[Feature]: Text-generation failure popups should name the attempted model and link to Settings #17995

Thanks — this is a real gap on top of #12651 / #12652, and the code matches the proposal.

Not a duplicate. #12651 is the fallback choosing the wrong slug. #12652 only changes what the Claude CLI error says. Neither names the attempted model or points at Settings. #5359 is the silent title/branch half of the same family.

What happens today

Commit / PR generation already reaches the client: GitActionsControl turns any failure into a generic "Action failed" toast with TextGenerationError.message (Text generation failed in generateCommitMessage: …). No model name, no action.

That error type is only { operation, detail, cause } (packages/contracts/src/git.ts). Drivers receive modelSelection but do not put the slug on the error.

resolveTextGenerationProvider() / fallbackTextGenerationProvider() in apps/server/src/serverSettings.ts do know they rewrote an unset/unusable selection to DEFAULT_TEXT_GENERATION_MODEL_BY_PROVIDER (e.g. claude-haiku-4-5). They then overwrite textGenerationModelSelection in place, so GitManager / the drivers later see a concrete selection and cannot tell it was a fallback unless we thread that through. The Settings picker can also display a catalog default via resolveAppModelSelectionState() without persisting textGenerationModelSelection — which is how #12651 stayed hidden.

Title / branch generation still only Effect.logWarnings in ProviderCommandReactor (maybeGenerateThreadTitleForFirstTurn, maybeGenerateAndRenameWorktreeBranchForFirstTurn, regenerate-title). Same failure, no popup at all — that is #5359.

What already exists for the Settings jump

• Picker: Settings → General → Text generation model (#text-generation-model)
• Source-control override: #source-control-text-generation (used when sourceControlWriterModelSelection is set — GitManager prefers that for commit/PR)
• Hash navigation: command palette / settings search already do navigate({ to: "/settings/general", hash: "text-generation-model" })
• Toast "Settings" CTA pattern: ProviderUpdatePrimaryNotification

Suggested shape

1. Put the attempted instance + slug on TextGenerationError (or the toast payload). If the resolver fell back, say so explicitly: no text generation model was configured, so the provider default (claude-haiku-4-5) was used.
2. On commit/PR failure toasts, add a Settings button to /settings/general#text-generation-model (or #source-control-text-generation when the writer override was the one that ran).
3. When [Bug]: Thread title and branch-name generation fail silently when the text generation provider is unhealthy #5359 surfaces title/branch failures, reuse the same copy + CTA so the whole one-shot family is consistent.

Smaller useful slice if we split it: always name the attempted model + Settings button, then add the "this was a fallback" origin once the resolver exposes it.

Accepting as a QoL enhancement. No extra info needed. Keep #12651 (fallback slug) and #12652 (CLI detail) separate — this is presentation only.