Research: ThinkingContent Integration for Reasoning Models

Research - ThinkingContent Integration for Reasoning Models

Date: 2026-02-17 Owner: agent Phase: Research

Goal

Research how to integrate tinyAgent v1.2.0's ThinkingContent type and reasoning effort levels into tunacode's TUI. The goal is to cleanly separate reasoning/chain-of-thought from final response text, improving display for reasoning models like DeepSeek R1.

Findings

tinyAgent v1.2.0 API (New Features)

Key Pattern - Type Guards:

def is_thinking_content(block: AssistantContent | None) -> TypeGuard[ThinkingContent]:
    return block is not None and block.get("type") == "thinking"

Block Separation:

content = message.get("content") or []
thinking_blocks = [b for b in content if is_thinking_content(b)]
text_blocks = [b for b in content if is_text_content(b)]

Current Tunacode State

Critical Gap: Line 527-528 in main.py:

if assistant_event.get("type") != "text_delta":
    return  # <-- thinking_delta silently dropped!

UI Patterns Available

Key Patterns / Solutions Found

1. Event-driven streaming: tinyAgent emits typed events with thinking_start/delta/end - tunacode just needs to handle them
2. Canonical type exists: ThoughtPart already in type system, adapter conversion works
3. Panel metadata pattern: Return tuple[RenderableType, PanelMeta] for styled content
4. Collapsible sections: Could use Rich Collapsible or custom widget with .thinking-panel CSS

Integration Plan

Phase 1: Stream Event Handling

File: src/tunacode/core/agents/main.py

1. Add thinking_callback parameter to RequestOrchestrator
2. Modify _handle_message_update() to handle thinking_delta:

ev_type = assistant_event.get("type")
if ev_type == "text_delta":
    # existing code
elif ev_type == "thinking_delta" and self.thinking_callback:
    delta = assistant_event.get("delta")
    if delta:
        await self.thinking_callback(delta)

Phase 2: UI Display

New file: src/tunacode/ui/renderers/thinking.py

def render_thinking(content: str, collapsed: bool = True) -> tuple[RenderableType, PanelMeta]:
    """Render thinking/reasoning content with subdued styling."""
    meta = PanelMeta(
        css_class="thinking-panel",
        border_title="[#808080]reasoning[/]",
    )
    # Either collapsed or truncated display
    ...

CSS: src/tunacode/ui/styles/panels.tcss

.thinking-panel {
    outline: solid $muted;
    padding: 0 1;
}

Phase 3: App Integration

File: src/tunacode/ui/app.py

1. Add #thinking-output widget (collapsible Static)
2. Add _thinking_callback() method
3. Wire to process_request() call

Phase 4: Model Configuration

File: src/tunacode/configuration/models_registry.json

Add reasoning field to DeepSeek/Kimi models:

{
  "id": "deepseek/deepseek-r1",
  "reasoning": true
}

Knowledge Gaps

1. Reasoning effort levels: How should user configure low/medium/high? Config setting or per-request?
2. Thinking display: Collapsed by default? Expandable? Or just truncated?
3. Token counting: Should thinking tokens count toward context display?
4. History: Should thinking be persisted in session or ephemeral?

References

• tinyAgent v1.2.0 release: https://pypi.org/project/tiny-agent-os/1.2.0/
• tinyAgent GitHub: https://github.com/alchemiststudiosDOTai/tinyAgent/releases/tag/v1.2.0
• Example reasoning code: tinyAgent/examples/example_reasoning.py
• Tunacode streaming: src/tunacode/core/agents/main.py:473-536
• Canonical types: src/tunacode/types/canonical.py:51-56
• Panel system: src/tunacode/ui/renderers/panels.py