Gemini 2.5 Flash kullanıldığında bot cevap vermiyor (sadece reaction atıyor) / Bot does not reply with Gemini 2.5 Flash / Gemini 2.5 FlashでBotが返信しない

[arikantr]

🇹🇷 Türkçe

Merhaba,
Misskey AI botunu Docker ile kurdum. Gemini API anahtarım config.json içinde doğru şekilde tanımlı.
Bot mention alıyor, loglarda <<< An message received ve API: notes/reactions/create görünüyor ama asla cevap üretmiyor, sadece like (reaction) atıyor.

Testlerimde fark ettim:

Kodda varsayılan endpoint gemini-2.0-flash-exp olarak sabitlenmiş.

Benim hesabımda aktif olan model ise gemini-2.5-flash.

curl ile 2.5-flash çağırınca cevap dönüyor ama bot hala 2.0-exp’i çağırıyor ve bu model artık çalışmıyor.

Öneri:

Varsayılanı gemini-2.5-flash yapmak
veya

config.json’daki "geminiModel" parametresini okuyacak şekilde esnek hale getirmek.

Teşekkürler 🙏

🇬🇧 English

Hello,
I installed the Misskey AI bot with Docker. My Gemini API key is correctly set in config.json.
The bot receives mentions, and logs show <<< An message received and API: notes/reactions/create, but it never generates a reply – it only sends a like (reaction).

From my tests:

The code currently hardcodes the endpoint as gemini-2.0-flash-exp.

On my account, the available model is gemini-2.5-flash.

Calling 2.5-flash via curl works fine, but the bot still tries to call 2.0-exp, which no longer responds.

Suggestion:

Update the default to gemini-2.5-flash,
or

Make it configurable via "geminiModel" in config.json, so users can choose newer models when available.

Thanks 🙏

🇯🇵 日本語

こんにちは。
Misskey AI BotをDockerで導入しました。GeminiのAPIキーはconfig.jsonに正しく設定済みです。
Botはメンションを受け取り、ログには <<< An message received と API: notes/reactions/create が出ますが、返信を生成せずリアクションだけを返します。

テストの結果:

コードではエンドポイントが gemini-2.0-flash-exp にハードコードされています。

しかし私のアカウントで利用可能なのは gemini-2.5-flash モデルです。

curlで2.5-flashを呼び出すと応答がありますが、Botは依然として2.0-expを呼び出しており、このモデルはもう動作していません。

提案:

デフォルトを gemini-2.5-flash に更新する、
または

config.jsonの "geminiModel" パラメータからモデルを選択できるようにする。

よろしくお願いします 🙏

[arikantr]

Summary (AiChat module – Gemini replies not coming through)

Misskey commands (ping/dice) work, but Gemini chat replies don’t. I reviewed the aichat module and found a few issues that likely cause silent failures and/or prevent grounding from working.

Key Findings

Grounding tool name is wrong
The code uses google_search, but Gemini v1beta expects googleSearch (camelCase). As a result, grounding never triggers.

Model/endpoint is hard-coded
Default is always gemini-2.5-flash. If the API key doesn’t have access to this model (or quota), requests can fail silently. There’s no way to pick the model from config.

Error details aren’t logged
The Gemini call doesn’t log err.response?.statusCode or err.response?.body, so real causes (“model not found”, “API key invalid/restricted”, etc.) are invisible.

Minor: got usage & MeCab side effects

Using parseJson is unnecessary since got().json() already parses.

If keywordEnabled: true while MeCab paths aren’t set, some setups can emit runtime errors that add noise.

Proposed Fixes
A) Fix grounding tool naming

- if (aiChat.grounding) {
-   geminiOptions.tools = [{google_search:{}}];
- }
+ if (aiChat.grounding) {
+   // Gemini v1beta expects camelCase:
+   geminiOptions.tools = [{ googleSearch: {} }];
+ }

B) Allow model selection via config and keep &gemini-flash / &gemini-pro overrides

Create a small helper and use config.geminiModel as default:

- const GEMINI_20_FLASH_API = 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent';
- const GEMINI_15_PRO_API   = 'https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-pro:generateContent';
+ const makeGeminiApi = (model: string) =>
+   `https://generativelanguage.googleapis.com/v1beta/models/${model}:generateContent`;

In handleAiChat → TYPE_GEMINI branch:

- aiChat = {
-   question: question,
-   prompt: prompt,
-   api: GEMINI_20_FLASH_API,
-   key: config.geminiProApiKey,
-   history: exist.history,
-   friendName: friendName,
-   fromMention: exist.fromMention
- };
+ const modelFromConfig = (config as any).geminiModel || 'gemini-1.5-pro';
+ aiChat = {
+   question,
+   prompt,
+   api: makeGeminiApi(modelFromConfig),
+   key: config.geminiProApiKey,
+   history: exist.history,
+   friendName,
+   fromMention: exist.fromMention
+ };

Keep the inline overrides:

- if (msg.includes([KIGO + GEMINI_FLASH])) {
-   exist.api = GEMINI_20_FLASH_API;
+ if (msg.includes([KIGO + GEMINI_FLASH])) {
+   exist.api = makeGeminiApi('gemini-2.5-flash');
    reKigoType = RegExp(KIGO + GEMINI_FLASH, 'i');
- } else if (msg.includes([KIGO + GEMINI_PRO])) {
-   exist.api = GEMINI_15_PRO_API;
+ } else if (msg.includes([KIGO + GEMINI_PRO])) {
+   exist.api = makeGeminiApi('gemini-1.5-pro');
    reKigoType = RegExp(KIGO + GEMINI_PRO, 'i');
}

C) Improve error logging for the Gemini call

- } catch (err: unknown) {
-   this.log('Error By Call Gemini');
-   if (err instanceof Error) {
-     this.log(`${err.name}\n${err.message}\n${err.stack}`);
-   }
- }
+ } catch (err: any) {
+   this.log('Error By Call Gemini');
+   if (err?.response) {
+     this.log(`Status: ${err.response.statusCode}`);
+     this.log(`Body: ${err.response.body}`);
+   }
+   if (err instanceof Error) {
+     this.log(`${err.name}\n${err.message}\n${err.stack}`);
+   }
+ }

D) Simplify got usage

- res_data = await got.post(options,
-   {parseJson: (res: string) => JSON.parse(res)}).json();
+ res_data = await got.post(options).json();

E) Optional: guard keywordEnabled when MeCab is missing

If MeCab is not set up, either default keywordEnabled: false or soft-fail in the keyword module to prevent noisy errors.

Example config fields

Allow picking the model from config (with safe default):

{
  "geminiProApiKey": "AIza...REAL_KEY",
  "geminiModel": "gemini-1.5-pro"   // or "gemini-1.5-flash" / "gemini-2.5-flash"
}

Suggested test steps

Set a valid geminiProApiKey and (temporarily) remove API key restrictions.

Set "geminiModel": "gemini-1.5-pro" in config (commonly available).

Apply patches above, rebuild, restart.

Mention test in timeline: @ai hello &gemini-pro and @ai hello &gemini-flash.

Enable grounding test: @ai ggg what’s the latest about ? (should use googleSearch tool now).

If still failing, the enhanced logs will show concrete errors (status code + body).

[arikantr]

@syuilo I tried many times but couldn’t get it to work. Could you please handle this fix?

[arikantr]

18:02:27 [AiOS]: API: users/show
18:02:27 [AiOS]: <<< Bir mesaj alındı: acdke9lguve00280
18:02:28 [AiOS]: API: notes/reactions/create
18:05:01 [AiOS]: API: users/show
18:05:01 [AiOS]: <<< Bir mesaj alındı: acdkhi8duve0028l
18:05:01 [AiOS]: [aichat]: AiChat requested
18:05:01 [AiOS]: API: notes/conversation
Uncaught exception: Response code 400 (Bad Request)
HTTPError: Response code 400 (Bad Request)
    at Request.<anonymous> (file:///ai/node_modules/got/dist/source/as-promise/index.js:86:42)
    at Object.onceWrapper (node:events:634:26)
    at Request.emit (node:events:531:35)
    at Request._onResponseBase (file:///ai/node_modules/got/dist/source/core/index.js:603:22)
    at process.processTicksAndRejections (node:internal/process/task_queues:105:5)
    at async Request._onResponse (file:///ai/node_modules/got/dist/source/core/index.js:645:13) {
  input: undefined,
  code: 'ERR_NON_2XX_3XX_RESPONSE',
�
  timings: {
    start: 1757268301861,
    socket: 1757268301861,
    lookup: 1757268301888,
    connect: 1757268301889,
    secureConnect: 1757268301890,
    upload: 1757268301890,
    response: 1757268301922,
    end: 1757268301923,
    error: undefined,
    abort: undefined,
    phases: {
      wait: 0,
      dns: 27,
      tcp: 1,
      tls: 1,
      request: 0,
      firstByte: 32,
      download: 1,
      total: 62
    }
  },
  options: Options {
    _unixOptions: undefined,
    _internals: {
      request: undefined,
      agent: [Object],
      h2session: undefined,
      decompress: true,
      timeout: [Object],
      prefixUrl: '',
      body: '{"i":"Pk6o1NINOwVdxx56pgKUV6nVNlVH4cV9","noteId":"acdkhi8duve0028l"}',
      form: undefined,
      json: undefined,
      cookieJar: undefined,
      ignoreInvalidCookies: false,
      searchParams: undefined,
      dnsLookup: undefined,
      dnsCache: undefined,
      context: {},
      hooks: [Object],
      followRedirect: true,
      maxRedirects: 10,
      cache: undefined,
      throwHttpErrors: true,
      username: '',
      password: '',
      http2: false,
      allowGetBody: false,
      headers: [Object],
      methodRewriting: false,
      dnsLookupIpVersion: undefined,
      parseJson: [Function: parse],
      stringifyJson: [Function: stringify],
      retry: [Object],
      localAddress: undefined,
      method: 'POST',
      createConnection: undefined,
      cacheOptions: [Object],
      https: [Object],
      encoding: undefined,
      resolveBodyOnly: false,
      isStream: false,
      responseType: 'text',
      url: URL {},
      pagination: [Object],
      setHost: true,
      maxHeaderSize: undefined,
      signal: undefined,
      enableUnixSockets: false
    },
    _merging: false,
    _init: [ [Object] ]
  }
}

@syuilo