Bug: Provider returns 400 due to empty text content blocks in messages

[s6yoho]

Bug Description

When using a channel, the provider returns a 400 error because the messages array contains empty text content blocks.

Error Message

Provider any returned 400: Provider returned 400: Bad Request | Upstream: {"error":{"type":"new_api_error","message":"messages: text content blocks must be non-empty (request id: 20260211215402599063580RdzKE8eI) (request id: 20260211215350819939941BGRT2UD6)"},"type":"error"}

Root Cause

The assistant message contains an empty text content block ("text": ""). This happens when the model response includes thinking + tool_use but the text portion is empty. Claude Code Hub should filter out or skip empty text content blocks before forwarding the request upstream.

Example problematic message:

{
  "role": "assistant",
  "content": [
    {
      "type": "thinking",
      "thinking": "...",
      "signature": ""
    },
    {
      "type": "text",
      "text": ""
    },
    {
      "type": "tool_use",
      "id": "tooluse_oiIL11PVOn3Rv1A9WckX00",
      "name": "Skill",
      "input": {
        "skill": "api-relay-tester",
        "args": "..."
      }
    }
  ]
}

The {"type": "text", "text": ""} block is the problem - the Anthropic API rejects empty text content blocks.

Expected Behavior

Claude Code Hub should strip empty text content blocks (where text is an empty string) from messages before forwarding them to the upstream provider.

Environment

• Claude Code Hub version: v0.5.2

[github-actions]

Thank you for reporting this issue with a clear reproduction case.

Analysis:

I confirmed the root cause in the codebase. When an assistant message contains thinking + tool_use blocks with an empty text block ({"type": "text", "text": ""}), the empty text block passes through the entire proxy pipeline unfiltered and reaches the upstream Anthropic API, which rejects it with a 400.

Current state of the code:

The codebase already has a rectification pattern for similar Anthropic-specific issues in src/app/v1/_lib/proxy/thinking-signature-rectifier.ts. The rectifyAnthropicRequestMessage() function (lines 101-219) iterates through messages[*].content arrays and filters out problematic blocks:

• thinking blocks (lines 137-141)
• redacted_thinking blocks (lines 143-147)
• Blocks with stale signature fields (lines 149-155)

However, it does not filter empty text blocks (type === "text" with text === ""). This is the gap.

In src/app/v1/_lib/proxy/forwarder.ts (lines 862-1004), the rectifier is triggered on error detection and retries once with the cleaned request. The infrastructure for detect-rectify-retry is already in place.

Error rule coverage:

There is an existing error rule in src/repository/error-rules.ts (line 458) that matches "all messages must have non-empty content", but the error in this issue ("text content blocks must be non-empty") is a different Anthropic error message and is not currently matched by any detection logic or error rule.

What a fix would involve:

1. Extend rectifyAnthropicRequestMessage() in thinking-signature-rectifier.ts to also filter out content blocks where type === "text" and text === "" (or is falsy), following the existing pattern
2. Add a detection trigger in detectThinkingSignatureRectifierTrigger() (or a new dedicated detector) for the "text content blocks must be non-empty" error message
3. Optionally add an error rule in error-rules.ts for the "text content blocks must be non-empty" pattern for consistent error categorization
4. Add test cases alongside the existing tests in tests/unit/proxy/proxy-forwarder-thinking-signature-rectifier.test.ts

---

Automated response from Claude AI