Add dynamic per-task model selection during execution

[Copilot]

Tasks execute using a single global model today. This adds per-task model overrides so cheap/fast models handle simple work while capable models are reserved for complex tasks.

Changes

Schema

• Task.model?: string — optional per-task model ID; omit to fall back to the global model

Execution (executor.ts)

• executeTask() passes task.model to sendPromptSync, which threads it through to session creation (model ?? currentModel)
• New exported suggestModel(task) heuristic:
  • claude-sonnet-4 — >2 dependencies, description >200 chars, or >3 acceptance criteria
  • gpt-5-mini — everything else

Task Editor (task-editor.tsx)

• Added Model field; free-text input, blank = global default, hint shows gpt-5-mini / claude-sonnet-4 / gpt-4.1

Execute Screen (execute.tsx)

• Tasks with a custom model render the model ID in dimmed magenta next to the task title

sendPrompt / sendPromptSync (copilot.ts)

• Added model?: string option; overrides currentModel at session creation—no session reuse required

Example

import { suggestModel } from './services/executor.js';

const task = createTask({
  id: 'auth',
  title: 'Implement OAuth2 flow',
  description: '...',          // long
  dependsOn: ['db', 'session', 'config'],
});

task.model = suggestModel(task); // → 'claude-sonnet-4'

Original prompt

---

This section details on the original issue you should resolve

<issue_title>[enhancement] Add dynamic model switching during task execution</issue_title>
<issue_description>## Background

The Copilot SDK v0.1.30 (released March 3, 2026) introduces session.setModel() for mid-session model switching, implemented by Patrick Nikoletich in commit bd98e3a.

This new API allows changing the model mid-conversation without destroying the session, preserving conversation history while taking advantage of different model capabilities.

Proposal

Implement dynamic model selection during the Execute phase to optimize cost and performance:

1. Add model configuration to Task schema - Allow users to specify which model to use per task (e.g., gpt-4.1 for complex planning, gpt-5-mini for simple file operations)

2. Implement model switching in executor.ts - Before dispatching each task, call session.setModel(task.model) to switch to the appropriate model for that specific task

3. Add UI controls in Refine screen - Allow users to assign models to tasks during the refinement phase (similar to how dependencies are edited)

4. Smart defaults - Automatically suggest models based on task complexity:

   • Complex tasks (with many dependencies, long descriptions) → gpt-4.1 or claude-sonnet-4
   • Simple tasks (single file edits, basic operations) → gpt-5-mini
   • Parallel batch optimization → use faster models for independent simple tasks

Benefit

• Cost optimization: Use cheaper/faster models for simple tasks, reserve expensive models for complex planning
• Performance improvement: Faster model = quicker execution for straightforward tasks
• Flexibility: Users can fine-tune the model selection strategy for their specific workload
• Session preservation: No need to recreate sessions when switching models, maintains conversation context

Acceptance Criteria

• Task model in src/models/plan.ts includes optional model?: string field
• src/services/executor.ts calls session.setModel() before dispatching tasks with custom models
• Task editor in src/screens/refine.tsx allows selecting model from dropdown
• Default model selection logic suggests appropriate models based on task complexity heuristics
• Session events display model_change events in the Execute screen
• Tests verify model switching behavior in src/services/executor.test.ts
• Documentation in README explains model selection strategy

AI generated by Weekly Enhancement Suggestions

Comments on the Issue (you are @copilot in this section)

• Fixes [enhancement] Add dynamic model switching during task execution #28

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.