---

type: adr id: adr-0084 status: accepted created: 2026-06-16 updated: 2026-06-16

ADR-0084 — Boundary-safe prepare verbs; narrow the board-mutating close to scaffold-only

Context

The W1 surface map (corpus-works AUDIT-corpus-cli-surface) passed the whole candidate corpus-cli surface through the reconcile-only boundary (ADR-0077) and found exactly one boundary-eroding idea: a status.md-mutating close (a corpus finish/corpus close that flips a board row to a pass-like state). It crosses ADR-0077 Decision 8 — the human owns the review result (Pass / Fail / Unverified / Blocked); the CLI routes, it never adjudicates — because writing a board row to a pass-like state is adjudicating that human-owned verdict. ADR-0077 Decision 1 enumerates a "gated close" as a reconcile capability; that phrasing is precisely what the open DECIDE #1.2 contests, and it is the clause this ADR narrows.

Two loop verbs remain unbuilt and are boundary-safe (prepare engine, ADR-0077 D1): corpus pull (snapshot a ticket into intake/) and a finding scaffold (offered at Close — scaffold a finding file from a finished task). They prepare a new file; they read or derive no board state.

Decision

1. corpus pull <ref> writes exactly one verbatim intake snapshot (intake/<slug>.md with source/url/captured), and never a spec — normalizing a ticket into requirements is judgment work, not transcription. Prepare engine (D1); level: toolable.
2. The finding scaffold (corpus promote) writes one candidate finding file (findings/<slug>.md from the template, pre-filling from:), and never asserts a learning and never writes the board. It is offered at Close; the human fills and accepts it. Prepare engine (D1); level: toolable. This reverses future-cli.md's "promote folds into close, not its own engine" — the W1 map re-derived it as its own boundary-safe scaffold verb; recorded here.
3. The board-mutating close stays PARKED (DECIDE #1.2). No corpus-cli command writes status.md; the board is hand-edited. A future corpus close/corpus finish that mutates the board is a reversal of this clause and needs its own ADR — it is not shipped by drift. This narrows ADR-0077 D1's enumerated "gated close" to scaffold-only. Level: convention (the hard line), an invariant a boundary regression test asserts (no Core use-case writes status.md). Reopen trigger: write that ADR only on demonstrated adopter demand for the review-gate teeth — generation-volume / convenience demand does not qualify (refuted; corpus-works #1.2, and the #9/#11 measure-first verdict). This trigger is the durable home for the #1.2 decision; the GitHub feedback issue need not stay open as a separate tracker.

No checks.yaml rule, no contract-version bump — the prepare verbs touch no check; the drift-guard is untouched.

Alternatives considered

Alternative	Why weaker
Ship a board-mutating corpus close/finish now	Crosses ADR-0077 D8 (the CLI would adjudicate the human-owned result by writing a pass-like board row); it is the open DECIDE #1.2, deliberately parked, not shipped by convenience. Rejected.
Fold promote into close (not its own verb) — future-cli.md's prior framing	The board-mutating close is parked, so there is no close engine to fold into; the W1 map re-derived the finding scaffold as a standalone boundary-safe prepare verb. Rejected in favor of promote.
corpus pull auto-drafts a spec from the ticket	Normalizing a ticket into requirements is judgment, not transcription; auto-drafting smuggles authoring into a prepare verb. Rejected — pull writes the snapshot only.
Do nothing (leave the loop verbs unbuilt)	The loop verbs are the validated, boundary-safe widening of the CLI (gh #10); leaving them out is not a boundary win, just an incomplete loop. Rejected.

Consequences

Accepted. corpus pull and corpus promote ship as prepare verbs that scaffold one new file each and mutate no board; the board-mutating close is parked behind a future ADR; a boundary regression test (no Core use-case writes status.md) makes the no-mutation property an invariant, not a convention. Refines ADR-0077 (narrows the "gated close" clause wherever it appears — D1's reconcile list and D5's Supercharge list — to scaffold-only); does not re-open D8 (it reaffirms it). future-cli.md's pull/close/promote rows converge accordingly (the CONT rider). The implementation is cut and reviewed from corpus-works (SPEC-corpus-cli-m3-prepare). Negative: two more commands to maintain. Neutral: a team that wants board mutation must wait for (or propose) the parked-close ADR.