Summary
Rewrite the onboard mode as an agent-driven research → interview → customize cycle that produces project-tailored config and templates.
Problem
- Current onboard mode asks generic questions but doesn't research the codebase
- Templates are identical across all projects regardless of framework/architecture
- Step instructions say "search for relevant files" instead of referencing actual project patterns
- Interview questions are generic (UI design questions for a CLI tool)
Proposed design
Two-step setup:
kata setup — creates skeleton with seed templates copied from package (usable immediately)
kata enter onboard — agent researches project, interviews user, customizes seed files
Onboard phases:
- P0: Research — explore codebase, detect stack (Next.js? Express? CLI?), read package.json, directory structure, key source files, existing configs
- P1: Interview — targeted questions based on research findings (not generic)
- P2: Customize — patch seed templates/config with project-specific content:
- Rewrite step instructions to reference actual project files and patterns
- Tailor interview categories to the domain
- Add project-specific global_rules and per-mode rules
- Fill in project.* fields from detected values
- P3: GitHub — gh CLI, auth, labels
- P4: Validate —
kata doctor to verify everything works
Key principle: Agent customizes EXISTING seed templates, does not generate from scratch. Modifies behavioral parts (instructions, rules) while leaving structural parts (phases, dependencies, step ids) untouched. This preserves upstream merge compatibility — kata update can merge structural improvements while preserving the agent's customizations.
Example output
For a Next.js + tRPC + Drizzle project, the agent would customize:
global_rules with "This project uses pnpm", "Run pnpm build after changes — route types generated at build time"
- Template instructions with "Check src/server/routers/ for existing router patterns"
- Interview architecture questions with "Which services does this touch? (Gateway / Auth / Multi-service)"
- No UI design interview category for a headless API project
Depends on
Research
See planning/research/2026-03-31-setup-overhaul-upstream-merging.md
Summary
Rewrite the onboard mode as an agent-driven research → interview → customize cycle that produces project-tailored config and templates.
Problem
Proposed design
Two-step setup:
kata setup— creates skeleton with seed templates copied from package (usable immediately)kata enter onboard— agent researches project, interviews user, customizes seed filesOnboard phases:
kata doctorto verify everything worksKey principle: Agent customizes EXISTING seed templates, does not generate from scratch. Modifies behavioral parts (instructions, rules) while leaving structural parts (phases, dependencies, step ids) untouched. This preserves upstream merge compatibility —
kata updatecan merge structural improvements while preserving the agent's customizations.Example output
For a Next.js + tRPC + Drizzle project, the agent would customize:
global_ruleswith "This project uses pnpm", "Run pnpm build after changes — route types generated at build time"Depends on
Research
See
planning/research/2026-03-31-setup-overhaul-upstream-merging.md