Anthropic wire: interrupted-turn replay drops thinking signature, causes hard 400 "Invalid signature in thinking block"

Start by reading the streaming consumers, finalizeReasoning(), toWireMessages(), and assistantParts() described in the issue, then compare how the gateway path handles thinking metadata with consumeFullStream. Reproduce an interrupted turn and inspect the replayed assistant message. Done means the replay preserves the thinking signature and redacted data, and the interrupted request no longer fails with the reported 400.

Summary

Anthropic-wire sessions fail with a hard 400 whenever the user sends a new message while the model is still streaming. The error is unavoidable in that scenario and the only recovery is continue or restarting the session:

Error: 400 Invalid_request_error Upstream request failed: [invalid_request_error]
messages.1.content.0: Invalid signature in thinking block. The block is bound to
a different conversation. Remove the block, or set
`thinking.block_binding.prefix_mismatch_behavior` ...

Trace ID: 907c498bc0028b25be7ae3dc058394f7

Environment

• Command Code 1.87.3 (dist/cli.mjs)
• Linux
• Custom BYOK provider in ~/.commandcode/providers.json with "api": "anthropic-messages"
• Model: claude-haiku-5-5

Steps to reproduce

1. Configure a BYOK provider using the Anthropic Messages wire:

   {
     "provider": {
       "opencode": {
         "baseURL": "https://opencode.ai/inference/go/anthropic",
         "models": {
           "claude-haiku-5-5": { "api": "anthropic-messages" }
         }
       }
     }
   }
   

2. Start a session on that model and let it produce a turn containing real thinking.

3. While the turn is still streaming, type and submit a new message.

4. The request fails with the 400 above, every time.

Root cause (located in dist/cli.mjs, v1.87.3)

Anthropic thinking blocks carry a signature that is bound to one specific assistant prefix. Command Code has two streaming consumers and only one of them round-trips that signature:

• consumeFullStream (AI SDK path, used by runAiSdkStream for custom providers): finalizeReasoning pushes {type:"thinking", thinking:c, signature:u} where u comes from readReasoningMetadata() — the real signature is preserved and replayed. This path works.
• consumeStream (Command Code gateway /alpha/generate path): finalizeReasoning hardcodes {type:"thinking", thinking:u, signature:""}. Additionally toWireMessages() serializes the thinking block as only {type:"reasoning", text:n.thinking} — the signature and redactedData fields are dropped entirely, so the gateway never receives them.

When a turn is interrupted mid-stream, withResumedAssistantTurn() rebuilds the partial assistant turn via assistantParts(). That helper only emits a reasoning part when r.signature || r.redactedData is truthy. On the gateway path the signature is always "", so the replayed assistant turn goes out with a thinking block that has no signature (or an empty one). The upstream rejects it as "bound to a different conversation".

Verification performed

I ran a local fake Anthropic endpoint and captured the actual request bodies with cmd -p --resume:

• Plain --resume continuation on a custom anthropic-messages provider: the 64-character signature is fully replayed (messages[1] thinking block carries it), and the call succeeds. So ordinary multi-turn history is fine.
• The failure is specific to the interrupted-turn replay / gateway path, where the signature is lost as described above.

This confirms the defect is in the interrupted-replay serialization, not in the BYOK provider, the proxy, or normal history handling.

Relationship to existing issues

This looks like the same class of defect as #877 and #887 (BYOK thinking/reasoning content dropped from replayed history). The fix referenced in #877 (PR #5283) does not appear to be present in the published 1.87.3 build — toWireMessages() still omits signature completely.

Note that @ai-sdk/anthropic 3.0.135 (already a dependency of 1.87.3) supports the escape hatch:

"thinking": { "blockBinding": { "prefixMismatchBehavior": "drop_block" } }

which automatically adds the thinking-binding-controls-2026-08-01 beta and makes the upstream drop a mismatched block instead of erroring. Command Code's thinkingHook() never sets this, so the option is currently unreachable from the client.

Expected behavior

Either:

1. Preserve and replay signature / redactedData on the consumeStream gateway path (match what consumeFullStream already does), or
2. Set thinking.blockBinding.prefixMismatchBehavior: "drop_block" for Anthropic-wire models so a stale signature degrades gracefully instead of producing a hard 400.

Option 1 is the correct fix; option 2 would at least stop the session from dying.

Workaround for affected users

• Do not send messages while a turn is still streaming; wait for it to finish.
• If you must interrupt, start a fresh process with cmd -p "..." --resume <sessionId>, which takes the signature-preserving path.