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
Choose an agent setup from the new tab menu - #34728
The desktop + tab menu now lets you search for an agent setup, such as “opus high,” and launch it immediately. Clicking an agent row or pressing Enter launches with the current defaults. Only the chevron opens setup, where you can choose a model, effort, Speed/Fast and permissions, edit Arguments, and launch Chat or Terminal. Untouched launches preserve the existing defaults. Arguments and their removable permission/Fast badges stay in sync, and the setup panel stays anchored and scrolls within a narrow window.
This PR covers the desktop + tab menu. PR 2 in this lane brings the same picker to Create worktree, continuation, AgentCombobox, Source Control settings/Customize, automations and phone controls; those entry points keep their current surfaces here, tracked in #35802. The shared launch core is tracked in #34275 and has merged. Existing dispatchers and hosts continue to own execution and launch failures. There is no new RPC, saved field or minimum host version.
Differences from the common pattern:
Intended: clicking a row launches; the chevron alone opens setup. Controls show resolved defaults, while untouched launches send no replacement choices. This preserves current local, SSH and paired-host routing.
Intended: free-form Arguments apply to Terminal. Deliberate permission and Fast/Speed choices also reach Chat. Settings files and opaque Arguments remain literal; dependent controls stay disabled when their values cannot be established safely, while ordinary launches remain available.
Intended: setup uses the whole row as its anchor, caps its width to the available side space, falls below the row when necessary, and observes window/row resizing only while open.
Real Claude and Codex ordinary Chat/Terminal launches and supported searched launches produced replies. Configured Codex Chat and Terminal confirmed High, priority and Ask in provider settings. Final targeted checks on 3571c9c2 verified Claude inline settings through actual argv and a real reply, malformed permission editing without duplicate text, Arguments toggling, Codex Speed choices, reopened defaults, anchored placement, narrow widths and inner scrolling. Claude's launched settings were one merged object preserving language English with Fast requested. Narrow 700/600 widths and shrink while open were verified by viewport emulation in the real app, alongside the PR's real-Popper geometry tests; physical OS window resizing was not exercised. The measured 0.5px edge overrun was accepted as rounding.
Claude Fast-on was not exercised: it is unavailable with the QA load-balancer sign-in. Both the picker and the existing composer toggle returned Standard. The picker's Fast Arguments and typed Chat options were verified; actual Fast-on is not claimed. Malformed typed-Chat permission readback and mutation of model/effort defaults were not exercised; the malformed text check was a UI check with no invalid launch. Codex Chat launches with no explicit effort override intentionally remain sparse.
CI passes on reviewed head 2e676b48 (the main merge preserves the QA'd picker code from 3571c9c2); fresh correctness and mechanism reviews are clean. The reported Sonnet Fast loss was refuted through the final command compiler: 4 tests covering 16 paths retained Fast true or deliberate false and unrelated JSON.
Deferred P3 polish, tracked in #35802: six setup labels use English fallback text in non-English locales; a menu with no agents and no query can show an extra search-empty line; and the separate Fast chip spells out a second --settings argument even though launch correctly merges one settings object. The chip display should make that merged meaning clear.
Final Claude setup (the Fast chip display is the deferred cosmetic above):
[Medium impact] The PR appears safe to merge; no new blocking issue was established.
Summary
Adds searchable agent setups to the desktop + tab menu.
The desktop + menu searches agents and setups, then launches the chosen option.
The setup panel keeps deliberate Chat and Terminal choices separate.
The setup panel uses defaults and model choices from its current host.
Diagram
%%{init: {'theme': 'neutral'}}%%
flowchart TD
A[Open + tab menu] --> B[Read agent rows and host defaults]
B --> C[Search agents and setups]
C --> D[Click row or press Enter]
C --> E[Open chevron setup]
E --> F[Choose model, effort, permissions or Speed]
E --> G[Edit Terminal Arguments]
F --> H[Choose Chat or Terminal]
G --> H
D --> I[Existing launch dispatcher]
H --> I
Declare the picker entry parameter in the runLaunch prop type.
AgentLaunchPickerMenu calls runLaunch(row, draft, entry). agentPickerLaunchInputs needs that entry for two things: the agentPickerArgumentsKnown authority check and the inherited entry.terminal.permissionMode. This prop type drops the third parameter. The current caller still works only because launchAgentRow happens to accept a third argument at runtime. A future caller that matches this declared type would silently drop entry. Its configured launches would then skip the unknown-Arguments guard and lose the inherited permission. Declare the full signature so the compiler enforces the contract.
Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 2 remain after this review.
The + tab setup panel differed from the approved layout, and Arguments could overlap its chips, widen the panel or disappear after dismissal. It now uses the requested compact 416px panel, shared Permissions menu and borderless configure chevrons with left dividers. Long Arguments wrap within the field, and the complete authored model/effort/permission/Fast/Arguments draft stays in memory when the panel closes or a launch finishes. A different owner starts fresh; untouched launches still use the agent’s own default.
Scoped review covered the layout/copy/focus delta after 2e676b48, plus the long search-result chevron fix in fbca0a68. Fixes include single-panel ownership, replacement focus through deferred close effects, shared-component opt-in layout variants, and explicit initially-unselected model/effort coverage. The conflict-forced main merge a5e2894c preserves both sets of host-preference tests and the approved performance assertions. Four affected merge test files passed (73 tests); the earlier broad renderer run passed 188 tests across 11 files, and the final focus/panel run passed 40 tests. The focus regression now waits for the outgoing Radix scope’s actual unmount event instead of a timing sleep (881c7a81; 18 tests passed). Two final bot findings were repaired in e8b1456d: the no-match message is outside the hidden result group and the setup dialog has its existing translated Configure-agent accessible name (41 renderer tests passed). The post-merge remote-equivalence test fixture now registers/restores the existing catalog store (152e55ce; all 29 endpoint-equivalence tests passed). Production launch behavior and assertions are unchanged. Brennan’s merged shared Permissions controls were integrated in the conflict-forced main merge 956e0f93: the + tab dropdown now uses the same icons, descriptions, Full access color and separated own-settings choice as Chat, while retaining inherited/raw Arguments behavior and the compact layout. Eight explicit permission/settings/panel test files passed (170 tests). Later owner-requested Arguments/416px/draft-retention work is in 1a7523c3 (95 tests across seven named files). Review found that caching the alternate launch route could make reopened defaults disagree with the primary action; 9d1cd674 now caches authored edits only, with both preferred-route regressions failing before and passing after (88 tests across six files). Main merge 5b39fae0 carries the diagnosed main lint fix; 3d85b561/b966f232 implement the real 24px borderless chevron/divider and keep muted styling on the glyph (50 final tests across two files). The same reviewer sealed accumulated source/tests and exercised visual evidence on a2294059. The final tests-only catalog fixture repair a2294059 adds the required default marker without changing production code (21 menu tests passed). CI passes on a2294059 (run), including static analysis/typecheck, all 12 unit shards, package, mobile and WSL checks.
Screenshots below show real Claude and Codex catalogs, with global Chat UI enabled and each provider’s actual Chat or Terminal preference. The earlier general-layout captures are on fef2ff52 at 440px with bordered configure buttons; shared Permissions captures are on 956e0f93. They retain their actual default-preference/interaction provenance. New current-head 416px/Arguments/chevron screenshots below supersede those older pixels for the changed appearance. Delayed keyboard replacement, unassisted typing, Escape, hover, arrows, Tab, outside click and compact permission dropdown were exercised in the hidden app. Current model rows measured 29px with 13px names, 6px list inset and 1px gap; settings measured 10/12/4px padding with 10px gaps and 28px rows; footer padding is 8/10px. In a short window the footer initially needs scrolling; actual scrolling makes both buttons fully visible, enabled and hittable, and scrolling back reaches the search.
Intended differences from the original mock, following the October 29 rulings: no footer summary, 28–30px model rows, larger model names, “Only applies to terminal (CLI) launches”, and Launch labels. Codex retains its previously approved catalog-driven Speed Standard/Fast choices. Temporary polish tracked in #35802: no NEW badge because model data has no such flag, translated-copy fallbacks and earlier disclosed display polish; two remaining minor size gaps are now explicitly included there (12px search text versus the mock’s 13.5px, and 24px buttons/41px footer versus 28px/45px). These are disclosed gaps, not a claim of pixel identity.
This follow-up is a UI/tests-row check with branch screenshots, carrying the earlier real-agent runtime evidence below. Live Legacy models were unavailable in the observed catalogs; its existing source/tests remain the evidence. The supplemental screenshot driver timed out after saving corrected default/scroll artifacts; those specific artifacts were inspected rather than treating its exit as a blanket pass. The isolated app and helper processes were stopped gracefully, and real config hashes were unchanged.
The current shared Permissions dropdown was exercised and captured for both agents on 956e0f93: icons/descriptions, warning-colored Full access, own-settings separator, compact trigger, Escape containment and selection/restoration passed. No agents were launched for this targeted screenshot check. Raw/custom values and comparison with the composer remain source/test evidence; no composer was opened for a comparison screenshot. The isolated copy stopped gracefully, real config hashes were unchanged and admission3340 is released. These images supersede the Permissions appearance in the older layout captures.
Claude Permissions
Codex Permissions
Current 416px Arguments and configure controls were exercised on b966f232; the only later commit, a2294059, changes tests. The same reviewer independently inspected all eight PNGs, source, retention records, all 82 evidence-file hashes, cleanup and final CI. The 271-character unbroken token wraps inside the field, with model/settings and both buttons visible. Outside-click dismissal followed by reopening the same ordinary Claude row preserved the exact 342-character normalized draft, visible residual, chips, Opus/High/Accept edits/Fast-on values. The actual 24×24 configure buttons have a 1×16 left divider, transparent borderless rest state, ghost hover feedback, and six-pixel search-panel anchor gap. Current Permissions containment and Escape-to-parent passed.
The first targeted stage stopped on the visible --settings '{}' residual. Independent source comparison confirmed this is the already deferred FINAL-QA-CHIP-1 display issue in #35802, unchanged from the previously reviewed mechanism. The continuation preserved that observation and compared the whole normalized draft; it does not claim the cosmetic issue fixed or a new runtime argv pass. Exact original inherited-versus-explicit QA-draft restoration was not established: displayed choices were restored, then the isolated app closed. The retention proof above compares the independently concordant authored before/after state. Real config hashes were unchanged; apps, brokers and drivers stopped gracefully, and admissions 3373/3380 are released.
The live Codex catalog offered six efforts (Low/Med/High/XHigh/Max/Ultra), all on one line; the exact terminal hint also fits one line. Seven-choice live layout was not exercised. The lane owner approved disclosing the actual-six screenshot plus the existing seven-label/source proof: actual Geist font metrics measured 349.06px for the seven-label row including Effort, within 390px available. This is a measurement/source result, not seven-choice live screenshot evidence. No seventh catalog option was fabricated and no agent was launched in these screenshot checks.
For Brennan’s review (current appearance supersedes the older 440px/bordered captures below):
Long Arguments
Same draft reopened
Current Codex (six actual efforts)
Current Permissions
Ordinary configure controls
Search result controls
Search-row setup anchor:
Readable chip/residual layout, with the disclosed existing empty-settings display issue:
For Brennan’s review (earlier approved-mock/default-preference comparison):
Agent
Approved mock
Chat default
Terminal default
Claude
Codex
Short-window scrolling, before and after:
Search results with long model names now truncate within their row and keep the configure chevron fully inside and hittable:
A nonmatching query keeps its existing feedback visible. This presentational capture uses the actual production panel; the dialog also reuses its existing Configure-agent accessible name:
Earlier unchanged runtime verification and disclosed limits (historical panel images)
The final review is complete on 2e676b48, with the bounded Claude/Codex QA below and its disclosed limits. No open P0/P1 findings remain.
The desktop + tab menu previously launched only the default agent setup. It now offers agent/model/effort search and a chevron setup panel, with Chat and Terminal actions; clicking an ordinary row still preserves current defaults. The setup controls derive current host facts, keep Arguments and permission/Fast badges in sync, and remain anchored within narrow windows.
Fresh final correctness, performance/readiness and whole-mechanism comparison completed on 2e676b48f436058cd0e727d6546eee8295bd3396. The main merge preserves all reviewed picker/shared code from 3571c9c292a0f0a880cb7f6371e0c50ebd997442; only existing main locale updates were integrated. CI passes on the reviewed head (run).
The owner follow-up fixes stale Chat defaults on reopen, fixture typing, Arguments pointer focus, selectable declared Codex Speed tiers, known/unknown value authority, native row anchoring, width/fallback/scrolling, Claude inline settings recognition, malformed permission text, and Fast badge preservation without extra Source Control mode controls. An independent final-compiler check refuted the reported Sonnet Fast loss: 4 tests/16 paths preserve Fast true, deliberate false and unrelated JSON.
Deferred P3 polish, tracked in #35802: six setup labels use English fallback in non-English locales; an empty agent list with no query can display an extra search-empty line; the Fast chip can read as a second effective --settings argument beside literal inline settings. The actual launch correctly merges one object; follow-up should clarify the chip display. Completed catalog retention remains the previously settled concern without a demonstrated wrong-value input. There are no open P0/P1 correctness findings.
Intended differences from the common pattern: row activation launches and only the chevron configures; unknown current defaults disable dependent edits while ordinary launches stay available; untouched launch intent stays sparse; arbitrary Arguments remain Terminal-only, with deliberate permission/Fast choices shared with Chat; opaque/file settings stay literal. Other launch surfaces temporarily retain their current controls, tracked in #35802; automation work also follows #34332.
Final targeted QA on 3571c9c292a0f0a880cb7f6371e0c50ebd997442 is complete. The only later commit is main merge 2e676b48f436058cd0e727d6546eee8295bd3396; preservation review and CI cover it. No QA rerun is needed for that merge. Final evidence was inspected first-hand, including actual provider argv/runtime and screenshots; the sealed archive and all 214 member hashes match.
Flow
Verified result and provenance
F1 ordinary Chat/Terminal
Real Claude and Codex replies on 78c3cd86, carried across unchanged launch flows. Sparse Codex Chat default effort is intended, not measured Low.
F2 supported search
Real Opus/High and 6.1-Sol/High replies on 78c3cd86; main/branch comparison retained below.
F3 configured runtime
Codex Chat/Terminal High, priority and Ask plus replies on 56ab3425; final Claude inline settings on 3571c9c2 produced one effective {"language":"English","fastMode":true} object, preserved verbose/High/Ask and returned PICKER_QA_OK.
F4 controls
On 3571c9c2, Arguments toggle/reopen/default freshness and permission/Fast/Speed coupling passed bounded UI checks for both agents. Malformed literal text remained once after one Ask selection; no invalid launch.
F5 placement/scrolling
On 3571c9c2, both panels anchored without menu overlap; real-app viewport emulation covered 700/600 widths and shrink while open; actual inner scrolling reached Arguments/footer. PR real-Popper geometry tests complement that evidence. The 0.5px edge overrun is accepted within rounding.
Actual Claude Fast-on was not exercised: the QA load-balancer sign-in returned Standard from both picker and existing composer. Arguments and typed options were verified, without claiming Fast entitlement or a waiver. Physical OS window resizing was not exercised because the CDP method was unavailable; viewport coverage is the accepted evidence. Malformed typed-Chat policy readback and mutation of model/effort defaults were not exercised. This is a bounded flow verdict, not a blanket pass of unavailable provider behavior. QA admissions 3081 and 3172 are released.
For Brennan's review
Screenshots retain their actual flow/head provenance. Final controls/layout screenshots are from 3571c9c2; earlier replies cover unchanged flows.
Flush model search and list, one settings section with label-left rows
(Effort, compact Permissions menu, Fast mode), a fixed Arguments
disclosure row with its hint, and a summary footer with Launch chat /
Launch terminal. Settled defaults with nothing chosen keep the model
list and Effort pickable. Hovering another menu row no longer closes
the open panel.
Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 5 remain after this review.
The reason will be displayed to describe this comment to others. Learn more.
Actionable comments posted: 1
Caution
Some comments are outside the diff and can’t be posted inline due to GitHub limitations.
⚠️ Outside diff range comments (1)
Major · Project terminal permission on every terminal launch, not only… · AgentLaunchSetupPanel.tsx:44-65
src/renderer/src/components/agent/AgentLaunchSetupPanel.tsx:44-65 🔒 Security & Privacy | 🟠 Major | 🏗️ Heavy lift
Project terminal permission on every terminal launch, not only untouched drafts.
A draft that changes only the model or effort still has no permission choice and no Arguments. For a terminal launch, the launcher then derives permissions from the renderer settings. Those settings can be full-access while the panel shows the host's Ask for approval. The agent starts with broader permissions than the panel displays.
Gating the correction on a wholly untouched draft does not cover this path. Permission must be handled as its own inherited field.
Review info
Run configuration
Source mode: Auto
Repository: Unavailable
Branch: Unavailable
File: No configuration file from this repository applied
Commit: Unavailable
Layers: Repository UI
Review profile: CHILL
Plan: Advanced
Run ID: 77408610-698b-4855-bb9d-8a77c844e264
Commits
Reviewing files that changed from the base of the PR and between 40aae06 and efc43c0.
Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 3 remain after this review.
Project inherited terminal effort under the effort option ID.
For a Codex terminal entry such as gpt-5.6-terra, the picker displays the inherited effort value. The current sparse projection omits that value when arguments is undefined, so terminal launch can fall back to the renderer default. options.effort is the correct key: the Codex catalog maps it to -c model_reasoning_effort=<value>, and Claude maps it to --effort.
Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 1 remain after this review.
The reason will be displayed to describe this comment to others. Learn more.
Greptile has paused reviews on this repository — it used its 2000 free open-source review credits for this billing period. Reviews resume automatically on November 28. To continue before then, an organization admin can keep reviews running past the free credits — those bill as normal usage.
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
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.
The desktop + tab menu now lets you search for an agent setup, such as “opus high,” and launch it immediately. Clicking an agent row or pressing Enter launches with the current defaults. Only the chevron opens setup, where you can choose a model, effort, Speed/Fast and permissions, edit Arguments, and launch Chat or Terminal. Untouched launches preserve the existing defaults. Arguments and their removable permission/Fast badges stay in sync, and the setup panel stays anchored and scrolls within a narrow window.
This PR covers the desktop + tab menu. PR 2 in this lane brings the same picker to Create worktree, continuation, AgentCombobox, Source Control settings/Customize, automations and phone controls; those entry points keep their current surfaces here, tracked in #35802. The shared launch core is tracked in #34275 and has merged. Existing dispatchers and hosts continue to own execution and launch failures. There is no new RPC, saved field or minimum host version.
Differences from the common pattern:
Real Claude and Codex ordinary Chat/Terminal launches and supported searched launches produced replies. Configured Codex Chat and Terminal confirmed High, priority and Ask in provider settings. Final targeted checks on
3571c9c2verified Claude inline settings through actual argv and a real reply, malformed permission editing without duplicate text, Arguments toggling, Codex Speed choices, reopened defaults, anchored placement, narrow widths and inner scrolling. Claude's launched settings were one merged object preserving language English with Fast requested. Narrow 700/600 widths and shrink while open were verified by viewport emulation in the real app, alongside the PR's real-Popper geometry tests; physical OS window resizing was not exercised. The measured 0.5px edge overrun was accepted as rounding.Claude Fast-on was not exercised: it is unavailable with the QA load-balancer sign-in. Both the picker and the existing composer toggle returned Standard. The picker's Fast Arguments and typed Chat options were verified; actual Fast-on is not claimed. Malformed typed-Chat permission readback and mutation of model/effort defaults were not exercised; the malformed text check was a UI check with no invalid launch. Codex Chat launches with no explicit effort override intentionally remain sparse.
CI passes on reviewed head
2e676b48(the main merge preserves the QA'd picker code from3571c9c2); fresh correctness and mechanism reviews are clean. The reported Sonnet Fast loss was refuted through the final command compiler: 4 tests covering 16 paths retained Fast true or deliberate false and unrelated JSON.Deferred P3 polish, tracked in #35802: six setup labels use English fallback text in non-English locales; a menu with no agents and no query can show an extra search-empty line; and the separate Fast chip spells out a second
--settingsargument even though launch correctly merges one settings object. The chip display should make that merged meaning clear.Final Claude setup (the Fast chip display is the deferred cosmetic above):
Per-flow screenshots and final review.