, DATA_DIR }) {
   const byId = new Map();
@@ -309,27 +242,12 @@ function listOutputStyleEntries({ PLUGIN_ROOT, DATA_DIR }) {
     let files = [];
     try { files = fs.readdirSync(dir).filter((name) => name.toLowerCase().endsWith('.md')).sort(); } catch {}
     for (const name of files) {
-      const style = readOutputStyleEntry(path.join(dir, name), source);
-      if (style) byId.set(style.id, style);
+      const raw = readOptional(path.join(dir, name));
+      const meta = raw ? outputStyleMetaFromMarkdown(raw, name) : null;
+      if (meta) byId.set(meta.id, { ...meta, source, raw });
     }
   }
-  return [...byId.values()].sort((a, b) => {
-    const ai = OUTPUT_STYLE_ORDER.indexOf(a.id);
-    const bi = OUTPUT_STYLE_ORDER.indexOf(b.id);
-    if (ai !== bi) return (ai < 0 ? 999 : ai) - (bi < 0 ? 999 : bi);
-    return a.label.localeCompare(b.label, 'en', { sensitivity: 'base' });
-  });
-}
-
-function findOutputStyleEntry(value, styles) {
-  const id = normalizeOutputStyleId(value);
-  const compact = outputStyleCompactKey(value);
-  if (!id && !compact) return null;
-  return styles.find((style) => {
-    if (style.id === id || outputStyleCompactKey(style.id) === compact) return true;
-    if (outputStyleCompactKey(style.label) === compact) return true;
-    return style.aliases.some((alias) => alias === id || outputStyleCompactKey(alias) === compact);
-  }) || null;
+  return sortOutputStyles([...byId.values()]);
 }
 
 /** Shared format partial: a user copy in DATA_DIR overrides the built-in one. */
@@ -343,14 +261,14 @@ function loadOutputStyle({ PLUGIN_ROOT, DATA_DIR }) {
   // `agent.outputStyle` key is dropped by config canonicalization.
   const configured = String(readUnifiedConfig(DATA_DIR).outputStyle || '').trim() || DEFAULT_OUTPUT_STYLE_ID;
   const styles = listOutputStyleEntries({ PLUGIN_ROOT, DATA_DIR });
-  const style = findOutputStyleEntry(configured, styles)
-    || findOutputStyleEntry(DEFAULT_OUTPUT_STYLE_ID, styles)
+  const style = matchOutputStyle(configured, styles)
+    || matchOutputStyle(DEFAULT_OUTPUT_STYLE_ID, styles)
     || styles[0];
   if (!style) return '';
   const selected = stripFrontmatter(style.raw);
   if (!selected) return '';
   const common = style.keepSharedFormat ? readSharedFormatPartial({ PLUGIN_ROOT, DATA_DIR }) : '';
-  return [`# Output Style: ${style.label}`, common, selected].filter(Boolean).join('\n\n');
+  return [`# Output Style: ${style.label}`, OUTPUT_STYLE_ANCHOR, common, selected].filter(Boolean).join('\n\n');
 }
 
 function buildSharedToolContent({ PLUGIN_ROOT, omitTools = [] } = {}) {

package/src/output-styles/common.md

@@ -7,48 +7,32 @@ partial: true
 
 ## Shared Output Format
 
-- Lead with the answer or action. Write for a person, not as a console log. Match
-  the request's language, tone, and apparent expertise; use complete grammatical
-  sentences and explain unfamiliar terms without talking down to the reader.
-- Select content according to the active depth variation before formatting it.
-  Formatting must never restore information that variation discarded.
-- Choose the presentation from the content, the reader, and the task. Markdown is
-  available, not mandatory. Use paragraphs, line breaks, bullets, numbered steps,
-  headings, tables, or code in any combination that makes the particular answer
-  easier to understand; no response shape is the default or a required mapping
-  from question type to format.
+- Lead with the answer or action, then the context needed to understand it, then
+  lower-priority detail. Write for a person, not a console log: complete
+  sentences, explained terms, and a register matched to the reader's language,
+  tone, and apparent expertise.
+- Select content by the active depth variation before formatting it. Formatting
+  never restores information that variation discarded.
+- Choose the shape from the content and the reader. Markdown is available, not
+  mandatory; prose, bullets, numbered steps, headings, tables, and code may
+  combine in any way that makes this particular answer easier to read, and no
+  response shape is a default or a required mapping from question type.
 - Output renders as GitHub-flavored Markdown — tables, task lists, footnotes,
   fenced code with a language tag, and `$$…$$` math — in a proportional-width
-  surface. Space-aligned columns and ASCII art do not line up there; use a real
-  table or a fenced block instead.
-- Keep simple answers visually simple. For longer answers, add only the structure
-  that creates useful grouping, order, contrast, or emphasis. Do not rotate
-  formats for variety, force a recurring template, label the opening answer or
-  closing paragraph, create headings for tiny blocks, or nest lists without need.
-- Never produce an essay-shaped wall of text. When a block contains several
-  separable ideas or becomes hard to scan, split it or choose a clearer structure.
-  Do not overcorrect into fragments or excessive sections when a compact answer
-  already reads cleanly.
-- Use tables only for short enumerable or quantitative facts when they genuinely
-  improve comparison. Keep explanatory reasoning outside table cells.
-- Keep the flow linear so each sentence builds on what came before without
-  semantic backtracking. Keep related ideas together and give each paragraph or
-  list item one main idea. Split or list a block as soon as it serves multiple
-  roles. Avoid essay-shaped walls of text, fragments, excessive dashes,
-  symbol-heavy shorthand, dense paragraphs, and decorative emphasis.
-- Use an inverted pyramid when useful: answer first, then the context needed to
-  understand it, then lower-priority detail. State each material point once; do
-  not repeat the opening conclusion in a `Summary`, `Conclusion`, or `In short`
-  ending, and never force a recurring response template.
-- Keep established facts separate from plausible inference. Do not invent a
-  mechanism, value, consequence, or certainty to make an explanation feel
-  complete.
-- Use user-facing terms without unexplained internal shorthand, routine process
-  narration, raw output dumps, or emojis unless requested. Show code and file
-  references only when their exact form helps the reader.
-- Reader understanding outranks terseness, but every retained sentence must earn
-  its place. Do not restate the request or announce compliance with constraints
-  such as language or tool use. Remove preamble, filler, obvious statements, and
-  repetition. Once every requested point is covered, stop; do not append a recap,
-  solution, recommendation, next step, or offer of more help unless requested or
-  necessary to prevent a misleading answer. Preserve exact technical literals.
+  surface, so space-aligned columns and ASCII art never line up; use a real table
+  or a fenced block instead. Use tables only for short enumerable or quantitative
+  facts, and keep explanatory reasoning outside the cells.
+- Keep simple answers visually simple and add structure only where it creates
+  useful grouping, order, or contrast. Never produce an essay-shaped wall of
+  text, and never overcorrect into fragments, decorative emphasis, tiny headings,
+  needless nesting, or a recurring template.
+- Give each paragraph or list item one idea and keep the flow linear, so the
+  reader builds meaning forward without re-parsing what came before.
+- Keep established facts separate from plausible inference. Never invent a
+  mechanism, value, or certainty to make an explanation feel complete, preserve
+  exact technical literals, and show code or file references only where their
+  exact form helps.
+- State each material point once and stop when the request is covered. Do not
+  restate the request, close with a recap, or append an unrequested
+  recommendation, next step, or offer of help; no emojis, raw output dumps, or
+  narration of your own process unless asked.

package/src/output-styles/detailed.md

@@ -9,8 +9,8 @@ aliases: verbose, full
 
 Detailed — concrete explanation.
 
-- Information depth: retain the answer plus every established explanatory layer
-  that materially improves understanding—mechanism, evidence, trade-offs,
-  implications, and useful examples. Deepen the requested subject without
-  filling evidentiary gaps or broadening into fixes, recommendations, or adjacent
-  lessons unless asked.
+- Retain: the answer plus every established explanatory layer that materially
+  improves understanding — mechanism, evidence, trade-offs, implications, and
+  useful examples. Deepen the requested subject rather than widening it.
+- Omit: speculation that fills an evidentiary gap, and any expansion into fixes,
+  recommendations, or adjacent lessons that were not requested.

package/src/output-styles/extreme-minimal.md

@@ -2,16 +2,16 @@
 name: extreme-minimal
 title: Extreme Minimal
 description: Final decision or direct answer in one or two sentences
-aliases: extreme, oneline, one-line, one line, mono
+aliases: extreme, extreme-simple, one-line, mono
 ---
 
 ## Depth Variation
 
 Extreme minimal — final decision or answer.
 
-- Information depth: the entire response is one or two short sentences containing
-  only the final decision, result, or direct answer. Include no heading, label,
-  list, explanation, supporting fact, scope, caveat, recap, or follow-up, even
-  when asked to explain or report. For a cause question, the direct answer is the
-  highest-level cause, not its concrete causal chain. Never pack discarded detail
-  into a longer sentence.
+- Retain: only the final decision, result, or direct answer. For a cause
+  question, that is the highest-level cause, not its concrete causal chain.
+- Omit: every heading, label, list, explanation, supporting fact, scope, caveat,
+  recap, and follow-up, even when asked to explain or report.
+- Shape: one or two short sentences. Never pack discarded detail into a longer
+  sentence.

package/src/output-styles/minimal.md

@@ -2,15 +2,16 @@
 name: minimal
 title: Minimal
 description: Conclusion, core cause, and one decisive fact
+aliases: brief, short
 ---
 
 ## Depth Variation
 
 Minimal — conclusion and core cause.
 
-- Information depth: retain the conclusion, the single core cause that determines
-  it, and one or two decisive concrete facts that make the cause clear. Include
-  the immediate consequence when it completes the answer, then stop before a
-  full explanation. Do not add secondary evidence, deployment state, unaffected
-  scope, broader impact, exceptions, remediation, or process unless that category
-  is the requested answer.
+- Retain: the conclusion, the single core cause that determines it, one or two
+  decisive concrete facts that make the cause clear, and the immediate
+  consequence when it completes the answer. Stop before a full explanation.
+- Omit: secondary evidence, deployment state, unaffected scope, broader impact,
+  exceptions, remediation, and process, unless that category is the requested
+  answer.

package/src/output-styles/simple.md

@@ -9,9 +9,10 @@ aliases: default, concise, handoff
 
 Simple — concise summary.
 
-- Information depth: preserve enough context for the answer to stand on its own.
-  Retain the main answer, the essential causal sequence, strongest supporting
-  evidence, direct consequences, and material scope or caveat. Prefer a complete
-  explanation over compression that makes the reader infer a missing step. Omit
-  only secondary evidence, exhaustive implementation detail, examples, edge
+- Retain: the main answer, the essential causal sequence, the strongest
+  supporting evidence, direct consequences, and any material scope or caveat.
+  Preserve enough context for the answer to stand on its own, and prefer a
+  complete explanation over compression that makes the reader infer a missing
+  step.
+- Omit: secondary evidence, exhaustive implementation detail, examples, edge
   cases, adjacent implications, alternatives, and unrequested remediation.

package/src/rules/lead/01-general.md

@@ -11,4 +11,7 @@
   recoverability.
 - Mid-task: replacement supersedes; addition folds in; status gets a brief
   answer while work continues. After compaction, resume the summary.
+- Auto-compact owns context management: never propose stopping work to the
+  user and never stop on your own judgment; resume and continue the work
+  after every compaction.
 - Final text ends the turn only when done.

package/src/rules/shared/10-tool-workflow.md

@@ -18,3 +18,5 @@
 - Treat failure as new evidence and repeat steps 1–4 only for affected facets.
   Report a blocker when no deterministic next action remains.
 - Use only named tools present in the current tool surface.
+- Deferred tools auto-load on a direct call; when their exact arguments are
+  not visible, call `load_tool` first and read the surfaced schema.