Repository navigation
feat(browser): give each conversation its own built-in browser tabs - #4434
Merged
Merged
Conversation
Contributor
|
The latest updates on your projects. Learn more about Vercel for GitHub. 4 Skipped Deployments
|
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
![diff-warden[bot]](https://avatars.githubusercontent.com/u/118544361?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Warden security clearance: clear. No blocking findings (security or desktop↔den sync) in this diff (fb940de88f0acef07a99cf2641f7215d9170a518). Automated clearance satisfies the required-review gate only — a human still reviews and merges. Analysis run
reachjalil
enabled auto-merge (squash)
September 4, 2026 20:14
A conversation working in the background opened its browser tab in whatever conversation the user was reading: the native browser kept one global tab list with no owner, the agent's session id was dropped before the UI command bridge, and the side panel mirrored every native tab into whichever conversation was mounted. Tabs now belong to the conversation whose agent opened them. The bridge stamps UI commands with the requesting session, the control bus hands that origin to the action, and the Electron browser keeps an owner per tab. Only the on-screen conversation's tabs may take the screen; a tab opened for a background conversation loads silently there and is waiting when the user switches to it. Hidden tabs stay real pages for the agent driving them: a one-pixel on-window presence keeps Chromium painting while a main-process DevTools session emulates a full viewport and focus, so reads, clicks, typing, and screenshots all work before anyone looks. Both are undone when the tab returns to the screen. The ownership, surfacing, and background-rendering policy lives in the new @openwork/browser-tabs package, shared by the main process and the renderer and tested on its own.
reachjalil
force-pushed
the
feature/thread-owned-browser-tabs
branch
from
September 4, 2026 20:16
c1ff92b to
9d19c1b
Compare
The Den API image copies apps/desktop/package.json for version metadata and the gateway image builds apps/app; both now depend on @openwork/browser-tabs, so the pruned workspace installs need the package present.
reachjalil
force-pushed
the
feature/thread-owned-browser-tabs
branch
from
September 4, 2026 20:32
9d19c1b to
6b4a3c0
Compare
Contributor
There was a problem hiding this comment.
Warden security clearance: clear. No blocking findings (security or desktop↔den sync) in this diff (6b4a3c01d4a370b057046f1f973593221de81edb). Automated clearance satisfies the required-review gate only — a human still reviews and merges. Analysis run
…tabs journey The claim that a person can switch to the background conversation and find its page waiting was driven through the app's control rail. Click the conversation in the sidebar instead, as a user does, and give the first conversation a stable title so the click has a target.
Member
Author
Test evidence — a-background-conversation-s-agent-browses-silently-and-its-page-is-waiting-when- — ✅ passed✅ passed · 0 expectations passed · 0 failed · 0 pending
[world] desktop(daytona) · workspace(/tmp/openwork-builtin-browser-1788555078931) · session("New task") · [seed:raw] evalIn() Test run created 2026-09-04T20:49:42.394Z · Source: |
benjaminshafii
approved these changes
Sep 4, 2026
benjaminshafii
left a comment
benjaminshafii
left a comment
Member
There was a problem hiding this comment.
Reviewed the conversation ownership registry, origin stamping, and latest journey changes. The outstanding switching comment is addressed: the test now clicks named conversations through user.click, and the fresh published Daytona run passes all four steps. Cleared that resolved finding; merge remains subject to required checks.
auto-merge was automatically disabled
September 4, 2026 20:55
Pull Request is not mergeable
Contributor
There was a problem hiding this comment.
Warden security clearance: clear. No blocking findings (security or desktop↔den sync) in this diff (68254accdefff4668c86846c2d775e44eb5e86b7). Automated clearance satisfies the required-review gate only — a human still reviews and merges. Analysis run
Member
Author
|
Heads-up from the stacked PR #4452: after this branch moved onto newer |
This was referenced Sep 5, 2026
yomgui
added a commit
that referenced
this pull request
Sep 11, 2026
… person's tab, not agent browser control (#4779) * fix(app): seed a created session's snapshot so its composer can send immediately A freshly created session had no cached snapshot, so deriveSessionRenderModel reported "switching" while the first snapshot round trip ran and the composer disabled Run task. On Daytona that window is ~0.9 s and a click or Enter inside it was silently dropped with the draft intact. Seed the snapshot cache from the create response (stale, so the real fetch still runs as a background refresh). * fix(app): let a send that settled with a server failure recover instead of wedging A prompt_async answered with 5xx/408 is an uncertain admission (#4635), but the hold it created could never end: the acceptance check treated absence as no evidence, Stop refused while the admission was unknown, and the failure body was dropped, so the person saw a paused composer with no explanation. - Keep the settled response body on PromptAdmissionUnknownError and show it as the session error card right away; the hold still prevents any resend. - readPromptAdmission: after the exact-message GET misses, list the idle conversation; when native lists it without the message the POST is settled and did not admit, so the drain receives admission_rejected and halts like any definite failure (explicit retry only, never an automatic resend). - On release the pending row is removed and the submitted composer is kept as an unsent message to restore; on confirmed acceptance the stale failure card is cleared. * test(evals): read an empty Lexical composer as an empty draft Lexical keeps one empty paragraph in an empty editor, which innerText reports as "\n". readComposerState now drops that trailing line break so draftText matches what the person sees; user.see("composer", { text }) already trimmed. * test(evals): open the fresh conversation through the sidebar's New session button The sidebar button was renamed from "New task" to "New session" in #4491 (Sep 5) and the "new" scenario added in #4631 (Sep 8) still targeted the old label, so it could never find the button. Same user action, current label. * test(evals): open the saved-app verification return tab as a person's new tab returnVerification drove the agent browser entry point (browser.openUrl, the IPC behind browser.open_url) with no requesting conversation. Since #4434 the task host refuses that, and it never accepted an about:blank destination. The tab only stands in for the person's own browser so the openwork:// return link can be navigated without OS protocol registration, so create it the way a person opens a new tab and find its CDP target by diffing the page targets.
yomgui
added a commit
that referenced
this pull request
Sep 11, 2026
…re the spec acts on it (#4794) * fix(app): seed a created session's snapshot so its composer can send immediately A freshly created session had no cached snapshot, so deriveSessionRenderModel reported "switching" while the first snapshot round trip ran and the composer disabled Run task. On Daytona that window is ~0.9 s and a click or Enter inside it was silently dropped with the draft intact. Seed the snapshot cache from the create response (stale, so the real fetch still runs as a background refresh). * fix(app): let a send that settled with a server failure recover instead of wedging A prompt_async answered with 5xx/408 is an uncertain admission (#4635), but the hold it created could never end: the acceptance check treated absence as no evidence, Stop refused while the admission was unknown, and the failure body was dropped, so the person saw a paused composer with no explanation. - Keep the settled response body on PromptAdmissionUnknownError and show it as the session error card right away; the hold still prevents any resend. - readPromptAdmission: after the exact-message GET misses, list the idle conversation; when native lists it without the message the POST is settled and did not admit, so the drain receives admission_rejected and halts like any definite failure (explicit retry only, never an automatic resend). - On release the pending row is removed and the submitted composer is kept as an unsent message to restore; on confirmed acceptance the stale failure card is cleared. * test(evals): read an empty Lexical composer as an empty draft Lexical keeps one empty paragraph in an empty editor, which innerText reports as "\n". readComposerState now drops that trailing line break so draftText matches what the person sees; user.see("composer", { text }) already trimmed. * test(evals): open the fresh conversation through the sidebar's New session button The sidebar button was renamed from "New task" to "New session" in #4491 (Sep 5) and the "new" scenario added in #4631 (Sep 8) still targeted the old label, so it could never find the button. Same user action, current label. * test(evals): open the saved-app verification return tab as a person's new tab returnVerification drove the agent browser entry point (browser.openUrl, the IPC behind browser.open_url) with no requesting conversation. Since #4434 the task host refuses that, and it never accepted an about:blank destination. The tab only stands in for the person's own browser so the openwork:// return link can be navigated without OS protocol registration, so create it the way a person opens a new tab and find its CDP target by diffing the page targets. * test(evals): temporary Z2 diagnostics around dashboard delete route change * test(evals): saved-apps world.open returns once the destination page root has rendered * test(evals): remove temporary Z2 diagnostics from saved-app-creation
yomgui
added a commit
that referenced
this pull request
Sep 11, 2026
… route-commit wait in world open() (#4810) * fix(app): seed the delegated-task eval with a running parent thread The eval.task_activity.seed hook rendered an in-flight child under a parent whose thread status stayed "ready": it only marked the activity store busy. Since #4635 MessageList reads that as a stopped parent and the unobserved child shows "Waiting for task result" instead of the Working shimmer, so task-activity-shimmer and parent-child-permission-approval lost their [data-subagent-activity="shimmer"] row. A delegating parent is mid-run; mark the eval thread streaming like the chat-loading seed already does. * fix(app): seed a created session's snapshot so its composer can send immediately A freshly created session had no cached snapshot, so deriveSessionRenderModel reported "switching" while the first snapshot round trip ran and the composer disabled Run task. On Daytona that window is ~0.9 s and a click or Enter inside it was silently dropped with the draft intact. Seed the snapshot cache from the create response (stale, so the real fetch still runs as a background refresh). * fix(app): let a send that settled with a server failure recover instead of wedging A prompt_async answered with 5xx/408 is an uncertain admission (#4635), but the hold it created could never end: the acceptance check treated absence as no evidence, Stop refused while the admission was unknown, and the failure body was dropped, so the person saw a paused composer with no explanation. - Keep the settled response body on PromptAdmissionUnknownError and show it as the session error card right away; the hold still prevents any resend. - readPromptAdmission: after the exact-message GET misses, list the idle conversation; when native lists it without the message the POST is settled and did not admit, so the drain receives admission_rejected and halts like any definite failure (explicit retry only, never an automatic resend). - On release the pending row is removed and the submitted composer is kept as an unsent message to restore; on confirmed acceptance the stale failure card is cleared. * test(evals): read an empty Lexical composer as an empty draft Lexical keeps one empty paragraph in an empty editor, which innerText reports as "\n". readComposerState now drops that trailing line break so draftText matches what the person sees; user.see("composer", { text }) already trimmed. * test(evals): open the fresh conversation through the sidebar's New session button The sidebar button was renamed from "New task" to "New session" in #4491 (Sep 5) and the "new" scenario added in #4631 (Sep 8) still targeted the old label, so it could never find the button. Same user action, current label. * test(evals): create the empty-session workspace at its declared tmp path The emptySession world declared /tmp/openwork-empty-session-<ts> but seed.workspace() adopted the existing first-launch default workspace instead ("OpenWork Chat" under the dev profile home). On Daytona that profile lives inside the repo checkout, so the engine merged /workspace/.opencode/opencode.json ("permission": "allow") over the workspace's own catch-all and composer-run-mode never saw the workspace MCP rule take effect. * test(evals): name the server error when a run-mode read is not 200 * fix(app): reuse the running OpenWork server when the selected workspace is server-owned Workspaces created from the app live only in the server registry, so a renderer reload reaches startServerWithoutDesktopWorkspace and restarted the embedded server (and its engine) although it was healthy and already serving that workspace. Reuse it when it is ready with the preferred remote-access mode; only restart when nothing usable is running. * test(evals): open the saved-app verification return tab as a person's new tab returnVerification drove the agent browser entry point (browser.openUrl, the IPC behind browser.open_url) with no requesting conversation. Since #4434 the task host refuses that, and it never accepted an about:blank destination. The tab only stands in for the person's own browser so the openwork:// return link can be navigated without OS protocol registration, so create it the way a person opens a new tab and find its CDP target by diffing the page targets. * fix(app): re-read pending questions when a session reaches a terminal edge An interrupted question tool leaves the engine's pending list through its finalizer without a question.rejected event, so Stop on a delegated child left its question (and option buttons) in the parent view. Reconcile questions alongside permissions on terminal edges, for both engines. * Revert "fix(app): re-read pending questions when a session reaches a terminal edge" This reverts commit 9e70982. * fix(app): withdraw the stopped tree's unanswered questions after Stop OpenCode v1 runs tools on detached fibers: aborting a session marks the question tool part as aborted but leaves its request in GET /question, so a parent that stopped a delegated child kept showing the child's question and its option buttons. After Stop confirms the tree idle, reject the pending questions owned by the stopped sessions; other roots are left untouched. * test(evals): temporary Z2 diagnostics around dashboard delete route change * test(evals): saved-apps world.open returns once the destination page root has rendered * test(evals): create the permission-stop-recovery workspace at its declared tmp path Test 2 of parent-child-permission-approval ran bash without asking although the world wrote permission.bash="ask" to the workspace opencode.json. The world called seed.workspace() without create, so it adopted the first-launch default workspace under the dev profile, which on Daytona lives inside the repo checkout; the engine merges /workspace/.opencode/opencode.json ("permission": "allow") after the workspace file and the agent ruleset ends as [* allow, bash ask, * allow] (last match wins; permissions/effective reported shell allow, source engine, workspace file .../home/OpenWork Chat/opencode.jsonc). Same trap and same remedy as #4747 for emptySession; opt-in so the other splitPaneQuestions worlds keep their arrangement. * fix(app): withdraw the stopped tree's unanswered permission requests after Stop OpenCode v1 runs tools on detached fibers: aborting a session marks the tool part as aborted but leaves its Permission.ask pending in /permission, exactly as #4786 found for questions. The stopped tree would keep showing Allow once / Deny in every pane and after reload although nothing can consume the answer. After stopForegroundTree confirms every stopped session idle, also list pending permissions and reject the ones owned by the stopped tree; other roots are never touched and each withdrawal kind is best-effort on its own. * test(evals): remove temporary Z2 diagnostics from saved-app-creation * test(evals): make locate() misses list every candidate of the requested role plus route and page roots A miss used to quote only the first 8 visible buttons in DOM order, which are always the shell rail, so every miss read as "the page has no buttons" (#4794's diagnosis detour). The browser-side report now lists every rendered element of the requested role (buttons/links for role-less targets), the current route/hash, and whether the [data-app-header] / [data-dashboard-page] roots are mounted. The Node side caps the list at 40 and says how many the page really had. Matching is unchanged. * test(evals): run the serialized locate() miss callback against a fake page DOM Proves the browser-side branch the way Chromium will execute it: the report lists every rendered menuitem on a dashboard whose shell rail has nine buttons, names the route and mounted page roots, and a role-less miss lists all nine buttons instead of the first eight.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
You read one conversation while another conversation's agent needs a web page. The browser pops open in the conversation you are reading, showing a page that has nothing to do with it, and the working conversation shows nothing. The built-in browser kept one global tab list with no owner,
openwork_executedropped the agent's session id before the UI command bridge, and the side panel mirrored every native tab into whichever conversation was mounted. Any navigation of a hidden tab also auto-selected it onto the screen.What changes
origin, set from the tool context, never by the agent), the control bus hands that origin to the action, and the Electron browser keeps an owner per tab. Each side panel shows only its own tabs.panel-opened/panel-closedname the owner.@openwork/browser-tabspackage holds the ownership, surfacing, and background-rendering policy, shared by the main process and the renderer and unit-tested on its own.What stays the same
Tabs opened without a conversation (no session on screen) stay shared and behave exactly as before. The visible conversation's own tabs surface as before.
browser_*tools and the CDP handle contract are unchanged;open_urladditionally returnsowner_session_idandvisible.Proof
New journey
evals/specs/browser-tabs-owned-by-thread.e2e.test.ts: a background conversation opens a page — the on-screen conversation's tab, panel, and viewport are untouched; the hidden page lays out at 1280×800, reports focus, accepts a dispatched click and typed text, and screenshots at full size; switching to the owner shows its tab at the panel's viewport; switching back restores the first conversation's own tab only.OPENWORK_EVAL_REF=fb940de88f0acef07a99cf2641f7215d9170a518 pnpm evals:e2e browser-tabs-owned-by-thread—placement: daytona (daytona CLI authenticated), checkout resolvedfb940de88f0a, 1 passed / 0 failed / 0 skipped.pnpm evals:e2e browser-panel-viewport-recovery --local— 1 passed (the existing built-in browser journey, same world).packages/browser-tabs: 12 policy tests;apps/desktop/electron/browser-panel.test.mjs: 9 (5 new, stubbedWebContentsView);apps/serverplugin tests: 26 (1 new: origin stamping cannot be spoofed by the agent).@openwork/app,openwork-server,typecheck:electron,check:electron,evals lint:layers, boundary and channel ratchets: clean.