You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
fix(provider-muse): enable Auto mode for Muse threads - #38587
Restore Auto mode for Muse threads and delegates. Verified locally on Linux and macOS against the same base and PR head.
Problem
Muse delegates from an Auto thread fall back to Supervised and ask for command approval. Muse already has an automatic reviewer, but T3 Code does not offer Auto or select that reviewer.
Change
Offer Auto for Muse threads and delegates. Map it to Muse's :auto-review profile with promptUnmatched and the sandbox enabled. Add Auto to the existing session start/resume test table and update both permission guides.
Scope
Fixes#20739. Replaces closed, unmerged #23232. This restores expected Auto behavior for an existing provider. The promptUnmatched mapping follows the maintainer's stated preference when closing #35351.
Verification
The base reproduces the bug on both Linux and macOS. The PR head uses Auto, approves routine commands through Muse's reviewer, keeps protected edits subject to human approval, and preserves Auto after a server restart.
Both platforms were checked on head a6d1385ae06c7299dbc1ab9cb32f2d1f72ede2ba against base 66bc6145fe2b117a842d2172d3806474c2b16488. Each revision used a separate isolated T3 home, a disposable Git repository, and a real, signed-in Muse provider.
Check
Linux
macOS
Reproduce the bug on base
Auto default stores approval-required; git status --short asks for human approval
Omit runtimeMode; child inherits Auto and completes a command approved by the reviewer
Same result
Claude-to-Muse delegation
Covered on earlier revision fb7fc4d8df; not repeated on this head
Omit runtimeMode; Muse child inherits Auto and returns CLEAN after a command approved by the reviewer
Reject a protected .git/config edit
Pending request rejected through T3's RPC API; file stays unchanged
Pending request rejected with Reject in the T3 UI; file hash stays unchanged
Resume after server restart
Same saved thread keeps Auto and completes another command approved by the reviewer
Same result; also confirmed the native Muse session ID is unchanged
Auto start/resume regression
Added test row fails on base (:ask-me instead of :auto-review) and passes on head
Same result
On macOS, vp run check:changed 66bc6145fe2b117a842d2172d3806474c2b16488 passed formatting, lint, both unused-code checks, 10 affected package typechecks, and 72 tests across four files. vp run build:desktop passed all five tasks. Earlier Linux build and check results are recorded below with their tested revisions.
The human approval check used a protected write (protectedWrite:true, judgeEscalated:false). Escalation by the automatic reviewer (judgeEscalated:true) was not reproduced. Windows, native mobile, API-key sign-in, remote connections, packaged desktop installation/launch, and scenarios:prove were not tested.
Screenshots
macOS
Captured from the tested base and PR head. Before: Auto falls back to Supervised and the command waits for human approval. After: Auto is available and the command completes.
Before — base 66bc6145fe2b
After — head a6d1385ae0
A protected .git/config edit still waits for human approval in Auto. The test then clicked Reject and confirmed that the file hash was unchanged.
The model-catalog warning appeared on both revisions; the tests selected Spark explicitly. The failed Claude thread in the sidebar is the initial attempt with an expired token; the retry after sign-in passed.
Linux
These Linux screenshots predate the rebases. Their Muse runtime-mode code is unchanged on this PR head.
Before (373ec12be6)
After (971523cd9b)
Protected edit waiting for human approval in Auto:
Test environment and setup notes
Both platforms: Muse Code 1.4.23-R8977.1, SDK 1.4.6, and Muse Spark 1.3 at low effort, using existing Muse sign-in.
Linux: development server, CLI, MCP delegation, and T3's normal RPC API.
macOS: macOS 27.0 on Apple Silicon, Node 26.13.1, real T3 development app and Browser panel. Cross-provider parent: Claude Code 2.1.126, Haiku 4.5. Claude's expired OAuth token was refreshed before the successful retry.
Muse returned an empty model catalog on macOS on both base and head. Live tests selected Spark explicitly; it was also added as a custom model for the head UI. The catalog warning is unchanged by this PR.
For the base regression check, only the Auto test-table row was added. Base production code stayed unchanged.
Native events and saved thread state were inspected. These were live provider turns, not replay turns. No credentials are included in the evidence.
These local results address verification that the worker VMs could not perform without Muse sign-in. The stopped pipeline has not been restarted.
Linux: sanitized native events and saved thread state
{
"pr": "https://github.com/pingdotgg/t3code/pull/38587",
"head": "a6d1385ae06c7299dbc1ab9cb32f2d1f72ede2ba",
"recordedAt": "2026-11-18T13:03:12.192108+00:00",
"environment": {
"os": "Linux",
"surface": "isolated development server, CLI and normal RPC API",
"muse": "1.4.23-R8977.1",
"sdk": "1.4.6",
"model": "muse-spark-1.3",
"reasoningEffort": "low",
"auth": "existing signed-in Muse Power account"
},
"base": {
"commit": "66bc6145fe2b117a842d2172d3806474c2b16488",
"appState": {
"runtimeMode": "approval-required",
"modelSelection": {
"instanceId": "muse",
"model": "muse-spark-1.3",
"options": [
{
"id": "reasoningEffort",
"value": "low"
}
]
}
},
"startedAt": "2026-11-18T12:57:13.627687+00:00",
"sessions": [
{
"source": "base-parent",
"model": "muse-spark-1.3",
"approvalMode": "promptUnmatched",
"profiles": [
":ask-me"
]
}
],
"approvalEvents": [
{
"source": "base-parent",
"loggerTime": "2026-11-18T12:58:06.766Z",
"method": "approval/requested",
"toolName": "bash",
"judgeEscalated": false,
"protectedWrite": false,
"approvalRef": "0ed63a93bd9f",
"command": "git status --short"
},
{
"source": "base-parent",
"loggerTime": "2026-11-18T12:59:10.445Z",
"method": "approval/resolved",
"decision": "abort",
"resolvedBy": "user",
"approvalRef": "0ed63a93bd9f"
}
],
"delegationRequests": []
},
"fixed": {
"commit": "a6d1385ae06c7299dbc1ab9cb32f2d1f72ede2ba",
"appState": {
"runtimeMode": "auto",
"modelSelection": {
"instanceId": "muse",
"model": "muse-spark-1.3",
"options": [
{
"id": "reasoningEffort",
"value": "low"
}
]
}
},
"delegatedChildren": [
{
"runtimeMode": "auto",
"modelSelection": {
"instanceId": "muse",
"model": "muse-spark-1.3",
"options": [
{
"id": "reasoningEffort",
"value": "low"
}
]
}
}
],
"startedAt": "2026-11-18T12:52:50.969070+00:00",
"sessions": [
{
"source": "head-parent",
"model": "muse-spark-1.3",
"approvalMode": "promptUnmatched",
"profiles": [
":auto-review"
]
},
{
"source": "head-child",
"model": "muse-spark-1.3",
"approvalMode": "promptUnmatched",
"profiles": [
":auto-review"
]
}
],
"approvalEvents": [
{
"source": "head-parent",
"loggerTime": "2026-11-18T12:52:59.049Z",
"method": "approval/reviewing",
"toolName": "bash",
"judgeEscalated": false,
"protectedWrite": false,
"approvalRef": "d59a78554b90",
"command": "git status --short"
},
{
"source": "head-parent",
"loggerTime": "2026-11-18T12:53:02.054Z",
"method": "approval/resolved",
"decision": "approved",
"resolvedBy": "llmJudge",
"approvalRef": "d59a78554b90"
},
{
"source": "head-parent",
"loggerTime": "2026-11-18T12:53:50.711Z",
"method": "approval/reviewing",
"toolName": "mcp__t3_code__delegate_task",
"judgeEscalated": false,
"protectedWrite": false,
"approvalRef": "6c6225987bc2"
},
{
"source": "head-parent",
"loggerTime": "2026-11-18T12:53:53.755Z",
"method": "approval/resolved",
"decision": "approved",
"resolvedBy": "llmJudge",
"approvalRef": "6c6225987bc2"
},
{
"source": "head-parent",
"loggerTime": "2026-11-18T12:55:50.888Z",
"method": "approval/requested",
"toolName": "edit_file",
"judgeEscalated": false,
"protectedWrite": true,
"approvalRef": "463436b23795"
},
{
"source": "head-parent",
"loggerTime": "2026-11-18T12:56:30.565Z",
"method": "approval/resolved",
"decision": "abort",
"resolvedBy": "user",
"approvalRef": "463436b23795"
},
{
"source": "head-parent",
"loggerTime": "2026-11-18T12:58:09.031Z",
"method": "approval/reviewing",
"toolName": "bash",
"judgeEscalated": false,
"protectedWrite": false,
"approvalRef": "34e6c2c2c35d",
"command": "git status --short"
},
{
"source": "head-parent",
"loggerTime": "2026-11-18T12:58:12.533Z",
"method": "approval/resolved",
"decision": "approved",
"resolvedBy": "llmJudge",
"approvalRef": "34e6c2c2c35d"
},
{
"source": "head-child",
"loggerTime": "2026-11-18T12:53:58.235Z",
"method": "approval/reviewing",
"toolName": "bash",
"judgeEscalated": false,
"protectedWrite": false,
"approvalRef": "248b55b6fc87",
"command": "git status --short"
},
{
"source": "head-child",
"loggerTime": "2026-11-18T12:54:00.498Z",
"method": "approval/resolved",
"decision": "approved",
"resolvedBy": "llmJudge",
"approvalRef": "248b55b6fc87"
}
],
"delegationRequests": [
{
"runtimeModeProvided": false,
"target": {
"providerInstanceId": "muse",
"model": "muse-spark-1.3",
"options": {
"reasoningEffort": "low"
}
},
"mode": "wait",
"isolation": "shared"
}
]
},
"protectedApproval": {
"runtimeMode": "auto",
"pendingKind": "file-change",
"fileUnchangedBeforeDecision": true,
"decision": "decline via normal T3 RPC API",
"fileUnchangedAfterDecision": true,
"nativeOutcome": "abort, resolvedBy:user",
"protectedWrite": true,
"judgeEscalated": false
},
"restart": {
"sameIsolatedHome": true,
"sameSavedThread": true,
"result": {
"status": "completed",
"reply": "CLEAN",
"error": null
},
"persistedMode": "auto"
},
"regression": {
"base": "Expected failure: received :ask-me, expected :auto-review; only Auto table row applied to unchanged production base.",
"head": "Same focused start/resume test passed (1 case, 48 unrelated cases skipped)."
},
"limitations": [
"LLM-judge escalation (judgeEscalated:true) was not reproduced; the native human gate used a protected write.",
"This run used CLI/RPC, including API denial; menu and UI captures in the PR come from the earlier revisions explicitly labelled there.",
"No claim is made that a worker VM can reproduce this without its own signed-in Muse installation.",
"That Linux repeat did not rebuild the current-head desktop or repeat broad checks; the later macOS current-head build/check results are recorded below."
]
}
macOS: sanitized app state, native events, delegates, and check results
{
"recordedAt": "2026-11-18T21:20:31.785877+00:00",
"environment": {
"os": "macOS 27.0",
"arch": "arm64",
"node": "26.13.1",
"muse": "1.4.23-R8977.1",
"sdk": "1.4.6",
"museModel": "muse-spark-1.3",
"museEffort": "low",
"claude": "2.1.126",
"claudeModel": "claude-haiku-4-5"
},
"base": {
"commit": "66bc6145fe2b117a842d2172d3806474c2b16488",
"requestedDefault": "auto",
"storedMode": "approval-required",
"uiModes": [
"Supervised",
"Full access"
],
"command": "git status --short",
"humanApproval": "requested, then rejected through UI: abort / user",
"regression": "Expected failure: received :ask-me, expected :auto-review; only the Auto test-table row was applied to unchanged production code."
},
"head": {
"commit": "a6d1385ae06c7299dbc1ab9cb32f2d1f72ede2ba",
"storedMode": "auto",
"uiModes": [
"Supervised",
"Auto",
"Full access"
],
"parentCommands": {
"storedMode": "auto",
"reviewer": "llmJudge",
"commands": [
{
"command": "git status --short",
"exit_code": 0,
"terminal_status": "completed",
"output": ""
},
{
"command": "git status --short",
"exit_code": 0,
"terminal_status": "completed",
"output": ""
}
]
},
"delegation": {
"museToMuse": {
"storedMode": "auto",
"reviewer": "llmJudge",
"commands": [
{
"command": "git status --short",
"exit_code": 0,
"terminal_status": "completed",
"output": ""
}
],
"runtimeModeProvided": false,
"target": {
"model": "muse-spark-1.3",
"options": [
{
"id": "reasoningEffort",
"value": "low"
}
],
"providerInstanceId": "muse"
},
"mode": "wait",
"isolation": "shared"
},
"claudeToMuse": {
"storedMode": "auto",
"reviewer": "llmJudge",
"commands": [
{
"command": "git status --short",
"exit_code": 0,
"terminal_status": "completed",
"output": ""
}
],
"runtimeModeProvided": false,
"target": {
"providerInstanceId": "muse",
"model": "muse-spark-1.3",
"options": {
"reasoningEffort": "low"
}
},
"mode": "wait",
"isolation": "shared"
}
},
"protectedApproval": {
"tool": "edit_file",
"protectedWrite": true,
"judgeEscalated": false,
"t3RequestKind": "file-change",
"decisionSurface": "Reject button in the real T3 web UI",
"decision": "abort",
"resolvedBy": "user",
"fileUnchangedWhilePendingAndAfterRejection": true
},
"restart": {
"sameIsolatedT3Home": true,
"sameSavedT3Thread": true,
"sameNativeSession": true,
"storedMode": "auto",
"reviewer": "llmJudge",
"reply": "CLEAN"
},
"checks": {
"changedFiles": 5,
"format": "pass",
"lint": "pass",
"knip": "both checks pass",
"affectedPackageTypechecks": "10 pass",
"selectedTests": "72 pass across 4 files",
"desktopBuild": "5/5 tasks pass"
}
},
"notes": [
"Claude initially failed with an invalid OAuth token; reauthentication and the retry succeeded.",
"Native Muse model/list returned an empty model catalog on this Mac, also on the base. Live tests explicitly selected Muse Spark 1.3; it was added as a custom model for the head UI.",
"No production source changes, commits, or force push were needed.",
"No claim of judgeEscalated:true, Windows, native mobile, packaged release installation, remote connections, or scenarios:prove."
]
}
Earlier Linux validation (revisions labelled)
On fb7fc4d8df, real app checks covered Auto review, Claude-to-Muse and Muse-to-Muse inheritance, rejection of a human approval request through the UI, and resume after server restart.
vp run check:changed 6a29d58d14 passed formatting, lint, both unused-code checks, typechecks for 10 affected packages, and 73 selected tests. The pipeline later removed the extra status assertion and child-test parameterization; this PR keeps the profile regression case.
vp run build:desktop passed all five tasks on that revision.
The historical RightPanelTabs flick/swipe test passed once on base 6a29d58d14 and once on head fb7fc4d8df with the same pinned Chromium. This does not rule out intermittent failures.
Original patch by Victor Vogel with Claude Opus 5.5. Plan reviewed by GPT-6.2 Astra and Muse Wonder; Muse Wonder was a reviewer only.
Implemented and verified with GPT-6.1 Sol in T3 Code.
Restore Muse Auto support using its native :auto-review profile while retaining promptUnmatched and sandbox enforcement. Cover start, resume, and advertised runtime modes, and update permission guidance.
Based on pingdotgg#23232.
Co-authored-by: Victor Vogel <satsdev@only21mil.xyz>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The parametrized workflow-child case and the status assertion on the
supported-modes constant duplicated coverage the profile table already gives.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Reviewing files that changed from the base of the PR and between a6d1385 and 93242b9.
You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.
Use the checkbox below for a quick retry:
Trigger review
No actionable comments were generated in the recent review.
Walkthrough
Muse now supports Auto runtime mode and maps it to the :auto-review permission profile. Tests cover profile selection and workflow-child approvals. The documentation describes Auto mode and the limits on Auto-accept edits.
Changes: Muse Auto Mode
Layer / File(s)
Summary
Declare and apply Auto mode packages/provider-muse/src/server/status.ts, packages/provider-muse/src/server/sdk.ts, packages/provider-muse/src/server/status.test.ts, packages/provider-muse/src/server/adapter.test.ts, docs/user/providers-muse.md, docs/user/permission-modes.md
Muse lists Auto as a supported runtime mode and maps it to :auto-review. Tests cover the mapping on session start and resume, and verify workflow-child approval behavior in Auto and approval-required modes. The documentation describes Auto mode and states that Auto-accept edits is unavailable.
pingdotgg/t3code#23232: Maps Muse Auto to :auto-review, adds Auto to supported runtime modes, and updates the same documentation.
pingdotgg/t3code#35351: Adds Muse Auto with :auto-review and updates the same supported-mode declaration and documentation. Its description uses onRequest approval, while this PR tests promptUnmatched.
Muse threads and delegates can now use Auto mode through Muse's auto-review profile. Saved Auto-accept edits threads still run in Supervised, as the updated guides state. No concrete merge-blocking issue was found; behavior on older Muse hosts and on other platforms was not tested.
Pre-merge checks: 4
Check name
Status
Explanation
Title check
The title "fix(provider-muse): enable Auto mode for Muse threads" matches the main change. The PR adds Auto as a permission mode for Muse threads and delegates, maps it to Muse's :auto-review profile,…
Linked Issues check
[#20739] requires Auto for Muse threads and inherited Muse delegates, mapped to Muse's :auto-review profile. The reviewed head implements this. In packages/provider-muse/src/server/sdk.ts, `musePe…
Out of Scope Changes check
All changed files relate to the objective. Code changes are limited to sdk.ts (profile mapping) and status.ts (supported modes, comment). Test changes in adapter.test.ts and status.test.ts cov…
Description check
The description follows the repository template closely. It has a clear problem statement, a change section, a scope section that links issue #20739 and explains why no new feature is involved, and a …
Full details: Title check
Explanation
The title "fix(provider-muse): enable Auto mode for Muse threads" matches the main change. The PR adds Auto as a permission mode for Muse threads and delegates, maps it to Muse's :auto-review profile, and updates the related tests and docs. It is a conventional commit title in plain language.
Full details: Linked Issues check
Explanation
[#20739] requires Auto for Muse threads and inherited Muse delegates, mapped to Muse's :auto-review profile. The reviewed head implements this. In packages/provider-muse/src/server/sdk.ts, musePermissionProfile returns :auto-review for auto, :unrestricted for full-access, and :ask-me otherwise. In status.ts, MUSE_PRESENTATION.supportedRuntimeModes now includes auto, so Muse offers Auto in the mode picker. museApprovalMode already returns promptUnmatched for non-full-access modes, which matches the PR's promptUnmatched setting. adapter.test.ts asserts the :auto-review profile with promptUnmatched on session start and resume, and status.test.ts asserts auto is supported. The PR reports native Linux verification of reviewer approval, delegate inheritance, and resume after restart. LLM-judge escalation was not reproduced, and that gap does not block this PR's stated scope. Delegate inheritance depends on existing delegation logic, which this diff does not change. The native verification report supports that behavior. No broader-than-parent mode is introduced.
Full details: Out of Scope Changes check
Explanation
All changed files relate to the objective. Code changes are limited to sdk.ts (profile mapping) and status.ts (supported modes, comment). Test changes in adapter.test.ts and status.test.ts cover the new mode mapping, resume behavior, and child approval handling. Documentation changes in docs/user/permission-modes.md and docs/user/providers-muse.md describe the new Auto mode for Muse. No unrelated change was found in the reviewed diffs. The RightPanelTabs test mentioned in the PR is not in the changed-file summary, so it is not assessed as a change in this PR.
Full details: Description check
Explanation
The description follows the repository template closely. It has a clear problem statement, a change section, a scope section that links issue #20739 and explains why no new feature is involved, and a detailed verification section covering Linux and macOS with tested revisions, commands, results, and limitations. The title-like opening line is present but the template's conventional commit title is not formally separated from the body. The verification section is extensive and states what was not tested. The description is mostly complete, so this check passes.
Finishing Touches
Generate unit tests (beta)
Create a new PR
Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts
Reviewing files that changed from the base of the PR and between 2c4622e and fb7fc4d.
Files selected for processing: 6
docs/user/permission-modes.md
docs/user/providers-muse.md
packages/provider-muse/src/server/adapter.test.ts
packages/provider-muse/src/server/sdk.ts
packages/provider-muse/src/server/status.test.ts
packages/provider-muse/src/server/status.ts
Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 9 remain after this review.
The automated pipeline stopped on this pull request and will not merge it:
the fix is written and reviewed, but proving it needs something no worker has:
Permanent blocker: An installed, signed-in Muse provider is required and is explicitly unavailable on every VM. Auto selection, inherited Auto delegation, automatic review of routine shell commands, pending human approvals, and persistence after server restart remain unverified on both commits.
Native verification passed locally on the current head, a6d1385ae0. Auto review, inherited Auto delegation, pending approval and denial, and resume after a server restart all worked. On base 66bc6145fe, the same command reproduced the fallback to Supervised.
The PR description now includes sanitized events and the test details. The approval check used a protected write; LLM-judge escalation itself was not reproduced.
Could a maintainer review this evidence and advise how to proceed with the stopped pipeline?
Full macOS app verification is complete on current head a6d1385ae06c7299dbc1ab9cb32f2d1f72ede2ba. The PR description now combines the Linux results with this full macOS run and sanitized evidence.
Actual base 66bc6145fe2b: Auto fell back to Supervised, the menu lacked Auto, and git status asked for human approval. The added Auto regression row failed with :ask-me instead of :auto-review.
PR head: the UI offers Auto, the thread stores auto, and the native reviewer approves git status (llmJudge, exit 0).
Real Muse-to-Muse and Claude-to-Muse delegates both omitted runtimeMode; both Muse children inherited Auto and completed reviewer-approved commands, returning CLEAN.
A protected .git/config edit waited for human approval in the real T3 UI. Clicking Reject resolved it as abort / user; the file remained unchanged.
After a full T3 server restart, the same saved thread resumed the same native session in Auto and completed another reviewer-approved command.
Changed-file checks passed: format, lint, both unused-code checks, 10 package typechecks, and 72 tests across four files.
Desktop build passed: all five tasks on macOS.
Environment: macOS 27.0 / Apple Silicon, Node 26.13.1, Muse 1.4.23-R8977.1, SDK 1.4.6, Spark 1.3 at low effort; Claude Code 2.1.126 / Haiku 4.5 for the cross-provider parent. Separate isolated T3 homes and a disposable Git repository were used. Claude's expired OAuth token was refreshed before the successful retry. Muse's native model catalog was empty on both base and head; live tests explicitly selected Spark, with a custom model configured for the head UI.
This run exercises the real app, MCP delegates, UI denial, and T3 server recovery. The human gate was a protected write (judgeEscalated:false); packaged release installation/launch, Windows, native mobile, remote connections, API-key sign-in, and scenarios:prove are not claimed. Full sanitized state and native outcomes are in the PR description.
No code changes or force push were needed. This existing macOS comment was updated in place.
The macOS before/after and protected-approval screenshots are now embedded in the PR description, visible by default alongside the Linux captures.
The PR description now presents the Linux and macOS verification together, with a single results table and the detailed evidence below it. This replaces the earlier Linux-only presentation; the existing macOS comment still records that run's details.
The description distinguishes checks on the current head (a6d1385ae0) from checks and screenshots on earlier revisions. The code and PR head are unchanged.
The automated pipeline stopped on this pull request and will not merge it:
the fix is written and reviewed, but proving it needs something no worker has:
* Permanent blocker: An installed, signed-in Muse provider is required and is explicitly unavailable on every VM. Auto selection, inherited Auto delegation, automatic review of routine shell commands, pending human approvals, and persistence after server restart remain unverified on both commits.
It stays open for a person to finish or close.
in the other issue #23232 (comment), i have already used signed-in muse provider and doing my best to show things work for multiple times across linux and windows. i suppose what you mean by VM is virtual machine and i guess that's something in the github workflow that none of us, the community member/contributer, can touch... just really don't want the pull requests on the same issue getting blocked by the some issues that only the repo maintainer can fix.
I sure hope it's possible to get Muse fixes in here somehow :) Let me know if I can do anything to help.
This branch has not been deployed
No deployments
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
size:S10-29 changed lines (additions + deletions).takeover:blockedTakeover blocked; reason in the PR bodyvouch:unvouchedAuthor is not yet trusted in the VOUCHED list.
3 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.
Restore Auto mode for Muse threads and delegates. Verified locally on Linux and macOS against the same base and PR head.
Problem
Muse delegates from an Auto thread fall back to Supervised and ask for command approval. Muse already has an automatic reviewer, but T3 Code does not offer Auto or select that reviewer.
Change
Offer Auto for Muse threads and delegates. Map it to Muse's
:auto-reviewprofile withpromptUnmatchedand the sandbox enabled. Add Auto to the existing session start/resume test table and update both permission guides.Scope
Fixes #20739. Replaces closed, unmerged #23232. This restores expected Auto behavior for an existing provider. The
promptUnmatchedmapping follows the maintainer's stated preference when closing #35351.Verification
The base reproduces the bug on both Linux and macOS. The PR head uses Auto, approves routine commands through Muse's reviewer, keeps protected edits subject to human approval, and preserves Auto after a server restart.
Both platforms were checked on head
a6d1385ae06c7299dbc1ab9cb32f2d1f72ede2baagainst base66bc6145fe2b117a842d2172d3806474c2b16488. Each revision used a separate isolated T3 home, a disposable Git repository, and a real, signed-in Muse provider.approval-required;git status --shortasks for human approvalauto; Muse's reviewer approvesgit status --short; command succeedsruntimeMode; child inherits Auto and completes a command approved by the reviewerfb7fc4d8df; not repeated on this headruntimeMode; Muse child inherits Auto and returnsCLEANafter a command approved by the reviewer.git/configedit:ask-meinstead of:auto-review) and passes on headOn macOS,
vp run check:changed 66bc6145fe2b117a842d2172d3806474c2b16488passed formatting, lint, both unused-code checks, 10 affected package typechecks, and 72 tests across four files.vp run build:desktoppassed all five tasks. Earlier Linux build and check results are recorded below with their tested revisions.The human approval check used a protected write (
protectedWrite:true,judgeEscalated:false). Escalation by the automatic reviewer (judgeEscalated:true) was not reproduced. Windows, native mobile, API-key sign-in, remote connections, packaged desktop installation/launch, andscenarios:provewere not tested.Screenshots
macOS
Captured from the tested base and PR head. Before: Auto falls back to Supervised and the command waits for human approval. After: Auto is available and the command completes.
66bc6145fe2ba6d1385ae0A protected
.git/configedit still waits for human approval in Auto. The test then clicked Reject and confirmed that the file hash was unchanged.The model-catalog warning appeared on both revisions; the tests selected Spark explicitly. The failed Claude thread in the sidebar is the initial attempt with an expired token; the retry after sign-in passed.
Linux
These Linux screenshots predate the rebases. Their Muse runtime-mode code is unchanged on this PR head.
373ec12be6)971523cd9b)Protected edit waiting for human approval in Auto:
Test environment and setup notes
1.4.23-R8977.1, SDK1.4.6, and Muse Spark 1.3 at low effort, using existing Muse sign-in.26.13.1, real T3 development app and Browser panel. Cross-provider parent: Claude Code2.1.126, Haiku 4.5. Claude's expired OAuth token was refreshed before the successful retry.Linux: sanitized native events and saved thread state
{ "pr": "https://github.com/pingdotgg/t3code/pull/38587", "head": "a6d1385ae06c7299dbc1ab9cb32f2d1f72ede2ba", "recordedAt": "2026-11-18T13:03:12.192108+00:00", "environment": { "os": "Linux", "surface": "isolated development server, CLI and normal RPC API", "muse": "1.4.23-R8977.1", "sdk": "1.4.6", "model": "muse-spark-1.3", "reasoningEffort": "low", "auth": "existing signed-in Muse Power account" }, "base": { "commit": "66bc6145fe2b117a842d2172d3806474c2b16488", "appState": { "runtimeMode": "approval-required", "modelSelection": { "instanceId": "muse", "model": "muse-spark-1.3", "options": [ { "id": "reasoningEffort", "value": "low" } ] } }, "startedAt": "2026-11-18T12:57:13.627687+00:00", "sessions": [ { "source": "base-parent", "model": "muse-spark-1.3", "approvalMode": "promptUnmatched", "profiles": [ ":ask-me" ] } ], "approvalEvents": [ { "source": "base-parent", "loggerTime": "2026-11-18T12:58:06.766Z", "method": "approval/requested", "toolName": "bash", "judgeEscalated": false, "protectedWrite": false, "approvalRef": "0ed63a93bd9f", "command": "git status --short" }, { "source": "base-parent", "loggerTime": "2026-11-18T12:59:10.445Z", "method": "approval/resolved", "decision": "abort", "resolvedBy": "user", "approvalRef": "0ed63a93bd9f" } ], "delegationRequests": [] }, "fixed": { "commit": "a6d1385ae06c7299dbc1ab9cb32f2d1f72ede2ba", "appState": { "runtimeMode": "auto", "modelSelection": { "instanceId": "muse", "model": "muse-spark-1.3", "options": [ { "id": "reasoningEffort", "value": "low" } ] } }, "delegatedChildren": [ { "runtimeMode": "auto", "modelSelection": { "instanceId": "muse", "model": "muse-spark-1.3", "options": [ { "id": "reasoningEffort", "value": "low" } ] } } ], "startedAt": "2026-11-18T12:52:50.969070+00:00", "sessions": [ { "source": "head-parent", "model": "muse-spark-1.3", "approvalMode": "promptUnmatched", "profiles": [ ":auto-review" ] }, { "source": "head-child", "model": "muse-spark-1.3", "approvalMode": "promptUnmatched", "profiles": [ ":auto-review" ] } ], "approvalEvents": [ { "source": "head-parent", "loggerTime": "2026-11-18T12:52:59.049Z", "method": "approval/reviewing", "toolName": "bash", "judgeEscalated": false, "protectedWrite": false, "approvalRef": "d59a78554b90", "command": "git status --short" }, { "source": "head-parent", "loggerTime": "2026-11-18T12:53:02.054Z", "method": "approval/resolved", "decision": "approved", "resolvedBy": "llmJudge", "approvalRef": "d59a78554b90" }, { "source": "head-parent", "loggerTime": "2026-11-18T12:53:50.711Z", "method": "approval/reviewing", "toolName": "mcp__t3_code__delegate_task", "judgeEscalated": false, "protectedWrite": false, "approvalRef": "6c6225987bc2" }, { "source": "head-parent", "loggerTime": "2026-11-18T12:53:53.755Z", "method": "approval/resolved", "decision": "approved", "resolvedBy": "llmJudge", "approvalRef": "6c6225987bc2" }, { "source": "head-parent", "loggerTime": "2026-11-18T12:55:50.888Z", "method": "approval/requested", "toolName": "edit_file", "judgeEscalated": false, "protectedWrite": true, "approvalRef": "463436b23795" }, { "source": "head-parent", "loggerTime": "2026-11-18T12:56:30.565Z", "method": "approval/resolved", "decision": "abort", "resolvedBy": "user", "approvalRef": "463436b23795" }, { "source": "head-parent", "loggerTime": "2026-11-18T12:58:09.031Z", "method": "approval/reviewing", "toolName": "bash", "judgeEscalated": false, "protectedWrite": false, "approvalRef": "34e6c2c2c35d", "command": "git status --short" }, { "source": "head-parent", "loggerTime": "2026-11-18T12:58:12.533Z", "method": "approval/resolved", "decision": "approved", "resolvedBy": "llmJudge", "approvalRef": "34e6c2c2c35d" }, { "source": "head-child", "loggerTime": "2026-11-18T12:53:58.235Z", "method": "approval/reviewing", "toolName": "bash", "judgeEscalated": false, "protectedWrite": false, "approvalRef": "248b55b6fc87", "command": "git status --short" }, { "source": "head-child", "loggerTime": "2026-11-18T12:54:00.498Z", "method": "approval/resolved", "decision": "approved", "resolvedBy": "llmJudge", "approvalRef": "248b55b6fc87" } ], "delegationRequests": [ { "runtimeModeProvided": false, "target": { "providerInstanceId": "muse", "model": "muse-spark-1.3", "options": { "reasoningEffort": "low" } }, "mode": "wait", "isolation": "shared" } ] }, "protectedApproval": { "runtimeMode": "auto", "pendingKind": "file-change", "fileUnchangedBeforeDecision": true, "decision": "decline via normal T3 RPC API", "fileUnchangedAfterDecision": true, "nativeOutcome": "abort, resolvedBy:user", "protectedWrite": true, "judgeEscalated": false }, "restart": { "sameIsolatedHome": true, "sameSavedThread": true, "result": { "status": "completed", "reply": "CLEAN", "error": null }, "persistedMode": "auto" }, "regression": { "base": "Expected failure: received :ask-me, expected :auto-review; only Auto table row applied to unchanged production base.", "head": "Same focused start/resume test passed (1 case, 48 unrelated cases skipped)." }, "limitations": [ "LLM-judge escalation (judgeEscalated:true) was not reproduced; the native human gate used a protected write.", "This run used CLI/RPC, including API denial; menu and UI captures in the PR come from the earlier revisions explicitly labelled there.", "No claim is made that a worker VM can reproduce this without its own signed-in Muse installation.", "That Linux repeat did not rebuild the current-head desktop or repeat broad checks; the later macOS current-head build/check results are recorded below." ] }macOS: sanitized app state, native events, delegates, and check results
{ "recordedAt": "2026-11-18T21:20:31.785877+00:00", "environment": { "os": "macOS 27.0", "arch": "arm64", "node": "26.13.1", "muse": "1.4.23-R8977.1", "sdk": "1.4.6", "museModel": "muse-spark-1.3", "museEffort": "low", "claude": "2.1.126", "claudeModel": "claude-haiku-4-5" }, "base": { "commit": "66bc6145fe2b117a842d2172d3806474c2b16488", "requestedDefault": "auto", "storedMode": "approval-required", "uiModes": [ "Supervised", "Full access" ], "command": "git status --short", "humanApproval": "requested, then rejected through UI: abort / user", "regression": "Expected failure: received :ask-me, expected :auto-review; only the Auto test-table row was applied to unchanged production code." }, "head": { "commit": "a6d1385ae06c7299dbc1ab9cb32f2d1f72ede2ba", "storedMode": "auto", "uiModes": [ "Supervised", "Auto", "Full access" ], "parentCommands": { "storedMode": "auto", "reviewer": "llmJudge", "commands": [ { "command": "git status --short", "exit_code": 0, "terminal_status": "completed", "output": "" }, { "command": "git status --short", "exit_code": 0, "terminal_status": "completed", "output": "" } ] }, "delegation": { "museToMuse": { "storedMode": "auto", "reviewer": "llmJudge", "commands": [ { "command": "git status --short", "exit_code": 0, "terminal_status": "completed", "output": "" } ], "runtimeModeProvided": false, "target": { "model": "muse-spark-1.3", "options": [ { "id": "reasoningEffort", "value": "low" } ], "providerInstanceId": "muse" }, "mode": "wait", "isolation": "shared" }, "claudeToMuse": { "storedMode": "auto", "reviewer": "llmJudge", "commands": [ { "command": "git status --short", "exit_code": 0, "terminal_status": "completed", "output": "" } ], "runtimeModeProvided": false, "target": { "providerInstanceId": "muse", "model": "muse-spark-1.3", "options": { "reasoningEffort": "low" } }, "mode": "wait", "isolation": "shared" } }, "protectedApproval": { "tool": "edit_file", "protectedWrite": true, "judgeEscalated": false, "t3RequestKind": "file-change", "decisionSurface": "Reject button in the real T3 web UI", "decision": "abort", "resolvedBy": "user", "fileUnchangedWhilePendingAndAfterRejection": true }, "restart": { "sameIsolatedT3Home": true, "sameSavedT3Thread": true, "sameNativeSession": true, "storedMode": "auto", "reviewer": "llmJudge", "reply": "CLEAN" }, "checks": { "changedFiles": 5, "format": "pass", "lint": "pass", "knip": "both checks pass", "affectedPackageTypechecks": "10 pass", "selectedTests": "72 pass across 4 files", "desktopBuild": "5/5 tasks pass" } }, "notes": [ "Claude initially failed with an invalid OAuth token; reauthentication and the retry succeeded.", "Native Muse model/list returned an empty model catalog on this Mac, also on the base. Live tests explicitly selected Muse Spark 1.3; it was added as a custom model for the head UI.", "No production source changes, commits, or force push were needed.", "No claim of judgeEscalated:true, Windows, native mobile, packaged release installation, remote connections, or scenarios:prove." ] }Earlier Linux validation (revisions labelled)
On
fb7fc4d8df, real app checks covered Auto review, Claude-to-Muse and Muse-to-Muse inheritance, rejection of a human approval request through the UI, and resume after server restart.vp run check:changed 6a29d58d14passed formatting, lint, both unused-code checks, typechecks for 10 affected packages, and 73 selected tests. The pipeline later removed the extra status assertion and child-test parameterization; this PR keeps the profile regression case.vp run build:desktoppassed all five tasks on that revision.RightPanelTabsflick/swipe test passed once on base6a29d58d14and once on headfb7fc4d8dfwith the same pinned Chromium. This does not rule out intermittent failures.Original patch by Victor Vogel with Claude Opus 5.5. Plan reviewed by GPT-6.2 Astra and Muse Wonder; Muse Wonder was a reviewer only.
Implemented and verified with GPT-6.1 Sol in T3 Code.