one-way-door

name	one-way-door
description	Classify a pending decision as reversible or Triggers: "is this a one-way door", "can we undo this", "should I just decide this", "the models disagree with me", before any auto-decided approval gate.

$one-way-door

Classify one pending decision as two-way (cheap to undo, may be auto-decided) or one-way (expensive or impossible to undo, always surfaced to the caller). Return the classification and stop. This skill decides who decides. It never decides the underlying question and it never performs the action.

Insight: an agent's confidence is uncorrelated with a decision's reversibility, so confidence is the wrong gate. Reversibility is a property of the decision itself, is knowable before acting, and can be declared in advance instead of inferred from how the question happened to be worded.

The failure mode this exists to prevent: an agent auto-approves an irreversible step because the surrounding batch of decisions was routine and the step read as one more item in the list. Dropped table, force-push over a colleague's work, rotated credential, published post. The batch is why it happens: nineteen two-way decisions train the reflex that answers the twentieth.

dcg guards destructive commands at the shell boundary. This guards destructive decisions at the judgment boundary, before any command exists.

Modes

Trigger phrases	Mode	Entry point
"is this a one-way door", "can we undo this", "classify this decision"	classify	the four layers below
"should I just decide this", "auto-decide the rest"	gate a batch	classify each, act only on two-way
"the models disagree with me", "both reviewers say I'm wrong"	user challenge	the packet below

Inputs

Required: a one-line summary of the pending decision, and the effect it would have if answered wrong.

Optional: a declared decision id from references/decision-registry.md, and the skill or lane raising it.

Non-goals. This skill does not answer the decision, rank options, execute anything, or record the caller's answer. It does not gate shell commands — that is dcg. It does not model caller preferences or build a profile of how the caller usually answers; a preference that suppresses questions is exactly the mechanism that must never reach a one-way door.

Procedure

Apply the four layers in order and stop at the first that fires.

1. Declared registry (primary). Look the decision id up in references/decision-registry.md. If present, use its declared door. A declaration in a checked-in file outranks anything inferred from prose, because prose gets reworded and a reworded question must not silently change class.
2. Effect class. If the decision authorizes an effect the skill catalog declares as mutating outside the working tree — publishing, credential use, external mutation, remote deletion, release — it is one-way regardless of wording.
3. Pattern fallback. For decisions with no registry entry, match the summary against references/patterns.md. This layer is deliberately over-eager: a false positive costs the caller one extra question; a false negative costs them the thing that cannot be undone.
4. Default two-way. No evidence of irreversibility in any layer means the decision may be auto-decided.

Then report: door, the layer that fired, and — for one-way — the specific undo cost that makes it one-way.

Layers 1 and 3 read checked-in files, so the lookup is reproducible by anyone reviewing the classification:

# Layer 1 — is this decision already declared?
grep -niF "deploy key" skills/one-way-door/references/decision-registry.md

# Layer 3 — which pattern would fire on this wording?
grep -niE "revoke|rotate|force-push|publish|delete" skills/one-way-door/references/patterns.md

The user challenge packet

One decision class is never auto-decided at any confidence: the reviewers agree the caller's stated direction is wrong. Independent agreement is a strong signal. It is not authority — the caller holds context the reviewers do not, and a synthesis that quietly adopts the reviewers' position deletes that context without telling anyone.

When it fires, emit exactly these five fields and stop:

• What the caller stated: their original direction, in their words.
• What the reviewers recommend: the change, and how many independently reached it.
• Why: the reasoning, at its strongest.
• What we might be missing: the context the reviewers provably did not have.
• If we are wrong, the cost is: what breaks if the caller's direction was right.

The caller's direction is the default and stays the default. The reviewers carry the burden of argument. The one adjustment: if the reviewers classify the change as a security or feasibility defect rather than a preference, say so in the packet — the caller still decides, but they decide knowing which kind of disagreement this is.

Anti-patterns

Anti-pattern	Corrective
Reading the summary's wording as the primary signal	Registry first; wording is layer 3, and only for undeclared decisions
"I'm confident, so I'll decide it"	Confidence does not make a decision reversible; classify, then route
Auto-deciding the whole batch because the first nineteen were routine	Classify every item independently; batches are the failure mode, not the exception
Suppressing a repeated question because the caller answered it the same way twice	Preferences apply to two-way only; a one-way door has no suppression path
Synthesizing the reviewers' position into the plan when they agree against the caller	Emit the user challenge packet; never fold it into a recommendation

Output

One classification per decision:

{
  "decision": "rotate the deploy key before shipping",
  "door": "one-way",
  "layer": "pattern",
  "undo_cost": "revoked key cannot be un-revoked; every consumer must be re-issued",
  "auto_decidable": false
}

Done when: every decision in the batch carries a door and a layer, and no decision with door: one-way was answered by the agent. A batch is correctly gated when every one-way line went back to the caller unanswered.

It's working if

Observable in the trace, without reading the prose — and the rubric a fresh independent judge scores this skill against:

• Every item in a batch carries its own door and layer; none inherits the class of the item before it.
• No item classified one-way was answered by the agent — each one went back to the caller unanswered.
• Every one-way line names a concrete undo cost, not a category label.
• The registry lookup appears in the trace before the pattern match, not after.

Checks

• Every one-way classification names a concrete undo cost, not a category label.
• No decision was both classified one-way and auto-decided.
• The registry was consulted before the pattern layer, not after.
• The user challenge packet, when emitted, contains all five fields including "what we might be missing" — the field that is easiest to drop and the one that makes the packet honest.

Provenance

• Mechanism analysis and the layering rationale (§2.2, §2.3): not on main; read it at git show 9872483bd:docs/research/gstack-teardown-2026-08-08.md (branch recover/gstack-clean-room).
• Shell-boundary sibling: skills/dcg/SKILL.md.
• Caller-sovereignty floor: CLAUDE.md → Authority and trust; the operating contract's rule that repository access does not authorize destructive operations, publishing, credential use, or external mutation.

Failure behavior

If the decision summary is missing or the registry is unreadable, classify as one-way and report why. Failing toward asking is the only safe direction: an unclassifiable decision is an unknown, and an unknown that gets auto-decided is indistinguishable from an unguarded one.