← Home

The closing page — the loop the approval used to leave open

An approval used to be the end of the story. The agent got its green light and vanished: nobody knew what it did with it, and the human who approved a four-line summary had no way to check the outcome. "I approved a push" and "a push happened" are two facts, and Beeper only knew the first.

A closure is the other half of the loop. The agent proves what it did, the pager of the person who decided learns the result, and the proof is journalled against the page it closes.

agent ──▶ POST /v1/pages  ──▶ Central ──▶ pager:  310 a PUSH main>origin #07c2
                                              │
                                         human decides (signed)
                                              │
        ◀────────────────────── decision ─────┘
        │
        └──▶ does the work
              │
              ▼
        POST /v1/closures { page_id, summary, evidence }
              │
              ▼
        Central: page decided? approved? not closed yet? ──▶ 810 a DONE commit 1a2b3c4 #9f2e
                                              │
                                              ▼
                                        the SAME pager rings again

What is composed, and what is quoted

The honesty rule does not pause for receipts. The line of a closing page is composed from the evidence — a commit sha, a PR number, a file list — and the agent's own words land on the last detail screen, behind AGENT SAYS::

810 a DONE commit 1a2b3c4 PR #87 #9f2e
DIFF: …
AGENT SAYS: trois commits poussés, CI verte, rien d'autre touché

You decide on the fact, and you read the narrative as the narrative. An agent that sends summary without evidence gets a receipt that says closed #07c2 — visibly thin, and that is the point: a closure you cannot verify is a closure that looks like one.

Three deliberate differences

an approval page a closing page
asks a decision nothing
options / callback seeded replies, callback_url none, none
lifetime the rules' TTL one hour, then it expires inert
attention budget charged (green/amber) never charged
who sees it the rota's on-call the pager that decided, alone

The exemption from the budget is what makes the loop close: if a receipt were charged, a busy hour would hold the proofs behind the questions — the one thing an operator wants to read would be the last thing to arrive.

Because family 8 is exempt, a source cannot claim it: POST /v1/pages with code: 8xx is refused with reserved_code_family. Only the closure path writes an 8xx page, and only against a page the same station paged. Without that lock, declaring a code would be an unlimited pass on your attention.

The four refusals

A closure is a claim, and a claim can be wrong. Each case is a different answer, and each is journalled as closure.rejected:

Case Status Why it is not a silent no-op
page never decided 409 not_decided an agent cannot close a page it only asked for
page denied 409 denied_page a refusal is not an authorisation to proceed
page of another station 404 unknown_page you cannot close what you cannot see
page already closed 200, the first receipt a retry is not a second closure

The last row is the important one: closures.page_id is the primary key, so one page has one closure at the database level, not by convention. A retried request lands on the receipt that already exists and never overwrites it — an agent cannot rewrite its own proof by retrying harder.

On the wire

# The agent closes what it was approved to do.
curl -s -X POST -H "authorization: Bearer $STATION" https://…/v1/closures \
  -d '{"page_id":"pg_…","summary":"trois commits poussés","evidence":"commit 1a2b3c4 PR #87"}'
# → {"page_id":…,"closure_page_id":"pg_…","origin_fingerprint":"07c2…",
#    "closed_at":1763121600000,"page":{… the 810 line …}}

# The human reads the receipt of a page they approved.
curl -s -H "authorization: Bearer $STATION" https://…/v1/pages/pg_…/closure

origin_fingerprint ties the receipt to the exact payload that was displayed: the closure's own fingerprint covers the page id and the origin fingerprint, so a receipt cannot be detached from what it closes and re-attached somewhere else.

What it is not

Not an execution log, not a stdout. The evidence is one verifiable handle — something you can go look at — not a paste of 400 lines. Not an attestation either: the Central records what the station says it did; proving what happened would need the runner to sign its own output, which is a different feature on a different trust root. What it does is make silence impossible: after this, an approved page that never reports back is visibly, audibly, permanently unanswered.