gemma

name: gemma description: "로컬 LM Studio와 Google AI Studio(Gemini API)를 통해 Gemma 모델에 프롬프트를 전달한다. variant별 자동 라우팅(e2b/e4b→로컬, 26b/31b→원격)과 LM Studio 미가용 시 Gemini 폴백을 지원한다. gemma, gemma4, gemma로 요약해줘, gemma로 번역해, lm studio로 돌려줘, gemini api로 보내줘, 로컬 LLM, 오프라인 AI, 로컬로 처리해, 클라우드로 돌려줘, Gemma 호출 요청 시 사용한다. 민감 정보 오프라인 처리, 긴 컨텍스트 요약, 다국어 번역, 초안 생성 등에 적합." group: llm model: sonnet allowed-tools: Bash(bash:), Bash(lms:), Bash(op:), Bash(brew:) argument-hint: "[--local|--cloud] [variant] prompt"

Gemma (LM Studio + Google AI Studio)

Dispatch a prompt to Gemma via local LM Studio (OpenAI-compatible API) or Google AI Studio / Gemini API. Variant selects the backend by default, with automatic fallback to the remote API when LM Studio is unavailable.

The routing, HTTP calls, and 1Password integration are implemented in Rust under tools/gemma/ (Cargo workspace). The scripts/*.sh files are thin launchers that cargo run the gemma binary — same CLI surface, same stderr info:/warn:/error: format, same exit codes as the previous bash scripts.

How to invoke

All invocations go through scripts/query.sh. Use the Bash tool:

# Default: variant e4b on LM Studio, auto-fallback to Gemini if LM Studio down.
bash /Users/ujuc/.claude/skills/gemma/scripts/query.sh "이 문단을 3줄로 요약해줘: ..."

# Explicit variant.
bash /Users/ujuc/.claude/skills/gemma/scripts/query.sh e4b "hello"

# Force remote (Gemini API).
bash /Users/ujuc/.claude/skills/gemma/scripts/query.sh --cloud 31b "복잡한 추론 문제: ..."

# Force local; fail if LM Studio unavailable (privacy-strict mode).
GEMMA_NO_FALLBACK=1 bash /Users/ujuc/.claude/skills/gemma/scripts/query.sh --local e4b "민감 데이터: ..."

stdout contains only the model response. stderr has a single info: line with backend=lmstudio|gemini model=<id> so you can surface the real model to the user (e.g., prefix Gemma (gemma-3n-e4b-it via LM Studio) 응답:).

Variant → Backend Routing

Override with --local / --cloud or GEMMA_BACKEND=lmstudio|gemini.

See references/models.md for the model ID matrix and resolution regex.

Fallback behavior

If a request targets LM Studio but GET /v1/models fails within 3 seconds, or no loaded model matches the variant, the script silently falls back to Gemini API. The switch is logged to stderr:

warn: LM Studio unreachable at http://localhost:1234
info: backend=gemini model=gemini-flash-latest (Gemma not available on API), fallback from LM Studio

To disable fallback (e.g., for privacy-sensitive prompts that must not leave the machine): set GEMMA_NO_FALLBACK=1 or pass --local. The script then exits with code 3 and prints setup hints.

Remote model auto-discovery

Google AI Studio publishes new Gemma variants on their own schedule, so IDs are resolved at runtime via scripts/list-gemini-models.sh (5-min cached). The script picks the highest-version Gemma matching the requested variant, falling back to gemini-flash-latest or gemini-pro-latest aliases if no Gemma match exists. No code change needed when Gemma 4 ships on the API — it will be picked up automatically.

Override with GEMMA_GEMINI_MODEL=<id> for a specific model.

Setup (first run)

Rust toolchain is required. Launchers defer to cargo run; install via https://rustup.rs. The first invocation compiles the gemma binary (release profile, ~6–30s); subsequent runs are instant via Cargo's incremental cache. curl/jq are no longer needed — the Rust build links its own TLS stack (rustls) and JSON handling.

On first invocation, scripts/query.sh still calls ensure-deps.sh if a backend dependency is missing:

• lm-studio cask (for local path — lms CLI ships inside the cask)
• 1password-cli cask (for remote path — API key lives in 1Password)

Prompt-less install: GEMMA_AUTO_INSTALL=1.

After install, you still need to:

1. LM Studio: open GUI once, download lmstudio-community/gemma-3n-E4B-it-MLX-4bit, run lms server start and lms load <model>. Details in references/backends.md.
2. 1Password: store the API key at op://key/gemini-key/credential (override via GEMMA_OP_REFERENCE), ensure op account list returns your account (Touch ID integration auto-signs).

Procedure

1. Identify the prompt body and (optional) variant from the user's request.
2. For long inputs, build the prompt with a clear instruction on top and the body in a single string (heredoc or quoted).
3. Call query.sh via bash with the chosen flags.
4. Surface the stdout response to the user with a header that names the actual backend and model (read from the stderr info: line). Never present Gemma/Gemini output as if it were Claude's own reply.
5. On script failure, relay the hint printed to stderr (install, server start, 1Password signin, etc.) to the user verbatim.

Environment variables

Full walkthrough in references/backends.md.

When to use Gemma vs Claude

Appropriate for:

• Summarizing or classifying sensitive data offline (use --local + GEMMA_NO_FALLBACK=1)
• Multilingual translation (140+ languages supported natively)
• Long-document summarization (128K+ context)
• Drafting notes, obsidian pages, initial outlines
• Structured JSON output or function-call style tool use

Better left to Claude:

• High-difficulty math/reasoning (AIME, GPQA Hard) — Claude wins decisively
• Large-codebase navigation — Claude Code's Agent/Explore is purpose-built
• Music/non-speech audio understanding — Gemma does not support it

Error handling

References

• references/models.md — variant table, MLX model IDs, Gemma 4 benchmarks.
• references/backends.md — LM Studio install, Google AI Studio key flow.
• references/delegation-guide.md — when to delegate to Gemma from Claude.

Eval Criteria

Binary checks the skill must pass when invoked in realistic conditions:

1. With LM Studio running and Gemma loaded, a default e4b call returns a non-empty response and logs backend=lmstudio on stderr.
2. With LM Studio stopped and GEMMA_NO_FALLBACK unset, the same call silently falls back to Gemini and returns a non-empty response.
3. GEMMA_NO_FALLBACK=1 with LM Studio stopped exits non-zero without contacting Google.
4. --cloud 31b resolves to a gemini-pro-latest (or newer Gemma 31b if the API exposes one) model and returns a response.
5. GEMMA_OP_REFERENCE pointing to a non-existent item produces a clear error (exit 4) rather than a Google 401.