Summary

• resolveActualContextLimit() now checks modelContextLimitsCache first for all providers before falling back to provider-specific logic
• Anthropic providers previously skipped the cache entirely, ignoring user-configured context limits from opencode.json

Fixes #2836

Problem

resolveActualContextLimit() had an early return for Anthropic providers that bypassed modelContextLimitsCache:

if (isAnthropicProvider(providerID)) {
  return getAnthropicActualLimit(modelCacheState) // ← cache never checked
}

This caused Claude Opus 4.6 (1M context, GA since March 2026) to report as 200K when used behind a proxy that doesn't return the context-1m beta header.

Fix

Check the config-populated cache first, then fall back to provider-specific defaults:

const cached = modelCacheState?.modelContextLimitsCache?.get(`${providerID}/${modelID}`)
if (cached) return cached

if (isAnthropicProvider(providerID)) {
  return getAnthropicActualLimit(modelCacheState)
}
return null

Changes

• src/shared/context-limit-resolver.ts — Reorder resolution: cache → provider-specific → null
• src/shared/context-limit-resolver.test.ts — Update test to validate cache is respected for Anthropic; add fallback test

Testing

• Updated existing test: cache entry for Anthropic now returns cached value (was asserting 200K)
• Added new test: no cache entry + 1M disabled → falls back to 200K default
• Existing tests for aliases and non-Anthropic providers pass unchanged

Summary by cubic

Respect cached model context limits by checking modelContextLimitsCache before provider fallbacks in resolveActualContextLimit(). Fixes Anthropic models incorrectly reporting 200K when a cached 1M limit is configured from opencode.json.

• Bug Fixes
  • Reordered resolution to: cache → provider-specific logic → null (for all providers).
  • Anthropic no longer bypasses the cache; tests updated for cached 1M and default 200K fallback.

^{Written for commit af5aaf7. Summary will update on new commits.}