#51 fix(opencode): strip trailing assistant messages to prevent prefill 400

Problem

rewriteRequestBody() never removed trailing assistant-role messages. A resumed or compacted session that ends on an assistant turn (e.g. after a failed tool round, or a session restored mid-response) reached Anthropic with assistant-message prefill. Claude Code OAuth models reject this:

This model does not support assistant message prefill.
The conversation must end with a user message.  (HTTP 400)

This recurs in practice on session resume/compaction and forces the user to manually send another message to recover.

Fix

Add stripTrailingAssistantMessages() and call it right after JSON.parse(body), so trailing assistant messages are popped before billing-header computation and signing. This mirrors the behavior of the built-in @ai-sdk/anthropic provider and the packages/pi conversion path, which already strip trailing assistant messages.

Only messages after the last user message are removed — the active user turn and any assistant message followed by a user message are always preserved. The alternative (no strip) is a hard 400 that returns nothing.

Tests

Adds 5 tests:

• strips single trailing assistant message
• strips multiple trailing assistant messages
• preserves assistant message followed by user message
• preserves assistant tool_use followed by user tool_result
• strips trailing assistant after tool_result + assistant

Updates existing fixtures that previously ended on assistant turns to end on user messages. All opencode tests pass (264), typecheck clean, biome clean.

---

^{Need help on this PR? Tag @codesmith with what you need. Autofix is disabled.}

---

Summary by cubic

Strips trailing assistant messages during request rewrite so conversations end with a user message, preventing Anthropic Claude Code OAuth 400 “assistant message prefill” errors. Aligns behavior with @ai-sdk/anthropic and packages/pi.

• Bug Fixes
  • Added stripTrailingAssistantMessages() and called it in rewriteRequestBody() right after JSON.parse.
  • Only removes assistant messages after the last user message; preserves assistant→user and tool_use→tool_result flows.
  • Added 5 tests for strip/preserve scenarios and updated fixtures to end on user messages.