fix(ai): use enable_thinking for Z.ai instead of thinking param

[okuyam2y]

Problem

The Z.ai thinking control in openai-completions.ts sends thinking: { type: "disabled" }, but Z.ai actually uses enable_thinking: false — the same format as Qwen.

Because Z.ai doesn't recognize the thinking parameter, it ignores the disable request and always runs with thinking enabled. This wastes reasoning tokens and adds significant latency (measured 1.8x slower on chat responses).

Fix

Merge the Z.ai and Qwen code paths since they use the same enable_thinking: boolean format.

Before:

if (compat.thinkingFormat === "zai" && model.reasoning) {
    (params as any).thinking = { type: options?.reasoningEffort ? "enabled" : "disabled" };
} else if (compat.thinkingFormat === "qwen" && model.reasoning) {
    (params as any).enable_thinking = !!options?.reasoningEffort;
}

After:

if ((compat.thinkingFormat === "zai" || compat.thinkingFormat === "qwen") && model.reasoning) {
    (params as any).enable_thinking = !!options?.reasoningEffort;
}

Testing

Verified with Z.ai GLM-4.7 and GLM-5 models:

• Before fix: all responses contain reasoning_content (thinking=True)
• After fix: no reasoning_content when enable_thinking: false (thinking=False)
• GLM-4.7 chat response: 3.2s → 1.8s (1.8x faster)
• GLM-5 chat response: 12.9s → 7.0s (1.8x faster)

🤖 Generated with Claude Code

[github-actions]

Hi @okuyam2y, thanks for your interest in contributing!

We ask new contributors to open an issue first before submitting a PR. This helps us discuss the approach and avoid wasted effort.

Next steps:

1. Open an issue describing what you want to change and why (keep it concise, write in your human voice, AI slop will be closed)
2. Once a maintainer approves with lgtm, you'll be added to the approved contributors list
3. Then you can submit your PR

This PR will be closed automatically. See https://github.com/badlogic/pi-mono/blob/main/CONTRIBUTING.md for more details.

[badlogic]

cheers

[okuyam2y]

Thanks for merging! 🙏