Using reasoning models with OpenRouter provider yields "provider error" due rewritten reasoning blocks.

[simonhj]

Description

Recently opencode become neigh-on unusable with OpenRouter for me. Most sessions end in a provider error.

It essentially breaks openrouter as provider for me, so its likely something related to my setup, but I can't figure out what.

From the logs the underlying error (using the repro below) is

  {
    "type": "error",
    "error": {
      "type": "invalid_request_error",
      "message": "messages.1.content.6: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified. These blocks must remain as they
  were in the original response."
    },
    "request_id": "req_011CYKKYMLyQjjKY8MRfvLHk"
  }

Me and Codex did some digging and it seems that tool Parts have duplicated reasoning attached to them in the session.

The following diff solves the issue for me, but it feels a bit like treating the symptom. If its a good approach, lemme know I can turn it into a PR.

diff --git a/packages/opencode/src/session/message-v2.ts b/packages/opencode/src/session/message-v2.ts
index 178751a22..b5d41b8dd 100644
--- a/packages/opencode/src/session/message-v2.ts
+++ b/packages/opencode/src/session/message-v2.ts
@@ -488,6 +488,24 @@ export namespace MessageV2 {
   })
   export type WithParts = z.infer<typeof WithParts>

+  function metadata(input: Record<string, any> | undefined): Record<string, any> | undefined {
+    if (!input) return input
+
+    const prune = ["reasoning", "reasoning_content", "reasoning_details"]
+    const result: Record<string, any> = Object.fromEntries(
+      Object.entries(input)
+        .map(([provider, value]) => {
+          if (typeof value !== "object" || value === null || Array.isArray(value)) return [provider, value]
+          const next = Object.fromEntries(Object.entries(value).filter(([key]) => !prune.includes(key)))
+          if (Object.keys(next).length === 0) return undefined
+          return [provider, next]
+        })
+        .filter((entry): entry is [string, unknown] => !!entry),
+    )
+    if (Object.keys(result).length === 0) return undefined
+    return result
+  }
+
   export function toModelMessages(input: WithParts[], model: Provider.Model): ModelMessage[] {
     const result: UIMessage[] = []
     const toolNames = new Set<string>()
@@ -608,7 +626,7 @@ export namespace MessageV2 {
             assistantMessage.parts.push({
               type: "text",
               text: part.text,
-              ...(differentModel ? {} : { providerMetadata: part.metadata }),
+              ...(differentModel ? {} : { providerMetadata: metadata(part.metadata) }),
             })
           if (part.type === "step-start")
             assistantMessage.parts.push({
@@ -645,7 +663,7 @@ export namespace MessageV2 {
                 toolCallId: part.callID,
                 input: part.state.input,
                 output,
-                ...(differentModel ? {} : { callProviderMetadata: part.metadata }),
+                ...(differentModel ? {} : { callProviderMetadata: metadata(part.metadata) }),
               })
             }
             if (part.state.status === "error")
@@ -655,7 +673,7 @@ export namespace MessageV2 {
                 toolCallId: part.callID,
                 input: part.state.input,
                 errorText: part.state.error,
-                ...(differentModel ? {} : { callProviderMetadata: part.metadata }),
+                ...(differentModel ? {} : { callProviderMetadata: metadata(part.metadata) }),
               })
             // Handle pending/running tool calls to prevent dangling tool_use blocks
             // Anthropic/Claude APIs require every tool_use to have a corresponding tool_result
@@ -666,7 +684,7 @@ export namespace MessageV2 {
                 toolCallId: part.callID,
                 input: part.state.input,
                 errorText: "[Tool execution was interrupted]",
-                ...(differentModel ? {} : { callProviderMetadata: part.metadata }),
+                ...(differentModel ? {} : { callProviderMetadata: metadata(part.metadata) }),
               })
           }

Plugins

None

OpenCode version

1.2.10

Steps to reproduce

In a git clone of the opencode repo the following repros the bug for me:

opencode run -m "openrouter/anthropic/claude-opus-4.6" "Read-only investigation. Use at least 10 tool calls (glob, grep, read) across packages
/opencode/src/session and packages/opencode/src/provider. Find every place where reasoning or provider metadata might be rewritten, dropped, trimmed, or remapped. Return a compact table with file, function, and risk. Do not edit files." --thinking true --variant high

Screenshot and/or share link

No response

Operating System

WSL2 on windows 11

Terminal

Windows Terminal

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• HuggingFace provider: reasoning_details not stripped from conversation history for Kimi-K2.5 #13002: Same root cause — reasoning/thinking metadata being passed back in conversation history causing API errors. That issue covers HuggingFace provider with reasoning_details, while this one affects OpenRouter with Anthropic thinking/redacted_thinking blocks. The proposed fix in this issue targets both.

[austinphilp]

Adding a repro from macOS (darwin) with OpenCode 1.2.10:

Model/Provider: openrouter/anthropic/claude-opus-4.6
Thinking: high
Mode: interactive TUI (fresh session)

Steps

1. Start TUI with Claude Opus 4.6 + thinking high
2. Run any skill (e.g. shortcut-api)
3. Next turn fails with invalid_request_error

Error
messages.1.content.4: thinking or redacted_thinking blocks in the latest assistant message cannot be modified

Notes

• The same prompt works via opencode run ... (one-shot).
• Disabling thinking makes the issue go away.

Sanitized log excerpt

ERROR service=llm providerID=openrouter modelID=anthropic/claude-opus-4.6 error={
  "error": {
    "name": "AI_APICallError",
    "statusCode": 400,
    "responseBody": "{\"error\":{\"message\":\"Provider returned error\",\"code\":400,\"metadata\":{\"raw\":\"{\\\"type\\\":\\\"error\\\",\\\"error\\\":{\\\"type\\\":\\\"invalid_request_error\\\",\\\"message\\\":\\\"messages.1.content.4: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.\\\"}\\\"}\"}}"
  }
}

I also parsed the request payload from the debug output: the previous assistant message is sent as a plain string (no thinking/redacted_thinking parts) even though the session export shows it included reasoning metadata. Looks like interactive history serialization is dropping the thinking block.

Happy to provide more logs/session export if helpful.

[han-nah-dev]

Same issue here. Same cause as well.