Parallel function calling fails with unrecoverable 400 error after multi-step MCP tools

[jw409]

Problem

After extended MCP tool usage (e.g., tools like sequential-thinking that make many tool calls with branching/revision patterns), subsequent parallel tool calls fail with:

400 error: "number of function response parts is equal to the number of function call parts"

This error is unrecoverable. Once triggered, the session is corrupted and ALL subsequent prompts fail - not just tool calls, but any user input. In some modes, the session may degenerate into a "mute" state where responses stop entirely. Users must /clear or restart, losing all conversation context and work in progress.

Reproduction

1. Configure Gemini CLI with an MCP server that makes many tool calls in complex patterns
   (e.g., sequential-thinking with its branching/revision features, or similar multi-call MCP tools)
2. Run a query that triggers extended tool usage
3. After completion, run parallel tool calls (e.g., multiple file reads)
4. Observe 400 error - session is now corrupted

Note: Discovered using a private fork of sequential-thinking with extended features, but the core behavior (many tool calls followed by parallel calls) is expected to reproduce with similar workloads.