Repository navigation
Upstream sync 6 (final): merge cloudflare-os up to 8060cd9f - #43
Merged
Merged
Conversation
…e#731) * feat: let a creation's approver configure it * refactor: share one disposeConfiguratorFrame * feat: host the creation configurator in the approval dialog * test: approve creations with and without a configuration
… when one changes (cloudflare#753) * Keep resource catalogs fixed in the system prompt, and tell the agent when one changes The system prompt lists each always-available resource with its catalog, such as the skills in the Context Library. The backend rebuilt that list on every turn, so a skill added or edited while a chat was open changed the system prompt. The next request then missed the prompt cache from the start, and on Claude models that drop thinking blocks after a prefix change, the chat lost its earlier reasoning. The chat now saves the list its first turn shows, and each compaction checkpoint saves its own. Before each model request, the agent compares the live catalogs with the ones it last saw and records a catalogChange message for each one that differs. History replay shows that message as a synthetic observeUserChanges call that lists the entries added and the ones no longer listed. A catalog that fails to load is not a change, so a failed load no longer costs two cache misses. * Sign Gemini's unsigned function calls, and share the synthetic observation builder Gemini 3 rejects a request whose current turn holds a function call without a thought signature. The catalog note's synthetic observeUserChanges call lands after the user's message and has none, so the first Gemini request after a catalog change failed. The model handle's onPayload chain now gives each unsigned Gemini function call skip_thought_signature_validator, the value Google documents for a call the model didn't make. pi can't send it, because it passes on only base64 signatures. Callback turns with pending user edits, which already sent an unsigned call there, get the same fix. The user's edits, reverts and catalog changes now build their synthetic call through one function, makeReplayObservation.
…loudflare#767) * fix(native): relay authentication popups through the system browser * fix: harden native auth handoff recovery and billing refresh * fix: refresh global billing picker after native auth * fix: allow cancelling native connector authentication * fix: preserve billing choices on focus and settle native expiry
…dflare#766) A chat fixed its always-available bindings on its first turn, so a gatekeeper the user turned on afterwards, such as Web Search, reached only chats started after that. The rule dates from numbered env[] capsules, where a new one would shift the numbers. Bindings have names now. On each turn, a chat that isn't spawned now takes in each always-available gatekeeper the workspace gained. It names it from the gatekeeper's suggested binding name, made unique in the chat's scope. The system prompt keeps the list the chat started with, and the catalogChange note tells the agent about the new binding. The evals can now add Agent Skills to the user's Skills library before a turn. Change Calendar's third turn gets its new rules that way.
…ch (cloudflare#795) PUBSUB_TOPIC becomes PUBSUB_CHAT_TOPIC (Workspace Events subscriptions) and PUBSUB_GMAIL_TOPIC (users.watch), so a deployment can deliver each through the topic that grants its publisher, chat-api-push@ or gmail-api-push@. Setting both to one topic is the old single-topic setup; a deployment upgrading sets both to its PUBSUB_TOPIC. The push endpoint, its service account and its routing by payload shape are unchanged: each topic gets its own push subscription to the same `/pubsub`. Each kind of hook is refused when its topic is unset. The README also names the push-authentication IAM Google requires: actAs on the push service account for whoever creates a subscription, and getOpenIdToken on it for the Pub/Sub service agent, which its default role already grants; Token Creator on the account only if not.
Bumps the github-actions group with 2 updates: [actions/upload-artifact](https://github.com/actions/upload-artifact) and [actions/download-artifact](https://github.com/actions/download-artifact). Updates `actions/upload-artifact` from 7.0.1 to 7.0.2 - [Release notes](https://github.com/actions/upload-artifact/releases) - [Commits](actions/upload-artifact@043fb46...cf430e0) Updates `actions/download-artifact` from 8.0.1 to 8.0.2 - [Release notes](https://github.com/actions/download-artifact/releases) - [Commits](actions/download-artifact@3e5f45b...9000827) --- updated-dependencies: - dependency-name: actions/upload-artifact dependency-version: 7.0.2 dependency-type: direct:production update-type: version-update:semver-patch dependency-group: github-actions - dependency-name: actions/download-artifact dependency-version: 8.0.2 dependency-type: direct:production update-type: version-update:semver-patch dependency-group: github-actions ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
* Let users mute a workspace's finished-task notifications
Kernel half. Each user's `notificationMute` lives on the workspace's
record in their User DO: `{ until: Date }` or `{ indefinite: true }`,
mutually exclusive, set via `AuthenticatedApi.setWorkspaceNotificationMute`
(null unmutes). While it lasts, `publishNotification` drops the
workspace's `taskCompleted` notifications before any tab or device sees
them. Permission requests always come through, since the task is stuck
until the user answers. An expired mute is simply ignored, so no timer.
* Add Mute / Unmute to workspace menus
Workspace menus (sidebar row, /workspaces list, the editor's phone menu)
offer "Mute notifications" for 1 or 3 hours, until local midnight, 1 or 3
days or until turned back on, and "Unmute notifications" while muted. The
bell-slash shows on muted workspaces, titled with when the mute ends.
* Fix blueprint gatekeeper connection setup A blueprint binding was marked ready whenever exactly one connected account matched the vendor and resource type, and its suggested resource was seeded into the configurator unverified. That marked a Gmail-only Google account ready for a Calendar binding it had no grant for, and seeded the author's calendar into a recipient's form even though the recipient could not use it. Gatekeeper bindings now always go through configuration. Eligibility requires the resource grant, and an account missing it can expand the grant in place; new connections request only the binding's resource types. The configurator is seeded only from a previously saved assignment, and switching accounts clears it. Saved assignments become unusable as soon as their account disconnects, expires, or loses the grant, so Create cannot submit a stale account. * Verify Calendar write access before seeding a suggested calendar The Calendar configurator seeded any calendar ID it found in a suggested resource URL and reported the form ready, so a blueprint author's own calendar looked selectable to a recipient who cannot write it. The configurator now asks the gatekeeper whether the connected account can write that exact calendar and leaves the field unselected when it cannot. A 404 means the calendar is not on the account's list; other failures are rethrown so a transient error is not mistaken for missing access. * Recognize limited Calendar writers as writable Google reports a calendar shared with edit rights but no private-event visibility as writerWithoutPrivateAccess. The picker and the suggested- calendar check treated it as read-only, so such a calendar could not be chosen at all. It now counts as writable for choosing a calendar. Observer admission is deliberately unchanged: a collaborator must still be a full writer or owner, because the workspace may already hold private event details read through the owner's account. The two predicates live side by side in calendar-api.ts so the difference is explicit. * Keep a requested Calendar availability mode when the calendar lookup fails The configurator runtime discards every seeded value when initialValuesFromResourceUrl throws. Checking a suggested calendar now makes an API call for every calendar ID, so a transient failure there silently turned a requested all-visible connection into a this-calendar one. The availability mode is read from the URL alone, so it is now kept, and a failed lookup only leaves the calendar for the user to choose. * Prefer a blueprint binding account that already holds the grant The connection dialog selected the first account with live credentials, so a Gmail-only Google account listed ahead of one with Calendar access was chosen and asked to grant access it did not need. It now prefers an account that can configure the binding as is. The dialog also matched the blueprint's gatekeeper name case-sensitively while the rest of the page did not. It now resolves the vendor once, case-insensitively, and uses its id for accounts and connecting. * Publish blueprint accounts only once their subscription has replayed A restarted accounts subscription replays each account before ready(), and the page published the list after every replayed account. Until the replay finished, saved connections for accounts not yet replayed read as unavailable, and Create reopened the configurator instead of creating the Gadget. The list is now published from ready() onward, so a restart keeps the previous list until the new one is complete. --------- Co-authored-by: Nathan Disidore <nathan@cloudflare.com>
* Reject malformed agent edits before buffering and trace persistence * Keep blueprint failure expectation with early transform validation * Leave replica-only permission import with the replica PR * Keep recoverable file validation separate from replica synchronization and timings * Preserve read knowledge until agent edits buffer successfully
…ntials (cloudflare#769) * Extract scoped native Git bootstrap for GitHub and GitLab * Unify Git pulls and native bootstrap around raw pack streams * Preserve Git fetch errors across Workers RPC * Clarify Git pull authority and consumer documentation * Cover overseer Git pull transfer validation * Cancel unclaimed Git pull streams on disposal
…ept Google Chat's copied links (cloudflare#758) * Match a trailing slash followed by a query in matchesResourceUrlPattern The trailing-slash retry toggled the slash at the end of the whole URL, so a URL with a query or fragment, like https://github.com/o/r/?tab=readme, never tried /o/r. Toggle it on the parsed path. * Accept Google Chat links that carry Chat's cls view parameter Chat's own "Copy link" adds ?cls=..., which names no resource, so a pasted link was refused. Ignore cls and keep refusing any other query parameter or fragment; describe() still reports the canonical URL. * Match pasted resource URLs the way the backend does The composer's resource picker had its own segment matcher that read a query string as part of the path, so a Google Chat link ending in ?cls=7 was offered a thread completion appended after the query, lowercased, and its account row showed a "Grant access" that did nothing until the placeholder was filled. - Decide full matches with matchesResourceUrlPattern, the URLPattern test the backend applies, and keep the segment logic only for completing half-typed paths. Completions keep the pattern's case, go before any query or fragment the URL already has, and offer a pattern's own query (NetSuite's ?id=:customerId) only to a URL without one. - Send the URL exactly as pasted: gatekeepers such as Confluence, Slack and Calendar read its query. - While the URL has a placeholder, account rows name it ("Fill in :threadId first") instead of offering access, and the overlay shows that Esc dismisses it. - When the gatekeeper refuses a URL, the toast carries its message. * Complete resource URLs from URLPattern's parsed path Completions split the raw pattern string, so Tab could fill in pattern syntax ({/:rest}*, regex groups, a stray "?") that left the user stuck on a placeholder. Read the host and path from URLPattern instead and complete only up to the first optional segment. Drop query completion: it only served hand-typed NetSuite URLs, and its "=:" placeholder rule blocked real links such as ?q=:open. Reserve the overlay footer's height in the list budget, so short windows no longer push the panel past the top. * Resolve Google Chat message links to their thread A message's Copy link is /room/{space}/{thread}/{message}, which matched no resource pattern, so pasting one offered nothing. Let the thread pattern take an optional message id ({/:messageId}?), as GitLab's and X's patterns take a trailing tab, and have parseChatUrl ignore it, as the thread configurator already did. The pattern names the resource type in stored state, so read a grant recorded under the old pattern under the new one. Still keyed on the old string: an admin's disabled-resource entry for Chat threads, and the typeUrlPattern of existing thread bindings and blueprints exported from them. * Complete a pattern's choice between words to its first word A parameter such as :kind(room|dm) lost its regex like any other, so the picker completed chat.google.com/ to /:kind/:spaceId, a placeholder asking the user to spell the choice, and treated a half-typed "ro" as its value. Complete it to its first word instead. A full link with another word, such as /dm/..., still matches through URLPattern. * Resolve Google Chat direct message links Chat links to a direct message under /dm/ rather than /room/, so pasting one, or a link to a message in one, matched no resource and parseChatUrl refused it, though the thread configurator and the agent's space lookup already accepted it. Let the conversation and thread patterns take either word (:kind(room|dm)) and parse /dm/ as /room/; the configurators still mint /room/. The space configurator now prefills from such a link, and from one carrying Chat's cls parameter, which it refused before. The conversation pattern is renamed like the thread pattern, so a grant recorded under it is read under the new name. As with threads, an admin's disabled-resource entry and the typeUrlPattern of existing conversation bindings and blueprints exported from them still name the old pattern. * Complete every word of a pattern's choice, and a bare scheme in plain placeholders A choice such as :kind(room|dm) completed only to its first word, so a half-typed /dm/ path got no suggestion until the link was whole. Keep each required segment as the words it accepts: any of them matches, and a completion spells the first. A scheme still being typed completed to the raw pattern, so Tab could insert regexes and optional groups (x.com/:username([A-Za-z0-9_]{1,15}), Chat's :kind(room|dm) and {/:messageId}?) that no filled-in URL could match. Complete it from the same required segments as the path. * Parse a resource URL once in matchesResourceUrlPattern Toggle the trailing slash on the parsed URL itself and test the URL as given, rather than keeping the parsed URL alongside a copy of it. * Refresh an account's stored description when nothing is left to grant The Workshop judges an account's grants from the description stored at its last connect or reconnect. After a gatekeeper renames a resource's urlPattern (Google's Chat conversation and thread patterns, on this branch), that copy names the old pattern, so the account shows "Grant access" for a resource it holds; asking the gatekeeper finds nothing to grant, and the button did nothing until a full reconnect. ensureAccountResources now refreshes the stored description when the gatekeeper returns no flow, which every caller reaches only after the stored copy showed a grant missing. It re-reads the record after describe(), so a disconnect while describe() ran is not undone. The refresh is best-effort: a caller such as the observer setup reads null as access confirmed and continues, which a failed describe() must not turn into an error. * Toggle the configurator prefill's trailing slash on the path The fallback that seeds configurator values from a resource URL toggled the trailing slash on the whole URL, so .../folders/FOLDER123/?usp=sharing became ...?usp=sharing/ and seeded nothing, though matchesResourceUrlPattern now accepts that URL and routes it to the configurator. Toggle it on the parsed path, as matchesResourceUrlPattern does. * Complete a resource URL's path in its own case The picker lowercased every typed segment, but URLPattern matches paths case-sensitively, so typing RO against :kind(room|dm), or i against a literal Issues, completed to a URL that matched nothing. Lowercase only the host, as URLPattern does the pattern's, and compare path segments exactly. * List a pasted URL's full matches in the order the gatekeeper declared them Sorting full matches by pattern string put the broadest first, so a GitLab merge request link was offered as a GitLab Project before the Merge Request row. * Complete a resource pattern's required query A pattern that requires a query, such as NetSuite's custjob.nl?id=:customerId, completed to its path alone, which can never match. Completions now end with the pattern's query, its values left for the user to type. * Explain why ensureAccountResources re-reads the account's description * Note which saved Chat resource names the rename doesn't translate * Complete to a pattern's query only a URL the rest of the pattern matches A URL with every required path segment was offered the pattern's query even with segments after them or the wrong scheme, so the completed URL still matched nothing. The query is now offered only when the URL matches the pattern with any query in its place. * refactor: simplify the logic
) * Search through the deployment's AI Gateway in Web Search The Web Search gatekeeper always searched through the account's "default" AI Gateway, so a deployment's own gateway, with its stored provider keys, limits and logs, never saw the searches. It now reads the gateway from CF_AI_GATEWAY, the same var and value as the backend's. The deploy service sets it to the instance's gateway; anywhere else it stays "default". * Keep CF_AI_GATEWAY out of the Web Search config Previews upload a backend secret named CF_AI_GATEWAY, and staging-config.test.ts fails on any config that declares a var under a secret's name. The gatekeeper now reads the var when the deploy service sets it and falls back to "default" in code, like the backend, which also leaves it undeclared.
…loudflare#754) * Type-check in the Lint job instead of before the tests `pnpm build` ran before `pnpm test` in the Build and test job, adding about a minute to the longest job. The test tasks declare the codegen they read, so they do not need it. It now runs in the Lint job, which is also a required check, and runs even when a lint step fails so one run still reports both kinds of error. Lint gets a 10-minute timeout, since it now runs the build. * Run the test suites in parallel jobs `Build and test` ran every package's tests in one job on a 4 vCPU runner, so the workshop-backend suite shared the CPU with every other suite and set the pace. Now workshop-backend, workshop-frontend and integration-tests each run in their own job, and the other packages share a fourth. Each job runs the same `test` task as `pnpm test`, selected with `vp run --filter`. A job that tests one package runs one suite at a time, so it gives that suite 3 vitest workers instead of the 2 that concurrency.ts budgets for 4 concurrent suites. main's ruleset requires a check named `Build and test`, so a job with that name waits for the four and passes only if all of them pass. No job restores the dependency cache any more: restoring it took longer than downloading the packages (setup p50 29 s against 20 s).
* Bump the editor-codemirror group with 4 updates Bumps the editor-codemirror group with 4 updates: [@codemirror/language](https://github.com/codemirror/language), [@codemirror/view](https://github.com/codemirror/view), [@lezer/highlight](https://github.com/lezer-parser/highlight) and [@codemirror/legacy-modes](https://github.com/codemirror/legacy-modes). Updates `@codemirror/language` from 6.12.4 to 6.13.1 - [Changelog](https://github.com/codemirror/language/blob/main/CHANGELOG.md) - [Commits](https://github.com/codemirror/language/commits) Updates `@codemirror/view` from 6.43.13 to 6.43.14 - [Changelog](https://github.com/codemirror/view/blob/main/CHANGELOG.md) - [Commits](https://github.com/codemirror/view/commits) Updates `@lezer/highlight` from 1.2.4 to 1.2.5 - [Changelog](https://github.com/lezer-parser/highlight/blob/main/CHANGELOG.md) - [Commits](https://github.com/lezer-parser/highlight/commits) Updates `@codemirror/legacy-modes` from 6.5.4 to 6.5.5 - [Changelog](https://github.com/codemirror/legacy-modes/blob/main/CHANGELOG.md) - [Commits](https://github.com/codemirror/legacy-modes/commits) --- updated-dependencies: - dependency-name: "@codemirror/language" dependency-version: 6.13.1 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: editor-codemirror - dependency-name: "@codemirror/view" dependency-version: 6.43.14 dependency-type: direct:production update-type: version-update:semver-patch dependency-group: editor-codemirror - dependency-name: "@lezer/highlight" dependency-version: 1.2.5 dependency-type: direct:production update-type: version-update:semver-patch dependency-group: editor-codemirror - dependency-name: "@codemirror/legacy-modes" dependency-version: 6.5.5 dependency-type: direct:production update-type: version-update:semver-patch dependency-group: editor-codemirror ... Signed-off-by: dependabot[bot] <support@github.com> * Deduplicate @lezer/common and @lezer/lr in the lockfile The bump moved @codemirror/language and @lezer/highlight onto @lezer/common 1.5.3 while @codemirror/lang-markdown and @lezer/markdown stayed on 1.5.2. NodeProp ids are allocated per module instance, so lang-markdown's headingProp collided with highlight's rule prop and highlighting `# Heading` threw "tags is not iterable", disabling syntax highlighting in the Workshop and Context Library editors. Resolve every consumer to @lezer/common 1.5.3 and @lezer/lr 1.4.11. --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Nathan Disidore <nathan@cloudflare.com>
Bumps the remaining-npm group with 2 updates: [isomorphic-git](https://github.com/isomorphic-git/isomorphic-git) and [motion](https://github.com/motiondivision/motion). Updates `isomorphic-git` from 1.42.2 to 1.43.1 - [Release notes](https://github.com/isomorphic-git/isomorphic-git/releases) - [Commits](isomorphic-git/isomorphic-git@v1.42.2...v1.43.1) Updates `motion` from 13.4.3 to 13.5.1 - [Changelog](https://github.com/motiondivision/motion/blob/main/CHANGELOG.md) - [Commits](motiondivision/motion@v13.4.3...v13.5.1) --- updated-dependencies: - dependency-name: isomorphic-git dependency-version: 1.43.1 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: remaining-npm - dependency-name: motion dependency-version: 13.5.1 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: remaining-npm ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
* Let a resource configurator ask for the access its next step needs A configurator runs in a sandboxed iframe that cannot open pop-ups, so it has no way to ask for an OAuth scope it lacks. ResourceConfiguratorHost gains ensureResources(resourceUrlPatterns): Workshop resolves true when the account the configurator was opened for already holds grants covering them, and otherwise opens the provider's consent pop-up and resolves false, leaving the configurator to re-check through its own gatekeeper capability. It must be called from a click in the frame so the browser allows the pop-up. configurator-ui exposes the call on the render context, and declares a Button and a Notice for the step that uses it. The Slack and Cloudflare configurator tests build a render context by hand, so they gain the member.
* Keep Google Chat link URLs in messages * Re-read Chat's stored markup when an edit's response omits it
* Await the share-link rejection assertions that never ran Four tests built an expect(...).rejects chain without awaiting it, so the assertion settled after the test had already passed. Vitest 5 fails such floating assertions; await them so they actually check something. * Move the test toolchain to Vitest 5 and @cloudflare/vitest-plugin @vitest/browser-playwright 5 needs Vitest 5, and @cloudflare/vitest-pool-workers never supported it: the package was renamed @cloudflare/vitest-plugin, which added Vitest 5 in 1.4.0. Rename it everywhere and move vitest to ^5.0.5. vite-plus goes to ^1.2.0, which bundles Vitest 5.0.4 rather than 1.0.0's 5.0.1. 1.3.0 bundles 5.0.5 and would remove the last duplicate copy, but is still inside the release-age window. The plugin keeps running on the current Wrangler/Miniflare stack: the cloudflare-toolchain bumps (cloudflare#785) make google-hooks-upgrade flaky (ECONNRESET, 3-5 failures in 8 runs against none on main) and are left for their own change. vitest-evals 0.17 still declares vitest <5 but works unchanged on it (harness, judges, reporter and JSON output checked), so a peerDependencyRules entry allows it until a release widens the range. Supersedes cloudflare#791 and cloudflare#786. * Run vp's vitest tasks without Vitest 5's filesystem module cache Vitest 5 turns fsModuleCache on by default, and every package's vitest shares one cache under the workspace root's node_modules. vp runs those tasks in parallel, so when one run invalidates the cache it deletes the directory while another is writing into it, and that run fails to start with ENOTEMPTY. vp already caches the whole task, so the builder appends --no-fsModuleCache to each vitest command, as it already pins the reporter. * Turn off Vitest 5's filesystem module cache for the backend's Workers suite vp's tasks already run without it, but a direct vitest run in the package still uses it, and under the Workers pool every run after the first fails: cached .txt modules come back with root-relative ids the textModules plugin cannot read, and the pool's WebSocket rejects the larger cached modules (1009, Message is too large).
cloudflare#797) * fix: resume decisions through existing durable agent records * fix: recover explicit compaction across durable object resets * fix: close action cursor before adopting legacy agent turns * fix: bound durable resume retries and preserve alarm delivery * fix: clarify resume authority and avoid shared fixture conflicts * fix: notify callers when recovering logged agent outcomes
* Project text attachments into independent chat worktrees * Support byte-preserving attachment worktrees * Preserve binary attachments and opaque file state in Workshop * Simplify blob attachments and reuse chat scratch worktrees * Preserve local blob reads and guard model image dimensions * Migrate legacy attachments and harden blob-backed worktree reads * Store attachments and binary writes as ordinary Git blobs * Address attachment review nits and rename races * Harden attachment migration and many-image replay * Use scratch worktree naming throughout attachment ingress
…flare#811) Reviews of gatekeepers tend to stop at general correctness. Add a review checklist to the write-gatekeeper skill covering the API's shape, common user stories, and whether bindings and sessions partition authority along the lines users grant, and point REVIEW.md at it.
Brings in the rest of upstream's first-parent history, cloudflare#241-cloudflare#268 (through "Ask gatekeeper reviewers to pressure-test the agent-facing API", 2026-10-17), which finishes the sync: uploaded files as ordinary workspace files in a chat's scratch worktree, durable resumption of approved agent turns, catalogs fixed in the system prompt with change notices, gadget exports on chat bindings, workspace notification mutes, gatekeeper user-selection tickets, a creation's approver configuring it, native-app sign-in fixes, Drive Office conversion on connect, and Vitest 5. Resolutions that keep Tyms behaviour: - Attachments: upstream's scratch-worktree ingress, plus the fork's Office conversion moved to it. An Office/OpenDocument upload stays byte-for-byte and gets a best-effort Markdown rendition beside it (report.docx.md), recorded as ChatAttachmentRef.renditionPath; the agent is told to read it. - Agent turns: upstream's ownership checks and durable resume, plus the fork's supersession check; the alarm keeps the fork's 10-minute hold and hand-over inside upstream's retrying, resuming alarm. Applying an action or a configured creation keeps the fork's apply deadline. Upstream's inline teardown replaces the fork's #tearDownAgentTurn. - Catalogs: upstream's per-turn catalogs with change notices; the fork's per-chat skill and Knowledge prompt-context snapshots are kept beside them. - User DO: the fork's single alarm (#armAlarm) also sweeps upstream's picker selections. - Pinning moves to AuthenticatedApi.setWorkspacePinned, as upstream did. - Frontend: lucide icons (BellOff for mutes), App/template wording, and the fork's failureTitle toasts with upstream's "Could not complete" wording. Tests: durable-decision-resume is skipped (it sets up per-user models), as are the per-user-model integration tests upstream added; new tests cover the Office rendition and its failure path. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
I have read the CLA Document and I hereby sign the CLA 2 out of 7 committers have signed the CLA. |
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
7 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.
Step 6, the last of the upstream sync. It merges the remaining upstream first-parent commits cloudflare#241–cloudflare#268, through
8060cd9fon 2026-10-17, and stacks on #42, which is already merged. That's 31 conflicted files. After this the fork is 0 commits behind upstreammain.What upstream brings:
Resolutions worth reviewing
report.docx.md), recorded asChatAttachmentRef.renditionPath.readFile.#tearDownAgentTurn; it was the same steps.AuthenticatedApi.setWorkspacePinned, as upstream did.BellOfffor mutes.failureTitletoasts, with upstream's "Could not complete…" wording.Tests
durable-decision-resume, 34 tests;useServerConfig, async logout).Verification (local)
configs:check.libasound2.main280 → 280). Warnings rose from 324 to 342, in upstream's code.After merging
The sync is complete. Next is bumping the
cloudflare-ossubmodule in tyms-os, then a careful deploy, starting with one tenant, because of the new Durable Object migrations:🤖 Generated with Claude Code