fix(translator): prevent doubled tool args in OpenAI-to-Claude response

[vishalrajv]

Problem

When using OpenAI-compatible models (e.g. mimo) with Claude Code clients, tool calls frequently fail with InputValidationError because the model's arguments arrive as doubled JSON (e.g. {query:x}{query:x}).

Root Cause

The OpenAI-compatible model sends duplicate SSE chunks with finish_reason set. The translator's finish block had no guard against this, so it ran twice — emitting tool arguments as two separate input_json_delta events. Claude Code concatenated both halves, producing invalid doubled JSON.

Changes

• openai-to-claude.js: Add !state.finishReason guard to prevent double finish emission
• openai-to-claude.js: Add deduplicateDoubledJson() as defense-in-depth in sanitizeToolArgs
• claude-to-openai.js: Guard string tool input on request side to prevent double-encoding
• openai-to-claude.test.js: Fix test state (add missing nextBlockIndex, finish_reason)

Testing

• All 12 translator unit tests pass
• Verified with live Claude Code + mimo-v2.5-free — tool calls now work correctly

[bloodf]

Heads-up: this guard regresses the multi-format pivot paths.

The !state.finishReason guard added here keys off state.finishReason, which is a shared translator-state field that upstream pivot stages also write. In Antigravity/Gemini → OpenAI → Claude (and any X → OpenAI → Claude pivot whose first stage sets state.finishReason), gemini-to-openai.js sets state.finishReason — for stream.js usage injection — before openaiToClaudeResponse runs. So by the time the finish chunk reaches openai-to-claude, state.finishReason is already truthy → the finish block is skipped → the buffered tool-call arguments are never flushed as input_json_delta, and content_block_stop/message_stop are never emitted. Tool calls then reach Claude clients with empty input.

Minimal repro — translateResponse(ANTIGRAVITY, CLAUDE, …) with a single functionCall emits only message_start + content_block_start (no input_json_delta, no stops).

A fix that preserves the dedupe intent without the cross-stage collision is to use a Claude-specific dedupe flag instead of the shared field:

// before
if (choice.finish_reason && !state.finishReason) {
// after
if (choice.finish_reason && !state.claudeFinishHandled) {
  state.claudeFinishHandled = true;

Keep state.finishReason = choice.finish_reason; inside the block for stream.js. This still fixes the mimo double-finish case while not being suppressed by an upstream pivot stage.

[vishalrajv]

Addressed the regression flagged in the review.

Fix: Replaced the shared !state.finishReason guard with a Claude-specific !state.claudeFinishHandled flag. This prevents the finish block from being suppressed when an upstream pivot stage (e.g. Antigravity/Gemini → OpenAI → Claude) sets state.finishReason before openaiToClaudeResponse runs.

• state.claudeFinishHandled = true is set inside the finish block (prevents double emission)
• state.finishReason = choice.finish_reason is kept inside the block for stream.js usage injection

All translator tests pass (golden-response-stream, format-roundtrip, bugs-antigravity, bugs-openai-bridge).

[vishalrajv]

Hi @decolua, just a gentle reminder about this PR when you get a chance. It fixes doubled tool arguments that cause InputValidationError when using OpenAI-compatible models with Claude Code clients. All translator tests pass. Happy to address any feedback or make adjustments. Thank you for your time!