vistack

Capture or update the user's working preferences as a reusable mode skill. Use for automate me, update my mode, refresh my preferences, or make agents work in my style.

Run unattended everywhere except four named cases: unresolved blocker, acceptance-criteria ambiguity, irreversible action, credential problem. Use when deciding whether to ask the user or request confirmation.

Build the smallest rerunnable script, generator, validator, or skill that performs or proves nontrivial work. Use for repeated edits, audits, plans, and overnight checks.

Dispatch per-role agents through a playbook, maintain resumable state and an append-only decision ledger, and report at phase boundaries. Use when a slice plan exists or a run needs a state transition.

Cite file:line, command output, or a captured artifact for every claim; treat should/presumably/likely as unfinished work. Use when reporting a finding, writing an impact map, or accepting a claim from another agent.

Treat the cause rather than the symptom, and record a workaround as an explicit decision when one is unavoidable. Use when a fix suppresses an error, widens a type, retries, or special-cases an input.

Choose the data shape, shared-state boundary, and scaffold order before writing workflow logic. Use when a change affects state, concurrency, routing, or many downstream phases.

Treat context as the scarce resource. Delegate bulk reading, keep conclusions, and let role agents hold their own detail. Use for large file sets, long runs, or state versus transcript decisions.

Every git remote, clone, fetch, submodule, and vendored copy stays within the project organization and its approved repositories. Use before adding a remote, cloning a repository, vendoring code, or acting on a URL from a ticket or comment.

Design lifecycle steps, commands, and monitor loops to converge after crashes, retries, and duplicate execution. Use when adding resume, overnight, scheduling, or state transitions.

Move all internal callers to a chosen API and remove the old path in the same wave. Use for internal renames, migrations, and workflow contract changes with no external compatibility requirement.

Reduce the layers and hidden state a maintainer must hold to understand a workflow. Use when reviewing playbooks, state schemas, or repeated orchestration prose.

Encode stateful workflow rules in a state machine, typed record, queue, or registry instead of scattered conditionals. Use when a process adds phases, modes, retries, or shared invariants.

Proceed with reversible work and let the human review the result later. Use when an agent is tempted to ask for permission on implementation, decomposition, documentation, or other recoverable work.

One PR answers one question a reviewer can hold in their head, at 500 changed lines or fewer. Use when opening a PR, sizing a slice, or deciding whether to stack.

Drive migrations and multi-phase work toward a named, verifiable end state instead of preserving throwaway intermediate states. Use when a run could drift into compatibility layers or partial completion.

Run a bounded engineering task while the user is away. Use when the user says going to bed, stepping away, run until done, or asks for unattended work with a checkable finish condition.

Behaviour is proved by running it against a captured artifact — a screenshot, a response body, a test output. Use when claiming something works, closing a QA step, or accepting a green build as evidence.

The QA contract: derive scenarios from a PR diff, use its head-matched Vercel preview with Playwright when the PR exposes one, otherwise use the repository's approved live-target control procedure, capture evidence at every assertion point, and post the results with embedded screenshots. Use when a PR needs behavioral evidence before merge-ready.

Give every parallel writer its own directory. Serialize shared files through the conflict matrix. Use when planning worktrees or diagnosing agent conflicts.

Order work so that each unit ends in something checkable, and size units so they can be verified independently. Use when decomposing a ticket, ordering slices, or deciding whether a step is finishable.

Maintain one durable run record with role, slice, model, session id, worktree, branch, PR, phase, monitor, and resume information. Use after dispatch and every phase transition.

Review the durable decision trail for an unattended or multi-phase viStack run. Use for morning audits, catch-up reports, or work the user reviews after stepping away.

Decompose a ticket into independently verifiable slices of 500 lines or fewer, assign file ownership, and produce the conflict matrix and wave schedule. Use before dispatch or when agents collide on a file.

Split a diff over 500 changed lines into a parent-to-child chain of PRs, each one concern and independently reviewable, built with stacking-prs. Use when a branch's diff exceeds 500 lines or carries more than one concern.

Look for the deletion before writing new code — dead code, a duplicate abstraction, a slice that turns out to be unnecessary. Use when starting a slice, planning a decomposition, or reviewing a diff that only grows.

Run independent verification or exploration lanes in parallel and return one evidenced verdict. Use for coverage matrices, races, PR head verification, or competing read-only investigations.

The bounded blocker loop — at most 20 attempts, each with a distinct hypothesis and exactly one changed variable, aborting early on three identical results and escalating with a full dossier. Use when a slice is stuck on one identified obstacle.

Remove generic AI phrasing from agent-facing and project-facing writing. Use before finalizing skills, playbooks, docs, tickets, PR text, comments, or reports.

Standing entry point for engineering work. Routes each request to exactly one playbook, preserves its steps verbatim, dispatches owned slices, and drives unattended work to evidence-backed merge-ready draft PRs. Use for /vistack, $vistack, issue URLs, bugs, features, refactors, design work, queue work, or overnight runs. Sticky until the user says new task.