[critical] Opus still limited to 200k tokens despite context-1m header

[randomm]

Description

After adding context-1m-2025-08-07 and adaptive-thinking-2026-01-28 beta headers (PR #152), Opus models still fail at ~200k tokens. The 1M context window is NOT working.

The body.permission: property 'body.permission' is unsupported warning also appears but is cosmetic - not the real problem.

Real Problem

Opus cannot exceed 200k token context despite the 1M header being set.

Possible Causes

1. Compaction bug (from upstream 1M tokens for Opus 4.6 anomalyco/opencode#12338): Token counting includes output/thinking tokens, triggering compaction too early
2. Model context limit: Models snapshot may have 200k hardcoded for Opus
3. Header only works for Sonnet: Upstream originally only applied context-1m for Sonnet models
4. SDK limitation: @ai-sdk/anthropic@2.0.58 may not support 1M context

Related

• Upstream issue: 1M tokens for Opus 4.6 anomalyco/opencode#12338
• Upstream PR: feat(provider): add adaptive thinking and 1M context support for Claude Opus 4.6 anomalyco/opencode#12342
• Our PR: feat(opencode): enable 1M context + adaptive thinking for Opus 4.6 (#150) #152

[randomm]

Closed — confirmed this is an Anthropic API account-level restriction (Usage Tier 4 required), not a code issue. Our code correctly sends the context-1m-2025-08-07 beta header and overrides limits to 1M. See epic #158 for ongoing fork strategy.