fix(anthropic): normalize sampling params under extended thinking (temperature/top_p)

[zhiru]

Problem

Claude models that run with extended thinking (e.g. Opus 4.8 via the Claude Code provider) return HTTP 400 when the request carries non-default sampling params:

[400]: `temperature` may only be set to 1 when thinking is enabled or in adaptive mode.
[400]: `top_p` must be greater than or equal to 0.95 or unset when thinking is enabled or in adaptive mode.

The VS Code Copilot "Ollama" BYOK provider sends temperature: 0.7 and top_p: 0.9, so every request to a thinking-enabled Claude model fails. It reproduces across all routing modes (grouped / raw / combo) because thinking is applied at the provider/model layer they all share.

Root cause

thinking is injected by per-model requestDefaults (applyProviderRequestDefaults) after the translator/constraint passes run. The existing enforceThinkingTemperature only runs inside the Claude Code bridge path (buildAndSignClaudeCodeRequest), which is not on the executor path these requests take — so it never fired. And even when it did fire, it pinned temperature but never handled top_p at all.

Fix

1. enforceThinkingTemperature now also drops top_p when thinking is enabled/adaptive (in addition to pinning temperature = 1).
2. Call it at the final dispatch chokepoint in BaseExecutor — before fingerprinting/CCH signing serialize the body — for claude and Claude-Code-compatible providers. Single point every Claude routing mode (grouped/raw/combo) and the native passthrough share. No-op when thinking isn't active.

Unit tests extended (tests/unit/claude-code-parity.test.ts), eslint + typecheck clean.

[gemini-code-assist]

Code Review

This pull request updates the Claude Code API constraints to drop the top_p parameter and pin temperature to 1 when extended thinking is active, preventing API errors from non-default sampling parameters. It also adds corresponding unit tests to verify this behavior. The review feedback suggests adding defensive checks in both the executor and the constraint utility to ensure that the request body is a valid, non-null object before attempting to access or modify its properties, preventing potential runtime crashes.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

If transformedBody is null, undefined, or not an object, casting it to Record<string, unknown> and passing it to enforceThinkingTemperature will cause a runtime crash (TypeError) when accessing body.thinking. We should add a defensive check to ensure transformedBody is a valid non-null object before calling enforceThinkingTemperature.

Suggested change

	if (this.provider === "claude" || isClaudeCodeCompatible(this.provider)) {
	enforceThinkingTemperature(transformedBody as Record<string, unknown>);
	}
	if (
	(this.provider === "claude" || isClaudeCodeCompatible(this.provider)) &&
	transformedBody &&
	typeof transformedBody === "object" &&
	!Array.isArray(transformedBody)
	) {
	enforceThinkingTemperature(transformedBody as Record<string, unknown>);
	}

[gemini-code-assist]

To make enforceThinkingTemperature robust and prevent potential runtime errors from unexpected or improperly typed inputs (especially when called with type assertions like as Record<string, unknown>), we should add a defensive guard clause at the beginning of the function to verify that body is a valid non-null object.

Suggested change

	export function enforceThinkingTemperature(body: Record<string, unknown>): void {
	const thinking = body.thinking as Record<string, unknown> | undefined;
	if (thinking?.type === "enabled" || thinking?.type === "adaptive") {
	body.temperature = 1;
	if (body.top_p !== undefined) {
	delete body.top_p;
	}
	}
	}
	export function enforceThinkingTemperature(body: Record<string, unknown>): void {
	if (!body || typeof body !== "object" || Array.isArray(body)) return;
	const thinking = body.thinking as Record<string, unknown> | undefined;
	if (thinking?.type === "enabled" || thinking?.type === "adaptive") {
	body.temperature = 1;
	if (body.top_p !== undefined) {
	delete body.top_p;
	}
	}
	}

[kilo-code-bot]

Code Review Summary

Status: No New Issues Found | Recommendation: Address existing review comments before merge

Files Reviewed (3 files)

• open-sse/executors/base.ts — new enforceThinkingTemperature dispatch in native Claude path
• open-sse/services/claudeCodeConstraints.ts — extended enforceThinkingTemperature logic (drops top_p + pins temperature)
• tests/unit/claude-code-parity.test.ts — 3 new parity tests

Open Review Items (2 prior comments, not re-raised here)

File	Line	Issue	Severity
open-sse/executors/base.ts	1128	transformedBody cast to Record<string, unknown> without null/object guard — potential TypeError	HIGH
open-sse/services/claudeCodeConstraints.ts	32	enforceThinkingTemperature lacks defensive guard for non-object input	MEDIUM

Other Observations (not in diff)

No observations outside the diff.

---

Fix these issues in Kilo Cloud

---

_{Reviewed by step-3.7-flash-20260528 · 1,190,061 tokens}

[diegosouzapw]

Merged into release/v3.8.24 — thank you, @zhiru! 🙌

Sharp root-cause: thinking is injected by per-model requestDefaults after the translator/constraint passes, so the bridge-path enforceThinkingTemperature never fired for executor-path requests — and even when it did, it never handled top_p. Moving the normalization to the final dispatch chokepoint in BaseExecutor covers grouped/raw/combo and the native passthrough in one place, and dropping top_p cleanly satisfies Anthropic's "≥0.95 or unset" rule while preserving the never-send-both invariant.

Validated locally: claude-code-parity 31/31, typecheck:core + eslint clean, clean merge with the release branch. Shipping in the next release.