feat(web,mobile): @-mention a model to name it for the agent#31207

Problem

To tell an agent which model to use, for example when delegating, you have to type a provider's exact model id from memory. Nothing helps you find it, and nothing in the message shows whether the id is real.

Change

Typing @ plus part of a model or provider name (@opus, @gpt) now offers matching models in the mention menu. Each typed word must start a word of the model's or provider's name, from two characters, so a short query like @a still lists only files. Picking one inserts model:<instance>/<model> into the prompt, for example model:cursor/deepseek-v41-flash-1017. It shows as a chip with the provider's icon in the composer and in sent messages, on web, desktop, iOS and Android. Tapping a chip shows its name, provider, and exact reference.

• Chips take the logo's color. Providers whose logo has its own color (Claude, Kiro, Muse, Antigravity) tint the chip with it; logos that swap between black and white per theme (OpenAI and others) take the text color, as the logo does. Web keeps the theme's contrast-solved chip lightness and takes the logo's hue; mobile runs the color through its existing chip contrast solver.
• It names a model; it does not switch one. The thread's model is unchanged, and mentions offer every ready provider's models, not only those the thread could switch to.
• The prompt text is the contract. No schema, server, or adapter changes. Agents receive plain text, and the copy actions keep the source text.
• Unknown references stay text. A mention becomes a chip only if its provider instance and model both resolve in that environment, so a half-streamed or invented slug is never presented as real. References inside math stay text too.
• Titles stay readable. A thread title seeded from the message uses the model's slug rather than the raw reference.
• Web and mobile share the logic. Ranking and parsing are in @t3tools/client-runtime/composer-model-mentions. modelPickerSearch moves to client-runtime for this. Its only code changes are exporting its type and using .sort on its own local array, because Hermes has no toSorted.

One mobile change this feature needed goes beyond the mention itself:

• Provider logos for native chips. Native chips take raster images, while provider logos are SVG, so sync-context-chip-icons.mjs now rasterizes them. Logos that swap between black and white per theme take the chip's accent, and colored logos keep their colors, as on web. ACP registry icons are SVG URLs a native chip can't decode, so those chips use a generic glyph.

The Android fix for taps on a chip's right half applies to every chip type and changes the native fingerprint, so it is now its own PR, #33126. This PR has no native changes.

The mobile mention menu is also now headed Mentions rather than Files when it lists anything other than files.

Provider status pushes do not rerender messages or the composer, because the chip atoms keep their identity until the labels or model list actually change.

Scope and approval

This is a feature, and it has no linked maintainer approval: I did not open an Ideas discussion first, so it does not meet the contribution guide as written. I run this in my fork and am offering it so you can see the working behavior. Please close it or redirect it to a discussion if that is the better route.

Maria (maintainer) took this PR over and trimmed it to the feature: the Android tap fix moved to #33126, the web rich-text builders keep main's signature, and tests that repeated another layer's coverage were dropped.

Verification

Current head c48c806c13, merged with current main (the web checks below ran on 2eb9556f62; later commits restore a mobile test from main and tint chips with the logo color, checked separately below):

• vp run check:changed passes: format, lint, knip, typecheck of 40 packages, 19 test files (483 tests) including the ChatMarkdown browser test, and React Compiler coverage.
• Web, by hand in a real client: @opus and @gpt list models with provider icons above files; picking one inserts a chip and leaves the thread's model unchanged; the popover shows the exact reference; backspace removes the space, then the chip; @a and @agent list only files; a message sent on the free Replay (dev) provider keeps the chip, and the thread title reads ask gpt-6.1-sol to review.
• Not re-run on this head: desktop, iOS and Android. The mobile code matches the earlier head below except for the @ ranking and how message chip atoms are keyed.

Earlier head 42d765c53e (by the original author). CI on that head: 23 passed, 0 failed.

The desktop script checks that it runs in Electron with the desktop bridge; @gpt offers the model; picking it inserts a chip with no raw model: text; backspace removes the trailing space and then the chip whole; the thread's model selector is unchanged; the title never shows raw model: text; the sent message renders the chip; and the popover shows the exact reference.

The device runs check the menu; the chip; the tap dialog; typing after a chip; backspace; references inside math; hand-typed and unknown references; the sent message; long-press copy and share; and the Mentions and Files headings.

Provider icons:

Dark mode was checked on iOS only.

An agent also acted on a mention: a real turn on GPT-6-Astra, told to delegate to the GPT-6.1-Sol chip, started a child on Codex 6.1 Sol. That run was stopped early, so it shows the delegation starting, not the work finishing.

Automated: vp run check:changed passes at the head. The full mobile (278 files) and client-runtime (151 files) suites pass, as do the browser tests ChatMarkdown and composer-list-split. Fixes for logic bugs each have a test that failed first: thread titles, composer tokens, chip details, math exclusion, backspace, and the SVG fallback.

Web on the current head

@gpt lists four Codex models with their icon above the files

picking a model inserts a chip with the provider icon; the model selector stays on 6 Astra

the chip popover shows the name, provider and model:codex/gpt-6.1-sol

@agent lists only files, no Claude models

a sent message keeps the chip and the thread title reads ask gpt-6.1-sol to review

Chip colors on the current head

Before, every model chip used the model accent:

before: the Claude Opus 5.5 chip uses the olive model accent

After, the Claude chip takes Claude's orange and the Codex chip takes the text color, like the OpenAI logo. The Claude chip's contrast on the composer is about 6.3:1 in light and 5.9:1 in dark:

after, light: Claude chip in burnt orange, Codex chip in the text color

after, dark: Claude chip in orange, Codex chip in the text color

Mobile chip colors are not yet seen on a device.

Web and desktop (earlier head)

Mobile (earlier head)

Not checked, and known limits

• Android backspace used adb key events, not an on-screen keyboard's delete key. One earlier run also saw 15 delete events remove 21 characters, which a smaller re-test did not reproduce and I have not explained. It may be the test tooling's helper keyboard.
• Screen readers: Android's composer exposes the raw model: text as its accessibility value, and sent chip labels include the raw reference on both platforms. Neither is visible on screen.
• Android cursor: tapping a composer chip can leave the cursor in front of it, so text typed right after closing the dialog would land there.
• Menu order: threads rank above models, so a thread title matching the query pushes the models down, and on mobile you may have to scroll.
• Slugs the syntax can't express: a custom model slug containing a space, comma or quote stays plain text. The agent still receives what you typed.
• Console warning: inserting a model chip from the menu logs React's "flushSync was called from inside a lifecycle method" warning on web. Nothing visible breaks; not yet traced to this PR or main.
• Code spans: a reference inside backticks followed by a space can still chip, which matches how $skill tokens already behave.

Implemented with Claude Opus 5.5 via Claude Code; takeover and trim with Claude Opus 5.5 in Claude Code via T3 Code.

🤖 Generated with Claude Code

Reviewed with sol-loop: 5 rounds with claude-opus-5-5 at unknown. Verdicts: approved, approved.