alignment

name	alignment
description	Interview the operator until shared understanding. Gated branches one at a time; independent discrete choices may share one structured-question turn. Use when stress-testing an idea, design, or requirement; resolving open questions before acting; or when the user says alignment, /alignment, or wants to get aligned.
disable-model-invocation	true

Alignment

Read-only. Do not edit files, apply to the cluster, create a worktree, or implement until alignment is reached and the operator explicitly asks to proceed.

Interview relentlessly until shared understanding (grill-me pattern). Walk gated branches one at a time; independent discrete choices may share one structured-question turn. Explore the tree when that answers the question; do not ask the operator to repeat what the repo already shows.

Question channel: .agents/context/questions.md. Harness adapters name the concrete tool (Cursor: question-format.mdc / .agents/rules/question-format.md).

When to run (default: early)

Run before writing issues, plans, or code when any of these apply:

Trigger	Examples
New app, feature, or wide change	Desired state unclear, multiple valid approaches, stop condition fuzzy
Issue needs scoping	find-work tier "needs scoping"; acceptance bullets not yet writable
Plan needed / new or extended plan	Rough idea; turning an issue into docs/plans/ checkboxes
Security / protected-path tradeoff	Anything touching talos, generic-app, bootstrap, .agents/
slice: hitl	Product, security posture, or protected-path tradeoffs
Implement gate	Manifest / code work while expectations still open

Skip when acceptance is already explicit (clear issue file, plan checkbox with verification). Operator "proceed" after alignment counts; do not skip when scope is still fuzzy and no alignment summary exists in the thread. Do not re-run alignment in the same thread if a summary already exists.

Typical flow

idea / gap / find-work row
        │
        ▼
   /alignment  (read-only; grill until shared understanding)
        │
        ├──► file-issue → docs/issues/
        ├──► project-planner → docs/plans/
        └──► implement-change (only after operator says proceed + dedupe)

Alignment output should be pasteable into an issue (Problem, Acceptance) or plan (Goal, slice checkboxes) without another discovery pass.

Next step (after proceed)

Suggested fork:

Next	Invoke
File the gap	file-issue
Plan the how	project-planner → docs/plans/
Implement a clear Launch brief	implement-change
Docs-only under docs/	docs worktree + documentation persona / reconcile-docs as applicable
Restore / deploy domain work	matching skill (helm-deployment, restore skills, …)
Cursor / .agents/ / hooks	chore/* worktree; protected paths per protected-paths
Still fuzzy	more alignment (or stop if unattended)

After proceed, ship via the hybrid model (development-loop.md#ship-model): unauthorized → draft-commit; authorized → ship-work then clock-out after merge. Alignment itself is read-only; later implement laps use worktrees (worktrees.md).

Routed from

• Issue and plan authoring — before file-issue or a new plan when scope is still fuzzy
• implement-change: alignment gate before worktree
• find-work: Launch brief for vague or multi-path items
• subagents.md and clarify-dont-guess.md: extended ambiguity
• project-planner: before heavy planning when several decisions depend on operator input

Output

Short summary: decisions made, open items, suggested next step (table above).