tesseract-agent 0.11.15 → 0.11.16

package/dist-bundle/cli.js

@@ -24422,6 +24422,13 @@ function schemaFieldRejected(status2, body) {
 function anthropicBaseUrl(env) {
   return env?.ANTHROPIC_BASE_URL?.trim() || process.env.ANTHROPIC_BASE_URL?.trim() || "https://api.anthropic.com";
 }
+function isAnthropicEndpoint(env) {
+  try {
+    return /(^|\.)api\.anthropic\.com$/i.test(new URL(anthropicBaseUrl(env)).hostname);
+  } catch {
+    return false;
+  }
+}
 var OAUTH_BETA_HEADER;
 var init_model_json = __esm({
   "../../packages/shared/dist/model-json.js"() {
@@ -24974,15 +24981,12 @@ function isAnthropicModelId(model) {
   const id = (model ?? "").trim().toLowerCase();
   if (!id)
     return false;
-  if (TIER_ALIAS.test(id))
-    return false;
   return /(^|[/.\-])(claude|opus|sonnet|haiku|fable)([/.\-]|$)/.test(id) || /^(opus|sonnet|haiku|fable)\b/.test(id);
 }
-var TIER_ALIAS, ONE_MILLION_FAMILIES, SUFFIX;
+var ONE_MILLION_FAMILIES, SUFFIX;
 var init_model_long_context = __esm({
   "../../packages/shared/dist/model-long-context.js"() {
     "use strict";
-    TIER_ALIAS = /^tsr-/i;
     ONE_MILLION_FAMILIES = [
       /(^|[/.])claude-sonnet-(4-6|4-7|4-8|5)\b/i,
       /(^|[/.])claude-opus-(4-6|4-7|4-8|5)\b/i,
@@ -26626,10 +26630,10 @@ function providerBase(key2, which2) {
     return fromServer;
   return which2 === "anthropic" ? gatewayAnthropicBaseUrl(key2.baseUrl) : which2 === "openai" ? gatewayOpenAiBaseUrl(key2.baseUrl) : gatewayGeminiBaseUrl(key2.baseUrl);
 }
-function aliasProviders(key2) {
+function slotProviders(key2) {
   const out = /* @__PURE__ */ new Map();
   for (const slot of Object.keys(TIER_PROVIDER)) {
-    const id = key2.models[slot];
+    const id = key2.models[slot]?.trim();
     if (id)
       out.set(id, TIER_PROVIDER[slot]);
   }
@@ -26638,25 +26642,26 @@ function aliasProviders(key2) {
 function tiers(engineId, key2, models) {
   const provider = gatewayProviderForEngine(engineId);
   const preference = provider ? TIER_PREFERENCE[provider] : void 0;
-  const byAlias = aliasProviders(key2);
+  const byModel = slotProviders(key2);
   const asked = (v) => {
     const t = v?.trim();
     if (!t)
       return void 0;
     if (t.includes("/"))
       return t;
-    return byAlias.get(t) === provider ? t : void 0;
+    return byModel.get(t) === provider ? t : void 0;
   };
   const pick3 = (slots) => {
     for (const slot of slots) {
-      const id = key2.models[slot];
+      const id = key2.models[slot]?.trim();
       if (id)
         return id;
     }
     return void 0;
   };
-  const main2 = asked(models.main) || preference && pick3(preference.main);
-  const fast = asked(models.fast) || preference && pick3(preference.fast) || main2;
+  const serverDefault = asked(key2.defaults?.model);
+  const main2 = asked(models.main) || serverDefault || preference && pick3(preference.main);
+  const fast = asked(models.fast) || serverDefault || preference && pick3(preference.fast) || main2;
   return { main: main2 || void 0, fast: fast || void 0 };
 }
 function planGatewayRun(engineId, key2, models, homeBase) {
@@ -26698,7 +26703,7 @@ var init_gateway_plan = __esm({
     };
     TIER_PREFERENCE = {
       // No fast rung of its own on the OpenAI side: Codex is one model plus a reasoning dial,
-      // and `tsr-low` remains the cheap fallback for anything that asks for one.
+      // and the nano model remains the cheap fallback for anything that asks for one.
       openai: { main: ["codex", "gpt", "high"], fast: ["low", "high"] },
       anthropic: { main: ["claude"], fast: ["claudeFast", "claude"] },
       "google-ai": { main: ["vision"], fast: ["vision"] }
@@ -33285,7 +33290,7 @@ var CAPABILITIES_DOC;
 var init_capabilities_doc_generated = __esm({
   "../../packages/agent-core/dist/capabilities-doc.generated.js"() {
     "use strict";
-    CAPABILITIES_DOC = '# Tesseract \u2014 what the product can do\n\nOne-page map of **every user-reachable capability**, derived from the 80 area briefs in\n[`functional-testing/`](functional-testing/). It is the source for the `tesseract_capabilities`\ntool, so the agent can answer "what can you do?" about *Tesseract*, not just about its own toolbelt.\n\n> Read this before telling a user something is or isn\'t possible. \xA718 lists what is deliberately\n> **not** built \u2014 promising any of it is the most common way to be wrong about this product.\n\n---\n\n## 1. What Tesseract is\n\nA desktop-first AI workspace: an always-on agent that can drive your machine, your browser and your\nfiles; plus a fleet layer that makes every machine you own reachable from any of them; plus a set of\ncreative, communication and security products built on the same transport. It ships as **eight\nseparate clients** and is not a single app.\n\n## 2. Clients (each a distinct product)\n\n| Client | What it is |\n|---|---|\n| **Desktop** (Electron, mac/Win/Linux) | The full product. Tab workspace \u2014 **no router**: a recursive split-pane tree of tab groups, **24 tab kinds**, **10 dockable panels** (dock left/right/bottom, float, detach), multi-window, native menu, command + search palettes (**31 stable command ids**), \u23181\u20130 tab assign/recall, per-machine keybinding overrides, memory governor + layout restore. |\n| **Web** (`dist-web`) | Same React renderer over a relay transport. **No local agent, no local engine** \u2014 it always runs on a remote host. Terminals are all remote and pinned to the opening host. Documented stubs: undo-delete, engines, runner configs, window resize; voice keys sit in `localStorage`. Always auth-gated. |\n| **Mobile** (Flutter, iOS/Android/macOS) | A **roaming client** \u2014 it never hosts an agent. Chat with streaming, team chat, present-follower, full password vault (incl. TOTP + native autofill), realtime voice, rdesk viewer, clipboard sync, push. Uniquely: **phone-as-host** (serve your photo library and files to your desktop) and a create-video screen. |\n| **CLI** `tesseract-agent` (npm) | Headless daemon, ~18 commands: `login/logout/whoami`, `run`, service `install/start/stop/restart/status/logs`, `link add`, `runner \u2026`, `net \u2026`, self-update. `--instance <id>` namespaces state dir, service name and device id. Docker-friendly (`login --token` then `run` as PID 1). |\n| **Browser extension** (Chrome MV3) | Claude sidebar (sessions, slash commands, todos, sources, DOM tree, run modes), element picker, vault/autofill popup with biometrics, QR pairing, fullpage + plan views. Talks to the daemon through a native-messaging companion host. |\n| **Remote Desktop** (`apps/rdesk`) | Slim standalone binary, ~\xBC the size \u2014 ships **no AI/browser payload**. Server/client/chooser roles, coexists with the full daemon (demotes itself to client rather than quitting). |\n| **Mesh** (`apps/mesh`) | Slim standalone binary for the private network only \u2014 "moves packets, not pixels". Create a mesh, enrol, list peers. No role chooser by design. |\n| **Thin installer** (Flutter, ~20\u201340 MB) | The default download. Resolves the release feed at run time, picks the arch-correct asset, streams + sha512-verifies, installs, launches. Version-agnostic. |\n\nPlus the **admin/operations console** (`admin.tesseract.dev`: gateway credentials, users and keys,\nerror inspector, remediation trigger, audit log) and the **landing site** with its waitlist.\n\n## 3. The agent\n\n- **Turn lifecycle**: stream, interrupt, rapid-follow-up auto-cancel (first seconds only), stall\n  watchdog, orphan-turn recovery, token compaction, usage-limit handling, cost ledger per session.\n- **Run modes**: **5 permission modes \xD7 6 effort levels**, thinking budget, per-turn engine/model pick.\n  Default permission mode is `bypassPermissions`.\n- **Interactive cards** (~9): permission request, plan review, elicitation, confirm, verification,\n  user dialog, skill UI, archive password, unattended report. "Deny & stop" also aborts the turn.\n- **Block renderers** (~9): mission, subagent, report, sources, provenance, event, notice, surface,\n  image grid \u2014 plus tool-run cards joined at render time with 6/8-line clamping.\n- **Composer** (~25 actions): attach, slash commands, @-mention, reverse search, screen-region\n  capture, element picker, voice, undo-last-turn, queued-message reorder.\n- **Sessions**: list, search (one host-side search across every engine\'s stored conversations),\n  resume, summarize, **fork**, merge, delete-message, keep-alive, focus; session mirror + transfer;\n  share a conversation by unguessable capability token.\n- **Subsystems**: reinforced verification, plan splitter, narrator (live status strip, capped at 24\n  model calls/turn), hook bus (9 events), option ratings, unattended mode.\n- **Also**: a playable **chess** surface (5 tools) and chat slash toys (`/shrug`, `/tableflip`, \u2026).\n\n## 4. Engines, runners, gateway\n\n- **4 engines \xD7 3 auth modes** (`oauth` / `apiKey` / `localCli`), a 24-capability engine matrix,\n  engine + model swap mid-session (model/permission-mode reconcile live; effort/cwd/MCP/system-prompt\n  force a respawn).\n- **Runner config pool** with rotation, cooldown and failover; auto-retry is bounded and only fires\n  when the turn did no work.\n- **Provider gateway** \u2014 a credential-injecting reverse proxy, **19 routes** across AGENT (anthropic,\n  openai, google-ai, groq, xai, deepseek, moonshot, openrouter, custom-openai, customer-litellm,\n  azure-openai), EMBEDDING, SEARCH (serpapi), VIDEO (heygen, gemini), VOICE, MEDIA (klipy, unsplash,\n  pexels). Virtual keys `tsk_live_*`, budgets, LiteLLM price map, spend ledger, credential pools with\n  health-based failover that never crosses a category. Model aliases: `tsr-low`, `tsr-high`,\n  `tsr-vision`, `tsr-claude`, `tsr-claude-fast`, `tsr-image` ($0.10/image flat).\n- **Framework cross-product gate**: the user\'s self-updating `claude` CLI \xD7 the pinned Agent SDK,\n  verified nightly across pinned CLI versions.\n\n## 5. Files, editors, viewers\n\n- **Finder-like browser**: list/grid/columns, breadcrumbs, bookmarks + OS favourites, **host\n  switcher** (every op can be remote), new/rename (F2)/move/duplicate/cut-copy-paste/delete + undo,\n  Get Info, send-to-machine, find-in-files, native drag in **and out** (macOS `NSFilePromiseProvider`\n  \u2014 drag a not-yet-downloaded remote file to Finder).\n- **~20 typed viewers/editors**: Monaco source editor, PDF (view + edit), Word/docx, spreadsheets,\n  presentations, markdown, data grid, hex, audio, video, font, HTML canvas, diff, info.\n- **Archives as folders**: ~20 formats (zip, 7z, tar, gz/bz2/xz/zst, rar, iso, wim, cab, cpio, lzh\u2026)\n  browsed, extracted and **edited in place** from the same tree. `rar`/`iso`/`cab` are read-only.\n  Encrypted archives prompt for a password and can remember it.\n- **Image editor**: layered canvas, 10 tools (marquee, lasso, move, crop, paint, eraser, fill,\n  eyedropper, text, zoom), 25 rail controls, **200+ AI effect styles**, layers, adjust/resize, markup,\n  and 3 AI ops (`prompt`/`inpaint`/`outpaint`), transparent-background output, animate\u2192video.\n- **Video editor**: EDL timeline (cut/trim/reorder/transitions/speed/crop/reframe/flip/reverse/colour\n  effects/audio tracks), burned-in captions, **18 MCP tools** \u2014 probe, render, frame extract, silence\n  detect, transcribe with diarization, extract/isolate audio, music generation, voice list/clone/TTS,\n  avatars, caption export, effects, upscale (ffmpeg + local Real-ESRGAN free; fal-Topaz paid).\n- **Undo & time travel**: agent-driven **file rewind** (undo what a turn wrote), file-tree \u2318Z with a\n  per-window trash, workspace snapshots and git checkpoints, canvas/design version history with\n  diff-and-restore, and message deletion with uniform semantics across engines.\n- **Native inspectors**: **Flutter** widget inspector (tree, on-device tap-to-pick, properties,\n  deterministic Dart source edits, hot reload) and **React Native** (source binding + AST edits).\n\n## 6. Create, Design Studio, Stage, Publish\n\n- **Create** \u2014 5 generative document formats (incl. infographic and poster), ~90 actions: outline\n  ops, 10 insert kinds, card generation, AI layouts, export.\n- **Design Studio** \u2014 a 10-stage wizard (pick \u2192 launch/browse \u2192 preview \u2192 target \u2192 inspect \u2192 brief \u2192\n  build \u2192 QA \u2192 deploy) over **254 templates across 19 industry verticals**, Figma import, in-studio\n  theme/token editing.\n- **Stage** \u2014 presentation engine: decks, slides, transitions, **slide audio** (narration/music/sfx\n  with music ducking under narration), a live assistant, presenter mode, zoom/blank. 8 present tools\n  let the agent drive the deck; followers converge from absolute-state snapshots.\n- **Export**: PowerPoint (.pptx), deck and card export, PNG.\n- **Publish** \u2014 private/org/public visibility, server-rendered public pages on a 192-bit token, live\n  data sources (SSRF-guarded), embeds; **community gallery** with explore/following/team/mine, like,\n  follow and clone.\n- **Deploy** \u2014 one-click hosting to `*.tesseract.dev` in per-user hardened containers (cap-drop ALL,\n  no-new-privileges, read-only, egress-locked), metered on compute/storage/egress. One of the few\n  genuinely entitlement-gated features. A separate **per-user cloud agent** container is provisioned\n  on demand so you have a machine in the fleet even with every laptop closed.\n\n## 7. Generative media\n\n| What | Notes |\n|---|---|\n| **Images** | 2 providers (OpenAI gpt-image-2, Google). Generate (4-up by default), edit, inpaint/outpaint, transparent background (inpaint + chroma-key). **$0.10/image**. Ungated, BYO key. |\n| **Stock images** | AI-ranked Unsplash/Pexels search as its own tab. Free tier, rate-limited. |\n| **Video** | Google Veo 3.1 \u2014 4/6/8s clips, 720p/1080p, 16:9 or 9:16, keyframe conditioning (first/last frame), native audio. Free on the bundled tier; otherwise cost-confirmed. Multi-scene storyboards chain clip-to-clip for longer videos. |\n| **Avatar video** | HeyGen talking presenter from a script. |\n| **Music / speech** | ElevenLabs music generation, TTS, and instant **voice cloning**. |\n| **Fonts & brand** | Prompt \u2192 AI glyph sheets \u2192 **vision QC** \u2192 vectorize \u2192 kerning/spacing \u2192 a real installable font. Plus logo/wordmark candidates, SVG lockups and a brand kit. Assembly/vectorize/build are local and free. |\n| **Trending effects** | One-tap video effects (hug/kiss, squish, bring-photo-to-life, age, anime restyle, 3D figurine, outfit change, consent-gated face-swap\u2026). |\n\n## 8. Voice\n\n- **3 realtime providers**: OpenAI Realtime, Gemini Live, ElevenLabs ConvAI \u2014 plus a non-realtime\n  **STT\u2192LLM\u2192TTS pipeline cascade**. Audio goes **client\u2194provider direct**; the daemon only mints the\n  ephemeral credential and runs the control loop.\n- Group calls, live captions, and the **\u2325Space voice widget**.\n- **Voice as a control modality**: the realtime model is the *operator*, Claude Code is the *worker*.\n  It decomposes a spoken request into one small dispatched step at a time, all to the same agent\n  session, and is explicitly forbidden from claiming an outcome the agent did not report \u2014 when\n  unsure it dispatches a check. Its own screen view is a deliberately low-res ~1 fps grab.\n\n## 9. Computer control\n\n- **Two distinct MCP servers**: `tesseract-computer` (**26 tools** \u2014 raw control: move/click/drag/\n  scroll/key/type, see_screen, read_text, read_ui (accessibility), clipboard, windows, apps, open\n  url/path, notify, annotate, draw) and `tesseract-screen` (**15 tools** \u2014 a guided\n  look \u2192 figure_out_next \u2192 perform_step \u2192 **verify** loop that can fail and retry).\n- Plus `tesseract-os-control`: processes, network, structured fs, AX tree + AX actions, preferred app\n  and usage profile, screen overlays (arrow/bubble/highlight/text/popup/choice), and\n  `request_user_input` / `confirm_action`.\n- **Click grounding** is AX-tree + a local Moondream vision sidecar (free); screenshots are\n  Retina-aware and downscaled to point coordinates so clicks line up.\n- **Off by default** \u2014 `computerControlEnabled` must be turned on.\n- **\u2325Space co-pilot HUD** + **13 daemon-owned windows** (spotlight, pastebar, publisher, annotate,\n  draw, overlay, input, choice, consent, control-bar, ai-cursor, rdesk call/client), the tray, screen-\n  region capture with multi-display number flash, and `tesseract://` deep links. The HUD is\n  deliberately **invisible to the AI\'s own screenshots**.\n\n## 10. Integrated browser\n\n- In-app browser tabs with a shared persist partition, per-tab zoom, OAuth popups, URL-safety and\n  proxy routing.\n- **Agent control over CDP**: navigate, snapshot, screenshot (viewport + full page), click (synthetic\n  and real trusted input), type, key, eval, cookies, raw CDP for any domain, buffered event draining.\n  Drives the **user\'s real logged-in browser** via the companion extension, or the built-in engine.\n- **Inspectors**: DOM tree, properties, native inspector; markup pen; click-to-source (JSX elements\n  resolve to their source line).\n- **6 Toolkit mission playbooks**: accessibility, GDPR/cookies, health, performance, SEO, UX review \u2014\n  each configures, runs, and emits a structured report.\n- **`appbrowser` collaboration**: presence, cursors, comments and edits on a visible tab.\n\n## 11. Web search\n\n`serp_*` over **121 engines / 1,726 parameters**, ~19 tools grouped by family: web, news, images\n(incl. reverse-image and Google Lens), video (incl. YouTube transcripts), maps/local, shopping,\nmarketplaces (Amazon/eBay/Walmart/Home Depot/Zillow), travel (flights/hotels), places & reviews,\nacademic (Scholar/patents), jobs, finance & trends, app stores, ads transparency, social.\nBrokered free tier: **100 searches/month signed in, 10/month anonymous** \u2014 the only hard per-user\nquota in the product.\n\n## 12. Multi-machine\n\n- **Host registry & fleet**: every machine you sign into registers itself; pick one and run the turn\n  there. Dual-role nodes are both client and host at once.\n- **Remote fs + terminals** over the relay (9 relayed channels: agent, signal, browser, frame, fs,\n  term, editor, rdesk, present). A remote driver may only launch an enumerated interpreter.\n- **File Provider (macOS)**: remote machines appear as native **Finder locations** \u2014 browse, read,\n  write, rename, move, Trash + Put Back, thumbnails, progress. Identity survives rename/move.\n- **Drops** \u2014 AirDrop-like sharing to your own devices or another account. E2E-sealed, blind fan-out\n  (first device to accept wins; the sender never learns which), receive policies + receive codes,\n  blocks and abuse reports that are deliberately **indistinguishable** from "nobody was there".\n- **Universal clipboard**: text/HTML/images fan out to your other machines; >512 KB syncs a thumbnail\n  + reference and pulls the bytes on demand. A privacy gate blocks password-manager and secure-input\n  content. Plus **paste history** (\u2318\u21E7V pastebar), clipboard over the rdesk channel, and mobile\n  text-only sync.\n- **Peer-link**: agent\u2194agent direct transport with a 3-tier ladder (LAN TLS \u2192 observed public IP \u2192\n  UDP hole-punch), cert-pinned, relayed only as a fallback.\n- **Mesh VPN**: own PKI with SAS verification, join tickets, recovery mnemonic, a policy bundle with a\n  fail-closed freshness clock, an authenticated Noise data plane, disco/path selection, a 25-pair NAT\n  punch matrix, and a privileged `netd` helper that **derives** every route itself (the IPC has no\n  "add route" field, by design) with journal-based teardown after a crash or reboot.\n- **Remote desktop**: screen share and control. 3 grant paths (same-account registry, cross-account\n  code, unattended password), 3 engines (native, VNC, RDP), display picker, quality, Ctrl+Alt+Del,\n  in-session voice on the same peer connection, side chat, and a phone viewer.\n- **Present & collaboration**: a presenter drives followers across devices; join codes; live cursors,\n  comments and captions.\n- **Co-view mirroring**: a second device with the same conversation open sees **26 allowlisted\n  events**. Permission prompts are deliberately *not* mirrored \u2014 a card you can\'t answer is a dead\n  button.\n- **Settings & session sync**: keys, tokens, active engine and model are account-wide; `localCli`\n  paths are machine-local and must never cross. Plus continuous folder mirroring (active link).\n\n## 13. Team chat & conferencing\n\nSlack-class: spaces, categories, channels, DMs, threads, mentions, reactions, custom emoji, stickers,\nattachments, forum channels, pins, presence, search (Postgres **or** Elasticsearch), link unfurl,\nGIF/meme, bots + a bot API, incoming webhooks, automations, governance, retention and audit.\n**Huddles** over a self-hosted LiveKit SFU with captions and a summary posted back to the channel;\ntoken minting and webhooks work without the SFU, and captions still work when LiveKit is unconfigured.\n\n## 14. Vault (password manager)\n\nZero-knowledge, Argon2id + AES-GCM, server holds only opaque ciphertext. **5 item types** (logins,\npasskeys, payments, secure notes, personal info), ~28 daemon ops, **4 unlock paths** (master password,\ndevice, biometric, **QR from your phone**), sharing sealed to recipient device keys, transfer,\nimport/export (Chrome, Safari, zip), a generator, TOTP, the **macOS AutoFill provider**, and browser\nautofill through the extension.\n\n## 15. Extensions platform\n\nFamilies: **skills, code plugins, bundle plugins, MCP connectors** (3 transports: stdio/http/sse),\nplus **theme packs**. Marketplace discovery with a security verdict that gates install; enable/disable;\nper-extension encrypted secrets; updates; uninstall. Extensions contribute **tabs, panels, settings\npages, status items, menu items and reskins**. First-party skills are free; marketplace ones are\ngated \u2014 `plugins` is the one entitlement actually enforced.\n\n## 16. App Control \u2014 the AI operating Tesseract itself\n\n**144 capabilities across 19 domains** (`settings, navigation, engines, runners, runtimes,\nextensions, hosts, account, daemon, providers, image, voice, permissions, project, privacy, system,\nfonts, present, tour`), exposed as 2 meta-tools + one executor per domain, each input validated\nagainst its own schema. **There is no confirmation gate \u2014 actions happen immediately, by product\ndecision.**\n\n## 17. Platform\n\n- **Accounts**: invite-only (there is **no public sign-up**; Google sign-in on an unknown email is\n  rejected), password + Google, password reset, device registration/claim/keys, a CLI device-code\n  flow, and an offline-tolerance contract \u2014 only an explicit 401/403 signs you out.\n- **Plans**: `free`, `standard` (shown as **"Pro"**), `pro` (shown as **"Team"**) \u2014 17 feature flags,\n  quotas on compute/storage/egress/video, seats, invites, orgs. Your effective plan is the best across\n  every org you belong to.\n- **Missions, cron & unattended**: schedule agent turns that run with nobody present, with catch-up\n  after downtime, background tasks and structured reports.\n- **Settings**: **23 sections, ~300 controls** \u2014 appearance, editor, general, runners, agent, voice,\n  image editing, search, extensions, background tasks, account, team, sync, privacy, file drops,\n  remote desktop, permissions, hooks (9 events), env, about, shortcuts (31 rebindable commands).\n- **Theming**: DTCG design tokens, theme packs with **8 independently switchable facets** (colors,\n  shape, fonts, background, decorations, sounds, intro, extraCss) and a 3-entry-point **Safe Mode** to\n  recover from a bad pack.\n- **Accessibility & motion**: light/dark \xD7 comfortable/compact density throughout, reduced-motion\n  and sound-mute as separate switches, and labelled roles on every interactive element. Animation\n  is a product requirement, not a preference \u2014 every show/hide, open/close and surface switch is\n  animated (140\u2013220ms), and native unstyled controls are not used.\n- **OS permissions**: 7 tracked (microphone, camera, screen recording, full-disk access, accessibility,\n  notifications, local network) with per-platform availability; three require a relaunch after grant.\n- **Observability**: crash/error reporting from 6 producers with separate crash and telemetry consent\n  (**opt-out \u2014 on unless explicitly disabled**), fingerprinting and grouping, an error inspector, an\n  alerting bridge, in-app issue submission, and a GDPR export-and-erase path.\n- **The always-on background service**: the desktop app *is* the daemon (merged process); the slim\n  rdesk and mesh builds and the headless CLI keep the classic separate-service model, installed via\n  launchd / systemd / Windows services. It self-heals silently, degrades from GUI to headless rather\n  than dying, and only one daemon owns a machine at a time. This is why clipboard sync, remote\n  access, scheduled turns and the \u2325Space HUD keep working with no app window open.\n- **Connectivity**: the relay is stateless and queues nothing \u2014 a dropped socket re-bootstraps rather\n  than replaying. In-flight turns survive a disconnect and stay interruptible.\n- **Updates**: check \u2192 download \u2192 stage \u2192 apply \u2192 relaunch, with an external reaper, log rotation on\n  version change and a What\'s New screen. System events (sleep/wake, lock, AC/battery, clock jump,\n  displays, interfaces) drive recovery.\n\n## 18. Deliberately NOT built \u2014 do not promise these\n\n- **Create multiplayer co-editing / comments-in-Create** \u2014 deferred; the stores exist but are unwired.\n- **Marketplace** as a product \u2014 brainstorm only.\n- **Team chat**: voice messages, async video clips, keyword alerts, scheduled send, `/remind`, saved\n  messages, guest/single-channel access, outgoing webhooks, AI catch-up/digest, public communities,\n  nested threads, Block Kit, workflow builder, read receipts, E2E encryption for team chat.\n- **Create/Gamma parity gaps**: todo blocks, GIF/QR/illustrations, media library, form embeds, Google\n  Slides export, theme gallery picker, share analytics, public generate API, custom domains, SEO\n  toggle, password-protected publish, sites manager.\n- **Hosting**: the managed DB/auth/storage layer, secrets manager, git sync + 1-click Publish.\n- **Mobile**: file browsing, in-app browser, hosting an agent, rendering canvas/surface blocks.\n- **Web app**: undo-delete, local engines, runner configs, the errors panel, open/save-to-code for\n  remote hosts.\n- **Mesh**: there is no VPN tunnel a user can turn on yet \u2014 PKI, protocol and data plane are built,\n  route execution is not a shipped user feature. No Windows/macOS netd hosting.\n- **Other**: Homebrew formula, Chrome Web Store listing (the extension installs unpacked), Deepgram\n  voice, Office (ZetaOffice) editing, AWS Bedrock and Google Vertex gateway routes, VNC clipboard,\n  Wayland-specific clipboard handling, RN live widget tree (blocked upstream), the voice tutorial.\n\n## 19. Honest caveats\n\n- **Entitlements are almost entirely unenforced.** Only `plugins` is gated in production, and the\n  desktop client is never gated at all. Both the API and the relay **fail open** on billing errors.\n- **Realtime voice has never run live against a provider** \u2014 treat first-run voice as unproven.\n- **HeyGen avatar video has no key configured** as shipped.\n- **The self-hosted web stack has no per-user data isolation** \u2014 auth there is a login wall over one\n  shared workspace root.\n- `standard` = "Pro" and `pro` = "Team" \u2014 the plan id and the display name are inverted.\n';
+    CAPABILITIES_DOC = '# Tesseract \u2014 what the product can do\n\nOne-page map of **every user-reachable capability**, derived from the 80 area briefs in\n[`functional-testing/`](functional-testing/). It is the source for the `tesseract_capabilities`\ntool, so the agent can answer "what can you do?" about *Tesseract*, not just about its own toolbelt.\n\n> Read this before telling a user something is or isn\'t possible. \xA718 lists what is deliberately\n> **not** built \u2014 promising any of it is the most common way to be wrong about this product.\n\n---\n\n## 1. What Tesseract is\n\nA desktop-first AI workspace: an always-on agent that can drive your machine, your browser and your\nfiles; plus a fleet layer that makes every machine you own reachable from any of them; plus a set of\ncreative, communication and security products built on the same transport. It ships as **eight\nseparate clients** and is not a single app.\n\n## 2. Clients (each a distinct product)\n\n| Client | What it is |\n|---|---|\n| **Desktop** (Electron, mac/Win/Linux) | The full product. Tab workspace \u2014 **no router**: a recursive split-pane tree of tab groups, **24 tab kinds**, **10 dockable panels** (dock left/right/bottom, float, detach), multi-window, native menu, command + search palettes (**31 stable command ids**), \u23181\u20130 tab assign/recall, per-machine keybinding overrides, memory governor + layout restore. |\n| **Web** (`dist-web`) | Same React renderer over a relay transport. **No local agent, no local engine** \u2014 it always runs on a remote host. Terminals are all remote and pinned to the opening host. Documented stubs: undo-delete, engines, runner configs, window resize; voice keys sit in `localStorage`. Always auth-gated. |\n| **Mobile** (Flutter, iOS/Android/macOS) | A **roaming client** \u2014 it never hosts an agent. Chat with streaming, team chat, present-follower, full password vault (incl. TOTP + native autofill), realtime voice, rdesk viewer, clipboard sync, push. Uniquely: **phone-as-host** (serve your photo library and files to your desktop) and a create-video screen. |\n| **CLI** `tesseract-agent` (npm) | Headless daemon, ~18 commands: `login/logout/whoami`, `run`, service `install/start/stop/restart/status/logs`, `link add`, `runner \u2026`, `net \u2026`, self-update. `--instance <id>` namespaces state dir, service name and device id. Docker-friendly (`login --token` then `run` as PID 1). |\n| **Browser extension** (Chrome MV3) | Claude sidebar (sessions, slash commands, todos, sources, DOM tree, run modes), element picker, vault/autofill popup with biometrics, QR pairing, fullpage + plan views. Talks to the daemon through a native-messaging companion host. |\n| **Remote Desktop** (`apps/rdesk`) | Slim standalone binary, ~\xBC the size \u2014 ships **no AI/browser payload**. Server/client/chooser roles, coexists with the full daemon (demotes itself to client rather than quitting). |\n| **Mesh** (`apps/mesh`) | Slim standalone binary for the private network only \u2014 "moves packets, not pixels". Create a mesh, enrol, list peers. No role chooser by design. |\n| **Thin installer** (Flutter, ~20\u201340 MB) | The default download. Resolves the release feed at run time, picks the arch-correct asset, streams + sha512-verifies, installs, launches. Version-agnostic. |\n\nPlus the **admin/operations console** (`admin.tesseract.dev`: gateway credentials, users and keys,\nerror inspector, remediation trigger, audit log) and the **landing site** with its waitlist.\n\n## 3. The agent\n\n- **Turn lifecycle**: stream, interrupt, rapid-follow-up auto-cancel (first seconds only), stall\n  watchdog, orphan-turn recovery, token compaction, usage-limit handling, cost ledger per session.\n- **Run modes**: **5 permission modes \xD7 6 effort levels**, thinking budget, per-turn engine/model pick.\n  Default permission mode is `bypassPermissions`.\n- **Interactive cards** (~9): permission request, plan review, elicitation, confirm, verification,\n  user dialog, skill UI, archive password, unattended report. "Deny & stop" also aborts the turn.\n- **Block renderers** (~9): mission, subagent, report, sources, provenance, event, notice, surface,\n  image grid \u2014 plus tool-run cards joined at render time with 6/8-line clamping.\n- **Composer** (~25 actions): attach, slash commands, @-mention, reverse search, screen-region\n  capture, element picker, voice, undo-last-turn, queued-message reorder.\n- **Sessions**: list, search (one host-side search across every engine\'s stored conversations),\n  resume, summarize, **fork**, merge, delete-message, keep-alive, focus; session mirror + transfer;\n  share a conversation by unguessable capability token.\n- **Subsystems**: reinforced verification, plan splitter, narrator (live status strip, capped at 24\n  model calls/turn), hook bus (9 events), option ratings, unattended mode.\n- **Also**: a playable **chess** surface (5 tools) and chat slash toys (`/shrug`, `/tableflip`, \u2026).\n\n## 4. Engines, runners, gateway\n\n- **4 engines \xD7 3 auth modes** (`oauth` / `apiKey` / `localCli`), a 24-capability engine matrix,\n  engine + model swap mid-session (model/permission-mode reconcile live; effort/cwd/MCP/system-prompt\n  force a respawn).\n- **Runner config pool** with rotation, cooldown and failover; auto-retry is bounded and only fires\n  when the turn did no work.\n- **Provider gateway** \u2014 a credential-injecting reverse proxy, **19 routes** across AGENT (anthropic,\n  openai, google-ai, groq, xai, deepseek, moonshot, openrouter, custom-openai, customer-litellm,\n  azure-openai), EMBEDDING, SEARCH (serpapi), VIDEO (heygen, gemini), VOICE, MEDIA (klipy, unsplash,\n  pexels). Virtual keys `tsk_live_*`, budgets, LiteLLM price map, spend ledger, credential pools with\n  health-based failover that never crosses a category. Models are named the way their makers\n  name them \u2014 `claude-sonnet-4-6`, `claude-haiku-4-5-20251001`, `gpt-6-astra`, `gpt-5.5`,\n  `gpt-5.4-mini`, `gpt-5.4-nano`, `gemini-2.5-flash`, `gpt-image-2` ($0.10/image flat) \u2014 and a\n  signed-in key may also name a model we do not host as `vendor/model`.\n- **Framework cross-product gate**: the user\'s self-updating `claude` CLI \xD7 the pinned Agent SDK,\n  verified nightly across pinned CLI versions.\n\n## 5. Files, editors, viewers\n\n- **Finder-like browser**: list/grid/columns, breadcrumbs, bookmarks + OS favourites, **host\n  switcher** (every op can be remote), new/rename (F2)/move/duplicate/cut-copy-paste/delete + undo,\n  Get Info, send-to-machine, find-in-files, native drag in **and out** (macOS `NSFilePromiseProvider`\n  \u2014 drag a not-yet-downloaded remote file to Finder).\n- **~20 typed viewers/editors**: Monaco source editor, PDF (view + edit), Word/docx, spreadsheets,\n  presentations, markdown, data grid, hex, audio, video, font, HTML canvas, diff, info.\n- **Archives as folders**: ~20 formats (zip, 7z, tar, gz/bz2/xz/zst, rar, iso, wim, cab, cpio, lzh\u2026)\n  browsed, extracted and **edited in place** from the same tree. `rar`/`iso`/`cab` are read-only.\n  Encrypted archives prompt for a password and can remember it.\n- **Image editor**: layered canvas, 10 tools (marquee, lasso, move, crop, paint, eraser, fill,\n  eyedropper, text, zoom), 25 rail controls, **200+ AI effect styles**, layers, adjust/resize, markup,\n  and 3 AI ops (`prompt`/`inpaint`/`outpaint`), transparent-background output, animate\u2192video.\n- **Video editor**: EDL timeline (cut/trim/reorder/transitions/speed/crop/reframe/flip/reverse/colour\n  effects/audio tracks), burned-in captions, **18 MCP tools** \u2014 probe, render, frame extract, silence\n  detect, transcribe with diarization, extract/isolate audio, music generation, voice list/clone/TTS,\n  avatars, caption export, effects, upscale (ffmpeg + local Real-ESRGAN free; fal-Topaz paid).\n- **Undo & time travel**: agent-driven **file rewind** (undo what a turn wrote), file-tree \u2318Z with a\n  per-window trash, workspace snapshots and git checkpoints, canvas/design version history with\n  diff-and-restore, and message deletion with uniform semantics across engines.\n- **Native inspectors**: **Flutter** widget inspector (tree, on-device tap-to-pick, properties,\n  deterministic Dart source edits, hot reload) and **React Native** (source binding + AST edits).\n\n## 6. Create, Design Studio, Stage, Publish\n\n- **Create** \u2014 5 generative document formats (incl. infographic and poster), ~90 actions: outline\n  ops, 10 insert kinds, card generation, AI layouts, export.\n- **Design Studio** \u2014 a 10-stage wizard (pick \u2192 launch/browse \u2192 preview \u2192 target \u2192 inspect \u2192 brief \u2192\n  build \u2192 QA \u2192 deploy) over **254 templates across 19 industry verticals**, Figma import, in-studio\n  theme/token editing.\n- **Stage** \u2014 presentation engine: decks, slides, transitions, **slide audio** (narration/music/sfx\n  with music ducking under narration), a live assistant, presenter mode, zoom/blank. 8 present tools\n  let the agent drive the deck; followers converge from absolute-state snapshots.\n- **Export**: PowerPoint (.pptx), deck and card export, PNG.\n- **Publish** \u2014 private/org/public visibility, server-rendered public pages on a 192-bit token, live\n  data sources (SSRF-guarded), embeds; **community gallery** with explore/following/team/mine, like,\n  follow and clone.\n- **Deploy** \u2014 one-click hosting to `*.tesseract.dev` in per-user hardened containers (cap-drop ALL,\n  no-new-privileges, read-only, egress-locked), metered on compute/storage/egress. One of the few\n  genuinely entitlement-gated features. A separate **per-user cloud agent** container is provisioned\n  on demand so you have a machine in the fleet even with every laptop closed.\n\n## 7. Generative media\n\n| What | Notes |\n|---|---|\n| **Images** | 2 providers (OpenAI gpt-image-2, Google). Generate (4-up by default), edit, inpaint/outpaint, transparent background (inpaint + chroma-key). **$0.10/image**. Ungated, BYO key. |\n| **Stock images** | AI-ranked Unsplash/Pexels search as its own tab. Free tier, rate-limited. |\n| **Video** | Google Veo 3.1 \u2014 4/6/8s clips, 720p/1080p, 16:9 or 9:16, keyframe conditioning (first/last frame), native audio. Free on the bundled tier; otherwise cost-confirmed. Multi-scene storyboards chain clip-to-clip for longer videos. |\n| **Avatar video** | HeyGen talking presenter from a script. |\n| **Music / speech** | ElevenLabs music generation, TTS, and instant **voice cloning**. |\n| **Fonts & brand** | Prompt \u2192 AI glyph sheets \u2192 **vision QC** \u2192 vectorize \u2192 kerning/spacing \u2192 a real installable font. Plus logo/wordmark candidates, SVG lockups and a brand kit. Assembly/vectorize/build are local and free. |\n| **Trending effects** | One-tap video effects (hug/kiss, squish, bring-photo-to-life, age, anime restyle, 3D figurine, outfit change, consent-gated face-swap\u2026). |\n\n## 8. Voice\n\n- **3 realtime providers**: OpenAI Realtime, Gemini Live, ElevenLabs ConvAI \u2014 plus a non-realtime\n  **STT\u2192LLM\u2192TTS pipeline cascade**. Audio goes **client\u2194provider direct**; the daemon only mints the\n  ephemeral credential and runs the control loop.\n- Group calls, live captions, and the **\u2325Space voice widget**.\n- **Voice as a control modality**: the realtime model is the *operator*, Claude Code is the *worker*.\n  It decomposes a spoken request into one small dispatched step at a time, all to the same agent\n  session, and is explicitly forbidden from claiming an outcome the agent did not report \u2014 when\n  unsure it dispatches a check. Its own screen view is a deliberately low-res ~1 fps grab.\n\n## 9. Computer control\n\n- **Two distinct MCP servers**: `tesseract-computer` (**26 tools** \u2014 raw control: move/click/drag/\n  scroll/key/type, see_screen, read_text, read_ui (accessibility), clipboard, windows, apps, open\n  url/path, notify, annotate, draw) and `tesseract-screen` (**15 tools** \u2014 a guided\n  look \u2192 figure_out_next \u2192 perform_step \u2192 **verify** loop that can fail and retry).\n- Plus `tesseract-os-control`: processes, network, structured fs, AX tree + AX actions, preferred app\n  and usage profile, screen overlays (arrow/bubble/highlight/text/popup/choice), and\n  `request_user_input` / `confirm_action`.\n- **Click grounding** is AX-tree + a local Moondream vision sidecar (free); screenshots are\n  Retina-aware and downscaled to point coordinates so clicks line up.\n- **Off by default** \u2014 `computerControlEnabled` must be turned on.\n- **\u2325Space co-pilot HUD** + **13 daemon-owned windows** (spotlight, pastebar, publisher, annotate,\n  draw, overlay, input, choice, consent, control-bar, ai-cursor, rdesk call/client), the tray, screen-\n  region capture with multi-display number flash, and `tesseract://` deep links. The HUD is\n  deliberately **invisible to the AI\'s own screenshots**.\n\n## 10. Integrated browser\n\n- In-app browser tabs with a shared persist partition, per-tab zoom, OAuth popups, URL-safety and\n  proxy routing.\n- **Agent control over CDP**: navigate, snapshot, screenshot (viewport + full page), click (synthetic\n  and real trusted input), type, key, eval, cookies, raw CDP for any domain, buffered event draining.\n  Drives the **user\'s real logged-in browser** via the companion extension, or the built-in engine.\n- **Inspectors**: DOM tree, properties, native inspector; markup pen; click-to-source (JSX elements\n  resolve to their source line).\n- **6 Toolkit mission playbooks**: accessibility, GDPR/cookies, health, performance, SEO, UX review \u2014\n  each configures, runs, and emits a structured report.\n- **`appbrowser` collaboration**: presence, cursors, comments and edits on a visible tab.\n\n## 11. Web search\n\n`serp_*` over **121 engines / 1,726 parameters**, ~19 tools grouped by family: web, news, images\n(incl. reverse-image and Google Lens), video (incl. YouTube transcripts), maps/local, shopping,\nmarketplaces (Amazon/eBay/Walmart/Home Depot/Zillow), travel (flights/hotels), places & reviews,\nacademic (Scholar/patents), jobs, finance & trends, app stores, ads transparency, social.\nBrokered free tier: **100 searches/month signed in, 10/month anonymous** \u2014 the only hard per-user\nquota in the product.\n\n## 12. Multi-machine\n\n- **Host registry & fleet**: every machine you sign into registers itself; pick one and run the turn\n  there. Dual-role nodes are both client and host at once.\n- **Remote fs + terminals** over the relay (9 relayed channels: agent, signal, browser, frame, fs,\n  term, editor, rdesk, present). A remote driver may only launch an enumerated interpreter.\n- **File Provider (macOS)**: remote machines appear as native **Finder locations** \u2014 browse, read,\n  write, rename, move, Trash + Put Back, thumbnails, progress. Identity survives rename/move.\n- **Drops** \u2014 AirDrop-like sharing to your own devices or another account. E2E-sealed, blind fan-out\n  (first device to accept wins; the sender never learns which), receive policies + receive codes,\n  blocks and abuse reports that are deliberately **indistinguishable** from "nobody was there".\n- **Universal clipboard**: text/HTML/images fan out to your other machines; >512 KB syncs a thumbnail\n  + reference and pulls the bytes on demand. A privacy gate blocks password-manager and secure-input\n  content. Plus **paste history** (\u2318\u21E7V pastebar), clipboard over the rdesk channel, and mobile\n  text-only sync.\n- **Peer-link**: agent\u2194agent direct transport with a 3-tier ladder (LAN TLS \u2192 observed public IP \u2192\n  UDP hole-punch), cert-pinned, relayed only as a fallback.\n- **Mesh VPN**: own PKI with SAS verification, join tickets, recovery mnemonic, a policy bundle with a\n  fail-closed freshness clock, an authenticated Noise data plane, disco/path selection, a 25-pair NAT\n  punch matrix, and a privileged `netd` helper that **derives** every route itself (the IPC has no\n  "add route" field, by design) with journal-based teardown after a crash or reboot.\n- **Remote desktop**: screen share and control. 3 grant paths (same-account registry, cross-account\n  code, unattended password), 3 engines (native, VNC, RDP), display picker, quality, Ctrl+Alt+Del,\n  in-session voice on the same peer connection, side chat, and a phone viewer.\n- **Present & collaboration**: a presenter drives followers across devices; join codes; live cursors,\n  comments and captions.\n- **Co-view mirroring**: a second device with the same conversation open sees **26 allowlisted\n  events**. Permission prompts are deliberately *not* mirrored \u2014 a card you can\'t answer is a dead\n  button.\n- **Settings & session sync**: keys, tokens, active engine and model are account-wide; `localCli`\n  paths are machine-local and must never cross. Plus continuous folder mirroring (active link).\n\n## 13. Team chat & conferencing\n\nSlack-class: spaces, categories, channels, DMs, threads, mentions, reactions, custom emoji, stickers,\nattachments, forum channels, pins, presence, search (Postgres **or** Elasticsearch), link unfurl,\nGIF/meme, bots + a bot API, incoming webhooks, automations, governance, retention and audit.\n**Huddles** over a self-hosted LiveKit SFU with captions and a summary posted back to the channel;\ntoken minting and webhooks work without the SFU, and captions still work when LiveKit is unconfigured.\n\n## 14. Vault (password manager)\n\nZero-knowledge, Argon2id + AES-GCM, server holds only opaque ciphertext. **5 item types** (logins,\npasskeys, payments, secure notes, personal info), ~28 daemon ops, **4 unlock paths** (master password,\ndevice, biometric, **QR from your phone**), sharing sealed to recipient device keys, transfer,\nimport/export (Chrome, Safari, zip), a generator, TOTP, the **macOS AutoFill provider**, and browser\nautofill through the extension.\n\n## 15. Extensions platform\n\nFamilies: **skills, code plugins, bundle plugins, MCP connectors** (3 transports: stdio/http/sse),\nplus **theme packs**. Marketplace discovery with a security verdict that gates install; enable/disable;\nper-extension encrypted secrets; updates; uninstall. Extensions contribute **tabs, panels, settings\npages, status items, menu items and reskins**. First-party skills are free; marketplace ones are\ngated \u2014 `plugins` is the one entitlement actually enforced.\n\n## 16. App Control \u2014 the AI operating Tesseract itself\n\n**144 capabilities across 19 domains** (`settings, navigation, engines, runners, runtimes,\nextensions, hosts, account, daemon, providers, image, voice, permissions, project, privacy, system,\nfonts, present, tour`), exposed as 2 meta-tools + one executor per domain, each input validated\nagainst its own schema. **There is no confirmation gate \u2014 actions happen immediately, by product\ndecision.**\n\n## 17. Platform\n\n- **Accounts**: invite-only (there is **no public sign-up**; Google sign-in on an unknown email is\n  rejected), password + Google, password reset, device registration/claim/keys, a CLI device-code\n  flow, and an offline-tolerance contract \u2014 only an explicit 401/403 signs you out.\n- **Plans**: `free`, `standard` (shown as **"Pro"**), `pro` (shown as **"Team"**) \u2014 17 feature flags,\n  quotas on compute/storage/egress/video, seats, invites, orgs. Your effective plan is the best across\n  every org you belong to.\n- **Missions, cron & unattended**: schedule agent turns that run with nobody present, with catch-up\n  after downtime, background tasks and structured reports.\n- **Settings**: **23 sections, ~300 controls** \u2014 appearance, editor, general, runners, agent, voice,\n  image editing, search, extensions, background tasks, account, team, sync, privacy, file drops,\n  remote desktop, permissions, hooks (9 events), env, about, shortcuts (31 rebindable commands).\n- **Theming**: DTCG design tokens, theme packs with **8 independently switchable facets** (colors,\n  shape, fonts, background, decorations, sounds, intro, extraCss) and a 3-entry-point **Safe Mode** to\n  recover from a bad pack.\n- **Accessibility & motion**: light/dark \xD7 comfortable/compact density throughout, reduced-motion\n  and sound-mute as separate switches, and labelled roles on every interactive element. Animation\n  is a product requirement, not a preference \u2014 every show/hide, open/close and surface switch is\n  animated (140\u2013220ms), and native unstyled controls are not used.\n- **OS permissions**: 7 tracked (microphone, camera, screen recording, full-disk access, accessibility,\n  notifications, local network) with per-platform availability; three require a relaunch after grant.\n- **Observability**: crash/error reporting from 6 producers with separate crash and telemetry consent\n  (**opt-out \u2014 on unless explicitly disabled**), fingerprinting and grouping, an error inspector, an\n  alerting bridge, in-app issue submission, and a GDPR export-and-erase path.\n- **The always-on background service**: the desktop app *is* the daemon (merged process); the slim\n  rdesk and mesh builds and the headless CLI keep the classic separate-service model, installed via\n  launchd / systemd / Windows services. It self-heals silently, degrades from GUI to headless rather\n  than dying, and only one daemon owns a machine at a time. This is why clipboard sync, remote\n  access, scheduled turns and the \u2325Space HUD keep working with no app window open.\n- **Connectivity**: the relay is stateless and queues nothing \u2014 a dropped socket re-bootstraps rather\n  than replaying. In-flight turns survive a disconnect and stay interruptible.\n- **Updates**: check \u2192 download \u2192 stage \u2192 apply \u2192 relaunch, with an external reaper, log rotation on\n  version change and a What\'s New screen. System events (sleep/wake, lock, AC/battery, clock jump,\n  displays, interfaces) drive recovery.\n\n## 18. Deliberately NOT built \u2014 do not promise these\n\n- **Create multiplayer co-editing / comments-in-Create** \u2014 deferred; the stores exist but are unwired.\n- **Marketplace** as a product \u2014 brainstorm only.\n- **Team chat**: voice messages, async video clips, keyword alerts, scheduled send, `/remind`, saved\n  messages, guest/single-channel access, outgoing webhooks, AI catch-up/digest, public communities,\n  nested threads, Block Kit, workflow builder, read receipts, E2E encryption for team chat.\n- **Create/Gamma parity gaps**: todo blocks, GIF/QR/illustrations, media library, form embeds, Google\n  Slides export, theme gallery picker, share analytics, public generate API, custom domains, SEO\n  toggle, password-protected publish, sites manager.\n- **Hosting**: the managed DB/auth/storage layer, secrets manager, git sync + 1-click Publish.\n- **Mobile**: file browsing, in-app browser, hosting an agent, rendering canvas/surface blocks.\n- **Web app**: undo-delete, local engines, runner configs, the errors panel, open/save-to-code for\n  remote hosts.\n- **Mesh**: there is no VPN tunnel a user can turn on yet \u2014 PKI, protocol and data plane are built,\n  route execution is not a shipped user feature. No Windows/macOS netd hosting.\n- **Other**: Homebrew formula, Chrome Web Store listing (the extension installs unpacked), Deepgram\n  voice, Office (ZetaOffice) editing, AWS Bedrock and Google Vertex gateway routes, VNC clipboard,\n  Wayland-specific clipboard handling, RN live widget tree (blocked upstream), the voice tutorial.\n\n## 19. Honest caveats\n\n- **Entitlements are almost entirely unenforced.** Only `plugins` is gated in production, and the\n  desktop client is never gated at all. Both the API and the relay **fail open** on billing errors.\n- **Realtime voice has never run live against a provider** \u2014 treat first-run voice as unproven.\n- **HeyGen avatar video has no key configured** as shipped.\n- **The self-hosted web stack has no per-user data isolation** \u2014 auth there is a login wall over one\n  shared workspace root.\n- `standard` = "Pro" and `pro` = "Team" \u2014 the plan id and the display name are inverted.\n';
   }
 });
 
@@ -38847,13 +38852,6 @@ async function anthropicCall(args) {
   }
   return { text: null, rateLimited: false };
 }
-function isDirectAnthropic(env) {
-  try {
-    return /(^|\.)api\.anthropic\.com$/i.test(new URL(anthropicBaseUrl(env)).hostname);
-  } catch {
-    return false;
-  }
-}
 async function buildClaudeNarrationLlm(env) {
   if (!await hasAnthropicCred(env))
     return null;
@@ -38943,7 +38941,7 @@ async function buildClaudeVisionLlm(env) {
     ]
   });
 }
-var execFileP, USAGE_URL, FETCH_TIMEOUT_MS, cache, TTL_MS, RATE_LIMIT_BACKOFF_MS, inFlight, rateLimitedUntil, schemaUnsupported, schemaKey, SMART_TIMEOUT_MS, VISION_TIMEOUT_MS;
+var execFileP, USAGE_URL, FETCH_TIMEOUT_MS, cache, TTL_MS, RATE_LIMIT_BACKOFF_MS, inFlight, rateLimitedUntil, isDirectAnthropic, schemaUnsupported, schemaKey, SMART_TIMEOUT_MS, VISION_TIMEOUT_MS;
 var init_session_usage = __esm({
   "../../packages/agent-core/dist/session-usage.js"() {
     "use strict";
@@ -38958,6 +38956,7 @@ var init_session_usage = __esm({
     RATE_LIMIT_BACKOFF_MS = 5 * 6e4;
     inFlight = null;
     rateLimitedUntil = 0;
+    isDirectAnthropic = isAnthropicEndpoint;
     schemaUnsupported = /* @__PURE__ */ new Set();
     schemaKey = (base, model) => `${base}|${model}`;
     SMART_TIMEOUT_MS = RV_GRADE_TIMEOUT_MS;
@@ -39087,7 +39086,7 @@ var init_stream_stall = __esm({
 // ../../packages/agent-core/dist/model-switch.js
 function modelFamily(model) {
   const id = (model ?? "").trim().toLowerCase();
-  if (!id || id.startsWith("tsr-"))
+  if (!id)
     return void 0;
   return FAMILIES2.find((family) => id.includes(family));
 }
@@ -39492,12 +39491,14 @@ function sessionStoreFor(mode) {
 function unattendedReportDirective(send5) {
   return unattendedModeFor({ optedIn: send5.unattended, origin: send5.origin }) !== "off" ? prompt("agent.system.unattended-report") : "";
 }
-function sdkModelId(model) {
+function sdkModelId(model, env) {
   const id = model?.trim();
   if (!id)
     return void 0;
   if (process.env.TESSERACT_DISABLE_1M_CONTEXT === "1")
     return id;
+  if (!isAnthropicEndpoint(env))
+    return id;
   if (/[/]/.test(id) || /^[a-z]{2,}\./i.test(id))
     return id;
   return withOneMillionContext(id) ?? id;
@@ -41317,7 +41318,7 @@ var init_agent_manager = __esm({
           return false;
         }
         try {
-          await live.query.setModel(sdkModelId(next));
+          await live.query.setModel(sdkModelId(next, live.endpointEnv));
         } catch (err2) {
           log(`[model] ${turnSid}: could not switch to ${next}:`, err2 instanceof Error ? err2.message : String(err2));
           return false;
@@ -42852,7 +42853,7 @@ ${lines}`;
           if (opts.permissionMode)
             await live.query.setPermissionMode(sdkPermissionMode(opts.permissionMode));
           if (opts.model !== void 0) {
-            await live.query.setModel(sdkModelId(opts.model));
+            await live.query.setModel(sdkModelId(opts.model, live.endpointEnv));
             live.model = opts.model || void 0;
             this.rememberSessionModel(live.sessionId, live.model);
           }
@@ -43081,7 +43082,7 @@ ${lines}`;
           live.windowId = params.windowId;
         live.pendingTitle = params.message.trim().slice(0, 80);
         if (params.model)
-          void live.query.setModel(sdkModelId(params.model)).catch(() => {
+          void live.query.setModel(sdkModelId(params.model, live.endpointEnv)).catch(() => {
           });
         if (params.permissionMode)
           void live.query.setPermissionMode(sdkPermissionMode(params.permissionMode)).catch(() => {
@@ -43398,6 +43399,7 @@ ${prompt("agent.system.long-running")}`;
       }
       async createLiveSession(params) {
         const spawnModel = this.modelForSpawn(params);
+        const spawnEndpoint = { ...params.env ?? {}, ...params.engineEnv ?? {} };
         this.rememberSessionModel(params.sessionId, spawnModel);
         this.rememberSessionModel(params.clientRequestId, spawnModel);
         const openProject = params.projectDir ?? this.projectDir ?? params.sources?.[0];
@@ -43559,7 +43561,7 @@ ${prompt("agent.system.long-running")}`;
             // `modelForSpawn`, not `params.model`: a resume that nobody sent a model with (a
             // wake-up after a background task, a daemon-side continuation) would otherwise open a
             // 1M conversation against 200k. See sessionModels.
-            ...spawnModel ? { model: sdkModelId(spawnModel) } : {},
+            ...spawnModel ? { model: sdkModelId(spawnModel, spawnEndpoint) } : {},
             // Clamp to what THIS CLI accepts — an unsupported level is a hard exit-1, not
             // a warning. See clampEffort.
             ...ultracodeEffort ? { effort: ultracodeEffort } : {},
@@ -43568,7 +43570,7 @@ ${prompt("agent.system.long-running")}`;
             // of surfacing an error the user has to act on.
             // Same for the fallback: it REPLACES the main model on an overload, so a bare id here
             // would quietly move a long conversation onto a 200k window mid-run.
-            ...params.fallbackModel ? { fallbackModel: sdkModelId(params.fallbackModel) } : {},
+            ...params.fallbackModel ? { fallbackModel: sdkModelId(params.fallbackModel, spawnEndpoint) } : {},
             ...params.enableFileCheckpointing ? { enableFileCheckpointing: true } : {},
             // ON BY DEFAULT. This is what turns a silent helper into one that says what it is
             // doing — a short present-tense line on each progress frame, written by forking the
@@ -44299,7 +44301,7 @@ ${prompt("agent.system.long-running")}`;
                 // rather than the number in the run-mode menu.
                 ...this.daemonCompactEnv(params.env),
                 // A GATEWAY MODEL ALIAS MUST NOT BE WINDOW-ENFORCED BY THE CLI. Routed through
-                // our gateway the id is an alias (`tsr-claude`) the CLI does not know. Harmless
+                // our gateway it may be a retired tier alias the CLI does not know. Harmless
                 // for an ordinary query; fatal for its INTERNAL compaction query, which never
                 // returns — no frame, 120s stall, killed as "connection went quiet", replayed
                 // by the client, and stalled again every two minutes for ever. MEASURED on a
@@ -47514,9 +47516,9 @@ function materializeClaudeAuthHome(contents) {
   });
   return home;
 }
-function withClaudeModels(env, main2, fast) {
+function withClaudeModels(env, main2, fast, longWindow = true) {
   if (main2) {
-    env.ANTHROPIC_MODEL = withOneMillionContext(main2) ?? main2;
+    env.ANTHROPIC_MODEL = (longWindow ? withOneMillionContext(main2) : main2) ?? main2;
   }
   if (fast)
     env.ANTHROPIC_SMALL_FAST_MODEL = fast;
@@ -47563,7 +47565,7 @@ function resolveRunnerEnv(spec) {
     if (!baseUrl.trim()) {
       return { CCB_RUNNER_CONFIG_ERROR: `Set an address for ${provider.label} in Settings \u2192 Models & accounts.` };
     }
-    return withClaudeModels({ ANTHROPIC_BASE_URL: baseUrl, ANTHROPIC_AUTH_TOKEN: key2, ANTHROPIC_API_KEY: "" }, main2, fast);
+    return withClaudeModels({ ANTHROPIC_BASE_URL: baseUrl, ANTHROPIC_AUTH_TOKEN: key2, ANTHROPIC_API_KEY: "" }, main2, fast, false);
   }
   const endpoint2 = routerEndpointSync({
     providerId,
@@ -47637,7 +47639,10 @@ function readGatewayKey() {
         baseUrl: raw.baseUrl,
         apiKey: raw.apiKey,
         models: raw.models,
-        ...raw.endpoints ? { endpoints: raw.endpoints } : {}
+        ...raw.endpoints ? { endpoints: raw.endpoints } : {},
+        // Per-audience defaults (engine + model) travel through too — `tiers()` in shared
+        // honours `defaults.model` on this headless spawn path exactly as the desktop does.
+        ...raw.defaults ? { defaults: raw.defaults } : {}
       };
     }
   } catch {
@@ -98523,7 +98528,7 @@ var GeminiVeoProvider = class {
 };
 
 // ../../packages/video-core/dist/tesseract.js
-var DEFAULT_MODEL4 = "tsr-video";
+var DEFAULT_MODEL4 = "veo-3.1-fast-generate-preview";
 var REQUEST_TIMEOUT_MS2 = 9e4;
 var DOWNLOAD_TIMEOUT_MS2 = 18e4;
 var withTimeout4 = (signal2, ms) => signal2 ?? AbortSignal.timeout(ms);
@@ -99734,8 +99739,8 @@ var VIDEO_PROVIDERS = [
   {
     id: "tesseract",
     label: "Tesseract \xB7 free (no key)",
-    defaultModel: "tsr-video",
-    models: ["tsr-video"],
+    defaultModel: "veo-3.1-fast-generate-preview",
+    models: ["veo-3.1-fast-generate-preview"],
     keysUrl: "",
     free: true
   },
@@ -100001,7 +100006,7 @@ var IN_TURN_WAIT_MS = 8 * 6e4;
 var HEARTBEAT_MS2 = 15e3;
 var FAST_MODEL = VIDEO_PROVIDERS.find((p) => p.id === "google").defaultModel;
 var STANDARD_MODEL = "veo-3.1-generate-preview";
-var FREE_MODEL = "tsr-video";
+var FREE_MODEL = "veo-3.1-fast-generate-preview";
 var clampDuration2 = (n) => [4, 6, 8].reduce((best, d) => Math.abs(d - n) < Math.abs(best - n) ? d : best);
 function creds2() {
   return readDaemonAgentSettings().videoGen;

package/dist-bundle/mesh-core/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@fineguide-tesseract/mesh-core",
-  "version": "0.11.15",
+  "version": "0.11.16",
   "main": "index.js",
   "types": "index.d.ts"
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tesseract-agent",
-  "version": "0.11.15",
+  "version": "0.11.16",
   "license": "UNLICENSED",
   "publishConfig": {
     "access": "public",