Summary

Polish pack for the mobile chat experience (iOS and Android): model picker favorites/recents, per-model reasoning-effort gating, a context-window usage indicator, opt-in link previews, math rendering, and streaming text reveal polish — shipped as a sequence of small independent PRs.

Problem to solve

The mobile chat surfaces are functional but lack several quality-of-life refinements users expect from modern chat clients:

• The model picker treats all models equally; frequently used models require scrolling on every switch.
• The thinking-level control is shown regardless of whether the selected model supports configurable reasoning effort, which is confusing when it has no effect.
• There is no visibility into how full the session context window is, so long sessions hit compaction or truncation without warning.
• Bare URLs in the transcript render as plain links with no preview affordance.
• LaTeX math in assistant responses renders as raw TeX source.
• Streamed prose appears in raw chunks without any reveal polish.

Proposed solution

A sequence of small, independently reviewable PRs, one per item and platform:

1. Model picker favorites + recents — pin/unpin favorites and recents-first ordering, persisted per platform (UserDefaults / DataStore).
2. Per-model reasoning-effort gating — hide or disable the thinking-level control when the selected model does not support it, driven by gateway model metadata; if the capability flag is missing from the model catalog schema, add it additively first.
3. Context-window usage indicator — a subtle composer indicator of tokens used vs. the model context window, based on what the gateway exposes per session; scope an additive RPC field first if nothing suitable exists.
4. Link previews in transcript — OpenGraph title/image cards for bare URLs, fetched client-side with a small bounded cache. Product call: previews load on tap-to-expand only — no ambient network fetches from message content (see Security: Indirect prompt injection via URL link preview metadata #22060 for the injection/exfiltration concern that motivates this default).
5. Math rendering (iOS/shared renderer) — LaTeX display/inline math in the shared chat renderer, extending the existing block-segmenter pipeline; Android math rendering is a noted follow-up, not in scope here.
6. Streaming text reveal polish (iOS/shared renderer) — word-paced fade-in for streaming prose; code and tables continue to render plain until complete; respects Reduce Motion.

Each PR carries tests along established seams, i18n sync when strings change, and license files for any newly bundled dependency.

Alternatives considered

• One large PR: harder to review and revert; the items are independent by design.
• Ambient (automatic) link-preview fetches: rejected for privacy and prompt-injection exposure; tap-to-expand keeps network activity user-initiated.
• Server-side rendering of previews/math: heavier gateway surface for what is client presentation concern.

Impact

• Affected: iOS and Android app users in chat
• Severity: quality-of-life; individually small, collectively a noticeable experience gap
• Frequency: daily (model switching, long sessions, links and math in responses)
• Consequence: extra taps to switch models, confusing no-op controls, surprise context compaction, raw TeX in transcripts

Evidence/examples

Comparable behavior in mainstream chat/LLM clients (favorites/recents in model pickers, context usage meters, tap-to-expand link cards, LaTeX rendering). Related: #22060 (link-preview metadata injection risk), #64607 (inline media display for channel surfaces — different surface, complementary).

Additional information

Gateway/protocol changes (capability flag, usage stats) are additive only, per the gateway protocol compatibility rules. Android math rendering is explicitly a follow-up.