Repository navigation
fix(server): integrate launch retry keys with current main - #22446
zainmitchell wants to merge 15 commits into
Conversation
| input.files.map(async (relativePath) => { | ||
| const result = await executeAtomQuery( | ||
| appAtomRegistry, | ||
| projectEnvironment.readFile({ |
There was a problem hiding this comment.
🟠 High lib/addProjectToMachine.ts:148
Selecting an ignored binary file always produces a failed copy: projectEnvironment.readFile rejects files containing NUL bytes, and this read failure is recorded in failed instead of transferring the file. Use a binary-safe transfer path or exclude unsupported files before offering them in the dialog.
🤖 Copy this AI Prompt to have your agent fix this:
In file @apps/web/src/lib/addProjectToMachine.ts around line 148:
Selecting an ignored binary file always produces a failed copy: `projectEnvironment.readFile` rejects files containing NUL bytes, and this read failure is recorded in `failed` instead of transferring the file. Use a binary-safe transfer path or exclude unsupported files before offering them in the dialog.
There was a problem hiding this comment.
This file is imported unchanged from main by the support merge. Git blob 367420c45e0f526f7ee6a878bb008abc59c8ee2f is identical at this PR's head ec2eba74, incorporated main f7f9e217, and main 5d4c202e (upstream source). It is outside both the launch-retry layer and the combined mode-limit/launch feature changes.
I am preserving the upstream file to keep this PR focused on integrating the existing stack. This comparison does not reproduce or dismiss the reported bug. Maintainer review of the merge remains required.
There was a problem hiding this comment.
Sorry, I'm unable to act on this request because you do not have permissions within this repository.
| @@ -0,0 +1,10 @@ | |||
| !macro customInstall | |||
| !ifndef DO_NOT_CREATE_START_MENU_SHORTCUT | |||
| ${IfNot} ${FileExists} "$newStartMenuLink" | |||
There was a problem hiding this comment.
🟡 Medium resources/installer.nsh:3
When $keepShortcuts is true, this recreates a Start Menu shortcut the user deliberately deleted on every update: customInstall treats the missing link as needing repair after the installer preserves its absence. Gate shortcut creation on $keepShortcuts so the user's choice is honored.
🤖 Copy this AI Prompt to have your agent fix this:
In file @apps/desktop/resources/installer.nsh around line 3:
When `$keepShortcuts` is true, this recreates a Start Menu shortcut the user deliberately deleted on every update: `customInstall` treats the missing link as needing repair after the installer preserves its absence. Gate shortcut creation on `$keepShortcuts` so the user's choice is honored.
There was a problem hiding this comment.
This file is imported unchanged from main by the support merge. Git blob 092f641c3359fb58bf2232132e9bd4ef0491ac56 is identical at this PR's head ec2eba74, incorporated main f7f9e217, and main 5d4c202e (upstream source). It is outside both the launch-retry layer and the combined mode-limit/launch feature changes.
I am preserving the upstream file to keep this PR focused on integrating the existing stack. This comparison does not reproduce or dismiss the reported bug. Maintainer review of the merge remains required.
There was a problem hiding this comment.
Sorry, I'm unable to act on this request because you do not have permissions within this repository.
| const scaledHeightDifference = previous | ||
| ? (previous.height - current.height) * current.scale | ||
| : 0; | ||
| if (Math.abs(current.height - win.innerHeight / current.scale) <= 1) { |
There was a problem hiding this comment.
🟠 High lib/softwareKeyboard.ts:48
On Firefox Android, opening the keyboard sets viewportOpen to false, so plain Enter can submit instead of preserving multiline input and the post-send keyboard blur is skipped. Firefox resizes both viewports together, making current.height equal win.innerHeight / current.scale here; navigator.virtualKeyboard is unsupported there, and the initial subscription check has the same limitation. Track keyboard visibility from a signal that does not require the layout and visual viewports to differ.
🤖 Copy this AI Prompt to have your agent fix this:
In file @apps/web/src/lib/softwareKeyboard.ts around line 48:
On Firefox Android, opening the keyboard sets `viewportOpen` to `false`, so plain Enter can submit instead of preserving multiline input and the post-send keyboard blur is skipped. Firefox resizes both viewports together, making `current.height` equal `win.innerHeight / current.scale` here; `navigator.virtualKeyboard` is unsupported there, and the initial subscription check has the same limitation. Track keyboard visibility from a signal that does not require the layout and visual viewports to differ.
There was a problem hiding this comment.
This file is imported unchanged from main by the support merge. Git blob e20d5fe00b4aa539792e0ee4a2d06658fc41c2cc is identical at this PR's head ec2eba74, incorporated main f7f9e217, and main 5d4c202e (upstream source). It is outside both the launch-retry layer and the combined mode-limit/launch feature changes.
I am preserving the upstream file to keep this PR focused on integrating the existing stack. This comparison does not reproduce or dismiss the reported bug. Maintainer review of the merge remains required.
There was a problem hiding this comment.
Sorry, I'm unable to act on this request because you do not have permissions within this repository.
ApprovabilityVerdict: Needs Human Approval Macroscope's review found this PR requires human approval — Despite the launch-retry title, the base-to-head tree is a 1,703-file merge containing production server orchestration, authentication, desktop/mobile/web behavior, release workflows, and Macroscope/static-analysis configuration changes. The unresolved High/Medium runtime findings further support keeping this change set with human reviewers. Needs human approval because:
Adjust the Minimum Blocking Severity for this repo — including turning it Off — in Settings. You can add or adjust custom eligibility rules. Learn more. |
An OAuth client on /mcp is capped by the access it was approved with. When another environment calls on behalf of one of its agents, that agent may run under narrower modes, and the target has no way to know. /mcp now reads `T3-Mode-Limit: <runtimeMode>/<interactionMode>` from OAuth clients and caps the request at the narrower of the two. The header can only narrow: a broader one changes nothing, and a malformed one is refused with 400. The client's interaction mode can now be capped too; before, it was always default. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A launch had no retry key, so an agent (or another environment calling on its behalf) that lost the response could only guess whether the thread was created, and a retry started a second one. t3_thread_launch now takes an optional clientRequestId. The command, thread and message ids derive from the caller and the key, and the launch service already replays a command it accepted, so a retry returns the first thread. Without a key every launch is new, as before. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… mode-limit layer
(cherry picked from commit c60c49f)
(cherry picked from commit 1fdb652)
(cherry picked from commit ebba131)
The launch section said launch has no retry key, but t3_thread_launch accepts clientRequestId and replays the first accepted launch for the same key within the caller's MCP session. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> (cherry picked from commit 452eb16)
452eb16 to
9621c64
Compare
|
Additional combined-stack runtime evidence is recorded on #22448: the reviewed integration was deployed on two Windows PCs with Nightly 3486, existing peer links reconnected after both server restarts, and native agent send/wait/read completed successfully across both machines. These results cover the combined Nightly integration, not this standalone PR head. This PR's code and target branch are unchanged. Exact source revisions, focused checks, model/harness attribution and verification limits are in the linked report. Maintainer direction/scope approval and review remain separate requirements. |
6152622 to
88ec543
Compare
An OAuth client on /mcp is capped by the access it was approved with. When another environment calls on behalf of one of its agents, that agent may run under narrower modes, and the target has no way to know. /mcp now reads `T3-Mode-Limit: <runtimeMode>/<interactionMode>` from OAuth clients and caps the request at the narrower of the two. The header can only narrow: a broader one changes nothing, and a malformed one is refused with 400. The client's interaction mode can now be capped too; before, it was always default. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
88ec543 to
3b46b4c
Compare
A launch had no retry key, so an agent (or another environment calling on its behalf) that lost the response could only guess whether the thread was created, and a retry started a second one. t3_thread_launch now takes an optional clientRequestId. The command, thread and message ids derive from the caller and the key, and the launch service already replays a command it accepted, so a retry returns the first thread. Without a key every launch is new, as before. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
3b46b4c to
1aa278b
Compare
Resolve the receipt-replay conflict in Orchestrator.ts by keeping the mode-ceiling check for replayed deferred launches and the target's queued-run.start replay. Drop a duplicate clientRequestId replay test that both sides added at different positions in handlers.test.ts. The target's launched-model lookup and this branch's accepted-launch lookup both stay: an accepted receipt still decides the replayed model first. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Conflict refresh: cdb82b0Merged the current target The resolution retains the receipt-validated launch retry and mode-ceiling recheck, upstream's new original-model lookup and selected queued-run replay, and the new upstream regression test. One byte-identical duplicate replay test was removed. Validation on upstream's required Node 26.12.1: 113 focused tests passed across six files; server typecheck and scoped format/lint passed, with only inherited lint warnings. Independent parent verification passed 15 launch-handler/peer-link tests on the #22447 candidate; the launch handler and orchestrator files are byte-identical between these two refreshed candidates. The parent command was The earlier two-machine runtime evidence remains evidence for its explicitly named older commits. It was not repeated on this new head. Maintainer scope approval and CI remain separate requirements. Resolution: Claude Opus 5.5 High through T3 Code. Independent source review and focused verification: GPT-6.1-Sol through Codex in T3 Code. |
A launch had no retry key, so an agent (or another environment calling on its behalf) that lost the response could only guess whether the thread was created, and a retry started a second one. t3_thread_launch now takes an optional clientRequestId. The command, thread and message ids derive from the caller and the key, and the launch service already replays a command it accepted, so a retry returns the first thread. Without a key every launch is new, as before. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
1aa278b to
071e661
Compare
The target's launch commit now skips new-launch checks when a retry finds its thread. Keep that for the worktree base ref and project defaults, and also skip the base ref for an accepted receipt as before. Keep checking existing-worktree membership on every launch: a retry may bind a workspace to an idle accepted launch, and this branch's tests require the refusal. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Launch-retry conflict refresh: 6d84a1eMerged the amended target Intentional resolution: retained Julius's earlier fix Upstream's new base-ref retry behavior and regression test are kept: a retry still succeeds after its original base ref disappears. Model precedence and the receipt-validated replay remain unchanged. Node 26.12.1 validation: 114 focused tests and server typecheck passed; scoped lint/format passed with two inherited warnings. Independently, the parent ran both named behavior tests on the combined #22448 candidate: the membership guard and deleted-base replay both passed. There is no new two-machine runtime claim for these heads. Resolution: Claude Opus 5.5 High through T3 Code. Independent source review and targeted verification: GPT-6.1-Sol through Codex in T3 Code. Maintainer scope approval and CI remain separate requirements. |
|
Documentation/reuse cross-link for launch-retry integration: The current review note's replay/worktree-membership trade-off remains explicit; stable evidence is not substituted for this head. The native messaging Ideas discussion and complete fork brief now bring together why/when/how, original/support PR dependencies, closed replacement history, later-layer scope, exact-head evidence and limitations. The corresponding core layer and selected fixes are represented in the version-pinned stable integration; it is not a copy of every current PR head. The stable source/download preview has its own manifest and evidence; its results do not establish this PR head's acceptance. This temporary fork is maintained until T3 supplies the capability or a better replacement. Original upstream authorship is retained. This is a documentation cross-link, not a new feature PR or merge request. Prepared with GPT-6.1 Sol through Codex in T3 Code. Implementation and historical model credits are stated in the linked brief. Factory catalog cross-reference — 2026-10-26. Support lineage for included launch identity: accepted launch receipts matter when a factory reconciles an ambiguous launch. This support PR is an alternative integration route for the original layer, not an extra required cherry-pick. Factory Ideas · Pinned fork source/guide · Source release. Version-pinned community experimental source preview; full lifecycle acceptance remains open. No upstream endorsement or current-PR-head test acceptance is claimed. Julius Marminge/T3 authorship is retained; later upstream heads are not acceptance evidence or a required cherry-pick sequence. Prepared with GPT-6.1 Sol through Codex in T3 Code. |
071e661 to
1c5e007
Compare
1c5e007 to
8a5f64e
Compare
The author rebased peer/launch-request-key again onto the updated mode-limit layer and a newer main (fe585bc); the layer itself is unchanged. This reapplies the previous resolution over that movement (merged against the previous target), so the support delta stays the same: replay before mutable preflights, worktree membership on every launch, accepted preparation and dispatch limits on replay, and the retry-key docs. - ThreadLaunchService.test.ts: the project mock keeps the support's getProject override and takes main's Effect.succeedSome; the support's MCP launch-retry test provides DocumentText the way the target's tests now do, since project launches now require it. - project/handlers.test.ts: the support's getAcceptedLaunch mock uses Effect.succeedNone, which the target's Effect diagnostics now require. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…#22448 The author rebased peer/forwarding again onto the updated peer-links layer and a newer main (fe585bc), adding DocumentText wiring, the succeedNone/succeedSome refactors and diagnostics comments. This reapplies the previous resolution over that movement (merged against the previous target), so the support delta is unchanged: bounded internal Stop retries with the global no-op Stop receipt sequence (Orchestrator and EventSink), child Stops carrying both stopDelegatedTasks: false and throughSequence, the internal thread.stop variant accepting stopDelegatedTasks, dequeued-run Stop targeting, the remote schedule secret-ref guard, forwarded tool validation and codecs, and the lower-layer launch and peer fixes. - Orchestrator.ts: main removed the unused layerUnavailable stub, so the support's readStoredEvents entry in it goes too. - PeerForwarding.test.ts: the target dropped its Option import for Effect.succeedNone, so the support's getAcceptedLaunch mock uses Effect.succeedNone as well instead of a now-undefined Option. - ThreadLaunchService.test.ts and project/handlers.test.ts: as in pingdotgg#22446, the support's MCP launch-retry test provides DocumentText the way the target's tests now do, and its getAcceptedLaunch mock uses Effect.succeedNone. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The author rebased peer/links again onto the updated launch-request-key layer and a newer main (fe585bc), adding Effect.succeedNone in PeerLinks and Schema.decodeEffect in PeerMcpClient. This reapplies the previous resolution over that movement (merged against the previous target); both refactors are kept and the support delta is unchanged: the pingdotgg#22446 launch fixes, event-stream answers matched by request id, sanitized link errors, and unlink removing only its own row's secret. Migrations keep main's 068/069 with PeerLinks at 070. ThreadLaunchService.test.ts and project/handlers.test.ts have the same inputs as pingdotgg#22446 and reuse its resolution (DocumentText for the MCP launch-retry test, Effect.succeedNone for the getAcceptedLaunch mock). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Summary
Keep launch retries tied to their accepted launch while integrating the updated
t3code/peer/launch-request-keybase at8a5f64e97.Preserves caller mode ceilings and checks existing-worktree membership on retries. Keeps upstream's current replay fallback, contracts and tool instructions. This updates the existing support PR; maintainer feature and workflow approval remain separate.
Evidence
b81fe3355.820e789c3is historical and does not verify this updated head.Merge Danger
Door: Two-way code integration.
Blast Radius: Server launch and replay behavior. No live app or profile changes; no new two-machine runtime pass for this head.
Original implementation: Julius Marminge. Conflict integration: Claude Opus 5.5 High through T3 Code. Independent review: GPT-6.1-Sol Medium through Codex in T3 Code.