Door

/door — Is This Decision a One-Way Door?

Most disasters are irreversible choices made casually, and most wasted time is reversible choices made ceremoniously. This skill sorts them — then spends effort only where it can't be refunded.

Invocation

/door uuid vs bigint primary keys
/door                              # triage the decision currently under discussion

Procedure

1. State the decision and the options. Two sentences max per option.

2. The sort question: if this turns out wrong, what does undoing it cost in 6–12 months? Answer in concrete units: a rename, a migration, a rewrite, a breaking API change for N consumers, a data backfill.

3. TWO-WAY DOOR (cheap to reverse) →

   • Pick now. Prefer the boring option. Write one line saying why. Move on.
   • Explicitly banned: further research, option matrices, asking for approval.

4. ONE-WAY DOOR (expensive/impossible to reverse) → answer all five, in writing:

   • Lock-in: exactly what becomes unchangeable (formats, contracts, published IDs, wire protocols, stored data shapes)?
   • Who pays: which future person/system eats the cost if wrong, and when?
   • Escape hatch: what would migration off this cost? Can a seam be added NOW to make it cheaper (adapter layer, versioned field, opaque ID)?
   • Shrink the door: can any part be deferred or made reversible? The best one-way-door answer is often "make it a two-way door first."
   • Kill-gate: what measurable signal, by when, tells us we chose wrong? (Hand this to /gate.)

5. Verdict line (goes in the commit / PR / decision log): DOOR: one-way | locked: <what> | escape: <cost> | gate: <signal> or DOOR: two-way | picked <X> because <one clause>.

Disagreement check (one-way verdicts)

For ONE-WAY doors, get a second, independent answer before committing: dispatch the same decision (options + lock-in facts, no shared reasoning) to a different model if one is available, otherwise an independent subagent with no shared context — and diff the two verdicts. Agreement → proceed. Disagreement is a mechanical proxy for "this is hard": queue it via /escalate instead of deciding inline.

Rules

• If you can't name the lock-in concretely, you don't understand the decision yet — that's the finding; go read the code/spec first.
• Deciding fast on two-way doors is not recklessness, it's the point.
• A one-way door decided under time pressure must at minimum get the escape-hatch seam, even if analysis is cut short.

Worked example (real ruling, 2026-07-09 — reproduced as a static transcript)

Decision: merge a PR wholesale vs cherry-pick only the target commits. The branch carried unrelated feature work alongside the target changes.

1. Options: merge whole (one revertable merge commit) vs cherry-pick (clean history, but hand-splitting 13 commits invites conflicts and silent loss).
2. Sort question: undoing a bad merge costs one git revert -m 1 — cheap.
3. Lock-in enumerated: no schemas, no published IDs, no API contracts touched — nothing permanent.
4. The stowaway work wasn't taken on faith: its commit trail (13 deliberate commits, including its own fix rounds) showed it was built and corrected, not dumped.
5. Merged whole, no option matrix, moved on.

DOOR: two-way | picked wholesale merge because a merge commit is one revert away and the stowaway work's own commit trail verified it

Composes with

• /gate — step 4's kill-gate is authored there.
• /premortem — run it on any one-way door before committing.
• spec-citation hook — one-way-door paths (schema/, migrations/, contracts) are exactly what you list in .claude/judgment.json protected globs.
• /precedent — run it first; a banked ruling can settle the door outright.
• /escalate — one-way doors with no precedent, or a failed disagreement check, queue there instead of being decided inline.
• /think — one-way doors deserve a second-order pass before committing.
• /verdict — every door ruling is logged there.