[Bug] OpenCode Go: kimi-k2.6 fails with HTTP 400 'cannot specify both thinking and reasoning_effort' — works via plain client

[AlfieChang]

Description

Every call to opencode-go/kimi-k2.6 fails with HTTP 400:

{"error": {"message": "Error from provider: cannot specify both 'thinking' and 'reasoning_effort'"}}

The client sends Moonshot's native thinking toggle together with top-level reasoning_effort, and the upstream rejects requests containing both. kimi-k3 on the same Go subscription works fine — it got the same class of fixes via models.dev#3331 (see #37552), but K2.6 never got that pass, even though it's still a paid model in the Go catalog.

One data point that shows the upstream itself is fine: a plain OpenAI-compatible client (sends only model + messages + token limits, no thinking/reasoning_effort/temperature/tools) hits the same endpoint https://opencode.ai/zen/go/v1/chat/completions with kimi-k2.6 and the same API key, and gets 200. So this is about the request envelope opencode builds for K2.6, not a model or gateway outage.

Related: #37771 (K2.6 also affected by the mcp/system field rejection), #27054, #38219.

Happy to test any patch.

Plugins

None

OpenCode version

N/A — reproduced via direct API calls to the Go endpoint

Steps to reproduce

1. opencode run -m opencode-go/kimi-k2.6 "hello" → HTTP 400 cannot specify both 'thinking' and 'reasoning_effort'
2. opencode run -m opencode-go/kimi-k3 "hello" → works
3. Direct call to https://opencode.ai/zen/go/v1/chat/completions with a minimal payload (model, messages only) for kimi-k2.6 → 200

Screenshot and/or share link

No response

Operating System

Windows 11

Terminal

Git Bash

[github-actions]

This issue doesn't fully meet our contributing guidelines.

What needs to be fixed:

• The issue does not use one of the three required templates (Bug Report, Feature Request, or Question). Please resubmit using the appropriate template.
• The description appears to be AI-generated. We do not accept AI-generated walls of text — please describe the problem in your own words with the essential details only.

Please edit this issue to address the above within 2 hours, or it will be automatically closed.

---

This issue might be a duplicate of existing issues. Please check:

• Cannot use Kimi or Qwen models in Opencode Go plan #38219: Kimi and Qwen models unusable on OpenCode Go plan (same symptoms)
• OpenCode Go: kimi-k3, kimi-k2.6, kimi-k2.7-code, glm-5.2, grok-4.5, qwen3.7-max, hy3-preview all fail with 'Upstream request failed' — non-standard 'mcp'/'system' fields rejected by strict provider validators #37771: kimi-k2.6 and other models failing with upstream request errors due to non-standard fields (directly overlaps with the mcp/system field rejection described here)

If you believe this was flagged incorrectly, please let a maintainer know.

[github-actions]

Thanks for updating the issue! It now looks good — the description is clear and specific, steps to reproduce are provided, and the required fields are filled in. We'll take a look.

[opencode-agent]

Closing as unable to reproduce on current dev (743f6410f2e5002723fc5e893039ac49fbfe0de8).

This cannot happen out of the box with the current model definition and request transforms:

• OpenCode Go uses the OpenAI-compatible transport: https://github.com/anomalyco/opencode/blob/743f6410f2e5002723fc5e893039ac49fbfe0de8/packages/opencode/test/tool/fixtures/models-api.json#L33606-L33610
• kimi-k2.6 has no configured reasoning options, so it does not generate a reasoningEffort variant: https://github.com/anomalyco/opencode/blob/743f6410f2e5002723fc5e893039ac49fbfe0de8/packages/opencode/test/tool/fixtures/models-api.json#L34133-L34141
• Reasoning-effort variants are only generated when the model advertises an effort option:

  opencode/packages/opencode/src/provider/transform.ts

  Lines 1635 to 1640 in 743f641

  	const effort = options.find((option) => option.type === "effort")
  	if (effort) return effortVariants(target, effort.values)
  	const toggle = options.some((option) => option.type === "toggle")
  	const budget = options.find((option) => option.type === "budget_tokens")
  	if (!budget) return toggle ? nonEmptyVariants(reasoningToggle(target)) : undefined

• Kimi thinking injection is guarded to Anthropic-compatible transports, which this model is not using:

  opencode/packages/opencode/src/provider/transform.ts

  Lines 1217 to 1226 in 743f641

  	// Moonshot's Anthropic-compatible API uses adaptive effort rather than token budgets.
  	// Request summaries so thinking content survives replay on subsequent turns.
  	if (
  	["@ai-sdk/anthropic", "@ai-sdk/google-vertex/anthropic"].includes(input.model.api.npm) &&
  	isKimiFamily(input.model) &&
  	input.model.capabilities.reasoning
  	) {
  	result["thinking"] = { type: "adaptive", display: "summarized" }
  	result["effort"] = "high"
  	}

The successful minimal direct API request only establishes that the endpoint accepts a minimal payload; it does not reproduce the alleged OpenCode request envelope. The report also omits the OpenCode version and the actual failing outbound JSON.

If this still occurs, please provide the exact OpenCode version or commit, complete command, selected variant, relevant model/provider configuration or overrides, and the sanitized exact outbound JSON body (with credentials and private content removed). We can reopen once there is an exact reproduction.