Onboard

Understudy Onboarding

The first thing a new user experiences. Goal: in a few minutes, leave them with (1) a small open model running locally on their own machine, (2) a clear sense of what Understudy is and why it matters, and (3) a saved profile so you never re-ask what you already learned.

Run this after install-agent-adapter. It follows the engagement doctrine in ../../docs/engagement-and-pacing.md: start the slow download first, then interview while it runs. Detail — profile schema, interview bank, tooling-detection table — is in reference.md.

Safety Gates

• Download approval + size cap. Name the exact model, quantization, and disk size, and get a quick yes before pulling weights. Default to the smallest verified American open model that gives a real onboarding win, and label it as a bootstrap model rather than a workload recommendation.
• Local-first, no upload. Profiling, interview answers, and the model run entirely on the machine. The profile is local; it holds preferences and detected tooling — never secrets, keys, or customer data.
• Gated weights (e.g. Gemma via Hugging Face) need license acceptance + an HF token; the Ollama path avoids this. Never print or commit a token.

Intake

Returning user? If ~/.understudy/profile.json exists, read it, greet them by where they left off, confirm nothing major changed, and skip straight to the work — do not re-run the full interview. Only first-timers get the full flow.

If the launch prompt came from install.sh --lower-my-ant-bill, treat the primary goal as lowering Anthropic/Claude API spend. Still do the local-first profile and quick proof, but keep the interview short and route the real work to ../lower-anthropic-bill/SKILL.md: inventory Anthropic call sites, re-baseline tokenizer risk, audit cache hits, and build an opportunity ledger before any code edits or provider calls.

Flow

1. Start the slow thing first (background). Detect the model runtime (mlx_vlm, mlx_lm, ollama, llama-server, lms). On Apple Silicon, the opinionated first out-of-box target is the smallest verified Gemma 4 local Understudy: the QAT-derived gemma-4-e2b-it-qat-mlx-vlm-understudy — google/gemma-4-E2B-it QAT weights converted by Understudy to MLX 4-bit at group_size=32 (matching Q4_0's block structure). The verified snapshot is stored at https://models.understudylabs.com/session?model=gemma-4-e2b-it-qat-mlx-vlm-understudy (R2 source: r2://understudy-model-snapshots/models/google/gemma-4-e2b-it/mlx-vlm-0.6.2/qat-understudy-4bit-g32/). It is about 3.6 GB on disk, runs at 3.9 GB peak memory (2.6x less than BF16), and is 4/4 certified (generation, OpenAI-compatible serving, logprobs, tool calls) at the prescribed decode. Announce the model, quantization, size, source, and ETA, get a quick yes, then route through ../manage-local-models/SKILL.md and run the CLI snapshot pull:

   understudy models pull gemma-4-e2b-it-qat-mlx-vlm-understudy
   

   This caches the first Understudy under ~/.understudy/models/gemma-4-e2b-it-qat-mlx-vlm-understudy and logs progress/ETA under ~/.understudy/agent-tools/logs/. If the MLX runtime is missing, the slow step is install MLX + pull — get one quick approval, then background it. Read understudy models snapshots before promising the pull. If the live catalog and bundled certified fallback disagree with the session endpoint, surface the mismatch and offer the CLI repair flow; do not silently install a vanilla or third-party model. Then immediately move on; do not watch the bar.

2. Profile the machine (while it downloads). Detect OS/chip (Apple Silicon vs CUDA), RAM / unified memory, free disk. State what fits locally. This is the hardware inventory from ../run-local-model-lab/SKILL.md, brought to the front.

3. Detect ML tooling → infer experience. Check for PyTorch, vLLM, Transformers, TRL/PEFT, MLX, llama.cpp, Ollama, the HF CLI, nvidia-smi (table in reference.md). Lots of ML libs ⇒ experienced; none ⇒ likely first-timer. Use this to pre-fill the interview, not to skip it.

4. Interview (one batched AskUserQuestion). Confirm the inference rather than interrogate: "I see PyTorch and vLLM here, so I'll assume you're comfortable with ML tooling — right?" or "No ML tooling yet — first time running models locally?" Capture: experience tier, primary goal (cost / latency / quality / learning / compliance), hard constraints (ZDR / local-only / approved providers), and preferred coaching depth. Question bank in reference.md.

5. Write the profile and agent card. Save ~/.understudy/profile.json (schema in reference.md): experience tier, detected tooling, hardware, goal, constraints, and the three meet-them-where-they-are dials — vocabulary, coaching depth, opinion strength. Append, don't overwrite, the history of workloads and decisions. For ~/.understudy/agent-card.json, check whether the Understudy desktop app is running first (read the card's app.running / app.pid, or hit app.base_url): the app is the canonical local daemon and maintains the card itself — server endpoint, warm models, shutdown state. Only refresh the card yourself as a fallback when the app isn't installed or running, recording live runtime facts: the local model, endpoint, serving process, companion status, and the exact command or URL for talking to the local Understudy. If ~/.understudy/companion.json points at a dead pid, clear it and record the stale pid in the card.

6. Land the quick win: show the local Understudy exists. Once the snapshot is cached, route through ../ladder/SKILL.md and start the onboarding climb:

   understudy run -- uv run --with mlx-vlm --with mlx-lm python skills/ladder/serve.py
   

   Then open http://localhost:8011/ladder.climb.html?task=sort-email&model=gemma-4-e2b or prove it headlessly:

   curl -N 'http://localhost:8011/run?task=sort-email&model=gemma-4-e2b'
   

   The ladder server loads the cached QAT snapshot for gemma-4-e2b, streams a scored local run, and can optionally compare against the billed gateway lane. Record the live runtime facts — endpoint, served_by, model path, and follow-along URL — in the agent card (step 5). Briefly teach the idea: an open-weight model is downloadable weights you run yourself; local is free and ZDR-safe; you iterate small and local, then graduate to a larger model in the same family via the gateway when you need the quality.

7. Profile the user's real workload. The main path after the local proof is not a model duel. Ask the user for a codebase, trace folder, dataset, eval runner, prompt file, or app route. If they point at a project, route to ../understand-workload/SKILL.md first: inspect prompts in situ, trace the request/response path through code, summarize the dataset or trace distribution, name the real task, and confirm that understanding with the user before any optimization. If there is already a real captured environment, skip the toy sandbox. Only use ../design-simulated-environment/SKILL.md when there is no resettable real workload yet. If the stated goal is lowering an Anthropic bill, route to ../lower-anthropic-bill/SKILL.md instead of asking for a generic problem.

8. Make head-to-head optional. A frontier-vs-local comparison is useful when the user needs to feel the quality gap, calibrate taste, or get buy-in. It is a side quest, not the default evidence path. If the user wants it, follow the ladder VS lane and disclose that the frontier side is billed. Otherwise keep going through workload understanding, capture evidence, and local evaluation against the actual task slice.

9. Route onward. Hand to the understudy orchestrator for the improvement loop. In the normal first-run path, the next worker is understand-workload: pick a real app, trace, prompt file, dataset, or eval and define the task before comparing or optimizing models. Use manage-local-models only to grow and organize the local model library, and run-local-model-lab only once there is a frozen real workload/eval to score.

Adapt everything to the profile: expand jargon and give first-timers one clear recommended path; stay terse and offer trade-offs to practitioners.

Output Standard

End with: runtime + model downloading (and ETA, or "cached"); hardware found and what fits locally; inferred experience tier and the dials set; the profile path written; the agent-card path refreshed; the quick-win result (local generation shown or pending); and one recommended next skill/command.

References

• reference.md — profile schema, interview bank, tooling map, experience→coaching dials.
• ../../docs/engagement-and-pacing.md — the background-first, fill-the-wait doctrine.
• ../../docs/open-model-spotlight.md — Gemma 4 & Nemotron 3 picks and hardware fit.