openclaw - 💡(How to fix) Fix [Bug] Anthropic Claude Opus 4.7 + thinking returns 400 invalid_request_error due to assistant message prefill [2 comments, 3 participants]

Summary

When using anthropic/claude-opus-4-7 (or any Opus 4.7+ / Sonnet 4.6 / Mythos Preview) with thinking enabled, every API request fails with HTTP 400 because OpenClaw's default request payload appends an assistant-role message prefill ({) to guide JSON output. Anthropic explicitly removed prefill support in these models as a breaking change (official docs).

This causes:

1. Every Telegram/main session request to fail
2. Failover decisions to mis-classify it as format error → trigger compaction → retry → infinite loop
3. Event loop saturation → server health-check timeout

Reproduction

Environment: OpenClaw 2026.4.25-beta.10, Node 24.14.1, Linux ARM64

Config:

{
  "agents": {
    "defaults": {
      "model": "anthropic/claude-opus-4-7",
      "thinking": "high"
    }
  }
}

Steps: Send any message via Telegram or any inbound channel.

Expected: Successful response with thinking-enhanced output.

Actual:

[agent/embedded] embedded run agent end: isError=true
  model=claude-opus-4-7 provider=anthropic
  error=LLM request rejected: This model does not support assistant message prefill.
        The conversation must end with a user message.
  rawError=400 {"type":"error","error":{"type":"invalid_request_error",
                "message":"This model does not support assistant message prefill.
                           The conversation must end with a user message."}}

Root Cause

In dist/extensions/anthropic/stream-wrappers.js, wrapAnthropicProviderStream does not check if the trailing message is role: "assistant" (prefill) before sending to Anthropic API. Anthropic's Opus 4.7 / Sonnet 4.6 / Mythos Preview reject any conversation that doesn't end with a user message.

Reference:

• https://platform.claude.com/docs/en/api/errors (search "prefilling assistant messages")
• https://platform.claude.com/docs/en/about-claude/models/migration-guide (search "Opus 4.7 breaking changes")
• https://platform.claude.com/docs/en/build-with-claude/extended-thinking (manual extended thinking removed in 4.7+)

Workaround (drop-in patch)

Add a wrapper that strips the trailing assistant prefill when payloadObj.thinking is set:

function createAnthropicPrefillFixWrapper(baseStreamFn) {
  const underlying = baseStreamFn ?? streamSimple;
  return (model, context, options) => {
    return streamWithPayloadPatch(underlying, model, context, options, (payloadObj) => {
      if (Array.isArray(payloadObj.messages) && payloadObj.messages.length > 0) {
        const last = payloadObj.messages[payloadObj.messages.length - 1];
        if (last && last.role === "assistant" && payloadObj.thinking) {
          log.warn("removing assistant prefill because thinking is enabled (Anthropic restriction)");
          payloadObj.messages.pop();
        }
      }
    });
  };
}

// Then in wrapAnthropicProviderStream, add as last wrapper:
return composeProviderStreamWrappers(
  ctx.streamFn,
  /* existing wrappers */,
  (streamFn) => createAnthropicPrefillFixWrapper(streamFn)
);

Verification (post-patch)

Suggested Fix (more robust than my workaround)

Recommend handling this at the model-capability layer:

1. Add a supports_prefill capability flag to Anthropic provider model registry
2. Default false for Opus 4.7+, Sonnet 4.6+, Mythos Preview
3. When false and last message role is assistant AND thinking is enabled → strip OR convert to a soft constraint in system prompt
4. Also consider adding output_config.format: "json" when supported (per Anthropic migration guide)

Related Issues

• #67710 (Opus 4.7 model resolver)
• #66462 (closed by #67024 — addresses compaction loop on similar pattern)
• #68241 (claude-opus-4-7 rejected as Unknown model)