mixdog 0.9.92 → 0.9.93

package/README.md

@@ -4,14 +4,17 @@
 ![node](https://img.shields.io/badge/node-%3E%3D22-brightgreen)
 ![license](https://img.shields.io/badge/license-MIT-blue)
 
-Standalone coding-agent CLI/TUI that runs an orchestrated, multi-provider
-agent workflow from one terminal — built for maximum performance at minimum
-cost.
+Standalone coding agent that runs an orchestrated, multi-provider agent
+workflow from one terminal — or from a full desktop workbench — built to
+get the same quality out of the same models with less time, cost, and
+context.
 
 Mixdog combines an Ink-based terminal UI, per-role model routing across
 providers, workflow agents, MCP/plugin/skill/hook support, lightweight
 memory, web search, channel integrations, and repo-native tools for reading,
-editing, testing, and reviewing code.
+editing, testing, and reviewing code. Mixdog Desktop wraps the same runtime
+in an Electron workbench with editor, git, terminal, file-explorer, media,
+and automation surfaces.
 
 ## Quick start
 
@@ -25,34 +28,59 @@ mixdog
 First run walks you through onboarding: provider auth, model pick, and
 workflow setup.
 
-## Terminal-Bench 2.1 — 89.9% (self-reported)
+## Terminal-Bench 2.1 — controlled full-run comparisons
 
-![Terminal-Bench 2.1 leaderboard with mixdog](https://raw.githubusercontent.com/tribgames/mixdog/main/benchmarks/terminal-bench-2.1/tb21-leaderboard.svg)
+**Same model, same quality — in a fraction of the time, context, and
+cost.** On the same 89 tasks, mixdog scored on par with both native
+harnesses — **78/89** vs Claude Code's **77/89** (within single-run noise)
+and **75/89** matching Codex CLI — while finishing faster, ending leaner,
+and costing less.
 
-Single-run score of **80/89 = 89.9%** (k=1) on
-`terminal-bench/terminal-bench-2-1`, using a cost-reduced per-role routing
-config:
+Each comparison matches the primary model and reasoning level on
+both sides, comparing each product as shipped: mixdog routes scoped
+read-only Explorer lookups to a smaller model, mirroring Claude Code's
+built-in Explore subagent (Haiku 4.5 by default in the 2.1.x baseline).
+Codex CLI ships no equivalent helper — the Sol-led mixdog run used GPT-5.6
+Luna for that scoped Explorer work. Results are self-reported single runs
+(`k=1`, 2026-08), not leaderboard submissions.
 
-| Role                | Model           | Effort      |
-|---------------------|-----------------|-------------|
-| Lead (orchestrator) | Claude Fable 5  | high        |
-| Explorer            | Claude Haiku 4.5| default     |
-| Worker              | Claude Opus 4.8 | medium      |
-| Heavy worker        | Claude Opus 4.8 | high        |
-| Reviewer            | GPT-5.5         | high (fast) |
+#### Claude Opus 5 vs Claude Code
 
-Against published model scores this places second overall — behind GPT-5.6
-Sol Ultra (91.9%), ahead of GPT-5.6 Sol (88.8%) and Claude Mythos 5 (88%),
-and well above the same primary model run standalone (Claude Fable 5,
-84.3%). Measured once (k=1) due to cost, so treat it as indicative — a
-max-effort k=5 run is planned and is expected to land higher. Per-task
-results, raw Harbor jobs, and the harness adapter live in
-`benchmarks/terminal-bench-2.1/`.
+![Terminal-Bench 2.1 comparison of mixdog with Claude Opus 5 and Claude Code](https://raw.githubusercontent.com/tribgames/mixdog/main/benchmarks/terminal-bench-2.1/tb21-opus-vs-claude-code.svg)
 
-Full transparency: 4 tasks refused by the primary model's safety layer were
-re-run routed to Claude Opus 4.8 (effort xhigh) and passed; timeouts and
-task resources were left unmodified. Raw Harbor `result.json`/`config.json`
-for every constituent run are included.
+#### GPT-5.6 Sol xhigh vs Codex CLI
+
+![Terminal-Bench 2.1 comparison of mixdog with GPT-5.6 Sol xhigh and Codex CLI](https://raw.githubusercontent.com/tribgames/mixdog/main/benchmarks/terminal-bench-2.1/tb21-sol-vs-codex.svg)
+
+- Speed: **1.43×** vs Claude Code, **1.27×** vs Codex CLI
+  (baseline elapsed agent time ÷ mixdog elapsed agent time)
+- Final context: **40–47% smaller** at task end (median tokens, measured
+  from both harnesses' session logs)
+- Priced cost: **29% lower** vs Claude Code, **at least 39.7% lower** vs
+  Codex CLI (mixdog $54.50–$58.84 vs Codex's recorded $97.54)
+
+Both sides run their standard single-agent loop. Anthropic cost includes
+measured cache writes; the archived OpenAI runs did not retain
+`cache_write_tokens`, so the Codex cost delta is a lower bound. Raw
+artifacts, the exact run commands, and the metric scripts that recompute
+every number above live under `benchmarks/terminal-bench-2.1/`.
+
+## Boot overhead — measured, not estimated
+
+Fixed cost of the default system prompt + tool schemas, measured with a
+minimal one-turn ping ("hi"). Every number below is what the provider's own
+usage accounting reported for the first API call — same tokenizer on both
+sides of each row, no offline estimates (2026-08).
+
+| Tokenizer | Baseline harness | First-call prompt | mixdog Lead | mixdog worker agent |
+|-----------|------------------|------------------:|------------:|--------------------:|
+| Anthropic — pristine containers, TB2.1 harness | Claude Code 2.1.220 | 19,149 tok (median, n=89) | **7,295 (−62%)** | 4,856 (−75%) |
+| OpenAI — same host, same model (`gpt-5.6-sol`) | Codex CLI 0.145.0 | 17,706 tok | **7,488 (−58%)** | 4,616 (−74%) |
+
+The Lead surface carries the full default rule set, skills manifest, and
+per-user profile; spawning a sub-agent costs a flat ~4.6–4.9k tokens. Low
+boot overhead compounds: it is re-read on every request, so it is the
+baseline of every cache write, compaction, and long-session turn.
 
 ## Why mixdog
 
@@ -68,9 +96,10 @@ for every constituent run are included.
 
 **Any provider**
 
+- Sign in with the subscriptions you already pay for: OAuth device flows for
+  Claude and ChatGPT/Codex accounts work alongside plain API keys.
 - Anthropic, OpenAI, Google/Gemini, xAI/Grok, DeepSeek, OpenCode Go,
-  OAuth-backed providers, OpenAI-compatible APIs, Ollama, and LM Studio/local
-  endpoints.
+  OpenAI-compatible APIs, Ollama, and LM Studio/local endpoints.
 - Live model catalog from provider `/models` endpoints, enriched with
   LiteLLM/models.dev metadata for context windows, output limits, pricing,
   tool support, reasoning, and recency.
@@ -81,20 +110,34 @@ for every constituent run are included.
 - Full-screen TUI with slash commands, provider setup, model/workflow
   pickers, statusline integration, and detailed tool cards — plus headless
   role mode for scripting.
-- Optional Discord/Telegram channels, webhook endpoints, cron schedules, and
-  voice-message transcription for remote/event-driven workflows.
+- Mixdog Desktop: a full agent workbench for Windows/macOS/Linux (see
+  below).
+- Web/mobile companion over relay pairing — scan a QR code to open your
+  running sessions in a phone browser and keep going from any network.
+- Optional Discord/Telegram channels, webhook endpoints, and cron schedules
+  with quiet hours for remote/event-driven workflows; channel voice messages
+  are transcribed locally with a managed Whisper server.
+- First-class Windows support: ConPTY terminals, PowerShell-aware shell
+  profiles, and a one-click desktop installer.
 
 **Memory**
 
-- Lightweight memory restores prior work context across sessions.
-- Important memories are automatically promoted — and demoted when stale.
+- Every session is ingested into a local memory store in the background, so
+  prior work, decisions, and fixes stay recallable across sessions via the
+  `recall` tool and `/memory`.
+- Semantic + lexical recall with local embeddings, time-window queries, and
+  project-scoped pools — multilingual, including Korean morphology.
+- A multi-pass consolidation cycle promotes important memories into a
+  compact core set — and demotes them when stale — so memory stays small
+  and current instead of growing without bound.
 
 **Agent-ecosystem compatible**
 
 - Skills, MCP servers, hooks, and plugins load through standard-compatible
   interfaces.
 - Workflow delegation through the `agent` tool and `/agents`: worker,
-  heavy-worker, reviewer, debugger, maintainer, and explorer roles.
+  heavy-worker, reviewer, debugger, maintainer, explorer, and
+  web-researcher roles.
 
 ## Run
 
@@ -112,20 +155,31 @@ mixdog
 # Start with an explicit route
 mixdog --provider anthropic-oauth --model claude-haiku-4-5-20251001
 
+# Start with a specific workflow active
+mixdog --workflow solo
+
 # Read-only tool surface
 mixdog --readonly
 
 # Enable remote/channel mode for this session
 mixdog --remote
+
+# Re-run the first-run setup wizard
+mixdog --onboarding
 ```
 
-Headless role mode is also supported:
+Headless role mode is also supported. It requires an explicit
+provider/model pair and runs with ephemeral config — host behavioral config
+and personal state are not loaded:
 
 ```bash
-mixdog worker "fix the failing test"
-mixdog reviewer "review the current diff"
+mixdog --provider anthropic-oauth --model claude-opus-5 worker "fix the failing test"
+mixdog --provider openai-oauth --model gpt-5.6-sol reviewer "review the current diff"
 ```
 
+Roles: `explore`, `worker`, `heavy-worker`, `reviewer`, `debugger`,
+`maintainer`, `web-researcher`.
+
 ## TUI basics
 
 Common slash commands:
@@ -135,15 +189,26 @@ Common slash commands:
 /model             choose the main provider/model (/effort, /fast tune it)
 /workflow          choose the active workflow
 /agents            show workflow agents and per-agent model overrides
+/project           switch working directory (project)
+/resume            resume a saved chat
+/usage             show total provider quota / balance
+/context           show the current context surface
+/memory            list and edit core memories
 /setting           open the runtime settings hub
 /mcp               manage MCP servers and tools
 /skills            choose a skill for the next request
 /channels          manage Discord, Telegram, and voice
 /compact           compact older conversation context
+/autoclear         reduce cache-miss cost after long idle gaps
+/theme             change the TUI color theme
 /clear             reset the conversation and screen
 /OutputStyle       show or switch Lead output style
+/update            check version and update mixdog
+/doctor            diagnose installation health
 ```
 
+Run `mixdog --help` for the full command and option reference.
+
 Use `/providers` first if no model is configured, then `/model` to pick the
 route. The model picker warms the provider catalog in the background and keeps
 Claude families such as Opus, Sonnet, Haiku, and Fable separate when filtering
@@ -155,6 +220,44 @@ directory (`workflows/<id>/`, `agents/<id>/`) and are edited on the desktop
 app's Workflows page. Schedules and webhooks are also managed in the desktop
 app.
 
+## Desktop app
+
+Mixdog Desktop (Electron) runs the same runtime as the CLI inside a full
+agent workbench. Installers are published on GitHub Releases (Windows
+one-click NSIS, macOS dmg/zip, Linux AppImage), and a guided onboarding
+wizard covers first-run setup. For development run `npm run dev` inside
+`apps/desktop`.
+
+- **Workbench shell** — VS Code-style activity rail and tab strip,
+  drag-and-drop tabs across pane groups, and unlimited splits that run
+  parallel agent sessions side by side — every pane hosts a live session
+  surface with its own draft and model controls — plus a command surface,
+  bottom panel, and problems view.
+- **Sessions and projects** — project-scoped session lists, resumable
+  sessions with per-pane route restore, live agent-activity indicators, and
+  usage dashboards in the sidebar.
+- **Editor and review** — Monaco editor pane with LSP integration, git and
+  inline diff viewers, and turn-by-turn review of agent edits with approval
+  cards.
+- **Source control** — git dock for staging/commits/branches,
+  auto-generated commit messages, GitHub CLI integration, pull-request
+  browsing, and a dedicated review pane.
+- **File explorer** — Windows-Explorer-grade folder pane: breadcrumbs and
+  path box, ribbon toolbar, places/drives/tree sidebar, grouped grid and
+  details views with shell icons and thumbnails, rubber-band selection,
+  clipboard and OS drag-and-drop, preview pane, and file properties.
+- **Terminal** — integrated terminal tabs on the native shell (ConPTY on
+  Windows) with shell-profile detection, isolated in a worker process so a
+  runaway shell never takes the app down.
+- **Studio** — media studio for image and video generation over
+  authenticated provider lanes, with a persistent local gallery, reference
+  images, and per-model resolution/aspect/duration controls.
+- **Automation** — visual editors for workflow and agent packs, cron
+  schedules, webhooks, and channel integrations.
+- **Settings hub** — provider auth, capability sweep, git identity, and
+  QR device pairing for the web/mobile companion, preloaded so every
+  category opens instantly.
+
 ## Scripts
 
 ```bash
@@ -206,6 +309,10 @@ src/
   agents/        # workflow agent definitions
   workflows/     # workflow definitions
   rules/         # Lead and agent instructions
+apps/
+  desktop/       # Mixdog Desktop — Electron workbench (main/preload/renderer)
+  mobile/        # mobile companion shell
+  relay/         # relay server for remote/web/mobile access
 scripts/
   smoke*.mjs     # smoke checks
   *test.mjs      # focused node:test checks
@@ -216,20 +323,9 @@ vendor/
 
 ## Published package contents
 
-The npm package is limited by `package.json#files` to:
-
-```text
-README.md
-scripts/
-src/
-vendor/
-```
-
-Test, smoke, and bench scripts are excluded from the tarball; they live in the
-repository only.
-
-`docs/` is not included in the published package unless `package.json#files` is
-changed.
+The npm tarball ships `README.md`, `src/`, `vendor/`, and runtime `scripts/`
+only (`package.json#files`); tests, smokes, benches, and `docs/` stay in the
+repository.
 
 ## License
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mixdog",
-  "version": "0.9.92",
+  "version": "0.9.93",
   "private": false,
   "type": "module",
   "description": "Standalone mixdog coding-agent CLI/TUI workspace.",
@@ -66,6 +66,7 @@
     "test:agent-job-views": "node --test scripts/agent-job-terminal-view-test.mjs",
     "test:agent-fanout": "node scripts/agent-parallel-smoke.mjs && node --test scripts/agent-route-batch-test.mjs scripts/execution-completion-dedup-test.mjs",
     "test:toolcall": "node --test scripts/toolcall-args-test.mjs",
+    "test:tool-batching": "node --test scripts/eager-patch-shell-order-test.mjs scripts/session-bench-batching-test.mjs",
     "test:shipmode": "node --test scripts/ship-mode-test.mjs",
     "test:shellhardening": "node --test scripts/shell-hardening-test.mjs scripts/shell-failure-diagnostics-test.mjs scripts/windows-hide-spawn-options-test.mjs",
     "test:placeholder": "node --test scripts/compacted-placeholder-scrub-test.mjs",
@@ -85,10 +86,10 @@
     "test:code-graph-clean-cache": "node --test scripts/code-graph-dispatch-test.mjs",
     "test:tui-queue": "node --test scripts/submit-commandbusy-race-test.mjs scripts/steering-drain-buckets-test.mjs scripts/abort-recovery-test.mjs scripts/execution-pending-resume-kick-test.mjs scripts/execution-resume-esc-integration-test.mjs scripts/pending-stale-injection-test.mjs",
     "test:tui-input-render": "node --test scripts/prompt-immediate-render-test.mjs",
-    "test:tui-streaming-window": "node --test scripts/streaming-tail-window-test.mjs",
+    "test:tui-streaming-window": "node --test scripts/streaming-tail-window-test.mjs scripts/tui-store-frame-batch-test.mjs && node scripts/tui-transcript-jitter-harness.mjs",
     "test:tui-ambiguous-width": "node --test scripts/tui-ambiguous-width-test.mjs",
     "test:release-assets": "node --check scripts/verify-release-assets.mjs && node --check scripts/verify-release-assets-test.mjs && node --check scripts/deploy-workflow-test.mjs && node --test scripts/verify-release-assets-test.mjs scripts/deploy-workflow-test.mjs",
-    "test:release-focused": "npm run test:release-assets && npm run test:tool-contracts && npm run test:placeholder && npm run smoke:patch && npm run test:patch-binary-cache && npm run test:providers && npm run test:deferred-tools && npm run smoke:compact && npm run test:compact && npm run test:context && node --test scripts/code-graph-root-federation-test.mjs scripts/code-graph-aggregate-cwd-test.mjs && npm run test:code-graph-dispatch && node --test scripts/code-graph-disk-hit-test.mjs && npm run test:shellhardening && npm run test:project-registry && npm run test:session && npm run test:workflow-editor && npm run test:embedding-runtime && node --test scripts/tui-transcript-perf-test.mjs",
+    "test:release-focused": "npm run test:release-assets && npm run test:tool-contracts && npm run test:tool-batching && npm run test:placeholder && npm run smoke:patch && npm run test:patch-binary-cache && npm run test:providers && npm run test:deferred-tools && npm run smoke:compact && npm run test:compact && npm run test:context && node --test scripts/code-graph-root-federation-test.mjs scripts/code-graph-aggregate-cwd-test.mjs && npm run test:code-graph-dispatch && node --test scripts/code-graph-disk-hit-test.mjs && npm run test:shellhardening && npm run test:project-registry && npm run test:session && npm run test:workflow-editor && npm run test:embedding-runtime && node --test scripts/tui-transcript-perf-test.mjs",
     "test:native-edit-wire": "node --test scripts/native-edit-wire-test.mjs",
     "test:patch-binary-cache": "node --test scripts/patch-binary-cache-test.mjs",
     "test:patch-parity": "node --test scripts/v4a-codex-parity-test.mjs",
@@ -150,10 +151,10 @@
   },
   "overrides": {
     "discord.js": {
-      "undici": "6.27.0"
+      "undici": "6.28.0"
     },
     "@discordjs/rest": {
-      "undici": "6.27.0"
+      "undici": "6.28.0"
     },
     "onnxruntime-node": {
       "global-agent": "^4.1.3"

package/scripts/code-graph-description-contract.mjs

@@ -85,7 +85,7 @@ const CODE_GRAPH_DESCRIPTION_MUTATION_CORPUS = [
     allPositiveProbes: true,
     mutate: (parts) => ({
       ...parts,
-      description: parts.description.replace(/keywords\s+(?:use|route through|select)/i, "keywords won't use"),
+      description: parts.description.replace(/keywords\s+(?:use|route through|select|via)/i, "keywords won't use"),
     }),
   },
   {
@@ -101,7 +101,7 @@ const CODE_GRAPH_DESCRIPTION_MUTATION_CORPUS = [
     allPositiveProbes: true,
     mutate: (parts) => ({
       ...parts,
-      symbolsDescription: parts.symbolsDescription.replace(/one symbols\[\] call/i, 'one files[] call'),
+      symbolsDescription: parts.symbolsDescription.replace(/one symbols\[\] call|one symbols\[\]/i, 'one files[] call'),
     }),
   },
   {
@@ -143,18 +143,16 @@ const CODE_GRAPH_DESCRIPTION_MUTATION_CORPUS = [
 function hasCodeGraphDescriptionContract({ description, modeDescription, symbolsDescription }) {
   return (
     hasPositiveClause(description, ['file modes', 'files[]'])
-    && hasModeClause(description, 'symbol modes', SYMBOL_MODES, 'symbols[]')
+    && hasPositiveClause(description, ['symbol modes', 'symbols[]'])
     && hasPositiveClause(description, ['exact identifiers', ...EXACT_MODES])
     && hasPositiveClause(description, ['keywords', ...KEYWORD_MODES])
     && !hasContradictoryTargetAssignment(description)
     && hasModeClause(modeDescription, 'file modes', FILE_MODES)
     && hasPositiveClause(modeDescription, ['symbols with files', 'files[]', 'file outline'])
-    && hasModeClause(modeDescription, 'symbol modes', SYMBOL_MODES)
-    && hasPositiveClause(modeDescription, ['fileless symbols', 'symbol_search', 'keywords'])
     && !hasContradictoryTargetAssignment(modeDescription)
-    && hasPositiveClause(symbolsDescription, ['exact identifiers', ...EXACT_MODES])
-    && hasPositiveClause(symbolsDescription, ['keywords', ...KEYWORD_MODES])
-    && hasPositiveClause(symbolsDescription, ['multiple exact symbols', 'one symbols[] call'])
+    && hasPositiveClause(symbolsDescription, ['exact identifiers'])
+    && hasPositiveClause(symbolsDescription, ['keywords'])
+    && hasPositiveClause(symbolsDescription, ['symbols[]'])
     && !hasContradictoryTargetAssignment(symbolsDescription)
   );
 }

package/scripts/tmp-cdp-errors.mjs

@@ -0,0 +1,41 @@
+import WebSocket from "ws";
+
+const list = await fetch("http://127.0.0.1:9342/json/list").then((r) => r.json());
+const page = list.find((t) => t.type === "page");
+const ws = new WebSocket(page.webSocketDebuggerUrl);
+let id = 0; const pending = new Map();
+function send(method, params) {
+  return new Promise((res) => { const mid = ++id; pending.set(mid, res); ws.send(JSON.stringify({ id: mid, method, params })); });
+}
+const seen = new Map();
+ws.on("message", (raw) => {
+  const msg = JSON.parse(raw);
+  if (msg.id && pending.has(msg.id)) { pending.get(msg.id)(msg.result); pending.delete(msg.id); return; }
+  if (msg.method === "Runtime.exceptionThrown") {
+    const d = msg.params.exceptionDetails;
+    const desc = d.exception?.description || d.text || "";
+    const key = desc.slice(0, 200);
+    seen.set(key, (seen.get(key) || 0) + 1);
+  } else if (msg.method === "Runtime.consoleAPICalled" && (msg.params.type === "error" || msg.params.type === "warning")) {
+    const text = msg.params.args.map((a) => a.value ?? a.description ?? "").join(" ");
+    const key = "[console." + msg.params.type + "] " + text.slice(0, 300);
+    seen.set(key, (seen.get(key) || 0) + 1);
+  } else if (msg.method === "Log.entryAdded") {
+    const e = msg.params.entry;
+    if (e.level === "error" || e.level === "warning") {
+      const key = "[log." + e.level + "] " + (e.text || "").slice(0, 300) + " @" + (e.url || "");
+      seen.set(key, (seen.get(key) || 0) + 1);
+    }
+  }
+});
+ws.on("open", async () => {
+  await send("Runtime.enable", {});
+  await send("Log.enable", {});
+  setTimeout(() => {
+    for (const [key, count] of [...seen.entries()].sort((a, b) => b[1] - a[1])) {
+      console.log(`x${count}  ${key.replace(/\n/g, " | ")}`);
+    }
+    if (!seen.size) console.log("(no errors captured in window)");
+    process.exit(0);
+  }, 12000);
+});

package/scripts/tmp-cdp-inspect.mjs

@@ -0,0 +1,41 @@
+import WebSocket from "ws";
+
+const list = await fetch("http://127.0.0.1:9342/json/list").then((r) => r.json());
+const page = list.find((t) => t.type === "page");
+if (!page) { console.error("no page target"); process.exit(1); }
+const ws = new WebSocket(page.webSocketDebuggerUrl);
+let id = 0; const pending = new Map();
+function send(method, params) {
+  return new Promise((res) => { const mid = ++id; pending.set(mid, res); ws.send(JSON.stringify({ id: mid, method, params })); });
+}
+ws.on("message", (raw) => {
+  const msg = JSON.parse(raw);
+  if (msg.id && pending.has(msg.id)) { pending.get(msg.id)(msg.result); pending.delete(msg.id); }
+});
+ws.on("open", async () => {
+  const expr = `(() => {
+    const gate = document.querySelector('.desktop-boot-gate');
+    const covers = [...document.querySelectorAll('.pane-surface-cover, .desktop-boot-cover')].map(c => ({
+      cls: c.className,
+      rect: Math.round(c.getBoundingClientRect().width) + 'x' + Math.round(c.getBoundingClientRect().height),
+      parent: (c.parentElement?.className || '').slice(0, 90),
+      hasSpinner: !!c.querySelector('.desktop-loading-spinner'),
+    }));
+    const gates = [...document.querySelectorAll('.pane-surface-gate[data-ready="false"], .stable-surface-switch[data-ready="false"], .stable-content-swap[data-ready="false"]')]
+      .map(g => g.className + ' | parent=' + (g.parentElement?.className || '').slice(0, 70));
+    return JSON.stringify({
+      bootGate: gate ? { ready: gate.dataset.ready, pending: gate.dataset.pending, timeout: gate.dataset.timeout } : null,
+      revealed: window.__mixdogDesktopRevealed,
+      shown: window.__mixdogWindowShown,
+      covers, gates,
+      metrics: (window.__mixdogBootMetrics || []).slice(-30),
+      recovery: !!document.querySelector('.desktop-recovery-screen'),
+      rootKids: document.getElementById('root')?.childElementCount,
+      url: location.href,
+    }, null, 1);
+  })()`;
+  const r = await send("Runtime.evaluate", { expression: expr, returnByValue: true });
+  console.log(r?.result?.value ?? JSON.stringify(r));
+  process.exit(0);
+});
+setTimeout(() => { console.error("timeout"); process.exit(1); }, 8000);

package/scripts/tui-transcript-jitter-harness.mjs

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
 import { build } from 'esbuild';
-import { readFile, rm } from 'node:fs/promises';
+import { rm } from 'node:fs/promises';
 import { dirname, join } from 'node:path';
 import { fileURLToPath, pathToFileURL } from 'node:url';
 
@@ -18,22 +18,6 @@ const inkAlias = {
   },
 };
 
-// Build-only probe: records the exact delta seen by the production helper
-// without adding another stateful estimator call from the harness.
-const growthProbe = {
-  name: 'streaming-tail-growth-probe',
-  setup(ctx) {
-    ctx.onLoad({ filter: /transcript-window\.mjs$/ }, async (args) => {
-      let source = (await readFile(args.path, 'utf8')).replace(/\r\n/g, '\n');
-      const target = '  return { tailRows: idEntry.rows, delta };';
-      const replacement = `  globalThis.__mixdogTailGrowthProbe = { live, baseline, delta, tailRows: idEntry.rows };\n${target}`;
-      if (!source.includes(target)) return null;
-      source = source.replace(target, replacement);
-      return { contents: source, loader: 'js' };
-    });
-  },
-};
-
 try {
   await build({
     entryPoints: [entry],
@@ -44,7 +28,7 @@ try {
     target: 'node22',
     jsx: 'automatic',
     packages: 'external',
-    plugins: [inkAlias, growthProbe],
+    plugins: [inkAlias],
     banner: {
       js: "import { createRequire as __mixdogCreateRequire } from 'node:module';\nconst require = __mixdogCreateRequire(import.meta.url);",
     },

package/src/rules/agent/00-core.md

@@ -1,7 +1,6 @@
 # Agent Constraints
 
-- Agent communication is English. One turn is one batch; include every
-  compatible read-only call.
+- Agent communication is English.
 - Call tools immediately: no preamble/progress; text only in final handoff.
 - Final handoff is fragments: outcome, key `file:line`, verification
   command+result, material risk/blocker. Never repeat the brief, process,

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

@@ -8,6 +8,7 @@
   validated target paths — never `~`, a root, or unresolved variables/globs;
   report material deletions with recoverability.
 - Act proactively; ask only for decisions.
+- Build only what the task requires; trust internal and framework guarantees.
 - Mid-task input: a replacement supersedes current work, an addition folds
   into it, a status question gets a brief answer while work continues; after
   context compaction continue from the summary — never restart or redo

package/src/rules/shared/01-tool.md

@@ -1,22 +1,18 @@
 # Tool Use
 
-- Before the first call, gather every known facet in one tool message; one
-  shortest route per facet: broad/uncertain→`explore` (roles without it:
-  `find`); partial path/name→`find`; verified root+wildcard→`glob`;
-  quoted/non-identifier literal or regex→`grep`; exact code identifier/
-  relation→`code_graph` before grep; known file/span→`read` directly without
-  `grep`; verified directory→`list`; known edit→`apply_patch` (span already
-  seen; else `read`/`grep` first); program/state change→`shell`; web/current
-  external info→`search`.
-- Shortest total calls, maximum batching — every turn: all independent calls
-  in one concurrent message (shell included); combine variants/symbols/
-  scopes/paths/queries per call; same-file regions as one real
-  `{path,offset,limit}` array; graph targets as arrays; `explore` facets in
-  one `query[]` (max 8, no rephrased duplicates); all new edits in one patch.
-  Distinct facets, not alternative routes; sequential singles only for a
-  step whose arguments require the previous result — fixed follow-ups
-  (pinned installs, known writes) go in the same batch; only apply_patch
-  executes in order.
+- Before the first call, gather every known facet — environment, capability,
+  artifact, and failure checks — in one bounded tool message. Use one shortest
+  route per facet: broad/uncertain→`explore` (roles without it: `find`);
+  partial path/name→`find`; verified root+wildcard→`glob`;
+  quoted/non-identifier literal or regex→`grep`; exact code
+  identifier/relation→`code_graph` before grep; known file/span→`read`
+  directly without `grep`; verified directory→`list`; known
+  edit→`apply_patch`; program/state change→`shell`; web/current external
+  info→`search`.
+- Shortest total calls, maximum batching — every turn: every determined call
+  in one concurrent message (mix `shell` in), merged per tool — one `shell`
+  chain, one `read`, one `apply_patch` carrying full verification in
+  `post_shell`; a later turn only for steps needing unseen output.
 - Verified paths: project root, session cwd, user-provided, tool-returned.
   `find` first for guessed path/name fragments; on ENOENT, find the basename.
   Retry `EXPLORATION_FAILED` once with changed tokens.
@@ -24,11 +20,12 @@
   nonzero `content_with_context` result is final — act on it (inspecting it
   via read/code_graph is valid); only zero/error results justify changed
   tokens or scope. Don't re-locate, re-verify, or reread returned spans.
-- `apply_patch` is the primary edit tool: send the patch as soon as target
-  path and new content are known. Hunk context comes verbatim from the newest
-  tool output of that span (`read`/`grep`/your own patch — post-patch content
-  after edits), never retyped from memory; one look-up beats a failed patch.
-  A same-turn shell after `apply_patch` runs once the patch lands.
+- Verify changes in proportion to risk with one decisive batched boundary
+  probe. A pass is final; on failure, fix and rerun only what failed. Keep
+  optional diagnostics non-fatal; report verified vs assumed.
+- `apply_patch` is the primary edit tool: once target path and new content are
+  known, include the patch in the current tool batch, hunk context verbatim
+  from the newest tool output of that span (post-patch content after edits).
 - After starting or receiving a background task, end the turn — its
   completion notification resumes the work. Never poll, sleep-loop, or block;
   explicit wait only for a result the current turn cannot proceed without.

package/src/runtime/agent/orchestrator/agent-runtime/cache-strategy.mjs

@@ -120,18 +120,31 @@ export function resolveCacheStrategy(agent, { autoClear } = {}) {
     if (isOneShotMaintenanceAgent(agent)) {
         return { tools: 'none', system: 'none', tier3: 'none', messages: 'none' };
     }
+    // Operator override for the BP4 (messages-tail) TTL. Short-lived
+    // rapid-turn deployments (bench-style: session dies in <15min) never
+    // benefit from a 1h tail across its premium window, so '5m' trades the
+    // 2x write premium (1h, $10/M) down to 1.25x ($6.25/M) and deactivates
+    // the 1h volatile-content anchor guard. Product defaults below stay
+    // untouched when the env is unset.
+    const envMessagesTtl = (process.env.MIXDOG_CACHE_MESSAGES_TTL || '').trim();
+    const applyEnv = (strategy) => {
+        if (envMessagesTtl === '1h' || envMessagesTtl === '5m' || envMessagesTtl === 'none') {
+            return { ...strategy, messages: envMessagesTtl };
+        }
+        return strategy;
+    };
     if (getHiddenAgent(agent)) {
-        return { tools: 'none', system: '1h', tier3: '1h', messages: '1h' };
+        return applyEnv({ tools: 'none', system: '1h', tier3: '1h', messages: '1h' });
     }
     if (agent && agent !== 'lead') {
         // Public (non-hidden, non-lead) agents keep the flat 1h tail — only
         // the Lead session's tail is linked to autoClear.
-        return { tools: 'none', system: '1h', tier3: '1h', messages: '1h' };
+        return applyEnv({ tools: 'none', system: '1h', tier3: '1h', messages: '1h' });
     }
     // Lead session (agent === 'lead', or no agent — raw/CLI callers default
     // to Lead behavior): message tail TTL is linked to autoClear (see
     // resolveLeadMessagesTtl).
-    return { tools: 'none', system: '1h', tier3: '1h', messages: resolveLeadMessagesTtl(autoClear) };
+    return applyEnv({ tools: 'none', system: '1h', tier3: '1h', messages: resolveLeadMessagesTtl(autoClear) });
 }
 
 /**