Commit Graph

  • 38a4b95450 [eric] resilience: one failed request now confirms with a probe instead of waiting for a second, and the reconnecting pill actually centres ciregenz 2026-08-11 01:05:47 -07:00
  • 210e8c7410 [eric] canvas: a stray pinch can no longer magnify the whole app out of view, and a non-finite camera is refused instead of bricking navigation ciregenz 2026-08-11 01:00:07 -07:00
  • 074b1a9bec [eric] release: experimental versions use the documented exp.N form so successive builds share one update channel v1.7.6-exp.2 ciregenz 2026-08-10 23:32:11 -07:00
  • 9efc0cb515 [eric] browser: a sign-in popup from a browser card keeps that card's user agent, so Google stops refusing it (ENG-238); reconnecting pill portals to body so it centres ciregenz 2026-08-10 23:19:27 -07:00
  • 8ceca9ac97 [eric] tool-ui: every ShowUI card needs a stable id, and one question can only be answered once across all its surfaces ciregenz 2026-08-10 22:46:12 -07:00
  • fe77b64d8c [eric] release: 1.7.6-exp2 carries tonight's sign-in, update, resilience and tool-ui fixes ciregenz 2026-08-10 22:07:34 -07:00
  • 3a2d1049c3 [eric] resilience: a stalled request probes the backend instead of waiting out its timeout, so a wedged port surfaces in seconds ciregenz 2026-08-10 21:45:23 -07:00
  • 3ea7b93428 arena: live-web episodes get one process each -- a single hang was poisoning whole rounds ciregenzandClaude Fable 5 2026-08-10 21:41:26 -07:00
  • 08d04d39f3 [eric] tool-ui: a press on an ask button in the collapsed pill reaches it, drag-capture no longer eats the click (ENG-232) ciregenz 2026-08-10 21:30:48 -07:00
  • c7b0982acc arena: central data store manifest -- one directory holds every episode, screenshot and raw log ciregenzandClaude Fable 5 2026-08-10 21:16:05 -07:00
  • b56e91b1eb [eric] resilience: api fetches ride backend restarts (bounded retry), a dead backend shows an honest reconnecting pill and self-heals ciregenz 2026-08-10 20:35:07 -07:00
  • c57c2431ee arena: audit pass -- fix HOLD-line history pollution, disclose metric asymmetry ciregenzandClaude Fable 5 2026-08-10 20:13:01 -07:00
  • c5218eee60 arena: v16 -- the four failure-pattern fixes, each traced to counted losses ciregenzandClaude Fable 5 2026-08-10 20:01:19 -07:00
  • 6a49713695 [eric] sign-in: deep links queue-and-drain (no lost race) with durable retry-until-identity-lands handoff ciregenz 2026-08-10 19:44:18 -07:00
  • bfd8920570 arena: opus-5 grid -- ours 82.4% (union 85.6%), browser-use DEGRADES on newer models ciregenzandClaude Fable 5 2026-08-10 18:51:44 -07:00
  • 69c34b7439 [eric] agents: CreateApp is offer-first; track/organize/plan asks answer in chat instead of scaffolding an app ciregenz 2026-08-10 17:55:10 -07:00
  • bceb407b56 [eric] updates: first check reads the experimental toggle from disk, ending the prerelease self-downgrade ping-pong ciregenz 2026-08-10 17:55:10 -07:00
  • fb71f3b686 [eric] tool-ui: posts, links and images open as browser cards on click, leaving fullscreen first (ENG-234) ciregenz 2026-08-10 17:32:53 -07:00
  • 91c273c8e3 [eric] tool-ui: question-flow answers deliver (onSelect/onComplete), respond dedupes, dead sessions never look answerable ciregenz 2026-08-10 17:32:53 -07:00
  • 49dfab9eb9 [eric] tool-ui: validation bounces return plain content, not isError, so they persist and dead asks can demote ciregenz 2026-08-10 17:32:53 -07:00
  • e433fdf95b arena: multi-model ledger -- ours leads the same-model pair on every axis at every tier tested ciregenzandClaude Fable 5 2026-08-10 16:51:25 -07:00
  • 4cf0971693 [eric] tool-ui: asks pair by tool_use_id, early clicks held, stops release waits, dead asks go quiet ciregenz 2026-08-10 14:45:11 -07:00
  • b0b7df3f61 [eric] agents: tool-ui default-on with media; WebFetch names its authed tier by-design ciregenz 2026-08-10 14:06:49 -07:00
  • 784b33c15a [eric] web: bing + brave join the keyless race behind ddg; bing went 50/50 on the burst that tripped both ddg and brave ciregenz 2026-08-10 14:02:42 -07:00
  • 321b1ef549 [eric] release: the experimental publish belt also sets the prerelease flag, which electron-builder dropped on both exp1 cuts ciregenz 2026-08-10 13:04:52 -07:00
  • d506cf1896 [eric] web: startpage's proof-of-work wall now covers POST too; name it a refusal so the breaker benches the dead engine honestly ciregenz 2026-08-10 13:01:38 -07:00
  • 785ee1301c [eric] boot: a failed first window load retries with backoff instead of stranding the splash forever, so a boot-time network-service crash stops reading as a bricked app v1.7.6-exp1 ciregenz 2026-08-10 10:38:36 -07:00
  • de1915e20f [eric] updates: opting out of experimental builds offers the current stable even though it is a downgrade, so a tester is never stranded on a prerelease ciregenz 2026-08-10 10:17:32 -07:00
  • 478e261a76 arena: v13 autopsy -- SoM overlay occludes micro-UIs and breeds coordinate-happiness, dropped ciregenzandClaude Fable 5 2026-08-10 10:14:10 -07:00
  • ade1207d86 [eric] crash: shutdown fuse hard-exits a wedged backend tree in 10s, uvicorn drains connections on an 8s budget, and three GPU deaths buy one boot of software rendering with an honest pill ciregenz 2026-08-10 09:02:31 -07:00
  • edd72376f6 [eric] browser: the routing classifier treats the browser as last resort; native integrations beat it, search beats a page visit, reading beats acting ciregenz 2026-08-10 08:47:13 -07:00
  • 7ba5a067d7 [eric] crash: browser-card teardown quiesces every webview to a committed about:blank before unmount, closing the live-surface destroy gap app cards already had covered ciregenz 2026-08-10 08:46:23 -07:00
  • d215facb94 [eric] release: version 1.7.6-exp1; an experimental build rides its base release notes and the Help agent names the running version ciregenz 2026-08-10 08:39:15 -07:00
  • 8aac6656c4 [eric] crash: app-card teardown waits for the about:blank commit event plus a settle frame instead of a 250ms fail-open timer, so a loaded machine stops destroying live GPU surfaces mid-composite ciregenz 2026-08-10 08:37:42 -07:00
  • df8af72125 [eric] crash: safe-mode loop breaker; two dirty exits in ten minutes boots with webviews parked as screenshots and app runtimes held until clicked, with a crash fingerprint from the newest minidump ciregenz 2026-08-10 08:36:47 -07:00
  • 2d5f35ad6e [eric] browser: agent commands restore the user's typing focus, so a running browser agent stops yanking the cursor out of the composer ciregenz 2026-08-10 08:27:12 -07:00
  • b91b271e19 [eric] ui: dead Create Custom connector entry removed, browser AI badge removed, and choice widgets show their question title and description ciregenz 2026-08-10 08:24:41 -07:00
  • 366f74f688 [eric] browser: the first-message fast path learns which integrations are natively connected, so a Slack ask stops opening a browser when Slack tools exist ciregenz 2026-08-10 08:18:59 -07:00
  • 263ce73b05 [eric] release: 1.7.6 release notes + the spawn-env test learns the auto-memory kill switch, so the pre-tag gates go green ciregenz 2026-08-10 08:15:38 -07:00
  • bcbb97d657 arena: definitive close -- v10 beats the real browser-use 75.2 to 69.6 at 8.5x speed, 0 false claims ciregenzandClaude Fable 5 2026-08-10 07:32:19 -07:00
  • 8d280afaa9 arena: v10 BEATS the real browser-use -- 75.2% vs 69.6% at 8.5x the speed, zero false claims ciregenzandClaude Fable 5 2026-08-10 05:12:33 -07:00
  • 027b3422e5 arena: v9 63.2% at 4.9s -- runway was the constraint; v10 fixes the last blindness ciregenzandClaude Fable 5 2026-08-10 04:13:17 -07:00
  • 5b93ca2e84 arena: v8 ties v7's 56.8 with a different shape -- drag and spatial now BEAT the real browser-use ciregenzandClaude Fable 5 2026-08-10 02:54:29 -07:00
  • 389c73fe00 arena: v7 lands multi-action + adaptive vision -- 56.8% at 4.8s, still zero false claims ciregenzandClaude Fable 5 2026-08-10 01:34:03 -07:00
  • 44d326e595 [eric] release: bump version to 1.7.6 ciregenz 2026-08-09 22:02:02 -07:00
  • 91058c0663 [eric] browser: a dashboard-less session that asks for a browser gets a one-line error instead of a hung tool call and an empty errored sub-session ciregenz 2026-08-09 20:24:03 -07:00
  • 704cada2a3 arena: final scoreboard -- all 8 arms complete on all 125 tasks, two verdicts stated plainly ciregenzandClaude Fable 5 2026-08-09 20:08:25 -07:00
  • 7297b0a8e3 [eric] agents: the CLI's own auto-memory is disabled on every spawn, so memories stop landing in per-chat buckets and MemoryRead/MemoryWrite on the shared store is the one memory surface ciregenz 2026-08-09 19:53:14 -07:00
  • 45b3a9de37 [eric] canvas: layout load keeps exactly one fullscreen owner, so a pre-1.7.6 profile with a second persisted fullscreen entry can no longer drag cards while fullscreen ciregenz 2026-08-09 18:31:04 -07:00
  • d7ed3cda2d [eric] apps: the template vite config loads as ESM (.mts) and deps ride a vendored lockfile, so a fresh app boots on the bundled node 20 instead of dying on ESM-only plugins ciregenz 2026-08-09 18:17:26 -07:00
  • 09df1c2d17 [eric] build: the webapp-template node_modules archive skips when already built for the digest, so an interrupted build stops re-running the 22s npm install every restart ciregenz 2026-08-09 17:04:58 -07:00
  • 0d98fd6bc7 [eric] model picker: groups render subscriptions first, API keys second, routers last with tier separators, never interleaved ciregenz 2026-08-09 17:01:41 -07:00
  • 97d0c7a8ef [eric] apps: an app nobody is editing serves its built bundle with no dev-server process (~270MB -> ~0), vite reboots the moment an agent binds or the dist goes stale ciregenz 2026-08-09 16:53:14 -07:00
  • 102ba3927a arena: v3/v4 ingest rounds + the dedupe bug those rounds exposed in the shipped ranker ciregenzandClaude Fable 5 2026-08-09 16:17:11 -07:00
  • 2476d4c03d [eric] linter: the file-size debt is reported (82 grandfathered oversized files, top offenders named) instead of hiding behind the enforcement zero ciregenz 2026-08-09 14:47:31 -07:00
  • 4804e232d9 [eric] canvas: NOTHING drags in fullscreen (untileForDrag chokepoint + guards on app and window cards; exit is the pill, green dot, or Escape) ciregenz 2026-08-09 14:47:31 -07:00
  • dfe819b095 arena: every browser stack scored by MiniWoB itself, plus the first two ingest wins ciregenzandClaude Fable 5 2026-08-09 14:41:02 -07:00
  • f10591f07c [eric] canvas: the legacy inline app picker is gone (Cmd+M and the composer route open the Applications window) ciregenz 2026-08-09 14:15:27 -07:00
  • b47ad9c60f [eric] agents: active time is stall-capped inter-event work, so one wedged turn can't book 54 hours of 'agent time' ciregenz 2026-08-09 14:08:34 -07:00
  • 5eca1df72d [eric] agents: mid-turn provider backoff shows a muted retrying pill instead of a dead card, and transient pills survive session status frames ciregenz 2026-08-09 14:01:11 -07:00
  • 8fdc1fa2f9 [eric] settings: the shortcut recorder owns the keyboard while armed (window-level capture, hotkeys suppressed), so rebinding no longer snaps back to fn ciregenz 2026-08-09 13:51:55 -07:00
  • d050517c4e [eric] canvas: fullscreen locks canvas panning (drag joined the wheel gate), thirds cut from the tile menus, Applications window gets always-on search + image-only tiles (screenshot > symbol > svg, letters never) ciregenz 2026-08-09 13:27:26 -07:00
  • 3c2aeb8fa7 [eric] apps: runtime status carries ready + serving_url, so 'spawned' and 'actually serving' stop being the same claim ciregenz 2026-08-09 12:51:12 -07:00
  • e4454a0b9e [eric] skills: the Add box resolves pasted README install commands (npx/pnpm/bunx, skills.sh URLs) through the parser that shipped without a client ciregenz 2026-08-09 12:51:02 -07:00
  • 14d4a71e15 [eric] agents: a TLS certificate failure is non-transient and says so, instead of burning 335s of retries on a broken cert chain ciregenz 2026-08-09 12:51:02 -07:00
  • c8c3e7be0e [eric] canvas: app card chrome carries no identity (bare strip); the dock tile is the personalized mark and prefers the app's real thumbnail over any glyph ciregenz 2026-08-09 12:13:05 -07:00
  • d013ac5f1f [eric] canvas: app chrome identity is the app's own monogram tile (no name), and the Workflows strip goes bare like Settings/Marketplace ciregenz 2026-08-09 11:56:27 -07:00
  • 95a41451b3 [eric] apps: grant identity is server-derived only (minted per-app tokens + serve-path referer), forged ids refused, grant dialogs queue ciregenz 2026-08-09 11:47:27 -07:00
  • c794706281 [eric] settings: the in-body X is gone (traffic lights are the close), onboarding cursor re-anchored to the red dot ciregenz 2026-08-09 11:47:27 -07:00
  • 99433fd399 [eric] canvas: window chrome homogenized (no separator, same surface as the body), app title centered, one exit affordance in fullscreen ciregenz 2026-08-09 11:47:27 -07:00
  • 54bdd42018 [eric] apps: card chrome goes slim (identity + reload/share/kebab), and fullscreen flips it to a dark edge-to-edge system bar with its own exit ciregenz 2026-08-09 11:03:41 -07:00
  • 5f6cf216ad [eric] canvas: the marquee inverts the PAINTED transform instead of a second camera copy, so a drawn-vs-painted mismatch can't exist ciregenz 2026-08-09 10:28:23 -07:00
  • 6bd3f93f07 [eric] chat: questions and tool permission asks render as the vendored tool-ui cards (question-flow + approval-card); QuestionForm extracted to kill an import cycle; registry loads the full question-flow union ciregenz 2026-08-09 10:01:02 -07:00
  • 0688c85607 [eric] apps: apps can call the user's connected tools behind a per-app grant gate (approval card, deny-by-default, Origin-derived identity) ciregenz 2026-08-09 08:52:36 -07:00
  • 0592437941 [eric] memory: agents get MemoryRead/MemoryWrite (atomic batch, overflow-consolidate protocol) and the prompt block freezes per session, hermes-agent style ciregenz 2026-08-09 08:52:30 -07:00
  • 31a5386186 [eric] voice: the mic parks warm for 5 minutes after each dictation, so the next fn press skips the ~2.5s device power-up ciregenz 2026-08-09 08:24:30 -07:00
  • 4ebf9a91f6 [eric] canvas: fullscreen can no longer strand a chromeless shell (ghost-id guard, one owner, layout-fetch prune, always-clickable exit pill) ciregenz 2026-08-09 07:48:30 -07:00
  • 02cc65632a [eric] tool-ui: repair the payload shapes agents actually emit (stat keys, step ids, receipt outcomes), so widgets render instead of quietly falling back ciregenz 2026-08-09 07:48:30 -07:00
  • 4ac57967ff [eric] agents: the last silent-quit nudge demands a text answer instead of more work, and an exhausted budget says so in the chat ciregenz 2026-08-08 22:35:41 -07:00
  • 18cc7c7417 [eric] voice: mic open overlaps the IPC, context and worklet setup, and the stop path no longer waits on the streaming shutdown it rarely needs ciregenz 2026-08-08 22:08:13 -07:00
  • 830a995c6a [eric] harness: MiniWoB via BrowserGym, the first externally-scored number here ciregenzandClaude Opus 5 2026-08-08 22:05:09 -07:00
  • c9840eb00b [eric] hardening: legacy pre-collapse tool ids canonicalize to core, and the docked mini gets a load listener plus a 1s belt so a silent layout shift can't strand it ciregenz 2026-08-08 20:48:47 -07:00
  • 16d6b5e378 [eric] agents: combined server answers tool calls on threads and survives a broken module, restoring the parallelism ten processes gave for free ciregenz 2026-08-08 14:35:35 -07:00
  • 4faba74d55 [eric] agents: every builtin tool server rides ONE process per chat, and the gate enumerates its grants instead of wildcarding ciregenz 2026-08-08 12:59:50 -07:00
  • 6dc6c0ecf2 [eric] docs: record every browser tool installed and run, including the ones that would not ciregenzandClaude Opus 5 2026-08-08 10:30:58 -07:00
  • 72c5effbff [eric] docs: measured competitive read of the browser-agent field ciregenzandClaude Opus 5 2026-08-08 10:06:01 -07:00
  • c0d33c0cfb [eric] browser: the docked mini re-measures when its slot announces a remount, retiring a 150ms forever-poll of layout reads ciregenz 2026-08-08 09:47:25 -07:00
  • 4d86b3377d [eric] agents: the three ungated meta MCP servers ride one process, cutting three idle interpreters per parked chat ciregenz 2026-08-08 09:41:45 -07:00
  • d60a99ddd1 [eric] boot: bootstrap fetches retry until the backend answers and heal on focus, so a lost boot race stops bricking first paint ciregenz 2026-08-08 09:34:13 -07:00
  • a982e3361f [eric] browser: an approval nobody can answer declines now instead of parking a turn for 5 minutes ciregenzandClaude Opus 5 2026-08-08 09:28:53 -07:00
  • 7a942fa9fa Merge remote-tracking branch 'origin/eric/dev' into eric/browser-merged ciregenz 2026-08-08 09:22:33 -07:00
  • 1db0a5d04e [eric] harness: vs_bench grades zero-card trials from the backend's own dryrun-report ciregenzandClaude Opus 5 2026-08-08 09:22:33 -07:00
  • 5c0d6d52b6 [eric] apps: render-health beacons ride the runtime log stream where they actually arrive, vite-plugin-terminal owns the page console ciregenz 2026-08-08 02:14:27 -07:00
  • ab847fe2c4 [eric] apps: a spawned agent lands on the most-recent dashboard when the app names none, so the card is visible not just real ciregenz 2026-08-08 02:14:27 -07:00
  • 9541d0cf09 [eric] apps: the warm node_modules cache proves completion with a sentinel and never npm-installs on the request path, which wedged the whole backend ciregenz 2026-08-08 02:14:26 -07:00
  • a0e7e719fa [eric] browser: delete the orphaned BrowserAgentInlineFeed component (missed in the action-feed removal commit) ciregenz 2026-08-08 01:22:54 -07:00
  • a12047ae66 [eric] apps: the full tool-ui widget set is vendored into every app workspace behind an @toolui alias, deps exact-pinned to the host's locked versions ciregenz 2026-08-08 01:22:43 -07:00
  • 610ba00a9d [eric] apps: the host SDK lands, apps call the user's models, fire workflows, and spawn positioned agents through pre-wired helpers ciregenz 2026-08-08 01:14:00 -07:00
  • 1f3c211146 [eric] apps: generated apps default to an Apple design language, SF-first type, system blue, iOS grays, hairline structure ciregenz 2026-08-08 01:05:40 -07:00
  • c764604a53 [eric] canvas: a rejected rect no longer scores as trusted, so the settle can fix a stale origin instead of parking fullscreen 52px off ciregenz 2026-08-08 01:00:32 -07:00