Voice: add speech-recognition language selection (Wispr Flow) (QUALITY-1066)

[warp-agent-staging]

Summary

Client half of QUALITY-1066 (voice language selection). Adds a "Speech Recognition Language" dropdown to Settings → AI → Voice so a user can pin the language Wispr Flow transcribes in. The selection flows from the client → the warp-server transcribe API → the Wispr Flow request body. "Auto-detect" (the default) sends no language, preserving today's behavior.

This is the client half; the server half (which forwards the language to Wispr) lives in the matching PR below. Both must ship for the feature to work end to end.

Changes

• app/src/settings/ai.rs: new VoiceInputLanguage enum setting — AutoDetect default, cloud-synced globally (SyncToCloud::Globally(RespectUserSyncSetting::Yes)), toml_path agents.voice.voice_input_language; plus code() (ISO 639-1) and display_name() helpers. Field added to AISettings.
• app/src/settings_view/ai_page.rs: renders the "Speech Recognition Language" dropdown immediately below the "Key for Activating Voice Input" dropdown, inside the same is_voice_input_enabled gate; adds the SetVoiceInputLanguage action + handler, the AISettingsChangedEvent::VoiceInputLanguage subscription, and enable/disable wiring.
• app/src/voice/transcriber.rs, app/src/server/voice_transcriber.rs: Transcriber::transcribe gains language: Option<String>, forwarded onto TranscribeRequest.
• app/src/editor/view/voice.rs and app/src/ai/blocklist/agent_view/agent_input_footer/mod.rs: read the setting and pass the language at both transcribe call sites.
• app/src/settings/ai_tests.rs: unit tests for the enum default/code() and TranscribeRequest language serialization.

Language list (v1, curated): Auto-detect + en, es, fr, de, it, pt, nl, hi, ja, ko, zh, ar, ru, pl, sv, tr, uk, id, vi, th.

Verification

• Unit tests: voice_input_language_default_is_auto_detect, voice_input_language_code_returns_iso_639_1_code, transcribe_request_omits_language_for_auto_detect, transcribe_request_includes_selected_language.
• ./script/format --check passes; ./script/check_no_inline_test_modules passes.
• cargo check --tests -p warp passes (the test crate, including ai_tests.rs, type-checks cleanly). Building/running the full warp lib test binary OOMs in this 15 GB sandbox (the single rustc --test process for the warp lib exceeds the memory limit); CI runs the full test suite.
• cargo clippy -p warp --all-targets --tests -- -D warnings passes with both default features and --features voice_input.

Visual proof (user-facing change) — captured from a signed-in, AI-enabled build

The Warp GUI was built from this branch (cargo build --bin warp --features voice_input, the internal/dogfood channel) and launched signed in via WARP_API_KEY against staging. The "Speech Recognition Language" dropdown was exercised on screen:

1. Voice section, Voice Input ON — both dropdowns visible:
   
2. Dropdown default = "Auto-detect" (closed dropdown):
   
3. Dropdown expanded — full 21-option list (Auto-detect; English; Spanish (Español); French (Français); German (Deutsch); Italian (Italiano); Portuguese (Português); Dutch (Nederlands); Hindi (हिन्दी); Japanese (日本語); Korean (한국어); Chinese (中文); Arabic (العربية); Russian (Русский); Polish (Polski); Swedish (Svenska); Turkish (Türkçe); Ukrainian (Українська); Indonesian (Bahasa Indonesia); Vietnamese (Tiếng Việt); Thai (ไทย)):
   
4. A different language selected — the closed dropdown label updates (here to "Indonesian (Bahasa Indonesia)"):
   
5. Voice Input toggled OFF — both the "Speech Recognition Language" and "Key for Activating Voice Input" dropdowns are hidden:
   

Screenshots are hosted on a throwaway draft release (voice-language-proof-QUALITY-1066, no tag created) purely to attach the images to this PR; it can be deleted after this PR merges.

Rework (review cycle 1)

• Addressed the review nit: hoisted the use crate::ai::voice::transcribe::TranscribeRequest; import from inside the two new test functions to the top of app/src/settings/ai_tests.rs (the file's top-level-import convention). No behavior change (commit 786d5e1).
• Addressed the blocking review concern (missing visual evidence): captured the signed-in visual proof above.

Matching PR

Server half: https://github.com/warpdotdev/warp-server/pull/12740

Originating thread: https://warpdotdev.slack.com/archives/C0BDQDW8V5E/p1784060704730339

Spec: agents/specs/QUALITY-1066: voice language selection.md (committed in this PR)

This PR was generated with Oz.

[warp-agent-staging]

Overview

Client half of QUALITY-1066 (voice language selection). Adds a VoiceInputLanguage enum setting (AutoDetect default, cloud-synced globally), a "Speech Recognition Language" dropdown in Settings → AI → Voice (gated on Voice Input enabled), and extends Transcriber::transcribe with language: Option<String> wired at both call sites (editor/view/voice.rs and the CLI agent_input_footer). The code is a faithful, correct mirror of the existing VoiceInputToggleKey pattern.

What I verified (the code is sound)

• VoiceInputLanguage mirrors VoiceInputToggleKey exactly: same derive list, same #[schemars(... rename_all = "snake_case")], same implement_setting_for_enum! shape (differing only in SyncToCloud::Globally vs Never, the toml path, and the description — all per spec). ISO 639-1 code() values are correct (en, es, … th).
• The AISettingsChangedEvent::VoiceInputLanguage variant is macro-generated by define_settings_group! from the new voice_input_language field on AISettings (confirmed in crates/settings/src/macros.rs) — so no explicit enum edit was needed.
• The dropdown creation, SetVoiceInputLanguage handler, render call, and event subscription all mirror the toggle-key equivalents, including the is_voice_input_enabled gate. Sensible additions: set_menu_max_height for the longer 22-item list, and correctly omitting the explicitly_interacted_with_voice one-time flag (per spec). LocalOnlyIconState::for_setting correctly shows no local-only icon for this globally-synced setting.
• Both transcribe call sites read voice_input_language.code().map(...) and pass Option<String>; AutoDetect → None → language omitted end-to-end. Unit tests cover default/code()/TranscribeRequest serialization (omits for AutoDetect, includes "nl" for Dutch).

Concern

• Missing visual evidence (blocking). This is a user-facing UI change (a new dropdown). The attached screenshot shows the Voice section with the Voice Input toggle disabled (AI requires a signed-in account) — it does not show the "Speech Recognition Language" dropdown rendered, the "Auto-detect" default, the option list, or the hide-when-Voice-Input-off behavior. Per this repo's review-pr-local guidance, author environment limitations (logged-out OSS sandbox) do not exempt user-facing changes from visual evidence. Please attach a screenshot or short recording captured in a signed-in, AI-enabled Warp environment showing: (1) the dropdown rendered below the key dropdown with "Auto-detect" selected by default; (2) a different language selected with the label updated; (3) the dropdown hidden when Voice Input is toggled off. The OSS sandbox can't sign in (this review sandbox has the same limitation), so capture it from a local desktop or a computer-use-enabled credentialed environment (see Warp's computer use docs). This corresponds to spec validation criteria #4 and #5, which are currently unmet.

Note

The matching server PR (#12740) has a separate review comment on the Wispr language field shape (the documented API expects a list ["en"], but the server sends a bare string). Resolving that is also required for the feature to work end-to-end.

Minor (non-blocking)

The new tests in ai_tests.rs embed use crate::ai::voice::transcribe::TranscribeRequest; inside two test functions; the file's convention is top-level imports (5 top-level use statements, 0 embedded elsewhere). Consider hoisting it to the top with the other imports.

Verdict

Found: 0 critical, 1 important, 0 suggestions

Request changes — the code is correct, but a user-facing dropdown change requires visual evidence of the rendered dropdown (Auto-detect default + option list + hide-when-off) captured in a signed-in, AI-enabled environment before merge.

[warp-agent-staging]

Rework (review cycle 1) — changes requested addressed.

Visual proof (the blocking concern): Captured from a signed-in, AI-enabled build. The Warp GUI was built from this branch (cargo build --bin warp --features voice_input, internal/dogfood channel) and launched signed in via WARP_API_KEY against staging. The "Speech Recognition Language" dropdown was exercised on screen — default "Auto-detect", the full 21-language option list, selection updates the label, and it hides when Voice Input is toggled off:

Full set + the expanded list are in the PR description. (Images are hosted on a throwaway draft release voice-language-proof-QUALITY-1066 — no tag created — purely to attach them here; deletable after merge.)

Nit (non-blocking): Also hoisted the use crate::ai::voice::transcribe::TranscribeRequest; import from inside the two new test functions to the top of app/src/settings/ai_tests.rs (786d5e1).

Note on the matching server PR (#12740): the Wispr language field is now sent as a single-element list []string{request.Language} per Wispr Flow's documented API (confirmed against the API docs), so a selected language flows end-to-end correctly.

[warp-agent-staging]

Re-review (rework cycle 1) — verdict: Approve

Blocker resolution verified

The blocking finding from the prior review (missing visual evidence for a user-facing change) is resolved. The PR body now embeds 5 screenshots captured from a signed-in dogfood build of this branch (--features voice_input, staging auth). This review downloaded and inspected the assets directly and validated them against the spec's acceptance criteria:

• Invariants #1/#2: the "Speech Recognition Language" dropdown renders in Settings → AI → Voice immediately below "Key for Activating Voice Input", with Auto-detect selected by default. ✔
• Full list: the expanded dropdown shows all 21 options (Auto-detect + 20 languages with native names). ✔
• Invariant #3: selecting a language updates the closed dropdown's label (shown with "Indonesian (Bahasa Indonesia)"). ✔
• Hide gate / criterion #5: with Voice Input toggled off, both dropdowns are hidden (not greyed out). ✔

The prior nit (test-only import hoisted to the top of ai_tests.rs) was addressed in commit 786d5e1 with no behavior change.

Rubric

Correctness, naming, comments, tests, and security all look good. The setting mirrors the existing VoiceInputToggleKey pattern; cloud sync choice (Globally(RespectUserSyncSetting::Yes)) matches the spec's rationale; the Transcriber trait change is wired at both transcribe call sites (editor + CLI agent footer) — a correct superset of the spec. The 4 unit tests cover default, code() mapping, and TranscribeRequest serialization for both Auto-detect and selected-language cases.

Notes for the human merger (not blocking)

• CI coverage: the Rust CI jobs (Formatting + Clippy, tests, release compilation) are skipped on this draft PR — they run when the PR is marked ready / enters the merge queue. Locally the author validated ./script/format --check, cargo check --tests -p warp, and cargo clippy -p warp --all-targets --tests -- -D warnings (both feature sets); the full test binary OOMs in the 15 GB sandbox. Confirm the full suite is green once marked ready.
• The throwaway draft release hosting the screenshots (voice-language-proof-QUALITY-1066) can be deleted after merge.

Verdict

Found: 0 critical, 0 important, 0 suggestions

Approve