[backend] Port 1M context / adaptive thinking for Opus 4.6 (upstream #12338) #150

Description

Port upstream PR anomalyco#12342 to enable 1M context window and adaptive thinking for Claude Opus 4.6.

Scope

This is a major AI SDK upgrade (v5→v6) with breaking changes. Not a simple cherry-pick.

Changes Required

Phase 1 - Minimal (just headers + adaptive thinking):

• Add context-1m-2025-08-07 beta header to Anthropic provider
• Add adaptive-thinking-2026-01-28 beta header
• Add adaptive thinking detection for Opus 4.6 in transform.ts
• Convert budgetTokens → effort for Opus 4.6 models

Phase 2 - SDK Upgrade (breaking changes):

• Upgrade ai 5.0.124 → 6.0.72
• Upgrade all @ai-sdk/* packages (18+ packages)
• Migrate LanguageModelV2 → LanguageModelV3
• Make toModelMessages async
• Rename tool factory functions
• Update test suites for new usage types
• Fix compaction token counting bug

Key Files

• packages/opencode/src/provider/provider.ts - Beta headers
• packages/opencode/src/provider/transform.ts - Adaptive thinking logic
• packages/opencode/src/session/compaction.ts - Token counting
• packages/opencode/src/session/llm.ts - Async migration
• packages/opencode/src/session/prompt.ts - Type migration
• packages/opencode/src/provider/sdk/copilot/responses/tool/* - Factory renames

Upstream Reference

• Issue: 1M tokens for Opus 4.6 anomalyco/opencode#12338
• PR: feat(provider): add adaptive thinking and 1M context support for Claude Opus 4.6 anomalyco/opencode#12342

Quality Gates

• Tests pass
• Type checking passes
• Manual test: Opus 4.6 with >200k tokens