Paperclip
Paperclip proposes slowing successful monitors while work waits for review
A Paperclip proposal makes monitor slowdown an explicit policy, combining successful checks and deployment confirmation with an optional review-age threshold. The inspected helper requires all configured gates; it does not support the body’s broader “any two” wording or establish the claimed savings as a measured result.
Paperclip opened a September 13 proposal to stop long-held review work from waking an agent at the same pace indefinitely. The author describes repeated acknowledgements from a monitor while its issue waits in review. Rather than asking the model to rediscover a slower schedule each time, the patch introduces policy and persisted state for a control-plane decision.
The source body is not internally consistent about the condition for slowing down. One test-summary paragraph describes combinations as any two of three gates, while the actual helper requires a configured green-check threshold, a positive slower cadence and confirmed deployment. If the operator also sets a review-tenure threshold, a valid timestamp old enough to meet it becomes another required condition. The source diff, not the looser prose, is the basis for this brief.
The facts
- The helper returns no slowdown when policy, state, the green threshold or a positive slowdown cadence is missing. - Deployment must be confirmed and consecutive successful checks must meet the configured threshold. - When slowdownAfterInReviewSeconds is configured, inReviewSinceAt must parse and its age must meet the threshold. - The patch adds the review anchor to shared validators and preserves it during execution-policy normalization. - The next-check helper uses the slower cadence only when the resolver returns one; otherwise it uses the current cadence. - The author reports 20 unit tests, but the claimed daily wake cost is not accompanied by independently inspected billing evidence.
Why it matters
The operator consequence is straightforward: waiting for a human should not automatically mean repeatedly buying another model turn. Paperclip is putting that pacing decision into the issue policy where it can be inspected and retained across writes. The important limitation is equally concrete. This is conditional slowdown, not a timer that falls quiet merely because an issue has spent an hour in review. The current code leaves existing monitors unchanged unless their policy and reported state satisfy the conditions.
Current
Inspected on 2026-09-14. The Paperclip stable-channel baseline is v2026.831.1 published 2026-09-02T04:56:26Z. The primary source was open (not merged) when captured. Mainline merges, open proposals and packaged releases are distinct availability states; the release baseline does not establish that a proposal has shipped.
Evidence
Primary evidence: paperclipai/paperclip PR #13368 (https://github.com/paperclipai/paperclip/pull/13368). Supporting context: Paperclip PR #13368 — inspected implementation diff (https://api.github.com/repos/paperclipai/paperclip/pulls/13368/files); Paperclip — official product README (https://raw.githubusercontent.com/paperclipai/paperclip/master/README.md). Public source bodies and linked documentation were inspected. Contributor test results remain attributed reports, not independently reproduced experiments.
Operator take
Measure actual monitor invocations and useful changes before setting a slower interval. Preserve a fast initial window where deployment checks still matter, then test missing anchors, invalid timestamps and a failed sample. Verify the heartbeat caller actually consumes the returned cadence on the deployed revision. Do not copy the proposal’s dollar estimate into a budget forecast: no workload mix, provider bill or before-and-after production measurement was validated here.
Caveat
The PR remained open. The author’s body and helper disagree about gate combinations, and this article follows the inspected helper rather than presenting the prose as settled behavior. Local unit-test claims do not prove production savings or complete heartbeat integration. This is an operator-policy proposal, not a release notice or a guarantee that waiting agents will become free.
The PR remained open. The author’s body and helper disagree about gate combinations, and this article follows the inspected helper rather than presenting the prose as settled behavior. Local unit-test claims do not prove production savings or complete heartbeat integration. This is an operator-policy proposal, not a release notice or a guarantee that waiting agents will become free.