feat: experimental keyless CLI backend with auto-fallback and ASR fallback

[dull-bird]

Summary

This PR adds an experimental, opt-in way to run Resource2Skill without Azure OpenAI or Gemini API keys by using locally authenticated agent CLIs (currently Claude Code) as the LLM backend, and by analyzing YouTube videos from yt-dlp subtitles + ffmpeg keyframes instead of Gemini native video understanding.

It now also auto-detects the keyless path when API keys are missing, and falls back to local ASR when a video has no native subtitles.

Changes

• "core/llm_cli.py": new CLI LLM backend for Claude Code. Supports multi-turn conversations via session resume ("--resume") and emulates OpenAI function calling with a strict JSON envelope protocol.
• "core/llm.py": dispatch to the CLI backend when "R2S_LLM_BACKEND=cli" and "R2S_CLI_TOOL=claude" are set. Existing API-key paths remain the default and are unchanged.
• "core/llm.py": auto-fallback to CLI when "AZURE_OPENAI_API_KEY" is missing and an agent CLI (claude, codex, kimi, omp) is on PATH.
• "core/analyzer_cli_video.py": new "R2S_VIDEO_BACKEND=cli" path. Fetches auto-generated VTT subtitles (with rolling-caption dedup), extracts key frames, and asks Claude Code to reason over the transcript + image paths.
• "core/analyzer.py": lazy-import "google-genai" so the CLI video path can run without the Gemini SDK installed, and auto-fallback to "R2S_VIDEO_BACKEND=cli" when "GEMINI_API_KEY" is missing.
• "core/analyzer_cli_video.py": ASR fallback when no native VTT is found. Supports "R2S_VIDEO_ASR=auto|faster-whisper|whisper" and "R2S_VIDEO_ASR_MODEL=tiny|base|small".
• "skill/Resource2Skill/SKILL.md": agent connector updated with skill discovery, default behavior, and ASR options.
• "README.md" / ".env.example": document the new no-key setup and ASR variables.

Verification

• Multi-turn memory test: model recalls a codeword across resume sessions.
• Tool-protocol test: model emits "tool_calls", harness executes, result is incorporated.
• "python cli.py domains" and "python cli.py validate-domain --domain web" run without API keys.
• Dry-run agent loop completes iterations without backend errors.
• Reproduced one skill from the official HF dataset ("microsoft/RESOURCE2SKILL") on a 65s CSS typewriter tutorial; the CLI subtitle+frame analysis produced the same core mechanism ("steps()" + "overflow:hidden" + blinking cursor) as the official Gemini-based skill.
• ASR fallback benchmark on a 60s English YouTube clip (M4 Macbook Air):
  • faster-whisper tiny: ~4s
  • openai-whisper tiny: ~21s
  • faster-whisper base: ~25s, noticeably more accurate than tiny
• Native VTT path still works; forced no-subtitle fallback produces a valid VTT.

Scope / Limitations

• Only "claude" is implemented. "codex", "kimi", and "omp" are reserved via "R2S_CLI_TOOL" but currently raise "not supported yet".
• "R2S_VIDEO_BACKEND=cli" image ingestion works best with Claude Code; kimi/codex/omp have not been verified for local image input.
• "python cli.py build" (embeddings) still needs a Gemini key unless keyword-only retrieval is acceptable.
• ASR fallback requires "faster-whisper" or "openai-whisper" installed; first model download is ~40MB (tiny) to ~150MB (base). Transcription speed depends on CPU.

Backward Compatibility

Default behavior is unchanged. Users can explicitly set "R2S_LLM_BACKEND=cli" and/or "R2S_VIDEO_BACKEND=cli" to use the new paths, or leave keys unset and let the system auto-fallback when a local CLI is available.

[dull-bird]

@microsoft-github-policy-service agree

[waltstephen]

Hi @dull-bird,

Thanks for contributing and sharing this keyless CLI path with the project. After narrowing the review, I think there are only two core issues to address before merge:

1. core/llm_cli.py: isolate conversations. Sessions are identified only from the first two messages, so two independent identical runs can reuse the same Claude session, with the second run sent as Continue.. Please scope the session to each run or use an explicit conversation identity.

2. The fallback should respect an explicitly supplied Gemini key. In core/analyzer.py and core/llm.py, backend selection checks the GEMINI_API_KEY environment variable but can ignore a valid per-call api_key= value and route the request to the CLI backend instead. Please include the explicit key when deciding whether fallback is needed.

One minor item: skill/Resource2Skill/SKILL.md currently fails the standard skill validator. Removing the unsupported author field and using a lowercase hyphen-case name should resolve it.

Thanks again for providing this contribution and making the work available to the community. Once these items are updated, I’ll take another pass.

[dull-bird]

Thanks for taking a look. I've pushed the fixes to feature/cli-backend-keyless-video (c9a9f77a).

I addressed three items:

• CLI sessions are now isolated by conversation_id, and the agent loop generates a fresh run id per task.
• Explicit api_key= now takes precedence over auto-fallback, so the code will not switch to the CLI backend by mistake.
• skill/Resource2Skill/SKILL.md now uses a validator-friendly header format.

Verified with python -m pytest -q test_keyless_cli_backend.py, python cli.py validate-domain --domain web/ppt/excel/reaper/blender, and python -m pytest -q domains/ppt/test_pptmaster_r2s_validator.py.