feat(protocol): carry reasoning, and show it in the desktop

[oratis]

The protocol carried reasoningTokens and nothing else. DeepSeek's reasoner produced its most distinctive output and every client dropped it — the CLI got this in #218, but the desktop couldn't, because there was nothing on the wire to render.

A new transient event, not a flag on item.delta

export interface ReasoningDeltaEvent {
  type: 'reasoning.delta';
  threadId: string; turnId: string; itemId: string; delta: string;
}

Separate on purpose: reasoning is not the answer, it is never persisted as a completed item, and a client that doesn't understand the type must be able to drop it rather than accidentally render it as assistant text. Gated by a new reasoningDeltas capability, so a client can tell whether silence means "no reasoning" or "server too old".

The app-server forwards the agent loop's thinking_delta events. VS Code and the LSP bridge forward protocol events unchanged, so they receive it without any change.

Desktop

Reasoning lands on the assistant turn as a distinct reasoning field — deliberately not concatenated into text, which is how it would get lost again — and renders as a collapsed block above the answer:

▸ thinking · 2 lines

Collapsed because reasoner output is long and is not the response. The line count is in the summary because while a turn is streaming it is often the only thing there is to look at. appendReasoningDelta can open a turn on its own, since reasoning usually arrives before any answer text.

Verification

4 unit tests on the stream reducer (opening a turn from reasoning alone, accumulating without touching the answer, starting a fresh turn after one finishes, no leakage into text), the capability assertion updated in the protocol runtime test, and a Playwright case that asserts the block is present, collapsed, and reveals the reasoning on click. Screenshotted the rendered result.

pnpm typecheck · lint · format:check clean; protocol 24 · core 750/16 skipped · desktop 94 · server 41 · cli 208 · vscode 12 · lsp 13 · scripts 21; playwright test → 6 passed.

🤖 Generated with Claude Code