A gate catches the intent whose work is live while its spec is still open

Summary

A change that delivers an intent says so in its commit trailer, and a merge-time gate refuses the change unless that intent reaches shipped/ in the same diff.

This is the symmetric twin of RS001. The issue ledger has been held honest by Resolves: iss-N plus a gate that refuses a trailer whose record does not move; the intent store has the same hole and no such gate, and 61 planned intents with open specs are what came of it.

Nothing infers that work is live. The author declares it, and the gate enforces the declaration — which is the whole reason this is buildable, since specs carry no task state and intents carry no delivering-commit stamp.

Scope

In. One trailer, one rule, and the refusal messages that make it actionable.

Out. intent ship and the intent supersede verb, held by the 2026-09-11 ruling until this gate has run a cycle. Clearing the standing 61, which this gate does not reach. Any change to the closed-spec half (staleIntents), which stays at release time and keeps working as it does.

Approach

The trailer

Delivers: itd-N

Matched exactly as RS001 matches its own, with the id pattern widened to accept both id shapes the store now holds:

DELIVERS_RE='^Delivers:[[:space:]]+(itd-[0-9]+)[[:space:]]*$'

itd-[0-9]+ already covers the sequential ids and the minted timestamp ids alike, so no second pattern is needed.

The verb is Delivers rather than Resolves because the acts differ: an issue is resolved by the change that fixes it, an intent is delivered by the change that makes its acceptance criteria true. Reusing Resolves for both would make the trailer ambiguous about which store to look in.

The rule

A sibling of scripts/check-issue-resolution.sh, sharing its structure so the two read as one convention:

1. Compute the set of intent ids entering shipped/ between base and head — the mirror of ids_entering_closed, over .abcd/development/intents/shipped/.
2. Walk every commit in the range and collect Delivers: trailers.
3. A declared id that is not in the entering set is a refusal, and the diagnosis is what the message carries.

The refusal shapes

RS001's lesson, learnt the hard way on a branch 235 commits behind main, is that one refusal text for several causes is a refusal nobody can act on. The same diagnoses apply here and are stated separately:

The last row matters because 20 planned intents carry spec_id: null and no verb mints one for them. A refusal that told those authors to run spec close would name a command with no argument.

Where it runs

Beside the issue gate, at merge time, in the record-lint job that already runs check-issue-resolution.sh. Not at the cut: the trailer is written by the author of the change, so the refusal belongs where that author is still holding it. This is scope condition 2 made concrete.

How this satisfies each acceptance criterion

1. A declared delivery must move the record. Step 3 above: a Delivers: id absent from the entering-shipped/ set is a refusal, and the planned/+open spec row names abcd spec close <spc-N>.
2. Every unshipped one is named. The walk collects all trailers before judging and emits one refusal per offending id, as RS001 does, rather than exiting on the first.
3. A change declaring nothing is refused nothing. The rule is keyed on the trailer: no trailer, no assertion to check. This is the gate's stated limit, and the criterion pins it so a later change cannot quietly widen the rule into refusing every planned intent.
4. A bad id says which. Rows 1-3 of the table: absent at head, absent everywhere, and already terminal are three distinct messages.
5. Same message shape and exit code as the issue rule. The rule reuses the issue gate's fail helper and its exit convention, and a test asserts the two refusals share their shape rather than a reviewer judging they feel alike.
6. The 61 are not refused. They carry no trailer, so nothing fires on them. The ## Scope section above states that clearing them is a separate act, so the criterion is met by what the gate does not do, and by the record saying so.

Risks

The gate is keyed on speech, and the measured problem is silence. Recorded on the intent as the mechanism's own falsifier: authors omitting the trailer to avoid the refusal would show the cause was reluctance rather than forgetting, and a voluntary declaration cannot reach someone who declines to declare. The cycle this gate runs before the verbs are reconsidered is what tests it.

Partial delivery. Scope condition 3: where an intent arrives across several changes, no single one finishes it. The trailer must therefore mean "this finishes it" rather than "this contributes to it", and the wording in CONTRIBUTING.md has to say so, or the gate will fire on the first of five pull requests and teach authors to stop writing it.

Open

Both questions were settled by the implementing session of 2026-09-23, recorded here because the record held no ruling on either:

• The rule's identifier is RS005. It joins the RS numbering rather than starting a series, because it lives in scripts/check-issue-resolution.sh, reuses that script's fail helper, exit convention and CI step, and is judged in the same pass over the same trailer lines as RS001. A sibling series would have claimed a second convention where the whole point is one.
• The trailer takes the intent id only. Delivers: spc-N is refused, not ignored: a line that reads as an attempt at the trailer and is not exactly Delivers: itd-N[, itd-M…] is an RS005 refusal that names the spelling. The intent is the thing delivered; a trailer that could name either store would be ambiguous about which to read, and a silently skipped spelling is a declaration its author believes armed. A line reads as an attempt when it begins with the word, in any case, followed by a colon, AND its value carries an id-shaped token ([A-Za-z]+-[0-9]+): Delivers: spc-7 and delivers: itd-7 are refused, while Delivers: the thing and a wrapped body line beginning deliver: that … are prose and pass. The acceptance criterion that a change declaring no delivery is refused nothing fixes that boundary; a rule refusing every line that begins with the word refused ordinary prose (main's own 182474f5 carries such a line).

Two consequences of the 1:n intent–spec rule (adr-2609151513118583) shaped the ordinary-case refusal: it names every spec still in open/ whose own intent: back-link names the intent, compared canonically (itd-007 is itd-7), rather than the intent's scalar spec_id; and a planned intent whose specs are all closed is refused with the stale-intent diagnosis the release cut uses. The standing backlog of planned intents with open specs is handed to iss-2608290808193471, which records that shape and against which the batch clearance of 2026-09-23 is filed; this gate does not reach it.