feat(provider): add GitHub Copilot auto model selection - #20136

Issue for this PR

Closes #10093

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Adds an "Auto (Best for task)" model to the GitHub Copilot provider that automatically selects the best available model for each request using GPT-5-mini (free on Copilot) as a semantic complexity classifier.

How it works:

1. User selects "Auto (Best for task)" from the Copilot model list
2. Before each LLM call, the auto model sends the user prompt to GPT-5-mini asking for a 1-5 complexity rating
3. The rating maps to model tiers: 1-2 = fast (mini/haiku/flash), 3 = standard (sonnet/gpt-4.1), 4-5 = reasoning (opus/codex)
4. The best available model from the matching tier is selected
5. The TUI message header shows the resolved model name (e.g., "Build · gpt-4.1 · 3s" instead of "Build · auto")

Why GPT-5-mini? We explored regex heuristics, claw-llm-router (15-dimension keyword scoring), fine-tuning bert-tiny on 113k labeled examples, RouteLLMs pre-trained BERT (278M params), and local LLMs (Qwen2.5-0.5B). GPT-5-mini outperformed all of them on coding prompts because it has the world knowledge to distinguish semantic complexity (e.g., "implement a linked list" vs "implement a lock-free concurrent hash map"). It is free on Copilot, requires no model files or dependencies, and adds 1-3s latency per request.

Design decisions:

• Model tiers built dynamically from provider metadata (capabilities, output limits) — no hardcoded model lists, works as new models are added
• Only the last user message is classified, not the full conversation history
• Prompts under 20 chars skip classification entirely (fast tier)
• Classification errors fall back to standard tier

Note: The Copilot server-side routing API (/models/session) used by VS Code is restricted to first-party OAuth clients. A chat.model plugin hook (see #10093 discussion) would allow this to be an external plugin instead of a core change.

Files changed:

• packages/opencode/src/provider/sdk/copilot/auto-model.ts (new) — CopilotAutoLanguageModel with GPT-5-mini classification and tiered routing
• packages/opencode/src/provider/provider.ts — intercepts modelID === "auto" in copilot custom loader; sorts "auto" first
• packages/opencode/src/plugin/copilot.ts — injects the "auto" model entry during auth loading
• packages/opencode/src/provider/sdk/copilot/index.ts — re-exports
• packages/opencode/src/session/processor.ts — updates message modelID from "auto" to resolved model name

How did you verify your code works?

Tested locally with bun dev and opencode run --model github-copilot/auto --print-logs:

• "hello" (6 chars) → skipped classification → fast → gpt-5.1-codex-mini
• "implement a LRU cache" → GPT-5-mini rated 3 → standard → gpt-4.1
• "design a distributed consensus algorithm similar to Raft" → GPT-5-mini rated 5 → reasoning → gpt-5.2-codex
• "write a function to add two numbers" → GPT-5-mini rated 1 → fast → gpt-5.1-codex-mini
• "debug why distributed cache returns stale data under high concurrency" → GPT-5-mini rated 4 → reasoning → gpt-5.2-codex
• TUI message header updates to show resolved model name
• Typecheck passes (bun turbo typecheck)

Screenshots / recordings

N/A — the only UI change is the message header showing the resolved model name instead of "auto"

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR