Interrupted tool calls leave invalid assistant message on retry

[undo76]

Summary

• When interrupting mid tool call (Ctrl+C) and then sending a new prompt, the provider rejects the conversation
  history because an unfinished assistant tool call is resent without a corresponding tool result.

Steps To Reproduce

• Start Crush with an OpenAI-style provider (e.g., DeepSeek/OpenAI/Azure OpenAI/OpenRouter).
• Prompt: “Use grep to find ‘TODO’ in the repo and summarize.”
• As soon as the tool starts, press Ctrl+C to interrupt.
• Send a new prompt (e.g., “hi”).

Actual Behavior

• Provider returns 400 invalid request (example from DeepSeek):
  • “Invalid assistant message: content or tool_calls must be set”
• Root cause: assistant message includes a tool call that wasn’t finished (no tool result), or becomes empty
  after filtering.

Expected Behavior

• Crush should sanitize history after interruption so the next request proceeds normally.

Environment

• Crush version: latest main (or specify)
• Provider: OpenAI-compatible (e.g., DeepSeek)
• OS: macOS/Linux/Windows
• Reproducible: always

Logs / Error

• 400 Bad Request from OpenAI-compatible API:
  • POST https://api.deepseek.com/v1/chat/completions
  • {"message":"Invalid assistant message: content or tool_calls must be set","type":"invalid_request_error"}

Proposed Fix

• In provider converters for assistant messages:
  • Include only finished tool calls (ToolCall.Finished == true) when serializing.
  • Skip appending an empty assistant message if both content is empty and there are no finished tool calls.
• Files/functions:
  • internal/llm/provider/openai.go: convertMessages (assistant case)
  • internal/llm/provider/anthropic.go: convertMessages (already skips empty; still filter unfinished calls)
  • internal/llm/provider/gemini.go: convertMessages (already skips empty; still filter unfinished calls)
• Optional: central “sanitize history” step prior to provider calls to remove unfinished tool calls.

Acceptance Criteria

• After interrupting mid tool call, sending a new prompt no longer produces provider errors.
• Works across OpenAI, Azure OpenAI, DeepSeek, OpenRouter; Anthropic/Gemini unaffected or improved.