Repository navigation
bug: Claude Sonnet 4.5 with thinking + ephemeral cache returns HTTP 400 #17883
Description
Description
When using claude-sonnet-4-5 via the Anthropic provider with extended thinking enabled, the API returns HTTP 400. The request includes thinking: { type: "enabled", budget_tokens: 31999 } and system messages with ephemeral cache_control type.
This was observed when using the model through a plugin (shop-plugin) that injects system prompts via experimental.chat.system.transform hook. The system array contains blocks with type: "ephemeral", which appears to conflict with the thinking configuration.
Error Details
service=llm providerID=anthropic modelID=claude-sonnet-4-5
error.name: AI_APICallError
error.url: https://api.anthropic.com/v1/messages
error.statusCode: 400
Request body:
model: "claude-sonnet-4-5"
max_tokens: 63999
thinking: { type: "enabled", budget_tokens: 31999 }
system: [{ type: "text", text: "..." }, { type: "ephemeral", ... }]
The same plugin works correctly with gpt-5-nano (no thinking enabled).
Steps to Reproduce
- Configure a plugin that uses
experimental.chat.system.transformhook to inject system prompts - Set model to
anthropic/claude-sonnet-4-5 - Enable thinking (happens automatically via oh-my-opencode config or OpenCode defaults)
- Send any message
Expected Behavior
The API call should succeed, either by:
- Stripping incompatible cache_control types when thinking is enabled
- Or validating the combination before sending to the API
Environment
- OpenCode version: 1.2.27
- Plugin: shop-plugin (system prompt injection via hooks)
- oh-my-opencode: 3.x
- OS: macOS Darwin 25.3.0
Related Issues
- Validate thinking level/variant against model capabilities before sending to API #17736 — Validate thinking level/variant against model capabilities (similar theme of pre-flight validation)
Metadata
Metadata
Assignees
Labels
No labels
Activity
cyberprophet commentedon Mar 17, 2026
Cross-reference from oh-my-openagent#2631
After investigating the oh-my-openagent codebase, this issue is confirmed to be on the OpenCode /
@ai-sdk/anthropicside, not oh-my-openagent.oh-my-openagent's role
oh-my-openagent applies
thinking: { type: "enabled", budgetTokens: 32000 }to all non-GPT models in its agent factories (e.g., Sisyphus, Oracle, Metis). This is a valid Anthropic API parameter for models that support extended thinking likeclaude-sonnet-4-5.The
ephemeralcache_control in system messages comes from the plugin layer (experimental.chat.system.transformhook) — oh-my-openagent's ownsystem-transform.tsis currently a no-op stub.Root cause
OpenCode's provider layer should validate that the combination of
thinking+ system messages withcache_control: { type: "ephemeral" }is compatible before sending to the Anthropic API. The Anthropic Messages API rejects this combination with HTTP 400.Suggested fix
Either:
cache_controlfrom system message blocks whenthinkingis enabled (pre-flight sanitization)Closing oh-my-openagent#2631 as this is an OpenCode-level concern.
ref: code-yeongyu/oh-my-openagent#2631
github-actions commentedon Jul 6, 2026
To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.