Bug

isReasoningTagProvider() in src/utils/provider-utils.ts does not match the google provider (used by gemini-api-key auth), only google-gemini-cli and google-generative-ai. This causes reasoning/thinking output to leak directly to the end user instead of being filtered.

Steps to reproduce

1. Authenticate with gemini-api-key (creates auth profile google:default with mode: "api_key")
2. Set default model to google/gemini-3-pro-preview (the value from GOOGLE_GEMINI_DEFAULT_MODEL)
3. Send a message with thinking enabled (e.g. thinking=low)
4. The user receives raw <think> reasoning output instead of only the <final> response

Root cause

The model string google/gemini-3-pro-preview yields provider=google. But isReasoningTagProvider only checks for google-gemini-cli (OAuth) and google-generative-ai (API type, not provider ID):

// src/utils/provider-utils.ts
if (normalized === "google-gemini-cli" || normalized === "google-generative-ai") {
  return true;
}

This means:

• reasoningTagHint is false — no <think>/<final> tag instructions in the system prompt
• enforceFinalTag is false — no output filtering for <final> blocks
• The model's thinking output passes through unfiltered to the user

Impact

• enforceFinalTag is not set → <final> tag filtering is skipped (get-reply-run.ts:495, agent-runner-utils.ts:146)
• reasoningTagHint is not set → system prompt omits <think>/<final> formatting instructions (system-prompt.ts:340)
• Affects both Telegram/channel delivery and embedded agent runs

Suggested fix

Add google to the provider check in isReasoningTagProvider:

if (normalized === "google-gemini-cli" || normalized === "google-generative-ai" || normalized === "google") {
  return true;
}

This aligns with:

• gemini-api-key auth creating profile with provider: "google" (auth-choice.apply.api-providers.ts:568)
• GOOGLE_GEMINI_DEFAULT_MODEL = "google/gemini-3-pro-preview" (google-gemini-model-default.ts:4)
• Model catalog tests using providers: { google: { api: "google-generative-ai", ... } }