dobby

Cut an npm release of @kvnwolf/dobby from the main checkout — inferred bump, raw-TS tarball gate, publish, tag, GitHub release with changelog.

Syncs documentation, authors the commit message + PR body, runs the gate and performs the git/gh ceremony directly (stage → `dobby check --fix` → commit → push → PR), then monitors the PR to a verdict. Use when committing code, finishing a task, pushing changes, or creating a PR.

One-pass migration of a consumer repo off vite-plus onto @kvnwolf/dobby — swap deps, thin `tsconfig.json` and delete the now-default tool configs (biome/vite/vitest/drizzle, kept only for real deltas), move files to canonical paths, regenerate dobby.config.json, drop .conductor, rewire CI + Vercel build, verify with `dobby check`. Run in a consumer repo, once, after updating the dobby plugin.

Set up a repo (greenfield or existing) for the work skills — interview, install @kvnwolf/dobby, scaffold the base files + thin tsconfig, write the dobby.config.json contract. Run once.

Disciplined diagnosis loop for hard bugs — build a fast deterministic feedback loop, rank falsifiable hypotheses, instrument one variable at a time. Use when something is failing, a bug is intermittent or non-obvious, or there's a performance regression.

Dispatch a scoped, ad-hoc task to a worker agent (or a few in parallel) and review what comes back — without the full /dobby:execute plan-and-waves ceremony. Use for a small fix or change, or a bounded investigation, when you don't need a STATE.md spec.

Build an approved plan's tasks — per task, separate agents implement → code-review → verify in a loop until both pass. Use to execute a task plan, standalone or as the execute stage after /dobby:spec.

Post-merge worktree teardown — use when the PR of the current goal's worktree is merged and you want to clean up and return to main.

Mine a marked session (a /dobby:mark indicator, or a raw path to a session .jsonl from any project) into concrete edits to dobby's skills, from how they actually behaved in the field. Use in the dobby repo when you have a session pointer to mine.

Emit a portable pointer ("indicator") to the CURRENT Claude Code session. Use when you hit friction with a dobby skill while working in a consumer project and want to flag this session for /dobby:learn — not fix the kit now.

Build a throwaway prototype to answer a design question before committing — an interactive terminal app for state/business-logic questions, or several UI variants on one route with a switcher. Use when the user wants to prototype something, compare UI variants, or sanity-check a data model or state machine; or when /dobby:interview or /dobby:research hits a question only answerable by playing with it.

Start a work session — normalize the goal (free-text prompt, or the configured tracker's issue — a GitHub `#123`/URL or a Linear `VON-123`/URL) and ground it in the codebase before interviewing or planning. Use at the start of any feature, fix, or refactor, or when handed a ticket to work on.

Turn an aligned task and its research into a concrete build plan with a vertical-slice task table and per-task verify recipes. Use after interviewing/researching a task, or to plan a feature and break it into tasks. No plan mode.

Quick-capture a follow-up, bug, or tech-debt item to the project tracker (GitHub Issues, Linear, or local). Use when you spot something worth tracking mid-work and want it logged, not triaged now.

Resolve an in-progress merge, rebase, or cherry-pick conflict by recovering each side's intent from history and PRs, then reconciling every hunk without losing either side's behaviour.

Triage an incoming issue or external PR — verify the claim, then write a behavioral agent brief or record the rejection in the out-of-scope KB.

Address external code-review findings on the current PR — triage with a human gate, delegate fixes, resolve threads, re-trigger review, report merge-readiness. Use when a review bot (Greptile, CodeRabbit) or human reviewer left comments on a PR and you want the feedback addressed and the review green, or to rebut a stale finding.

Creates or modifies agent skills, single- or multi-workflow (orchestrator). Use when the user wants a skill created, edited, or migrated between single and multi-workflow.

Recipe for client-side data fetching with TanStack DB — server function → Drizzle-derived query collection → LiveQuery. Use when adding data fetching, a new collection, or a list/table view.

The write-side conventions for this app — forms (useAppForm from @/shared, Zod validation, field + dialog anatomy) and mutation UX (submit-validated dialogs, optimistic in-place toggles, type-to-confirm, toasts). Use when touching any form, input, validation, submit flow, or data mutation. Write-side partner to /dobby:data-fetching.

Compact the current session into an ephemeral fork document a fresh Claude Code session can pick up.

Review the architecture of a module, subsystem, or repo and propose prioritized improvements. Use to assess structural health, find module-boundary smells, or plan a refactor — it reports opportunities, it doesn't change code.

Interview the user relentlessly to reach a complete, shared understanding of a task before it's planned or built. Use before planning a feature, change, or refactor; to stress-test a design; to resolve ambiguity; or when the user says "interview me" / "grill me".

Turn a loose, multi-session idea into a durable decision-map of dependency-linked investigation tickets, then drive it toward a plan one ticket per cycle.

Per-module file-type convention for this stack (TanStack Start + Drizzle/Neon + Better Auth) — which role file each piece of code belongs in (.server.ts · functions.ts · .browser.ts · schema.gen.ts), deep-path imports (no barrels), eager instances, env as single source. Use when creating or refactoring a module or feature slice, deciding where a piece of code goes (a server function, a server-only instance, a browser client), or co-locating a Drizzle schema.

Gather the technical context a plan needs before building — current docs, reusable skills/modules, resolved unknowns. Use after aligning on a task and before planning it, or when asked to research the tech or approach for an upcoming change. (For a one-off doc lookup, use find-docs directly.)

Learn a topic in-session — taught from trusted resources, verified with a tight feedback loop, recorded as evidence of understanding.

Generate an interactive bash wizard that guides a human through a manual procedure — opening URLs, capturing values, and writing .env files and GitHub Actions secrets.

Close out a work session — final human smoke test, doc/ADR reconciliation, optional skill packaging, then dispose the ephemeral STATE.md and hand off to /dobby:commit. Use when finishing a feature or session, or wrapping up before committing.