Repository navigation
UI Consistency: 2 issues found
Two findings, both about the new rule that hides the ACP driver from text-generation model pickers:
apps/web/src/components/settings/SettingsPanels.tsx—GeneralSettingsPanelreads from the filteredtextGenerationProvidersbut bothupdateSettingshandlers for the same setting still callresolveAppModelSelectionState(..., serverProviders)(~L2485, ~L2518). Its fallback can then select the hidden ACP instance and persist a text-generation selection whose driver only installsmakeUnsupportedTextGeneration, while the UI displays a different provider.apps/web/src/components/settings/SourceControlWritingSettings.tsx— the ACP exclusion is hand-rolled per component (here and inGeneralSettingsPanel) with no shared owner;ProviderSettingsPanel.tsx:493, which decides whether disabling an instance clearstextGenerationModelSelection, still resolves against the unfiltered provider list and can therefore act on a different instance than the settings UI shows.
Both were posted as inline review comments.
Details
Note
Your check run agent prompt is: .macroscope/check-run-agents/ui-consistency.md
More information about how Check Run Agents work can be found in our Help Center.
Files examined in scope: apps/web/src/components/chat/providerIconUtils.ts, apps/web/src/components/settings/SettingsPanels.tsx, apps/web/src/components/settings/SourceControlWritingSettings.tsx, apps/web/src/components/settings/providerDriverMeta.ts, apps/web/src/session-logic.ts.
Context used (not targeted): apps/web/src/modelSelection.ts (resolveAppModelSelectionState fallback semantics), apps/web/src/providerInstances.ts, apps/web/src/components/settings/ProviderSettingsPanel.tsx, apps/web/src/components/settings/AddProviderInstanceDialog.tsx, apps/web/src/components/Icons.tsx, apps/server/src/textGeneration/TextGeneration.ts, packages/contracts/src/model.ts.
Checked and found acceptable: ACPRegistryIcon exists and already routes color through cn("fill-black dark:fill-white", className), matching the other brand glyphs; the new PROVIDER_OPTIONS entry ("new" sidebar badge) and PROVIDER_CLIENT_DEFINITIONS entry ("Early Access" badge) follow the existing Cursor/Grok pattern; no new raw controls, Tailwind overrides, global CSS, theme variants, scroll masks, or environment-scoped renderer call sites were introduced by the web changes. Duplicate ../Icons import statements in providerIconUtils.ts and providerDriverMeta.ts were not reported: the repo's lint config enables no duplicate-import rule, so this is a cosmetic nit.
🔍 Agent activity — 21 tool calls
💭 I'm checking the in-scope web files—providerIconUtils, SettingsPanels, SourceControlWritingSettings, providerDriverMeta, session-logic—and want to verify whether the imported ACPRegistryIcon actually exists, since the diff doesn't seem to add it.
1. git_tools — ACPRegistryIcon
2. browse_code — apps/web/src/components/chat/providerIconUtils.ts
💭 I notice AddProviderInstanceDialog.tsx already references ACPRegistryIcon, but that file wasn't part of the diff, so it must predate this change. I want to check how ACPRegistryIcon is defined in Icons.tsx and compare it to how other icons handle className and currentColor.
3. browse_code — apps/web/src/components/Icons.tsx
4. browse_code — apps/web/src/components/settings/AddProviderInstanceDialog.tsx
💭 I notice there's already a "coming soon" ACP registry option in AddProviderInstanceDialog using ACPRegistryIcon, and now the new "acp" driver adds a real entry in PROVIDER_CLIENT_DEFINITIONS labeled "ACP Agent" with the same icon. I need to check whether COMING_SOON_DRIVER_OPTIONS still displays that placeholder entry, which could create a duplicate or conflict now that a real driver exists.
5. browse_code — apps/web/src/components/settings/providerDriverMeta.ts
6. browse_code — apps/web/src/session-logic.ts
💭 Now I'm checking SettingsPanels.tsx and SourceControlWritingSettings.tsx for UI consistency issues. I notice providerIconUtils.ts and providerDriverMeta.ts both have duplicate imports from the same module on consecutive lines, which is likely a lint violation rather than just harmless ordering. I also want to check whether filtering out "acp" from text-generation providers in SettingsPanels.tsx and the filter on useAtomValue in SourceControlWritingSettings.tsx creates new arrays every render, which could break memoization or effect dependencies — worth inspecting both files directly.
7. browse_code — apps/web/src/components/settings/SourceControlWritingSettings.tsx
8. browse_code — apps/web/src/components/settings/SettingsPanels.tsx
💭 I'm noticing the acp-filter logic is duplicated across SettingsPanels.tsx and SourceControlWritingSettings.tsx, so I want to check for other usages of resolveAppModelSelectionState and resolveSourceControlWriterModelSelection in the web codebase, plus any other UI surfaces (like thread title generation settings) that expose text-generation provider selection and might need the same acp filtering.
9. git_tools — resolveAppModelSelectionState|resolveSourceControlWriterModelSelection|textGenerationModelSelection
💭 I notice ProviderSettingsPanel.tsx line 493 also calls resolveAppModelSelectionState with unfiltered serverProviders, possibly rendering a text-generation model picker per provider. I should check the surrounding code around lines 480-560 and 780-830 to confirm.
10. browse_code — apps/web/src/components/settings/ProviderSettingsPanel.tsx
💭 I want to trace where textGenInstanceId is used within ProviderSettingsPanel, so I'm checking its usage around lines 780-830.
11. git_tools — textGenInstanceId
💭 I'm noticing that ProviderSettingsPanel doesn't apply the same filtering the PR introduces elsewhere, so its resolution could diverge from GeneralSettingsPanel's — picking an ACP instance in one place and a non-ACP one in another. I need to weigh whether this inconsistency is genuinely in scope, since the rule is to flag only violations the changed lines introduce or pre-existing issues the patch makes newly relevant.
12. browse_code — apps/web/src/modelSelection.ts
💭 I'm looking at candidate A more closely: the ACP filtering logic is duplicated inline in two components, but a third consumer in ProviderSettingsPanel that also resolves the text-generation selection wasn't updated, so it could resolve to a different ACP instance than the settings UI. I'm trying to determine if this is a real regression by walking through a scenario where a user has only ACP and codex enabled, checking whether disabling an instance mis-targets the clear logic. Since textGenerationModelSelection is typically stored explicitly by the picker, and users can't select ACP via the general panel now, the practical impact seems modest — so I should consider whether there's a stronger finding elsewhere. Another angle worth checking is whether the filter-duplication issue could be resolved with a shared helper like
supportsTextGeneration(provider)in modelSelection.ts, following the principle of the smallest durable contract. I also want to verify whether the filtered provider array in SourceControlWritingSettings causes unnecessary re-renders downstream, and what happens if filtering leaves an empty provider list — could that break the model resolution logic?
13. browse_code — apps/web/src/components/settings/SourceControlWritingSettings.tsx
14. browse_code — apps/web/src/modelSelection.ts
💭 I'm noticing the web client filters providers before resolving the model selection, while the server likely uses the full provider list for actual dispatch — this divergence means the UI could display a different model than what the server actually uses. Thinking about the smallest correct owner, the ACP exclusion probably belongs centralized around a provider capability check (like text generation support reported by the server) rather than scattered driver-kind checks in components. I want to check whether ServerProvider already exposes a capabilities flag for text generation support.
15. browse_code — In the t3code repo, does the ServerProvider contract (packages/contracts) expose any capability flags indicating whether a provider supports text generation (commit message / PR content / thread title generation)? Also, is there any existing shared helper in apps/web/src (e.g. modelSelection.ts, providerInstances.ts) that filters provider lists for text-generation eligibility? Answer with file paths and relevant type/field names.
💭 I'm confirming the ACP exclusion is hardcoded in two web components, with ProviderSettingsPanel left untouched since ACP is valid there. Now I'm checking other settings panels around line 2461-2520 to see where text generation model selection is resolved for the general settings panel. Specifically, I want to check whether that resolution passes the unfiltered
serverProviderslist — that would be a real inconsistency, since the picker itself displays a filtered list while the write handler resolves against the unfiltered one.
16. browse_code — apps/web/src/components/settings/SettingsPanels.tsx
💭 I'm noticing an inconsistency: the picker's display uses the filtered
textGenerationProviderslist, but the write handlers still pass the unfilteredserverProviderstoresolveAppModelSelectionState. This likely only matters in the fallback path, since a valid selected instanceId would resolve the same way regardless of which list is used. If the selected instance isn't found (e.g. disabled), the fallback grabs the first enabled entry from the unfiltered list, which could be the ACP instance—something the panel deliberately excludes since ACP can't do text generation. That would let a broken selection get persisted despite the display never showing it, creating a divergence between what's saved and what's shown. I'm also checking whether ACP instances would even be first in the fallback ordering, sinceresolveAppModelSelectionStatedoesn't apply the panel's sorting. This gives me a concrete inline finding: lines 2485 and 2518 should filter againsttextGenerationProvidersinstead of the full list. There's a related issue too—bothSourceControlWritingSettingsandSettingsPanelsduplicate the ACP-exclusion check inline, andProviderSettingsPanel:493still resolves against all providers without that filter, suggesting this logic should live in one shared helper rather than being copied across components. Though I realize I can only leave inline comments on lines that actually appear in the diff, so I need to check whetherProviderSettingsPanelis even a changed file before flagging it there. I also want to check whether the duplicate imports from "../Icons" inproviderIconUtils.tsandproviderDriverMeta.tswould trip any no-duplicate-imports lint rule.
17. browse_code — Does the t3code repo's lint setup (oxlint, eslint, biome config files at repo root or apps/web) enable a rule against duplicate imports from the same module (e.g. import/no-duplicates, no-duplicate-imports)? Show the lint config files and enabled rule sets.
💭 The duplicate import is just a cosmetic nit, not worth flagging. Checking the badge labeling for the new ACP provider across the picker and metadata—it's consistent with how Cursor and Grok use "new" plus "Early Access" together, so that's fine. Now I'm looking at how PROVIDER_ICON_BY_PROVIDER maps icons for the ACP provider. Actually the ACP driver and the "ACP Registry" coming-soon driver share the same icon in the Add Provider dialog, but that seems intentional since both relate to ACP—not really a strong finding. Mobile's ProviderIcon usage is out of scope here. I want to confirm ProviderDriverKind accepts "acp" as a valid branded string, which it should since make() likely accepts arbitrary strings, and I should also check whether PROVIDER_DISPLAY_NAMES in the web app needs an "acp" entry for the settings panel toast.
18. git_tools — PROVIDER_DISPLAY_NAMES
💭 I notice PROVIDER_DISPLAY_NAMES doesn't have an entry for "acp", so it would fall back to the generic label formatter, producing "Acp" instead of the expected "ACP Agent" used elsewhere—a UI inconsistency worth flagging, though the file this occurs in is outside the current scope. Actually, driverKindLabel lives in apps/web/src/providerInstances.ts, which is in scope but not touched by this PR, so comments would need to land on diff lines regardless. I want to trace how deriveProviderInstanceEntries actually uses driverKindLabel as a fallback, and check whether the server's presentation logic (AcpDriver setting displayName to "ACP Agent") overrides it before the label is ever shown in the picker or settings cards.
19. browse_code — apps/web/src/providerInstances.ts
💭 The missing "acp" entry in PROVIDER_DISPLAY_NAMES is minor and out of scope since it's contracts-related. I'm now checking whether other web surfaces have provider→driver maps like docs links or auth instructions that would also need an "acp" entry.
20. git_tools — ProviderDriverKind.make("cursor")
💭 I'm checking whether the acp driver needs an entry in PROVIDER_SETTINGS, but that list looks like legacy per-kind settings while the new schema-driven form likely already handles acp generically. I'm settling on one solid finding involving SettingsPanels.tsx. The write handlers use unfiltered serverProviders while the display reads from textGenerationProviders, meaning an ACP instance could get persisted as the text-generation model even though the ACP driver rejects text generation calls, breaking thread titles and commit messages -- the fix is to have the write path use textGenerationProviders too. Tracing whether this is actually reachable, I think the fallback selection logic could pick an enabled-but-unavailable provider and fall through to selecting the ACP instance instead, letting an invalid selection slip into the settings store even though the UI never intended to show it as a choice. I also notice the same class of issue in SourceControlWritingSettings: while that component filters serverProviders consistently within itself, the resolver call passes only the client-filtered providers to a function whose true resolution happens server-side against the full unfiltered list, so the UI could display a different writer model than what actually gets used. Since fixing that divergence properly would require server-side changes and is out of scope for this review, I'm leaning toward folding it into a broader comment about extracting a shared filtering predicate instead of duplicating the heuristic across components, including the stale resolution logic in ProviderSettingsPanel. I'll post this as a secondary, phrased-as-suggestion comment on the SourceControlWritingSettings diff, focused on the concrete risk that ProviderSettingsPanel resolves against an unfiltered list and could decide to clear or keep text-gen state based on a different instance than what the settings UI actually displays — this is directly relevant since the patch is what introduces the exclusion concept in the first place. Now I'm working out inline comment placement: GitHub only allows commenting on lines present in the diff, so for SettingsPanels.tsx I'll anchor the comment to the new
textGenerationProvidersconstant around lines 1880-1882 and flag that the twoupdateSettingshandlers further down still resolve against the unfilteredserverProviderslist, since those exact lines aren't in the diff and can't be commented on directly. For SourceControlWritingSettings.tsx I'll anchor to lines 47-49. I'm explaining the risk: if the selected instance isn't enabled or available, the fallback logic picks the first eligible entry from whatever list it's given, which could still be the ACP instance the panel is trying to hide, letting the panel persist a broken text-generation selection. PassingtextGenerationProvidersconsistently to both read and write handlers keeps them in sync, so I want to include a suggested fix, possibly as a unified diff snippet in the comment body since those lines fall outside the actual diff. Since the two related issues are in different locations, a single diff hunk won't cleanly cover both, so I'll skip the diff format and just describe the fix in prose. For the second comment on the settings component, I'm noting that the ACP exclusion logic is now duplicated across components and suggesting a shared helper function alongside the existing resolver so all text-generation consumers stay consistent — and flagging that another part of the panel still resolves the text-gen instance from the unfiltered list, which could cause the model selection to be cleared inconsistently. Since this is a genuine regression risk, I'll mark the review as a failure and post both comments.
21. modify_pr — Both findings concern the new "hide ACP from text-generation pickers" rule: the exclusion is applied to the read path only, and is duplicated per call site rather than living in one shared helper next to resolveAppModelSelectionState.
Agent Credits: 265 credits
Token usage: 62 uncached input · 1,578,865 cache-read input · 185,411 cache-write input · 22,694 output
Agent Credits may also include non-token charges from external tools such as web research.