Copilot Claude Opus 4.8 assistant prefill 400 after malformed pseudo tool-call text #31807

Description

github-copilot/claude-opus-4.8 can reject a follow-up request with:

APIError 400: This model does not support assistant message prefill.
The conversation must end with a user message.
isRetryable: false

After re-checking local SQLite metadata on OpenCode 1.17.7, this now appears to be downstream of the malformed pseudo tool-call issue in #31247.

The observed chain is:

1. Opus 4.8 emits pseudo tool-call text such as <invoke ...> as assistant text.
2. The assistant message is still marked finish=tool-calls.
3. There is no structured tool part in that assistant message.
4. OpenCode continues the tool loop.
5. The next provider request effectively ends with an assistant message.
6. Copilot Claude rejects that request as unsupported assistant prefill.

Steps to Reproduce

I do not yet have a small deterministic repro, but the failures occurred in normal OpenCode TUI sessions using github-copilot/claude-opus-4.8:

1. Start an OpenCode TUI session with github-copilot/claude-opus-4.8.
2. Perform a long tool-heavy task.
3. Continue through multiple automatic tool-call turns.
4. Hit a malformed Opus 4.8 turn where finish=tool-calls, there are zero structured tool parts, and the assistant text contains pseudo tool-call markup such as <invoke ...>.
5. Observe the next request fail with the assistant-prefill 400.

Expected Behavior

OpenCode should not send a Copilot Claude request whose final message is assistant.

More importantly, OpenCode should handle a malformed finish=tool-calls response with no actual tool call at the root, instead of continuing the tool loop with an assistant-ending history.

Actual Behavior

The provider rejects the follow-up request with:

This model does not support assistant message prefill.
The conversation must end with a user message.

Every sampled prefill 400 was immediately preceded by a claude-opus-4.8 assistant message with this structure:

finish: tool-calls
structured tool parts: 0
assistant text contains: <invoke ...>

Sampled chain:

Additional Evidence

Downstream symptom evidence:

https://gist.github.com/doomsday616/659c12bc0d0f02dff5a605e2ea6ec4cd

Root-cause evidence is in #31247:

https://gist.github.com/doomsday616/d3fb56acd0f14964a9797df35f1fcd96

Both gists contain only sanitized structural metadata from local SQLite rows. Private message text and local/private identifiers are not included.

OpenCode Version

1.17.7

Operating System

Windows 11

Terminal

Windows Terminal / PowerShell 7

Provider / Model

github-copilot / claude-opus-4.8

Related Issues

• Copilot Claude Opus 4.8 emits pseudo tool-call text instead of structured tool calls #31247 — likely root cause: Opus 4.8 pseudo tool-call text persisted as assistant text while finish=tool-calls and no structured tool part exists.
• This model does not support assistant message prefill / Github Copilot with Opus 4.6 #13768 — same/similar assistant-prefill rejection with GitHub Copilot + Claude Opus.
• Azure AI Claude rejects requests after tool abort: assistant message prefill not supported #19517 — Azure AI Claude via LiteLLM rejects assistant prefill.
• Bedrock Claude Opus 4.6 rejects assistant message prefill #29760 — Bedrock Claude Opus 4.6 rejects assistant prefill.

Sanitization

The original full session contains private content, so I cannot share a raw export publicly. The linked evidence includes only sanitized structural excerpts and provider error metadata.