Spec — AI thinking state

Hosted by chat-message-bubble.kmd e agent-step-trace.kmd. Streams via streaming-text.kmd (separate stream channel reasoning).

Princípios

1. Provider abstraction — gateway expõe reasoning_content separated from content; SDK detecta provider.
2. Collapsed by default — pra reduzir noise; user expande quando quer detalhes.
3. Auto-close on stream end — opcional, configurable; default close.
4. Preserve manual expand — user expandeu → não auto-close.

R1 — Anatomia

┌───────────────────────────────────────────┐
│ 💭 Thinking · 4.2s            [▼]         │  ← collapsed (default)
└───────────────────────────────────────────┘

┌───────────────────────────────────────────┐
│ 💭 Thinking · 4.2s · 234 words   [▲]      │  ← expanded
│ ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━        │
│ "I need to find recent docs about         │
│ Koder Stack architecture. The user is    │
│ asking about the Foundation layer..."    │
│                                           │
│ [reasoning trace continues...]            │
└───────────────────────────────────────────┘

Slots:

R2 — States

R3 — Provider abstraction

Gateway response includes optional reasoning_content:

{
  "content": [{"type": "text", "text": "..."}],
  "reasoning_content": "I need to think about this...",
  "reasoning_meta": {
    "model_supports": true,
    "duration_ms": 4200,
    "tokens": 234
  }
}

If reasoning_content is null/missing: do NOT render thinking block.

Provider support:

Gateway translates provider-specific to canonical shape above.

R4 — Auto-close

On done event of reasoning stream:

• If auto_close: true (default): collapse animated (smooth).
• If user already expanded manually: preserve expanded state.

User preference: stored per (koder_user_id, workspace_id).

R5 — Multiple thinking blocks (accordion)

Agent runs with reasoning per step (cross-link #108):

[Step 1] ▼ Thinking · 2s
[Step 2] ▶ Thinking · 1.5s
[Step 3] ▼ Thinking · 3s

Independent collapse/expand state per block.

R6 — Surface bindings

R7 — Acessibilidade

• Block: <details><summary> semantically (Web); equivalente Flutter.
• Toggle: keyboard Enter/Space.
• aria-live="polite" during streaming.
• Reduced-motion: no smooth collapse; instant.

R8 — i18n

T-suite

• T1 Mount streaming: response has reasoning_content streaming → block visible with "Thinking..." + breathing dot.
• T2 Done: stream complete → label becomes "Thought for 4.2s"; collapse animated.
• T3 No reasoning_content: response without field → block NOT rendered.
• T4 Manual expand: user clicks → expand; subsequent done does NOT collapse.
• T5 Multiple blocks (agent steps): each independent.
• T6 Reduced-motion: collapse instant.
• T7 Provider abstraction: OpenAI o-family + Claude responses both render correctly.

Cross-link

• Companion: chat-message-bubble.kmd, agent-step-trace.kmd, streaming-text.kmd
• Backend: services/ai/gateway/ (provider translation)
• Refs: MUI X Chat Reasoning, agents-kit Reasoning, Anthropic extended thinking