On a new forge a finished change goes live by itself: today it waits forever… #11668

A green agent pull in an app whose owner hasn't turned reviews on now lands by itself. That fixes the review wait only; four items in the issue are not done and the production probe has not run, so this should not close #11666.

I did not verify the cause on a probe forge. By the code, the crew's reviewer never needed to run: reviewed() held every agent pull until someone approved it, and nothing sent a review on a forge with no reviewer seated. An app's existing "Require an approving review" setting (repo.requireReview) defaults to off, so I used it as the "owner asked for reviews" switch.

What changed

• src/crew/autonomy.ts: reviewed() passes a green pull when requireReview is off. A verdict already on the head, including "changes requested", is still honoured first.
• src/crew/review.ts:

- sendToReview sends nothing when reviews are off, so no reviewer run costs credit. A person's explicit review request still goes through. - reviewWaitsFor no longer reports such a pull as waiting for a reviewer or a person's review.

• src/flow/next_actor.ts: an app with reviews off no longer reads "Next: reviewer, no review yet". It reads "green, landing by itself", or the check's own state while the check runs.
• Tests: a nextActor case, and a case in test/crew-now-review-waits.test.ts that fails without the change. I set that file's repo to requireReview on, because its two existing tests assumed the old gate.

Checks

• tsc --noEmit is clean and scripts/gate.ts printed nothing.
• I ran only the review, autonomy and landing tests, 49 in all. They passed, apart from the two tests above that needed requireReview on.
• scripts/fast.ts named 2021 test files, which is too broad, so I did not run them. The forge's check runs the full suite on the push.

Not done

• The "Your change is ready. Ship it" line on home and the app page, and a preview link on the pull.
• The Look panel's ERR_INTERNET_DISCONNECTED for a child forge's own app (#11651).
• One cost figure everywhere ($0.22 on the pull, $0.45 on the app page, "$2 of $2" on home).
• The production probe and its timeline.

I also did not add a per-app settings control for reviews, because the existing setting already does that.

To try it Leave requireReview off on an autonomous app. Once an agent pull's check passes, it lands without a review and shows no "waiting for reviewer" state. Turning the setting on restores the old gate.

🤖 Generated with Claude Code

Closes #11666

A green agent pull in an app whose owner hasn't turned reviews on now lands by itself. That fixes the review wait only; four items in the issue are not done and the production probe has not run, so this should not close #11666.

I did not verify the cause on a probe forge. By the code, the crew's reviewer never needed to run: reviewed() held every agent pull until someone approved it, and nothing sent a review on a forge with no reviewer seated. An app's existing "Require an approving review" setting (repo.requireReview) defaults to off, so I used it as the "owner asked for reviews" switch.

What changed

• src/crew/autonomy.ts: reviewed() passes a green pull when requireReview is off. A verdict already on the head, including "changes requested", is still honoured first.
• src/crew/review.ts:

- sendToReview sends nothing when reviews are off, so no reviewer run costs credit. A person's explicit review request still goes through. - reviewWaitsFor no longer reports such a pull as waiting for a reviewer or a person's review.

• src/flow/next_actor.ts: an app with reviews off no longer reads "Next: reviewer, no review yet". It reads "green, landing by itself", or the check's own state while the check runs.
• Tests: a nextActor case, and a case in test/crew-now-review-waits.test.ts that fails without the change. I set that file's repo to requireReview on, because its two existing tests assumed the old gate.

Checks

• tsc --noEmit is clean and scripts/gate.ts printed nothing.
• I ran only the review, autonomy and landing tests, 49 in all. They passed, apart from the two tests above that needed requireReview on.
• scripts/fast.ts named 2021 test files, which is too broad, so I did not run them. The forge's check runs the full suite on the push.

Not done

• The "Your change is ready. Ship it" line on home and the app page, and a preview link on the pull.
• The Look panel's ERR_INTERNET_DISCONNECTED for a child forge's own app (#11651).
• One cost figure everywhere ($0.22 on the pull, $0.45 on the app page, "$2 of $2" on home).
• The production probe and its timeline.

I also did not add a per-app settings control for reviews, because the existing setting already does that.

To try it Leave requireReview off on an autonomous app. Once an agent pull's check passes, it lands without a review and shows no "waiting for reviewer" state. Turning the setting on restores the old gate.

🤖 Generated with Claude Code

The earlier version of this change broke tests that expect the crew to review a pull. It treated "reviews off" as "never review", even where a reviewer is seated on the crew. The pushed fix keeps reviews running wherever a reviewer is seated.

I did not run the full suite or the forge's check here. I re-ran the suites behind the failures in the check log and they pass. bun x tsc --noEmit is clean. scripts/gate.ts printed nothing.

What changed

• Rule: an app whose owner has not turned on requireReview and that has no reviewer seated now lands a green pull on its own. No review is sent, nothing waits for one, and the pull's "next" line reads "queue" instead of "reviewer".
• Where: the gate in src/crew/autonomy.ts, reviewWaitsFor in src/crew/review.ts, and src/flow/next_actor.ts.
• Reverted: the sendToReview change from the first commit. A seated reviewer still gets a review sent whatever requireReview says.
• Pull page: with no reviewer and no review required, the page now says "no check yet" in place of "no review yet", so its first-view word budget still holds.

Tests changed

• Several tests described "an agent's pull waiting for a review" on a repo with the default setting (reviews off, no reviewer). That is the case this issue changes. I set require_review = 1 in those tests so they still check what they were written to check. They are in asks-counts, follow-sources, mcp, mcp-flow-why, path-steps-batch, pull-next-line, pull-next-tui, review-waits, review-gate and seat-slots.
• In review-waits, a person's own pull is now landed with reviews off, since a person's pull with reviews on needs an approval.
• Two tests that failed in the check passed when run alone: mcp-system-routes and the first-view budget test (attention). The terminal pull-screen test failed alone only because it expected the old review wording, and it now passes. I take the first two for load flakes and have not checked them under the full run.

Not done The other items in the issue are untouched: the Ship it line with preview link, the ERR_INTERNET_DISCONNECTED look panel, one cost figure everywhere, and finding why the reviewer never ran on a new forge. The production probe the issue asks for has not been done either. This pull is only the lands-by-itself rule, so it should not close the issue.

To try it Open an agent pull on an app with reviews off and no reviewer seated. Once its check is green it lands, and the pull page says it is landing by itself.

🤖 Generated with Claude Code

Seven tests failed in the check; I fixed the cause of five of them and pushed, and the other two were order-dependent. After the fix, mcp-pulls-pile, pipeline-path, crew-red-outside-diff, pages-budget and mcp-system-routes pass when run on their own. I did not re-run the full suite here, and red-main and crew-red-main were only checked in the runs below. The gate and tsc --noEmit are clean.

What changed

• Cause: earlier commits on this pull make a green pull land without a review when the owner has not asked for reviews and no reviewer is seated.
• Tests fixed: three tests assumed a review would wait. They now set require_review = 1 on their repository, as the previous commit did for its tests:

- test/mcp-pulls-pile.test.ts - test/pipeline-path.test.ts - test/crew-red-outside-diff.test.ts

• Pull page budget: the pull page does two more SQL statements to learn whether reviews are asked for. I raised its budget in test/pages-budget.test.ts from 110 to 112, with a comment. I did not cut the two queries.
• Order-dependent tests: the red-main lane-sentence failures and the mcp-system-routes failure did not reproduce on their own. red-main passed when run alone, and mcp-system-routes passed after the budget bump. Both looked like interference between test files sharing process state, so I left them alone.

No application code changed in this commit.

To try it: bun test test/mcp-pulls-pile.test.ts test/pipeline-path.test.ts test/crew-red-outside-diff.test.ts test/pages-budget.test.ts

This pull makes a green pull in an app with reviews off land by itself (#11666), and it changes the default for every app. I ran the tests this change touches and a sweep of the pull-, review- and path-related files. That sweep had three failures, and I fixed them below. I did not run the whole suite; the forge's check does that.

What it does. When an app's owner has not turned on requireReview (the default) and no reviewer is seated, a green pull lands with no review sent and nothing waiting for one. A conflict, a red check, a hold, the access wall and a "changes requested" verdict on the head still stop it. The new default is spelled out in docs/features/reviews-off-lands-on-green.md, with how to turn reviews on. Nothing was added to the README.

Which pipeline item this serves. None of #6534, #6611, #5276 or #3316. It serves #11666, which says a new forge's first change should go live by itself, and that issue names no pipeline item. If you want it held until it serves one, that is your call.

Addressing the review

• One rule. reviewsOff(repo, count) and reviewIsOff(db, repo) in src/crew/review.ts replace the three spellings. The gate (reviewed), the waits (reviewWaitsFor) and the next-actor line all use it.
• A drift bug the single rule exposed. stepsOfStored, the batch of pull steps, never asked the rule, so it said "No review yet" where stepOfStored said "landing by itself". It asks now, once per batch.
• Cost.

- nextOfPull takes reviewOff from its callers, so it no longer calls getRepoById. - I did not get the page's extra reviewer read down to zero. Whether a reviewer is seated is not on the pull page today, and seats are not a column on the repo, so it is one statement, skipped when reviews are on. - Because of that one statement, the pull page budget is 111 (it was 110, and 112 in my first push) and the batch cap in path-steps-batch is 6 (it was 5). Both are constant per request, not per row.

• Tests.

- next_actor.test.ts folds the reviews-off cases into the existing table, with no aliased imports. - review-gate and review-waits use afterEach helpers instead of setting a value and flipping it back by hand. - crew-now-review-waits uses a separate reviews-off repo (meadow) instead of a copied repo object. - The mass require_review = 1 edits in the other test files stay, one line each. They state that those tests assume reviews are on.

• New tests.

- A changes-requested verdict still holds a green agent pull with reviews off. - The terminal's pull screen shows "next: queue, green, landing by itself" with reviews off; the existing pull-next-tui case covers reviews on.

• Preview line. The gate asked for a Try: line in a commit message, so the last commit has one.

Not done. I did not exercise the preview.

Try it. Open a green agent pull of an app with reviews off. Its page, the terminal's pull screen and the MCP pull tool all read "Next: queue, green, landing by itself".

🤖 Generated with Claude Code

An app whose owner hasn't turned on reviews, and that has no reviewer seated, now lands a green pull by itself. This round answers the review on that; the policy itself is unchanged.

Changed

• One function (3). reviewsOff(db, repo, seated?) in src/crew/review.ts is now the only rule; reviewIsOff is gone. The gate reviewed reads the seated reviewers once and passes them in. On main it read them up to twice. reviewWaitsFor passes in the map it already reads.
• Read only when needed (4). nextAsking(input, off) in src/flow/next_actor.ts checks off() only when the answer would otherwise be "waiting for the reviewer". That is the one answer reviews being off can change. nextOfPull now takes the repo, and stepsOfStored asks at most once per batch. The batch cap in test/path-steps-batch.test.ts is back to 5.
• Imports (5). The imports in path.ts and why.ts now sit with the others. change.ts, mcp/tools/pull.ts and the TUI pull screen no longer import from crew/review.ts.
• Tests (6). askForReviews(db, repoId, on?) in test/helpers/reviews.ts replaces the scattered UPDATE repo SET require_review lines. New tests in test/next_actor.test.ts show off() is not called unless the pull would wait on the reviewer.
• Wording. The next-actor line now reads "green, reviews are off: it lands", which gives the reason. docs/features/reviews-off-lands-on-green.md now says how to turn reviews on: the Require review setting, or seating a reviewer.

Where I disagree

• 1 and 2 (an opt-in setting; direction). The owner wrote the policy in #11666: "Reviews are something an owner turns on, per app, when they want a second look; a new forge's first app has it off", and "a pull whose check passed lands by itself". A separate landOnGreen setting would reverse that. It would also need a migration and a settings control the issue doesn't ask for. The pull already narrows the change: a seated reviewer keeps reviews on, and a conflict, red check, hold, the access wall or "changes requested" still stops a pull. If the owner wants opt-in instead, that's a decision for the owner to make in the issue.
• 4 (pull page budget). The page budget stays at 111, up from 110. I traced every crew-seat read during the page walk, and the pull page makes none besides this one, so there is no loaded data to reuse. The read happens only when Require review is off and the pull would otherwise wait on a reviewer.

Checks

• bun x tsc --noEmit, bun scripts/gate.ts and bun scripts/design-lint.ts pass.
• I ran the 29 test files that cover this rule: 382 pass.
• test/guest-contribute.test.ts has one failure here, and the same failure on main in this sandbox: a git push over loopback exits 128. The forge's check runs it.
• I pushed the first commit of this round with pages-budget failing; the second commit fixes it.

Not done: 7 (preview). The preview is up and / returns 200, but it has no repositories, crew or check runner. I could not run an agent's green pull through it from the sandbox, so the preview was not exercised end to end.

To try it: in an app with Require review off and no reviewer seated, a green agent pull's page reads "Next: queue, green, reviews are off: it lands" and lands. Turn Require review on and it waits on a review again.

🤖 Generated with Claude Code