Thinking mode does not support this tool_choice

[augustwy]

模型执行结束后报错

opencode 模型配置如下
"opencodeProvider": "opencode-go",
"opencodeModel": "deepseek-v4-flash"

[lindixu6-hash]

我复核了两张截图和当前调用链，这里实际有两层问题：

1. 底层真实错误是 opencode-go / deepseek-v4-flash 的 thinking 模式拒绝 forced tool_choice。opencode-mem 走的是 OpenCode session.prompt + format: json_schema，provider wire payload 由 OpenCode 生成；本仓库的 manual OpenAIChatCompletionProvider.tool_choice 不参与这条路径，因此不应在这里硬编码 DeepSeek 模型名单或伪造 provider 参数。
2. 当前插件捕获底层错误后，无条件尝试可选 External API fallback；当用户没有配置 manual fallback 时，又用 External API not configured for auto-capture 覆盖了真实根因，所以最终 toast 误导。

我已提交 #258 修复第 2 层：没有完整 manual fallback 时保留原始 OpenCode provider 错误；若确实配置了 memoryModel + memoryApiUrl，原 fallback 行为不变。入口级回归证明 toast 会显示 Thinking mode does not support ...，不会再显示误导性的 External API 错误。

本地验证：目标 2/2、完整 408/408、typecheck、format、production build 全绿；mutation check 移除实现后会精确恢复原截图错误并使回归失败。

第 1 层需要 OpenCode/provider compat 修复：同一 opencode-go/deepseek-v4-flash + OpenAI Responses 错误已被其他客户端独立复现，正确策略是在 thinking 模型上省略 forced named selector、保留工具广告，而不是由插件自行重写供应商 payload。当前绕行是升级到包含该兼容策略的 OpenCode/provider 版本，或为 opencodeModel 选择支持 structured output 的非-thinking 模型。

#258 不声称修改上游 tool_choice，只确保用户能看到可操作的真实错误。

[lindixu6-hash]

Thanks for the report. To reproduce this, could you share:\n1. opencode-mem version (check the opencode plugin list)\n2. Your full opencode-mem config (redact API keys)\n3. Whether this happens on every request or only in thinking/reasoning mode\n4. The exact error text from the screenshot (it's hard to read at this resolution)\n\nThe tool_choice parameter may not be supported by all providers in thinking/reasoning mode. Knowing your provider and model version will help narrow this down.

[augustwy]

1. I am using the latest version of opencode-mem.
2. Please see the attachment.
3. It happens every time when used in opencode.
4. “Auto Capture Failed
   Summary generation failed: External API not
   configured for auto-capture”
   “Error from provider (Console Go): Upstream request
   failed: Thinking mode does
   not support this tool_choice”

opencode-mem.jsonc

[enyineer]

I investigated this and these are the findings with a repro you can run to verify on your side:

Root cause: opencode-mem forces tool_choice: "required", and DeepSeek V4 rejects forced tool calls while thinking is enabled.

Chain:

1. opencode-mem requests structured output with format: { type: "json_schema", schema } plus the StructuredOutput tool
2. opencode maps that to a forced tool call — packages/opencode/src/session/prompt.ts:1285:

   toolChoice: format.type === "json_schema" ? "required" : undefined,

3. the OpenAI-compatible chat protocol then sends tool_choice: "required" upstream
4. DeepSeek V4 (both deepseek-v4-flash and deepseek-v4-pro) with thinking mode on rejects it with 400 "Thinking mode does not support this tool_choice" — only "auto" / "none" are accepted while thinking

Repro script (Node 18+/Bun, no dependencies; set your key via env, nothing is printed except status + response body):

// repro.mjs — run:  KEY=sk-... node repro.mjs
// Endpoint: point at your gateway or DeepSeek direct (https://api.deepseek.com/chat/completions).
const ENDPOINT = process.env.ENDPOINT || "https://opencode.ai/zen/go/v1/chat/completions"
const KEY = process.env.KEY
if (!KEY) { console.error("set KEY env"); process.exit(1) }

const TOOL = { type: "function", function: { name: "StructuredOutput",
  description: "Return the final response in the requested structured format.",
  parameters: { type: "object", properties: { summary: { type: "string" },
    tags: { type: "array", items: { type: "string" } } }, required: ["summary", "tags"] } } }

const base = (model) => ({ model, stream: false,
  messages: [ { role: "system", content: "You MUST call the StructuredOutput tool with your answer." },
               { role: "user", content: "Say hello." } ],
  tools: [TOOL] })

async function probe(label, body) {
  const res = await fetch(ENDPOINT, { method: "POST", headers: { "Content-Type": "application/json",
    Authorization: `Bearer ${KEY}` }, body: JSON.stringify(body) })
  const text = (await res.text()).slice(0, 300).replace(/\n/g, " ")
  console.log(`\n=== ${label} ===\nHTTP ${res.status}: ${text}`)
}

await probe("A deepseek-v4-flash, tool_choice=required (opencode-mem path)",
  { ...base("deepseek-v4-flash"), tool_choice: "required" })

await probe("B deepseek-v4-flash, tool_choice=auto",
  { ...base("deepseek-v4-flash"), tool_choice: "auto" })

await probe("C deepseek-v4-flash, tool_choice=required, thinking disabled",
  { ...base("deepseek-v4-flash"), tool_choice: "required", thinking: { type: "disabled" } })

// Control: any non-DeepSeek model that accepts forced tool calls with reasoning
await probe("D gpt-5.6-luna, tool_choice=required (control)",
  { ...base("gpt-5.6-luna"), tool_choice: "required" })

Observed results (against the console-go gateway):

#	model	tool_choice	thinking	result
A	deepseek-v4-flash	required	(default, on)	400 Thinking mode does not support this tool_choice
B	deepseek-v4-flash	auto	(default, on)	200
C	deepseek-v4-flash	required	{ "type": "disabled" }	200
D	gpt-5.6-luna	required	(default, on)	200 via opencode-mem (OpenAI-style model accepts forced tools with reasoning)

A reproduces the exact failure. B proves the trigger is the forced tool_choice, not thinking per se. C proves the fix: disable thinking (thinking: { "type": "disabled" }, or reasoning_effort: "none" per DeepSeek V4 docs) and forced tool calls are accepted.

Why there's no config workaround on opencode 1.18.18: the fix belongs on the structured-output requests only, but there's no scoped config path — custom model aliases can't inject body overrides (the upstream model name is the config key), the plugin's own agent registration clobbers agent-level options, and the SDK session.prompt body has no options field.

Suggested fix: allow disabling thinking (e.g. thinking: { "type": "disabled" }) for the internal opencode-mem-structured agent / structured-output sessions, or send tool_choice: "auto" and validate the returned JSON instead of forcing the call.

[lindixu6-hash]

Thanks @augustwy and @enyineer — the new config details and direct provider probes confirm the original split diagnosis.

• Fixed in opencode-mem: fix: preserve opencode auto-capture errors #258 (included in v2.24.4 and later) preserves the real OpenCode/provider error instead of replacing it with External API not configured for auto-capture.
• Still upstream: OpenCode converts format: { type: "json_schema" } into a forced tool_choice: "required". DeepSeek V4 thinking mode rejects that request shape. The same behavior was tracked in anomalyco/opencode#15226. A proposed capability-aware change, anomalyco/opencode#29565, was closed without merge, so there is no upstream fix we can rely on yet.

opencode-mem cannot safely rewrite the provider wire payload because it delegates the authenticated request to OpenCode's session.prompt API. Changing the plugin's manual OpenAI provider would not affect this path.

Current workarounds:

1. configure opencodeModel to a non-thinking model that supports OpenCode structured output;
2. disable thinking for the selected model if your OpenCode/provider configuration exposes a supported model-level override;
3. configure a complete manual fallback (memoryModel + memoryApiUrl + credentials), in which case current opencode-mem versions will warn about the primary failure and use the fallback.

I am keeping this issue open as an upstream compatibility tracker. Once OpenCode supports capability-aware structured output (for example tool_choice: auto on thinking models), opencode-mem should inherit the fix without provider-specific model lists or payload rewriting.