Skip to content
JRS1986 edited this page May 6, 2026 · 8 revisions

FAQ

Is CodingScaffold a coding agent?

No. It is a scaffold that prepares a project for coding agents and local-first model routing.

Do I need an LLM for the first start?

No. Setup, hardware probe, credential templates, adapter generation, and tools select-model recommendations work without calling a model. You need an LLM only when a coding tool such as OpenCode or OpenClaude starts an actual agent session.

Does setup install tools?

Yes, when it is running interactively and the selected coding environment is missing. It asks before installing. You can also run coding-scaffold setup tool --tool opencode to validate the tool, or add --install to install a missing tool intentionally.

Does the scaffold install optional add-ons too?

Yes. Use coding-scaffold setup addon --addon llmfit, routellm, open-multi-agent, obsidian, or caveman-compression. Setup can also offer add-ons interactively. RouteLLM installs into the active Python environment, Open Multi-Agent installs into the target Node.js project, Caveman Compression is cloned under .coding-scaffold/tools/ as an optional external engine, and Obsidian remains manual on WSL because it is a desktop app.

Can setup configure the shared knowledge remote?

Yes. Use coding-scaffold setup knowledge --target . --backend obsidian --shared-remote <repo>. Setup can also ask for this during setup. The remote URL is metadata only; credentials and tokens stay local.

Can I refresh generated files later?

Yes. Use coding-scaffold setup update --target .. The command re-detects hardware and providers, recreates generated scaffold files, updates files that still match their last generated checksum, and writes .new files next to anything you edited locally.

Does it require cloud APIs?

No. It can work local-only. Cloud providers are used only when credentials are configured and the project privacy mode allows it.

Does it store secrets?

No. It writes ignored templates such as .coding-scaffold/.env.local, but it does not commit, print, or collect secret values.

Why not just use GitHub Copilot?

Copilot is useful for completion and chat. CodingScaffold focuses on agentic workflows: inspect, plan, edit, verify, review, and preserve reusable team habits.

Why Markdown for knowledge?

Markdown works in Git, GitHub, GitLab, editors, Obsidian, and memory tools. It is easy to review and easy to migrate.

Can CodingScaffold avoid context poisoning?

It can catch the easy cases early. Run coding-scaffold context budget --target . --source team to estimate whether shared knowledge, skills, policy, and agents are getting too large for a healthy session. Run coding-scaffold context compress --target . --source knowledge to create optional compressed sidecars with the built-in compressor. Use context budget --prefer compressed to estimate a sidecar-first session. Still use human judgment: narrow retrieval, keep policy uncompressed, and open a fresh session when history has become stale.

Should every team use RouteLLM or Open Multi-Agent?

No. Start with setup, OpenCode, skills, and knowledge. Add RouteLLM or Open Multi-Agent only after the team has a proven need.

Clone this wiki locally