github-copilot (Claude): 400 "text content blocks must contain non-whitespace text" from whitespace-only assistant message

[Tushar49]

Description

github-copilot Claude models hit a hard 400 from the Anthropic /v1/messages endpoint once the conversation history contains an assistant turn whose only text block is whitespace.

Error:

AI_APICallError: messages: text content blocks must contain non-whitespace text
POST https://api.githubcopilot.com/v1/messages

The offending entry in the request body is a single space:

{"role":"assistant","content":[{"type":"text","text":" "}]}

Root cause: ProviderTransform.message (packages/opencode/src/provider/transform.ts) already strips empty text parts before sending to Anthropic, but the @ai-sdk/anthropic branch filters text with part.text !== "" (exact empty), so a whitespace-only " " slips through. The reasoning filter right beside it already uses .trim() — so the two are inconsistent. A lone-space text block reaches Anthropic and is rejected.

Affected Claude models (e.g. claude-opus-4.7, claude-opus-4.7-1m-internal) are listed and selectable — they just 400 on the next request once a whitespace-only assistant turn is in history. This is easy to hit when a model-fallback setup replays an empty/whitespace response to Claude.

I have a fix + tests ready and will open a PR referencing this issue.

Plugins

model-fallback plugin in use, but the bug is in core ProviderTransform and is plugin-independent

OpenCode version

1.16.2

Steps to reproduce

1. Use a github-copilot Claude model (e.g. claude-opus-4.7).
2. Get an assistant turn into history whose only text content is whitespace (e.g. an empty/whitespace response that is replayed to Claude — common with a model-fallback chain).
3. Send another message. The request to /v1/messages fails with 400 "text content blocks must contain non-whitespace text".

Screenshot and/or share link

No response

Operating System

Windows 11

Terminal

Windows Terminal

[github-actions]

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

• Empty text parts in assistant message history cause 422 on strict providers #31046: Same root cause — empty/whitespace text parts in assistant message history survive into the wire payload and cause rejections from strict providers (Anthropic, Azure AI, Bedrock). That issue specifically covers the toModelMessagesEffect assistant message loop missing a guard equivalent to the user message loop's part.text !== "" check.
• Claude models behind OpenAI-compatible proxies get 400 prefill errors #20415: Same transform.ts filtering gap for Claude models routed through non-native SDK paths (GitHub Copilot, OpenAI-compatible proxies), where empty content stripping does not apply because api.npm !== "@ai-sdk/anthropic".

Your issue adds useful specificity (whitespace-only " " vs strictly empty "", and the inconsistency between the reasoning filter using .trim() vs the text filter using exact equality). If your upcoming PR addresses that gap, please reference all three issues so the fixes can be coordinated.

[rekram1-node]

Send the session this occurred in?

[rekram1-node]

opencode export > session.json