Opus 4.8, Agent Skills (+per-project toggle), Claude Design import, Nuxt generation, provider-aware Git + reliability & tests

[holoduke]

Summary

A broad capabilities + hardening pass on top of upstream: Claude Opus 4.8, per-project Agent Skills with a real on/off toggle, importing designs from Claude Design, full Nuxt app generation, a provider-aware Git layer (GitHub or Gitea), plus a reliability/security sweep and the repo's first automated tests.

Highlights

Build & agent

• Claude Opus 4.8 via the Agent SDK with adaptive thinking + an effort selector (Auto / Deep / Off) in the composer.
• Full Nuxt generation — new projects scaffold a multi-page Nuxt 4 + Nuxt UI app (layout with header/footer, shared navigation, Home / Pricing / About / Contact + a custom error page), not a single bare page.
• Provider-aware Git (`GIT_PROVIDER` = `github` | `gitea`) for repo create, push and reading CI deploy status; provider-agnostic "Git" wording in the UI.

Agent Skills

• Per-project + global skills, shown in settings and a quick-access modal (checkboxes + live search) in the chat toolbar.
• Hard per-project enable/disable: by default skills auto-load; the moment any skill is disabled the project switches to a staged model that loads only the enabled set (enabled project skills as dirs, enabled globals as symlinks) — a disabled skill is genuinely not loaded. Untouched projects behave exactly as before.

Import from Claude Design

• Upload a claude.ai/design `.zip`; the screens, fonts and assets are staged into a gitignored `design-reference/` folder (design-process noise filtered out, single wrapper directories handled), then the agent gets an editable instruction to port them into the app's own components/structure.

UI

• Resizable chat/preview split (draggable divider, width persisted).
• Publish flow shows the real CI deploy run (queued / running / success) plus the last deployed commit and relative time.

Reliability & security

• Encrypt provider tokens at rest (same crypto as env vars); legacy plaintext rows migrate on read.
• Preview dev servers: idle eviction + LRU so a small port pool can't be exhausted; stop leaking orphaned processes (group-kill of the detached tree); keep the server warm across refresh; fix a tight auto-start retry loop; clear deploy poll intervals on unmount; `no-store` on the status endpoint.
• Persist the encryption key so secrets survive a restart; DB indexes + crash recovery for stuck requests; keep prerender `failOnError` disabled so deploys don't break on a placeholder link.

Code quality & tests

• First automated tests: vitest + 59 unit tests (path utils, chat serializer, design-import, skill-id guards, formatting, tool-metadata parsers).
• Extractions / de-dup: pure skill-id module, agent system prompt to its own file, formatting helpers, and the tool-use metadata parsers now shared between the CLI adapters and the chat log (removing a triplicated copy).

Test plan

• vitest suite green (59 tests)
• Self-hosted deploy flow (repo create → push → CI build → live URL) verified end-to-end
• Skills: disable removes a skill from the staged set (not loaded); re-enable restores it; untouched projects unchanged
• Design import stages screens/fonts/assets (flat and wrapped zips), noise filtered
• Preview: warm-refresh reuse, idle eviction reclaims ports, no retry-loop
• Token encryption: stored value is ciphertext, reads decrypt, legacy migrates
• Resizable split drags + persists