Task: Adapt Agents Code-style voice composer to current trial

Identity

Task ID: 20261013-agent-composer-trial-d76b48c2
Mode: Feature
Branch: codex/20261013-agent-composer-trial-d76b48c2-agent-composer-trial
Worktree: /Users/chillishark/Makelore 麦洛/.codex-worktrees/makelore/20261013-agent-composer-trial-d76b48c2
Base commit: 7c51bb4a00
Owner: codex-root
Status: Implementing

Adapt the Agents composer onto running trial source: Main 7c51bb4a plus its exact current Renderer overlay 154a21a (JS 64b4a84 and timestamp CSS 154a21a). Preserve stream replies, avatar, attachment sheet, delivered-file privacy, newer account/settings, recovery and knowledge layout. Voice recognition, draft-only appending, Code sizing/keyboard behavior, multi-file/paste/drop upload are in scope.

User requests Code-style Agents chat input with voice; speech is application-level, independent of Learning. Main owns network/auth/files. Preserve newer trial UI and peer work; no provider/model configuration or automatic submission.
Concurrent Task Gate: isolated owned feature worktree registered and verified. Peer registry scopes were reviewed in preceding candidate task; edits here partitioned explicitly among root and three subagents. Primary foreign document remains untouched.
Planning Gate: passed after project entry, positioning, current-state, decision and architecture review. Main-baseline candidate 48c592f remains retained; this adaptation starts from exact current trial 7c51bb4a rather than removing newer product features. Only this task record and README are written as documentation.

Source implementation complete. Preserved peer-committed Renderer delta 7c51bb4a → 154a21a incrementally, including matching tests and README; only composer and typed upload changes differ in product tree from actual predecessor. Shared voice hook now resets mounted state during React StrictMode effect setup. No cloud model contract or automatic prompt submission changes. Current running app is not yet updated.

Current-baseline 253 related regression tests passed (11 files), 28 voice tests including StrictMode passed, and 31 new upload/lifecycle tests passed: 312 disjoint cases. Final typecheck, focused lint (0 errors; two existing CloudChat warnings), full Renderer/Main/Preload/utility Vite build and project-docs/diff gates passed. Independent review has no open blocker. 17 disjoint Electron scenarios passed: 12 in first run plus 5 targeted reruns after updating inherited timestamp/recovery assertions to the exact current trial contract. Published voice fixture validates WAV transcription, latest-draft append, no automatic prompts, 32px buttons and multi-line growth; layout covers narrow/wide windows, long messages, attachments and stream/stop. Voice screenshot visually inspected by root and QA. Final comparison moved the microphone beside send on the right, matching Code; a fresh full build/typecheck and two targeted Electron reruns passed, including 4px adjacency and 375px layout. Fixture audio never accesses the real microphone or paid provider. Real microphone/provider transcription remains distinct from fixture validation.

Task-owned compiled runtime is staged at .trial-runtime/agent-composer-20261013-d76b48c2; retains predecessor native shell, profile and Renderer origin. Before activation compare current pointer/compatibility link, inspect live UI for unsaved input, quit gracefully through CUA, then replace and verify visible composer. No profile contents read or copied by this update. The independent speech service fix ae9be8c5 remains undeployed.

Agents composer follows Code interaction and style; application-level speech deployment remains a server integration follow-up.