feat: add LLM provider detectors for other CLIs (Gemini, Codex, …)

[AhmadHammad21]

Context

We added a pluggable LLM backend detector registry in src/agent/llm/detectors/. The first provider, Claude Code, reuses the local CLI's subscription OAuth login as the LLM backend (no API key) via LiteLLM. Adding another provider is intentionally cheap: implement the LlmDetector protocol in a new module and append it to ALL_DETECTORS — no other code changes.

Feature

Add detectors for other locally-installed AI CLIs whose login maps to an LLM endpoint LiteLLM understands.

Candidates (priority order):

• Gemini CLI — strongest next target. OAuth creds in ~/.gemini/oauth_creds.json, free tier, LiteLLM gemini/*. Same pattern as Claude Code.
• OpenAI Codex — ChatGPT OAuth in ~/.codex/auth.json, LiteLLM openai/*. Feasible but subscription-OAuth-as-API is grayer than Anthropic's.
• Others (Copilot, Cursor, Goose, Qwen, Kimi, CodeBuddy): mostly agent CLIs, not raw LLM endpoints — low value for our architecture (we have our own agent loop).

Acceptance criteria

For each provider:

• Implements the LlmDetector protocol (name, label, provider, enabled, is_installed, resolve, status)
• Registered in ALL_DETECTORS
• Surfaced in the Settings → Environment and init-wizard backend cards
• Verified end-to-end with the real CLI installed: install → detect → live LLM call → confirm

Note

Do not build these blind. Each detector depends on the exact credential-file path, model-slot names, and OAuth token format of the target CLI — only verifiable with the CLI installed and logged in. Build each one only when its CLI is available to test.