Architect

Reach a visible robot result with the least new work, then customize. Inspect existing examples before researching or designing a new stack.

Without an established app, an outcome-only request such as trying interactive simulator examples, mapping an environment, or running a pretrained policy is a first-run request even without "new project", a robot brand, or "architect" in the prompt. Use this skill inline, not the optional heavy-research architect subagent. If a domain skill brought you here, select the baseline once and return to that skill for its mechanics; do not loop through architecture again for the same task.

Start from the outcome

• Inspect the current repository, the Robium app registry, and the closest working example before asking questions. Start with candidate READMEs, manifests, and launchers; exclude generated environments, caches, and model assets from discovery scans.
• Define what the robot should visibly do, where it will run, and what evidence would make the first slice credible.
• Honor the user's stated target platform and restrictions even when the agent runs on a different host. Ask about a compatible alternative, not for them to restate constraints they already gave.

Offer the shortest useful starting path

After a quick registry and candidate check, surface these paths briefly, mark the recommendation, and explain what would run first. Do this before broad research, architecture documents, cloning, or lengthy setup. Do not turn the four paths into a mandatory questionnaire: when intent is clear, explain and proceed within it; ask one short question only for a consequential choice.

For "just try it", default to the compatible example with the fewest new moving parts, not the most ambitious stack. Describe prerequisites and first-run work, not unmeasured speed promises. Read reuse paths only for adaptation, component reuse, or a fresh scaffold.

Inspect candidates and prove the selected slice

• Resolve the user's workspace with npx robium-ai workspace --json; its repo and apps paths are authoritative, not a hardcoded ~/robium or the installed plugin cache. For workspace discovery and an occasional quiet freshness check before a new example, read workspace updates. Checking is not updating.
• Discover candidates in the Robium app registry, using a sibling checkout when available. Match the task, host architecture, runtime requirements, and relevant verification evidence, not just the name or a stable label. Read the selected app's README and launcher; distinguish tested conditions from assumptions. For the homepage's first-run tasks, use first-run examples, then read that candidate's exact README and manifest without scanning the whole apps checkout.
• For a clear compatible match, explain the choice and proceed within the user's request: "This matches Robium's navigation example. I'll get it running first, then we can adapt it." Ask only when different candidates or an incompatible platform materially change the outcome. Honor an explicit request to build from scratch. Routine maintenance stays in the owning skill.
• Use the existing checkout without resetting or pulling over user changes; acquire a separate checkout when necessary. Record the source revision and any local changes. For whole-app reuse, run the baseline with its launcher, pinned environment, assets, and viewer; do not create a derivative, upgrade dependencies, or recreate infrastructure just to demonstrate it.
• Check prerequisites, credentials, and occupied ports before lengthy builds. Explain first-run downloads and reuse caches on later starts. Missing model access is a blocker, not permission to substitute a mock and call it live. Keep secrets out of prompts/logs; explain API charges and obtain authorization before paid calls. Do not provision cloud resources or move physical hardware as an implicit part of onboarding.
• Check the requested behavior at a meaningful milestone, not after each edit. Use the app's smallest relevant check, not a new suite or repeated UI tours. A low-risk visual check may be handed to the user with exact steps and an explicit unverified status; an open port is not proof the demo works. Stop only processes started for this run. If blocked, report the exact unmet prerequisite without inventing a new stack. Before a port, dependency overhaul, or prolonged troubleshooting, surface the blocker and a simpler compatible path; let the user choose the tradeoff. For try-only requests, offer one small customization after it works. For requested customization, proceed with the agreed change and keep the baseline.

Choose only the stack the slice needs

• When environment setup is needed, use environments early.
• For ROS-based mobile navigation, route through ros2, navigation, simulation, and the selected simulator and visualizer.
• For learned manipulation, route through lerobot, data, and huggingface; add a simulator only when the first slice needs one.
• Use integration when module boundaries or cross-process communication are themselves a design decision.
• Use testing when risk, uncertainty, or requested test work needs a verification decision; do not automatically author tests. Use test-assets for needed fixtures.
• Route provider mechanics to cloud-run or runpod only after deployment is part of the approved slice. A public session layer belongs to live-demo; publishing a finished app belongs to app-publishing.

When the stack is genuinely undecided, read stack-selection.md. Keep upstream project documentation as the source of truth for supported versions and hardware.

Keep the decision record light

• Keep decisions inline for tryouts and tiny demos. Use docs/architecture-brief.md when a new app or re-architecture benefits from a durable record, never as a gate before an agreed first visible result.
• Record the outcome, chosen direction, provisional assumptions, risks, next probe, and allowed pivots. It is a living decision record, not a contract.
• Read brief-template.md only when creating or revising that record. The filled example is illustrative, not a source of current compatibility facts.
• If no close app exists, read scaffold-patterns.md and prune the starting shape to the first slice.