Show the Kimi logo for Kimi models in the model picker

[warp-agent-staging]

Description

Every Kimi model in the model picker rendered the Warp glyph. model_leading_icon (app/src/ai/llms.rs) ends in llm.provider.icon().unwrap_or(Icon::Agent), and the server reports the Kimi models (kimi-k26-fireworks, kimi-k27-code-fireworks, kimi-k3-fireworks) as LLMProvider::Unknown — which has no logo — so they fell through to the Icon::Agent fallback while OpenAI, Anthropic, Google and xAI each resolved to their own mark.

This adds the Kimi mark and wires it up:

• app/assets/bundled/svg/kimi.svg — the Kimi monogram in the repo's bundled-icon convention (24×24 viewBox, single tintable path with the #FF0000 placeholder fill used by grok.svg / openai.svg / claude.svg). Bundled icons are rendered monochrome and coloured by the call site, so one asset covers light and dark themes at every size the picker uses.
• Icon::KimiLogo in crates/warp_core/src/ui/icons.rs (enum variant + path mapping), declared alongside ClaudeLogo / GeminiLogo / GrokLogo / OpenAILogo.
• is_kimi_model in app/src/ai/llms.rs, branched into model_leading_icon.
• ModelIconFlags::is_custom_endpoint, computed at both call sites from LLMPreferences::custom_llm_info_for_id like the existing flags.

Why match on the model id rather than the provider. LLMProvider is generated from the server's GraphQL enum (OpenAI | Anthropic | Google | Xai | Unknown); there is no Moonshot/Kimi variant, and adding one requires a server-side schema change. Following the existing is_custom_router_id / is_auto precedent, is_kimi_model matches the id (kimi exactly, or a case-insensitive kimi- prefix) rather than the display name, so an unrelated id that merely contains the letters — kimimaro-7b, gpt-kimi — does not match.

Custom-endpoint models are excluded. Both surfaces also list the user's configured custom/BYO endpoints, and those synthetic LLMInfos use the endpoint model's persisted config_key as their id (custom_llm_info_from, app/src/ai/llms.rs). A config_key carries no provider provenance, so an endpoint named kimi-anything would otherwise have been given Kimi's trademark. ModelIconFlags::is_custom_endpoint gates the Kimi branch off for those rows; they keep the generic fallback.

The new branch sits after the custom-router / auto / Bedrock / Gemini-Enterprise branches and before the provider fallback, so existing icon precedence is untouched: a Kimi model routed through Bedrock still shows the AWS mark, and an auto row still shows the generic agent glyph.

Out of scope: the input footer's model chip. render_model_section (app/src/terminal/profile_model_selector.rs) renders the model name plus a chevron and never calls model_leading_icon, so no model of any provider gets a logo there today. model_leading_icon has exactly two call sites — the inline picker (app/src/terminal/input/models/data_source.rs) and the model menus (app/src/ai/execution_profiles/model_menu_items.rs) — and both are fixed by this change.

Linked Issue

No tracked issue — requested directly in Slack: https://warpdev.slack.com/archives/C0BDQDW8V5E/p1786075128189149

• Where appropriate, screenshots or a short video of the implementation are included below.

Testing

Three unit tests in app/src/ai/llms_tests.rs, alongside the existing model_leading_icon precedence tests:

• kimi_models_are_matched_by_id_not_by_provider — kimi, kimi-k26-fireworks and KIMI-K3-Fireworks all resolve to Icon::KimiLogo while their provider is still Unknown; kimimaro-7b, gpt-kimi and gpt-oss-120b keep the Icon::Agent fallback.
• kimi_models_keep_the_existing_router_auto_and_host_precedence — a Kimi id still yields Dataflow / Agent / Aws / GeminiEnterpriseAgentPlatform under the corresponding flags.
• custom_endpoint_models_never_get_the_kimi_logo — a real build_custom_llm_infos row built from an endpoint whose config_key is kimi-anything keeps Icon::Agent, while the same LLMInfo without the flag still resolves to Icon::KimiLogo (so the flag, not the id shape, is what excludes it).

Checks run on this branch:

• ./script/format — clean.

• cargo clippy -p warp_core -p warp --all-targets --tests -- -D warnings — clean (exit 0).

• cargo nextest run -p warp -E 'test(llms)' — 36 passed, including all three new tests and the pre-existing auto/host precedence tests.

• I have manually tested my changes locally with a local build (cargo build --bin warp, run on Linux).

Manual verification. Captured on a locally built internal (--bin warp) client on Linux, driving the real /MODEL picker from an Agent Mode conversation. The model list is the real staging catalogue fetched by the signed-in client — nothing about the model list or the icon path was stubbed; available_model_menu_items → model_leading_icon → Icon::KimiLogo runs as production code.

• All three Kimi rows — kimi k2.6, kimi k2.7 code, kimi k3 — render the Kimi monogram (an angular K with a detached rounded pill at the upper right).
• The unfiltered list is the direct before/after. inkling, minimax 2.7, minimax 3, qwen 3.6/3.7 plus, deepseek v4 pro, gpt-oss 120b and muse spark 1.2 are all Unknown-provider models too, so they still render the Icon::Agent Warp glyph — which is exactly what the Kimi rows looked like before this change. The Kimi rows in the same list now render a visibly different mark.
• Provider precedence is unchanged: claude * rows still show the Anthropic starburst, gpt-5.* the OpenAI knot, grok * the Grok arc-and-slash, gemini * the Gemini mark, and the auto * rows the generic agent glyph.
• Selecting kimi k2.6 updates the footer chip to kimi k2.6 (fireworks). That chip shows no provider logo — expected, and unchanged: render_model_section never renders one for any model.

Disclosure on how the signed-in session was reached. This sandbox carries only a service-account WARP_API_KEY. The client's startup API-key sign-in is gated behind establishing staging IAP transit first (authenticate_user_after_iap_access), and the injected WARP_STAGING_IAP_BOOTSTRAP_JWT had already expired (15-minute lifetime), so IAP came up AccessUnavailable and sign-in never started — leaving the app on the logged-out onboarding screen with all AI surfaces gated. Re-minting that bootstrap JWT with oz federate issue-token and re-launching restored IAP transit, after which the client's own API-key sign-in completed normally. No source was patched and no auth predicate was forced; the branch diff is exactly the icon asset, the source files and the test file listed above, and the working tree was clean at push time.

Screenshots / Videos

Captured at df132cb. The follow-up commit only narrows which rows qualify for the logo (custom endpoints are now excluded), which no server-catalogue row in these captures is, so they still show the shipped behavior; that case is covered by custom_endpoint_models_never_get_the_kimi_logo.

Computer-use video recording

Opening Warp model picker and inspecting kimi model glyphs: Opens Warp's AI model picker, filters to kimi models, clears the filter to compare glyphs with other models, then selects a kimi model and returns to the input.

Computer-use screenshots (3)

Model picker filtered to kimi rows (kimi k2.6, kimi k2.7 code, kimi k3), each with a small capital-K-like glyph plus a detached dot at upper right to the left of the label.

Unfiltered model list showing kimi rows (K-glyph) alongside grok rows (Warp-independent circle-with-slash) and inkling / minimax rows (the unchanged Icon::Agent Warp glyph), allowing side-by-side comparison.

Agent input footer after selecting a kimi model — the model chip reads "kimi k2.6 (fireworks)". The footer chip renders no provider logo for any model; this surface is unchanged.

Agent Mode

• Warp Agent Mode - This PR was created via Warp's AI Agent Mode

CHANGELOG-IMPROVEMENT: Kimi models now show the Kimi logo in the model picker instead of the Warp logo.

Conversation: https://staging.warp.dev/conversation/a9dc6e6d-1e4f-4cff-9e77-3c75aa8f020e
Run: https://oz.staging.warp.dev/runs/019fda6a-16a0-7d3f-80a1-0140957d8966

This PR was generated with Oz.