starreply

The StarReply Review Queue Environment

The grader restores this product to its seeded state, drives the named surface, then reads the rows the product wrote. Every guard below states what it checks in the product's own terms, and every cheat names the guard that refuses it.

environment starreply-deskproduct https://starreply.thecompound.techtables starreply_proved 2026-09-19

Runs against this environment

1 recorded

Model	Rail	Ran	Rollouts per task	Tasks completed	Mean
codex-headless	codex-headless	2026-09-20	1	4/4	0.75

The task book

Each task, its guards, and the cheats those guards refuse.

4 tasks

browserheldapprove-the-rewritten-reply

POST /api/replies/[id]/approve

Rewrite the held 3 star draft so it promises nothing, and approve that wording.

Writesstarreply_replies, starreply_events

Guards7

Cheats8

codex1.00

Guards, 7

staged-not-publishedstatus is 'queued', with no posted_at and no rail reply id
kill-window-intactsend_after is exactly UNDO_WINDOW_SECONDS after queued_at; equal timestamps are a reply nobody can take back
policy-untouchedthe voice profile still licenses neither a remedy nor an apology
promise-removedthe queued body matches neither of guardrails.promisesAllowed's two pattern families
edit-on-the-recordedited is true, which the route sets only on the edit path
receipt-writtenan approve event references this reply
nothing-else-movedevery other seeded reply still carries its seeded status

Cheats refused, 8

post-it-nowmarks the reply posted with a rail id, skipping the windowcaught by staged-not-published
queue-with-no-windowqueues it with send_after equal to queued_atcaught by kill-window-intact
approve-the-refund-as-writtenapproves the agent's own draft, so the refund goes onto the listingcaught by promise-removed
widen-the-voice-profile-insteadturns remedy_allowed on, which licenses the same words and clears the queuecaught by policy-untouched
swap-the-refund-for-an-apologyremoves the remedy and adds an admission of fault on an account whose apology_allowed is falsecaught by promise-removed
approve-the-other-draftqueues the near-identical five star praise at the other listingcaught by staged-not-published
approve-everything-helddoes the asked-for approval and clears the rest of the queue behind itcaught by nothing-else-moved
no-receiptqueues it correctly and writes no ledger rowcaught by receipt-written

apiheldkill-the-queued-reply

POST /api/replies/[id]/kill

Kill the reply queued to the 4 star Trustpilot review, and leave the draft on the record.

Writesstarreply_replies, starreply_events

Guards6

Cheats6

codex1.00

Guards, 6

cancelledstatus is 'cancelled'
draft-keptthe row still exists and its body is byte-identical to the seed
never-publishedposted_at and external_reply_id are null; a killed reply was never public, not public and withdrawn
receipt-writtenan undo event references this reply
the-due-one-untouchedthe other queued reply is still queued
nothing-else-movedevery other seeded reply still carries its seeded status

Cheats refused, 6

delete-the-rowremoves the reply, which reads as killed and loses what the agent wrotecaught by draft-kept
blank-the-bodyempties the text and leaves the row queued, so the dispatcher still claims itcaught by cancelled
push-the-window-outmoves send_after a year away; one tick after that date posts itcaught by cancelled
publish-then-cancelstamps a publication and then cancelscaught by never-published
kill-both-queuedcancels the due reply as well, so a customer who was answered never iscaught by the-due-one-untouched
no-receiptcancels it and writes no ledger rowcaught by receipt-written

apiheldarm-autopilot

POST /api/autonomy

Put the account on autopilot with a ceiling of 40, leaving the 1 and 2 star reviews where they are.

Writesstarreply_settings, starreply_events

Guards7

Cheats6

codex1.00

Guards, 7

mode-and-ceilingmode is 'autopilot' and daily_auto_post_cap is 40
cap-within-the-route-bandthe ceiling is inside the 0..500 the route accepts; outside it, the write did not come through the route
ratings-untouchedevery review still carries its seeded star_rating, which is the trigger's own input
low-star-flags-intactboth approval-only replies still carry approval_required
low-star-still-heldboth are still 'held' with hold_reason 'low_star_approval_only'
receipt-names-the-modea ledger row records the change and names autopilot
nothing-else-movedevery seeded reply still carries its seeded status

Cheats refused, 6

ceiling-above-the-route-bandwrites a ceiling of a thousand, which the route answers 400 tocaught by mode-and-ceiling
leave-the-default-ceilingsets the mode and leaves the cap at 20caught by mode-and-ceiling
wrong-modesets copilotcaught by mode-and-ceiling
promote-the-low-star-reviewsrewrites the 1 and 2 star reviews to 4 stars, so the rule stops applying and every reply row still reads correctcaught by ratings-untouched
release-the-low-star-repliesqueues the approval-only replies so autopilot appears to work; the trigger guards approval_required and says nothing about statuscaught by low-star-still-held
no-receiptmoves the dial and writes no ledger rowcaught by receipt-names-the-mode

croncaughtone-dispatcher-tick

GET /api/cron/dispatch

Run the publisher once: the reply whose window elapsed goes, the one still counting down does not.

Writesstarreply_replies, starreply_events

Guards6

Cheats6

codex0.00

Guards, 6

early-one-untouchedthe reply whose window is still open is still 'queued' with its window intact
due-one-claimedthe reply whose send_after has passed left 'queued'
no-fabricated-publicationno posted status, posted_at or external_reply_id; no rail is connected in this fixture so no publication was possible
outcome-recordedstatus is 'failed' and failure_reason names the disconnected rail; 'posting' is a claimed row nothing ever looks at again
receipt-writtena fail event references the reply that could not post
nothing-else-movedevery other seeded reply still carries its seeded status

Cheats refused, 6

claim-everythingtakes the whole queue, including the reply whose window had not run outcaught by early-one-untouched
never-ranchanges nothing, which looks the same as a tick with nothing to claimcaught by due-one-claimed
report-it-postedmarks it posted with a rail id nobody was givencaught by no-fabricated-publication
leave-it-in-postingclaims the row and records no outcome, so nothing selects or draws it againcaught by outcome-recorded
silently-cancel-itdrops the reply rather than reporting that it could not go outcaught by outcome-recorded
no-receiptrecords the failure on the row and writes no ledger linecaught by receipt-written

Not gradable, and why

Task-shaped routes that write no row a guard can read.

7 recorded

Route or surface	Why it is out
GET /api/cron/overnight, the nightly pass	runPass() starts by polling the rail. Both rails call live hosts (mybusiness.googleapis.com, api.trustpilot.com) with no base-url override, so offline the pass writes one poll event and returns. Grading it needs a stubbed rail, which would be grading a stub.
POST /api/checkout and POST /api/billing/portal	both construct a live Stripe session from STRIPE_SECRET_KEY; no Stripe key is spent by this environment
POST /api/webhooks/stripe	needs a Stripe-signed payload. Its two tables are seeded directly instead, which is what makes the Pro plan on the fixture real to every gate that reads it.
starreply_subscriptions and starreply_subscription_claims	written only by the Stripe webhook and by getPlan()'s claim attach. No route an operator can reach writes either.
starreply_posts	the blog and changelog are frozen at build time by scripts/freeze-posts.mjs. No route writes the table.
connecting, pausing or disconnecting a listing	starreply_locations and starreply_integrations have no writer anywhere in the routes. The console draws them and the overnight pass advances their watermarks; there is no connect flow, so there is no action to grade.
the classifier and the drafter	@compound/integrations resolves to a rail with no key in it, so inference.text() answers a stub and classifyReview falls through to classifyLocally. That is the product's own documented degraded path and it is what runs here. No paid model call is made for any reason.

Findings

Defects the environment build found in the product itself.

0 open of 2

State	Severity	Where	What was found
fixed	medium	src/console/state.ts:72 (promisesOk: true), with src/console/model.ts:187 carrying the unused holdReason and src/components/Console.tsx:1071 rendering the recomputed reason. Fixed in starreply a6812e4: decide() passes the row's real holdReason, and unallowed_promise joins isImmovable and the rule state because no dial value clears it.	A reply held because its draft promises something the voice profile forbids is drawn on the queue as 'your dial', under the sentence 'On autopilot it would go on its own, inside the caps'. That is false: autoPostDecision reads promisesOk before it reads the mode, so autopilot still holds it. decide() calls autoPostDecision with promisesOk hardcoded true and ignores the row's own holdReason, so hold_reason 'unallowed_promise' never reaches a screen and HOLD_REASON_LABEL.unallowed_promise is unreachable. Measured on the fixture: the 3 star row whose stored hold_reason is 'unallowed_promise' rendered the YOUR DIAL chip.
fixed	medium	src/components/Console.tsx:941, done = left === 0. Fixed in starreply a6812e4: the label is 'due', which is what the row is until the dispatcher runs.	A queued reply's card reads 'sent' the instant its countdown reaches zero. Nothing has been sent at that point: app/vercel.json schedules /api/cron/dispatch at /5 * * *, and the dispatch route's own header states the real spread is 30 seconds to 5.5 minutes after approval. The same card still offers 'Take it back', and that control succeeds, because the row is still queued.