Skip to content
← Evidence archive

Public evidence summary

Public defect status and customer impact

Record date
Register updated through 12 October 2026; individual finding dates are shown.
Scope
Published register or policy snapshot; per-item scope and dates apply.
Current deployment
Not independently verified by this historical evidence. Current availability

Recorded result

The register updated through 12 October 2026 contains 322 findings: 30 open, 219 fixed in tests, 37 reported fixed in production and 36 closed. It preserves reported issues involving isolation, resumed-fork networking, cancellation, account deletion, usage attribution, replay billing, spending limits and test reliability. A test fix is distinct from a production fix. Some entries conflict with newer bounded demonstrations. Each finding retains its source status and update date. These are source-record claims, not a new independent test or a claim that every reported defect remains exploitable.

All recorded customer impacts and test limitations

Session status could be stale

fixed-in-test

Sessions could be shown as running after their virtual machine had stopped.

Result: Recorded status: fixed-in-test.

Date: 2026-08-25

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Help could start the demonstration

fixed-in-test

Requesting demonstration help unexpectedly executed the demonstration.

Result: Recorded status: fixed-in-test.

Date: 2026-08-25

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Session scratch storage could keep growing

fixed-in-test

Rebuildable session files were not collected, allowing storage use to grow.

Result: Recorded status: fixed-in-test.

Date: 2026-08-25

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Manual interruption was unavailable

fixed-in-test

The demonstration could not wait for a person to trigger the interruption.

Result: Recorded status: fixed-in-test.

Date: 2026-08-25

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Preservation date displayed incorrectly

fixed-in-test

A preservation record displayed an unexpanded date format.

Result: Recorded status: fixed-in-test.

Date: 2026-08-25

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Rapid restart can leave the service stopped

fixed-in-test

The service could exit when its listening address remained briefly occupied, so a too-fast restart could leave no running service. On 30 September 2026 the source changed this finding to fixed in tests: the service now waits a bounded few seconds for the address before giving up; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 25 August 2026.

Date: 2026-09-30

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

A source formatting inconsistency was recorded

closed

An internal source formatting inconsistency was recorded, with no customer impact stated. The finding is closed on 30 September 2026 because it had already been fixed on 15 September 2026, together with an automated formatting check.

Result: Recorded status: closed. Earlier recorded status: open on 25 August 2026.

Date: 2026-09-30

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Second-agent evidence can become stale

open

A test baseline prevented an evidence check from passing after core changes. As of 12 October 2026 the check passes again only because no core code has changed since the latest release; the problem returns with the next core change, and the issue remains open.

Result: Recorded status: open.

Date: 2026-10-12

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Claim-register validation inconsistency

open

The claim register failed its own validation check. As of 12 October 2026 a separate automated claim check runs on every change, but the original validation still awaits an update; the issue remains open. No current customer effect is established.

Result: Recorded status: open.

Date: 2026-10-12

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Usage recording could fail silently

fixed-in-test

The billing service could write usage into an unavailable table without reporting failure. The register records a correction in test.

Result: Recorded status: fixed-in-test.

Date: 2026-08-25

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Session state could be silently lost after storage replacement

fixed-in-test

Replacing the session database could leave subsequent changes unwritten. The record reports a test correction in v0.1.3; unavailable storage must refuse changes.

Result: Recorded status: fixed-in-test.

Date: 2026-08-25

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Storage replacement can cause explicit service refusals

fixed-in-test

Leftover database recovery files could make the gateway refuse requests with 503 responses until recovery or restart; this was an explicit availability failure, not evidence of silent loss or duplicate billing. On 30 September 2026 the source changed this finding to fixed in tests: the gateway now sets such leftover files aside, keeping them, and opens cleanly; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 25 August 2026.

Date: 2026-09-30

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Durability needs a fresh release-specific soak

fixed-in-test

Historical soak evidence does not automatically prove a later release. This register labels the finding fixed in test; older limitation text still calls it open.

Result: Recorded status: fixed-in-test.

Date: 2026-08-30

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Historical soak verdict could depend on changing evidence

closed

A historical soak test referred to a mutable record, making its verdict difficult to reproduce consistently. The finding is closed on 12 October 2026: the test now reads a pinned, hash-verified copy of the record and passed on 10 October 2026. It concerned only test evidence.

Result: Recorded status: closed. Earlier recorded status: open on 30 August 2026.

Date: 2026-10-12

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Old-checkpoint test selection was inconsistent

closed

The checkpoint selected for an aged-checkpoint test was not deterministic, limiting repeatability. The finding is closed on 12 October 2026: the test now pins its checkpoint by name and content hash, and a run on 10 October 2026 verified those pins before resuming it.

Result: Recorded status: closed. Earlier recorded status: open on 30 August 2026.

Date: 2026-10-12

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Diagnostic counts in verdict text can be stale

open

Printed diagnostic counts could disagree with the current check set; the recorded success decision used the process outcome. As of 12 October 2026 the text of one affected test has been corrected and two still await the same change; the issue remains open.

Result: Recorded status: open.

Date: 2026-10-12

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Leftover test files can invalidate a rerun

open

A dirty test destination can produce a failing second-agent rerun rather than a clearly aborted trial.

Result: Recorded status: open.

Date: 2026-08-30

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Managed inference onboarding remains unproven

open

A managed provider-key distribution dependency remains open, so managed inference should not be inferred from bring-your-own-key tests. As of 30 September 2026 a hardening change was recorded so that, while managed inference is unavailable, a call from an account without its own key is refused with a named error instead of being served and billed as managed; the issue remains open.

Result: Recorded status: open.

Date: 2026-09-30

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Hosted services did not survive host reboot

fixed-in-test

A historical service-lifecycle problem could stop hosted services after a reboot. The register records a fix in test.

Result: Recorded status: fixed-in-test.

Date: 2026-08-31

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Hosted interactive attachment status changed

fixed-in-production

The earlier register described hosted attachment as command-only, without an interactive terminal. On 30 September 2026 the source changed this finding to fixed in production. This status change is not a new independent test of all attachment behavior.

Result: Recorded status: fixed-in-production. Earlier recorded status: open on 3 September 2026.

Date: 2026-09-30

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Some staff-console data is unavailable

open

Some administrative views display unavailable because their data feeds do not exist. No missing value should be presented as zero.

Result: Recorded status: open.

Date: 2026-09-03

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Client distribution is incomplete

fixed-in-test

Windows binaries and some package and installation channels were outstanding, so availability had to be checked per platform. On 1 October 2026 the source changed this finding to fixed in tests: Windows builds and stricter package publishing were added to a client version that is not yet released; a test fix is not a production fix. Availability must still be checked per platform.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 3 September 2026.

Date: 2026-10-01

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Free-session controls have stated approximations

open

Automatic parking follows runtime rather than idle time; network grouping is approximate; separate virtual-machine classes are unavailable. As of 1 October 2026 an unused virtual-machine class setting was removed in tests, while runtime-based parking and approximate network grouping are unchanged; the issue remains open.

Result: Recorded status: open.

Date: 2026-10-01

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Running-session usage is not continuously reflected

fixed-in-test

Session-minute accounting included time only through the last recorded event, so an active span was accounted for at the next freeze. On 30 September 2026 the source changed this finding to fixed in tests: a running session's live span is now metered up to the present or the statement end, whichever is earlier; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 3 September 2026.

Date: 2026-09-30

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Automatic credit top-up is unavailable

open

The service does not expose automatic credit replenishment in this snapshot.

Result: Recorded status: open.

Date: 2026-09-03

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Operational status does not prove a session can start

fixed-in-test

The status response could report operational while the hosted service could not start a session. On 1 October 2026 the source changed this finding to fixed in tests: the status response now includes a session-host reachability component and reports degraded after two consecutive failed checks; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 3 September 2026.

Date: 2026-10-01

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Fork network startup was intermittently slow in the lab

closed

Forked-session network startup was intermittently slow on the laboratory platform. The finding is closed on 12 October 2026: its cause was addressed by a related fix recorded as fixed in production, and a repeat check then passed 8 of 8 runs in each tested configuration.

Result: Recorded status: closed. Earlier recorded status: open on 3 September 2026.

Date: 2026-10-12

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Forked commands could miss their network context

fixed-in-test

Command execution in a forked child could use the wrong networking context. The register records a fix in test.

Result: Recorded status: fixed-in-test.

Date: 2026-09-03

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

The usual laboratory environment was unavailable

closed

Some tests moved to hosted infrastructure while the normal laboratory environment was down; results retain the environment actually tested. The finding is closed on 30 September 2026: the laboratory environment was repaired on 17 September 2026 and tests passed there again. It concerned only the test environment.

Result: Recorded status: closed. Earlier recorded status: open on 6 September 2026.

Date: 2026-09-30

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Resume improvements depend on workload

closed

A measured resume optimization does not improve every workload, so the observed timings should not be generalized to all sessions. The finding is closed on 30 September 2026: it was a measurement rather than a defect, and no published claim relies on the unmeasured improvement.

Result: Recorded status: closed. Earlier recorded status: open on 6 September 2026.

Date: 2026-09-30

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Resume below five seconds is unproven

open

Further work is needed to reach the desired warm-resume time below five seconds. No completed test is recorded here.

Result: Recorded status: open.

Date: 2026-09-06

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Billing continuity after host loss required a separate test

fixed-in-test

Resume on another host alone did not prove carried spending, replay exclusion or budget enforcement. A later test recorded the correction; cross-region recovery and a recovery-time objective remain unclaimed.

Result: Recorded status: fixed-in-test.

Date: 2026-09-06

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Account-level erasure required verification

fixed-in-test

Account deletion needed to make stored content unrecoverable across backup copies. The register records a fix in test; separate production evidence and the retention gaps still apply.

Result: Recorded status: fixed-in-test.

Date: 2026-09-06

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Unsupported command options were silently ignored

fixed-in-test

A command accepted an unused option without warning. The test correction requires unsupported options to be refused.

Result: Recorded status: fixed-in-test.

Date: 2026-09-07

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Checkpoint timing drifted during long runs

fixed-in-test

The long-run test driver accumulated checkpoint timing drift. A corrected run records its actual timing violations; older results are not rewritten.

Result: Recorded status: fixed-in-test.

Date: 2026-09-07

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Cleanup could choose the wrong retained content

fixed-in-test

A cleanup reader did not interpret the preserved-content manifest correctly. The register records a fix in test; no safe customer workaround is specified.

Result: Recorded status: fixed-in-test.

Date: 2026-09-07

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

A reported erasure-recovery failure was a test error

fixed-in-test

A claimed request failure came from the test using the wrong request method. The recorded regression test verifies that an erased account cannot prevent another account from resuming.

Result: Recorded status: fixed-in-test.

Date: 2026-09-07

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Deleted accounts could leave parked-session residue

fixed-in-test

After account deletion, parked sessions could leave unusable local resources. A test correction is recorded; this does not close the later running-session deletion finding.

Result: Recorded status: fixed-in-test.

Date: 2026-09-08

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Timed-out checkpoints could be absent from the record

fixed-in-test

A timed-out checkpoint was not retried and left no failure entry, making a missed checkpoint hard to distinguish from an unattempted one. The register records a correction in test on 8 September.

Result: Recorded status: fixed-in-test.

Date: 2026-09-08

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Some production promises lacked direct canaries

closed

Payments, transactional email and self-service signup lacked direct production evidence when this entry was recorded; email and self-service signup later received production checks. The finding is closed on 12 October 2026 on a report that a real-card production payment test was done; no recorded run of the payment check backs this closure, so coverage must still be checked per claim.

Result: Recorded status: closed. Earlier recorded status: open on 8 September 2026.

Date: 2026-10-12

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Cleanup regression results are inconsistent

open

A storage-cleanup check alternated between success and failure without code changes, and the full suite had no defined run boundary.

Result: Recorded status: open.

Date: 2026-09-08

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Parking can leave networking resources allocated

fixed-in-test

Stopping after a checkpoint could leave a session network interface allocated, so every parked session leaked one. On 30 September 2026 the source changed this finding to fixed in tests: a confirmed stop now releases the session network, and a resume re-creates it; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 8 September 2026.

Date: 2026-09-30

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

An aged-checkpoint measurement was corrected

closed

Published copy said approximately 14.5 hours; the original measurement was 15.2 hours. The record was corrected rather than preserving the wrong figure. The finding is closed as of 16 September 2026.

Result: Recorded status: closed.

Date: 2026-09-16

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Customer credentials could alter credit improperly

fixed-in-production

The historical authorization defect allowed a customer to change credit through inappropriate routes. A production correction is recorded for 17 September. It does not prove the current deployment independently.

Result: Recorded status: fixed-in-production.

Date: 2026-09-17

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

A resumed fork child can lose its network

fixed-in-test

A high-severity durability finding recorded network loss after resuming a forked child session, so complete fork recovery could not be inferred from successful creation or isolation tests. On 1 October 2026 the source changed this finding to fixed in tests: a resumed fork child is now restored inside its own network environment. On 13 October 2026 the source recorded that this fix has been in the production engine since 6 October 2026 and has not yet been exercised there; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 16 September 2026.

Date: 2026-10-13

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Forked jobs cannot freely change provider or model

fixed-in-test

Provider and model routing for a forked child was limited, so the recorded Cruise version kept its model sequence within one family. On 30 September 2026 the source changed this finding to fixed in tests, with a forked child taking its parent's provider and model on its first call; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 17 September 2026.

Date: 2026-09-30

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

The test environment could not rerun foundational checks

closed

A broken laboratory environment blocked foundational test reruns. The record says it was repaired and the finding closed, with one negative-control rerun waiting for a separate cleanup amendment.

Result: Recorded status: closed.

Date: 2026-09-18

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Prepaid credit was not consumed by metered usage

fixed-in-production

The historical billing defect prevented usage from reducing prepaid credit, so exhaustion could not trigger parking. The register records a production correction. Production parking of an actually running session remained a separate incomplete test.

Result: Recorded status: fixed-in-production.

Date: 2026-09-18

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Test cleanup could prevent its own negative-control rerun

closed

A normal test left network resources behind, causing the following negative-control run to fail before measuring its intended behavior. The finding is closed on 1 October 2026: after the test's cleanup was amended, both runs passed in sequence with no leftover resources and no manual cleaning. It concerned only the test harness.

Result: Recorded status: closed. Earlier recorded status: open on 17 September 2026.

Date: 2026-10-01

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Settlement monitoring has an idle-period blind spot

fixed-in-test

A settlement worker that stopped while no usage was accruing could remain unnoticed until usage resumed, because no heartbeat was recorded. On 30 September 2026 the source changed this finding to fixed in tests: the worker records a heartbeat every cycle and monitoring alerts when it goes stale; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 18 September 2026.

Date: 2026-09-30

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Published signup credit was undergranted

fixed-in-production

For four days in September, a unit mismatch granted only a millionth of the published signup credit. The record states the affected customer was made whole and the production correction landed on 18 September.

Result: Recorded status: fixed-in-production.

Date: 2026-09-18

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

An old pricing override could survive a policy change

fixed-in-production

A value matching an earlier price book could become stale after a pricing change. A production correction is recorded on 18 September; the policy must still be checked against the deployed source.

Result: Recorded status: fixed-in-production.

Date: 2026-09-18

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Some signups did not receive the advertised credit

fixed-in-production

Signups without identity-provider account-age data, including Google signups, missed the published grant. A production correction is recorded on 18 September.

Result: Recorded status: fixed-in-production.

Date: 2026-09-18

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

A billing test used a superseded price assumption

closed

A test compared the signup grant with an old amount. The entry is closed after its expected value was tied to the price book; this is a test correction, not a new customer payment.

Result: Recorded status: closed.

Date: 2026-09-18

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

A payment test could not refresh its evidence

closed

A payment-notification test expected a field no longer returned after a console change, so the test could not pass from 2 September 2026 onward; the record states no customer payment path was broken. The finding is closed on 4 October 2026: the test was amended and passed in both its normal and negative-control modes in the test environment.

Result: Recorded status: closed. Earlier recorded status: open on 18 September 2026.

Date: 2026-10-04

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Parking could leave the host marked unclean

closed

A parked session could leave a networking resource allocated, causing health checks and later tests to report an unclean host. The finding is closed on 30 September 2026 as a duplicate of the related parking finding, whose fix, recorded as fixed in tests, releases the session network when a stop is confirmed.

Result: Recorded status: closed. Earlier recorded status: open on 19 September 2026.

Date: 2026-09-30

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Deleting an account can strand running sessions

fixed-in-test

An open finding recorded that account deletion made its running sessions unreachable through both customer and host control paths. On 7 October 2026 the source recorded this finding as fixed in tests; a test fix is not a production fix. A cleanup route for sessions left behind by deleted accounts was added and used once in production, but no production account deletion has yet exercised the fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 19 September 2026.

Date: 2026-10-07

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

A billing rehearsal could not be safely refreshed

fixed-in-production

The rehearsal required stopping the production gateway, so its evidence could become stale without a safe refresh path. On 10 October 2026 the source changed this finding to fixed in production: the rehearsal now runs beside the production service and passed there without disturbing it.

Result: Recorded status: fixed-in-production. Earlier recorded status: open on 19 September 2026.

Date: 2026-10-10

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

A test-only help control remains in shipped clients

fixed-in-test

The published register said a help-test control remained compiled into production clients. On 1 October 2026 the source changed this finding to fixed in tests: the control is built only into test builds of a client version that is not yet released; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 19 September 2026.

Date: 2026-10-01

Scope: Historical finding register; current deployment not independently verified.

Limit: Dates and statuses are those in the source register; a test fix is not a production certification.

Workaround: Only the safe recovery described above is recorded; otherwise no verified customer workaround is stated.

Client compatibility and structured logout output

fixed-in-test

The source reports two client defects: creating a duplicate session against an older server and returning more than one JSON document during logout. Both are marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Job handling depends on runner wording

open

Changes to a runner's refusal message can stop job orchestration from recognizing a refused model call. As of 30 September 2026 the issue remains open; a prepared change that reads the gateway's own refusal record has not been deployed.

Result: Recorded status: open.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

Recovery from uncertain task acceptance

fixed-in-test

The source records a queue hold and authorized release for uncertain task acceptance. Retrying the uncertain task itself remains unimplemented, despite the test-fix status.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Listing owned sessions in the client

fixed-in-production

The missing client command for listing a customer's own sessions is marked fixed in production on 30 September 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Managing provider keys in the client

fixed-in-production

The restriction that provider keys could only be added through the console is marked fixed in production on 30 September 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Cancelling an individual queued or held task

fixed-in-production

The absence of an individual-task cancellation command is marked fixed in production on 30 September 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Approval and agent session matching

fixed-in-test

The source reported that an approval outcome could be recorded without checking that the approval and agent belong to the same session. On 30 September 2026 the source changed this finding to fixed in tests: an outcome from an agent outside the approval's session, from another account, or from an agent that did not claim the approval is now refused. A test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 21 September 2026.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Structured explanation for a held task

fixed-in-test

A held task identified its blocker only in prose, making programmatic interpretation less reliable. On 30 September 2026 the source changed this finding to fixed in tests, with a held task now naming its blocking task and its hold as separate fields; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 21 September 2026.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Recording loss may not reach public status

fixed-in-test

The source reported that a recording-loss monitor wrote an internal alert but could not raise an incident on the public status feed, so public status alone could omit this failure. On 1 October 2026 the source changed this finding to fixed in tests; the change was not deployed, and a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 21 September 2026.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Starting the supported runner

fixed-in-test

The runner adapter could not start the pinned Claude Code version with its supplied options. The defect is marked fixed in tests; that status alone does not establish all deployed runner configurations.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Blocked prompts reported as successful

fixed-in-test

A prompt refused before submission was incorrectly recorded as a successful task. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Refused tools reported as successful work

fixed-in-test

A task for which every tool was refused could be reported as successful even though it accomplished no work. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Permission requests did not reach approvals

fixed-in-test

A missing runner protocol step prevented tool-permission requests from reaching the approval system. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Runner observations may be dropped

fixed-in-production

Some runner messages were silently ignored because their message shape differed from the adapter's expectation, losing observations. On 30 September 2026 the source changed this finding to fixed in production, and a line that still cannot be read is now treated as a lost observation rather than ignored.

Result: Recorded status: fixed-in-production. Earlier recorded status: open on 21 September 2026.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Conflicting implementation dependencies

open

The source records contradictory dependencies in an internal implementation plan. The customer impact is not specified, and the planning inconsistency remains open.

Result: Recorded status: open.

Date: 2026-09-21

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

Reading task lists and individual task state

fixed-in-production

The client lacked direct task-list and task-detail views; instruction state could only be read through recovery. On 30 September 2026 the source changed this finding from fixed in tests to fixed in production, recording that both views shipped in a released client version.

Result: Recorded status: fixed-in-production. Earlier recorded status: fixed-in-test on 30 September 2026.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Runner version reported incorrectly

fixed-in-production

The adapter read the runner version from the wrong field, causing its compatibility report to say that no version was reported. On 30 September 2026 the source changed this finding to fixed in production; the earlier report's incorrect sentence is kept and corrected in place.

Result: Recorded status: fixed-in-production. Earlier recorded status: open on 21 September 2026.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Approval timing is shorter than its recorded expiry

closed

The runner's permission timeout limited the useful approval window even when the approval record had a later expiry, so the approval record and the runner could disagree. On 12 October 2026 the source closed this finding: an approval now records the window in which a decision can still take effect, a later decision is refused and records nothing, and an approval left pending can be marked abandoned.

Result: Recorded status: closed. Earlier recorded status: open on 21 September 2026.

Date: 2026-10-12

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Runner platform mismatch in evidence

fixed-in-test

A compatibility record identified a macOS artifact where the tested runner was on Linux. On 1 October 2026 the source changed this finding to fixed in tests: the record now names the platform beside every artifact and records the runner artifact actually executed. The source notes that whether certification covers the same runner version on a different processor architecture is still undecided.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 21 September 2026.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Failed queue reads appeared empty

fixed-in-test

A failed read of held tasks was treated as an empty queue, so releasing a hold could leave instructions stranded. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Duplicate completion over an unresolved outcome

fixed-in-test

A second completion could record another outcome while the first outcome was unresolved. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Agent ownership was not checked at execution

fixed-in-test

The source reports missing agent-ownership checks at two execution boundaries. The defect is marked fixed in tests; current production authorization is not established by that status.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Work dispatched after a session or agent stopped

fixed-in-test

Instructions could still be dispatched to a saved session or a stopped agent. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Dispatch success without a confirmed state change

fixed-in-test

A dispatch could commit without checking that it had actually changed the intended record. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Gateway database contention caused immediate failures

fixed-in-test

Most pooled gateway database connections lacked a contention wait, causing immediate database-busy errors. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Incomplete session-store contention protection

fixed-in-test

The session store applied its wait for busy database operations to only one pooled connection; the source says no failure was observed from this gap and corrects its earlier statement that write-ahead logging was missing. On 30 September 2026 the source changed this finding to fixed in tests, with every connection now carrying both settings. On 13 October 2026 the source recorded that this fix has been in the production engine since 6 October 2026 and has not yet been exercised there; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 23 September 2026.

Date: 2026-10-13

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Restored attempts remained unclosed

fixed-in-test

A restored task was linked to a different attempt from the one that actually ran, leaving an attempt that could never close. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Cancelled queued tasks leave retry attempts open

fixed-in-test

Cancelling a queued task could leave its prepared retry attempt open indefinitely. On 30 September 2026 the source changed this finding to fixed in tests: every cancellation path now closes a prepared retry that never ran; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 23 September 2026.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Recovery ownership guard lacked a regression check

fixed-in-test

Tests did not detect removal of a guard used when adopting recovered work. The coverage gap is marked fixed in tests; it does not by itself establish production recovery behavior.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

One test run could disrupt another

fixed-in-test

Shared test-launcher resources allowed cleanup from one run to affect another active test environment. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Privacy lint produced a false positive

fixed-in-test

A privacy-related source check mistook a code identifier for a person and lacked a test for that false positive. The check is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Cancelled instructions returned after restore

fixed-in-test

Restoring a session could put an already cancelled instruction back into the queue. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Advice activity was labeled incorrectly

fixed-in-test

The advice path reported the wrong activity label, making the displayed state and its test evidence inaccurate. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Removed agents could still appear active

fixed-in-test

A removed agent could still be observed working. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Cancellation did not reach a terminal state

fixed-in-production

Tasks could remain in a cancelling state indefinitely. The defect is marked fixed in production on 30 September 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

A previously recorded defect recurred

fixed-in-test

The source records recurrence of a defect shortly after it was filed. The test-process issue is marked fixed in tests; no separate customer impact is specified.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Test environment could not exercise agent features

fixed-in-test

The private test environment could not use agent operations because the product capability check refused them. The test-coverage gap is marked fixed in tests, without establishing earlier production availability.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Later instructions lost conversational context

fixed-in-test

A second instruction to an agent started a conversation without the first instruction's context. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Conversation recovery depended on runner storage

fixed-in-test

The service could not re-establish a known continuation when the runner's own stored conversation was missing. The recovery gap is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Built adapter identity was not reproducible

fixed-in-test

An adapter artifact's digest could depend on unrelated local changes. The evidence-integrity defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Test liveness check did not establish availability

closed

A test environment's ping check did not establish that the relevant service was alive, so a cleanup check based on it would have been ineffective. On 30 September 2026 the source closed this finding after an automated check was added that rejects ping used as a liveness verdict in test scripts; it concerned only test tooling.

Result: Recorded status: closed. Earlier recorded status: open on 23 September 2026.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Test authentication failed without a clear diagnosis

fixed-in-test

The test launcher supplied authentication that the engine could not verify and did not clearly report the mismatch. The test-infrastructure defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Test configuration requires a rebuild

fixed-in-test

A test authentication configuration was fixed at build time, limiting testability; no customer impact is specified. On 30 September 2026 the source changed this finding to fixed in tests, adding an explicit override while keeping the previous default; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 23 September 2026.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A missing machine appeared to be running

fixed-in-test

Two customer-visible views reported a running session even when its machine did not exist. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Pause state was not measured

fixed-in-test

The pause path could report a state that was not observed. On 30 September 2026 the source changed this finding to fixed in tests: when the result cannot be read back, the session now reports as unavailable and accepts no new work; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 23 September 2026.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Saving and stopping had incomplete behavior

fixed-in-test

A save could be named only when it also parked the session, while stopping did not itself save it. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

State-transition coverage was incomplete

fixed-in-test

The permitted state-transition set could be widened without an automated test detecting it. On 30 September 2026 the source changed this finding to fixed in tests, adding a guard tied to the task-state rules and removing one unused transition those rules do not permit; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 23 September 2026.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A supposedly killed session could return

fixed-in-test

The client reported that a session was killed, but a later wake operation could bring it back. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Cancelled jobs could start another model attempt

fixed-in-test

A cancelled Cruise job could dispatch a new model attempt after cancellation. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Fork networking and egress isolation were incorrect

fixed-in-test

A forked child could not reach the control plane while other traffic remained broadly forwarded. The isolation defect is marked fixed in tests; this does not close the separate open fork-resume network-loss finding.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Guest access to host services was too broad

fixed-in-test

The source reports that a guest could reach host services that should have been isolated from it. The defect is marked fixed in tests, without independently establishing the current production boundary.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Fork connectivity ignored gateway configuration

fixed-in-test

A forked child's gateway connection depended on a fixed service port rather than the configured value; the source states the production port was unaffected. On 1 October 2026 the source changed this finding to fixed in tests; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 26 September 2026.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Agent image could not be built for the production fleet

fixed-in-test

The register reported that the agent image could not be built for the production fleet's processor architecture. On 1 October 2026 the source changed this finding to fixed in tests: the build now follows the image's own architecture, but the production-architecture path is proven only by self-tests, and a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 26 September 2026.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Test health checks rejected prepared images

fixed-in-production

A test health check failed after preserved images were introduced, preventing dependent tests from running. On 10 October 2026 the source changed this finding from fixed in tests to fixed in production, after the production health check was updated and passed all its checks.

Result: Recorded status: fixed-in-production. Earlier recorded status: fixed-in-test on 30 September 2026.

Date: 2026-10-10

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Uncertain effects were reported as known failures

fixed-in-test

A provider timeout or disconnect after tool execution could be recorded as a failed task with a known outcome even though acceptance was uncertain. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Corrupt checkpoint data could hang resume

fixed-in-test

A resume encountering corrupt checkpoint data could hang indefinitely instead of refusing it. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Test cleanup left network resources behind

closed

A checkpoint test left fork networking resources behind, causing the next health check to fail. On 1 October 2026 the source closed this finding after repeated test runs showed the next health check passing with no manual cleanup; it concerned only test tooling.

Result: Recorded status: closed. Earlier recorded status: open on 26 September 2026.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

A path check could execute unintended commands

fixed-in-test

A guest file-existence check could interpret supplied path text as a command. The isolation defect is marked fixed in tests; no current exploitability claim is made.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Guest error handling lost uncertain outcomes

fixed-in-test

Within a guest, the runner could discard a structured gateway error and report a known failure even when provider acceptance was uncertain. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Budget-refusal handling was unreachable

fixed-in-test

The adapter's dedicated budget-refusal behavior did not run with the pinned runner. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Concurrent consultations exceeded a token grant

fixed-in-test

Concurrent admissions for the same consultation could each pass its token limit, allowing the combined usage to exceed the grant. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Authorized members could not decide approvals

fixed-in-test

A project member with the project role that permits it could not decide an approval on a session they were authorized to act on. As of 30 September 2026 the authorization defect remains marked fixed in tests; a test fix is not a production fix.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Onboarding offered unavailable managed inference

fixed-in-test

The onboarding flow allowed selection of managed inference even though the service reported it unavailable. The defect is marked fixed in tests; customer-supplied provider keys remain required.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Jobs accepted disabled provider keys

fixed-in-test

A Cruise job could be accepted even when its only key for a required provider was disabled. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Advice delivered to a paused session

fixed-in-test

A consultation could be delivered to an adviser whose session was parked or being saved. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Advice accepted for a removed adviser

fixed-in-test

A consultation could be accepted after the adviser had been removed. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Advice questions were attributed as user instructions

fixed-in-test

An adviser's conversation displayed another agent's question as if a person had issued it. The attribution defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Candidate code could influence verification results

fixed-in-test

A candidate could alter the verification process so that incorrect work was accepted. This specific defect is marked fixed in tests; a broader open test-report integrity issue remains separately listed.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A paid test did not reliably enforce its spending stop

fixed-in-test

A paid test's spending limit used unapproved pricing and did not check continuously during each case. The test defect is marked fixed in tests; it does not establish a customer spending guarantee.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A paid test overstated its prerequisites

fixed-in-test

A paid test authorization could claim a verified provider limit without one and did not require an earlier keyless check. The evidence-quality defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

The service did not control the selected model

fixed-in-production

The service did not select the model called by the runner inside the machine. The missing control is marked fixed in production on 30 September 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Some required test cases could not run

fixed-in-test

Three cases in a private test suite could not run against the then-current candidate. The coverage defect is marked fixed in tests, without proving those earlier cases passed.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Readiness and retry checks used invalid assumptions

fixed-in-test

A test checked readiness before its prerequisite and retried by reusing a prior submission. The test defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Candidate code could replace the final test report

fixed-in-test

A candidate that had not fixed the problem could replace its final report and be accepted. This specific defect is marked fixed in tests; the broader report-integrity limit remains open.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Code under test can falsify its own report

open

The source retains an open high-severity issue in which code under test can fabricate a consistent report from inside the runner. A passing report alone does not independently prove that the work is correct. As of 30 September 2026 the issue remains open.

Result: Recorded status: open.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

Verification ran customer code with excessive host privilege

fixed-in-test

The source reports that verification ran customer project code with excessive privilege on the host. This critical isolation issue is marked fixed in tests on 26 September 2026; that status is not a production certification.

Result: Recorded status: fixed-in-test.

Date: 2026-09-26

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: critical. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Task lists silently omitted later records

fixed-in-test

The task list stopped after 200 records without indicating that more existed. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Deleting a session did not stop its machine

fixed-in-test

Session deletion promised to stop its machine but did not do so. The finding is marked fixed in tests; a separate dated agent-session production test also records successful stop and deletion for its single tested session.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Records within one second may sort incorrectly

fixed-in-test

Time values stored as text could order records written within the same second incorrectly. On 1 October 2026 the source changed this finding to fixed in tests, with every stored time now written in one fixed-width form and a one-time migration for existing records; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 26 September 2026.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Permission tests expected nonexistent output fields

fixed-in-test

A permission test looked for field names that the client did not produce. The test defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Accepted numeric input was refused later

fixed-in-test

Some valid numeric representations were accepted at job intake but then refused by the worker. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Denied verification tampering was not recorded

fixed-in-test

Attempts to alter protected verification inputs could be denied by the sandbox yet accepted without recording the attempt. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A test contaminated its own clean build inputs

fixed-in-test

A test wrote generated files into a candidate that was required to remain clean, preventing its image build. The test defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Leftover evidence was mistaken for current execution

fixed-in-test

A test treated another test case's leftover marker as proof that a tool was currently running. The false-positive defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

An instruction after approval was not acknowledged

closed

An instruction submitted immediately after an approved tool turn could receive no acknowledgement and make no model call. On 30 September 2026 the source closed this finding as environmental: the stall was caused by a credit safeguard in the test environment pausing the session, and the same sequence then passed in later certified runs.

Result: Recorded status: closed. Earlier recorded status: open on 26 September 2026.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Tool-running and retry checks were unreliable

fixed-in-test

A test inferred tool execution from a transient activity label and reused a fixed submission identity on retry. The test defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

An executing tool had the wrong activity label

fixed-in-test

During an approved tool operation, the displayed activity said general work rather than tool execution. The defect is marked fixed in tests on 26 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-26

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Provider keys could become unusable during a job

fixed-in-test

Required provider keys were checked only when the job was submitted, so later stages could fail after earlier model calls had already incurred usage. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Permission decisions outlived the request waiting for them

fixed-in-test

A permission decision could still be claimed after the runner had stopped waiting for it. The effects-accounting defect is marked fixed in tests on 26 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-26

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Test credit exhaustion interrupted the intended scenario

fixed-in-test

A test account lacked the credit needed for the planned run, so the funds control parked its machine mid-test. The test-setup defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Credit-paused sessions still appeared running

fixed-in-test

After the funds control parked a session, it still appeared to be running and related views did not identify the funds pause. The defect is marked fixed in tests on 26 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-26

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Tests could not distinguish a stopped machine from an unresponsive runner

fixed-in-test

A test lacked the evidence needed to distinguish machine shutdown from a runner that stopped answering. The diagnostic defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

The first task after pause and wake failed

fixed-in-test

After pause and wake, the first instruction failed because its recorded execution generation had changed. Pausing an idle adapter also waited unnecessarily. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-26

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A previous failed test left later retries held

fixed-in-test

A test retried behind a queue hold left by its own failed attempt, invalidating the later measurement. The test defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A crashed test was reported as successful

fixed-in-test

A test wrapper reported success after its scenario crashed; a related case also failed while reading the client's task count. The false-positive defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

An advice test inspected the wrong request

fixed-in-test

A consultation test examined the wrong request and failed without retaining enough evidence to diagnose the result. The test defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Releasing an uncertainty hold did not restore a worker

fixed-in-test

After uncertain acceptance, the adapter could stop permanently, so an authorized release of the hold reached no worker. The defect is marked fixed in tests on 26 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-26

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A stalled earlier task invalidated later tests

fixed-in-test

A test evaluated later cases while the adapter was stuck on an earlier closed task, and its diagnostic capture could not show that state. The test defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

An interrupted test left a closed task active

fixed-in-test

A test passed an interruption even though the adapter still held the closed task; its evidence parser also misread lists of open tools. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Unpriced earlier attempts prevented job resume

fixed-in-test

A Cruise job could not resume if earlier attempts had no recorded price. The defect is marked fixed in tests, but that does not establish an effective spending ceiling for unpriced usage.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Unpriced Cruise usage has no effective spending ceiling

open

While model costs are unavailable, a Cruise job's spending ceiling does not constrain it; the attempts limit is a separate bound. As of 30 September 2026 this high-severity money-related issue remains open: pricing support was prepared but stays inactive until model rates are ratified.

Result: Recorded status: open.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

Verification used files a candidate had removed

fixed-in-test

Verification could use files that the candidate had deleted, so the tested result did not match the candidate's final contents. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A wake test required readiness too early

fixed-in-test

A test required readiness after wake before the next instruction, although the supported sequence did not provide it then. The test defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A task test omitted the outcome it claimed

fixed-in-test

A test marked a task successful without supplying the outcome required by its own evidence contract. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Help-output ordering made a test unstable

fixed-in-test

A command's flag list appeared in unpredictable order, causing a test to pass or fail by chance. The test defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Several lists silently stopped after their first page

fixed-in-test

Session, agent, approval and proposal lists did not reliably offer later pages or indicate that additional records existed. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Request references could not be traced

fixed-in-test

Response request references were created too late to be sent through or logged with the underlying operation, making them unusable for tracing. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Agent-session refusals were not reachable through customer identifiers

fixed-in-test

The customer-visible session identifier did not reach the intended refusal for unsupported shell access, and older session routes returned not found. The routing defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-26

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Older session operations bypassed agent supervision

fixed-in-test

Older stop, resume and fork operations could reach an agent session without its normal supervision. The defect is marked fixed in tests on 26 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-26

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Concurrent distinct requests were unnecessarily refused

fixed-in-test

Request admission did not provide its documented atomic behavior, causing concurrent distinct requests to be refused. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Late consultation replies were accepted

fixed-in-test

A reply arriving after the consultation deadline could still be accepted as advice. The defect is marked fixed in tests on 26 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-26

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Host storage exhaustion stopped a restored session

closed

A restored session stopped answering after another session was provisioned because host storage ran out. The environmental finding is closed with a guard recorded; it is not a general availability guarantee.

Result: Recorded status: closed.

Date: 2026-09-26

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

New session network identities were not checked for collisions

fixed-in-test

Provisioning did not check a new session's network identity against existing sessions on the same host. The defect is marked fixed in tests on 26 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-26

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A retry replayed the previous task

fixed-in-test

A test retried the same instruction through the live view and received the prior task rather than a fresh attempt. The test defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A new-adviser test used invalid readiness and retry assumptions

fixed-in-test

A test required a new adviser to be ready before its first instruction and reused its name on retry. The test defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A consultation test passed without an advice turn

fixed-in-test

A test asked for consultation from an agent that could no longer take work and accepted a result without an actual advice turn. The false-positive defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A paid test confused exhausted budget with advice failure

fixed-in-test

A paid run spent its shared token allowance before a later consultation case and interpreted the resulting refusals as consultation or reply defects. The test defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Tools could alter the runner's control configuration

fixed-in-test

Tools and the runner shared an identity that allowed tools to change control configuration during execution. The isolation defect is marked fixed in tests on 27 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-27

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Customer-key attribution was missing

fixed-in-test

Usage admission could report no key source even though the session was bound to a customer-supplied key. The attribution defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-27

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Consultation payer and spending were not recorded

fixed-in-test

The service did not record who paid for a consultation, and spending attributed to advice could not be queried. The money-related defect is marked fixed in tests on 27 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-27

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Three tests lacked a negative control

open

Three test gates lacked a deliberately failing control, limiting the evidence that they would detect incorrect behavior. The coverage issue remains open.

Result: Recorded status: open.

Date: 2026-09-27

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

Advice-call attribution did not reach payment views

fixed-in-test

Per-call advice attribution did not reach the control plane, and payment views still described the capability as unavailable. The defect is marked fixed in tests on 27 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-27

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Provisioning showed misleading readiness guidance

fixed-in-test

A newly provisioned agent told the user to wait for readiness, although the supported runner could not become ready before its first instruction. The defect is marked fixed in tests.

Result: Recorded status: fixed-in-test.

Date: 2026-09-28

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

New agent sessions lacked a bound provider key

fixed-in-test

The client could create an agent session without binding a provider key, causing its first instruction to fail. The defect is marked fixed in tests on 28 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-28

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Idle agents always appeared stale

fixed-in-test

The client always labeled an idle agent stale, making the status misleading. The defect is marked fixed in tests on 29 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-29

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Metered model calls were absent from customer usage

fixed-in-test

A model call counted by the gateway was missing from the customer usage view. The finding is marked fixed in tests; the register separately says admission for agent-session calls is unverified.

Result: Recorded status: fixed-in-test.

Date: 2026-09-29

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Two token limits govern different call paths

fixed-in-test

The gateway budget constrained model calls from a session's machine, while a separate session cap applied only to service-originated calls. On 30 September 2026 the source changed this finding to fixed in tests, with each session now governed by one token budget fixed when it is provisioned; a test fix is not a production fix.

Result: Recorded status: fixed-in-test. Earlier recorded status: open on 29 September 2026.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Agent tools ran without the promised separate unprivileged user

fixed-in-test

Inside a session machine, an agent's runner and every tool it started ran with the privileges of the agent's controlling process instead of as a separate unprivileged user, although the published capability statement described that separation. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Free-tier limits were not applied to agent sessions from the newer API

fixed-in-test

Agent sessions provisioned through the newer API received the paid-tier token budget whatever the account's tier, and the free tier's one-running-session limit and capacity cap were not applied. The defect is marked fixed in tests on 30 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Session runtime billing could include time a session was parked

fixed-in-test

The engine's runtime meter measured from a session's first event to its last and ignored the statement window, so time a session spent parked before a later wake could be billed as runtime. The defect is marked fixed in tests on 30 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A funds pause could record a stopping state that was never measured

fixed-in-test

When a session was paused because funds ran out and the service could not read back the engine's answer, the session record showed it as stopping although that state had not been measured. The defect is marked fixed in tests on 30 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Approval actions were not tied to the agent that asked

fixed-in-test

Claiming, answering or abandoning a permission request did not check that the caller was the agent, session and account the request belonged to, so a holder of service credentials, including code inside another session, could act on an approval it did not own. The defect is marked fixed in tests on 30 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Advice after a failed wake named a check that did not re-measure

fixed-in-test

When a wake could not read back the session's state, the answer told the customer to check again with the session status command, but the stored session record was never measured again, so other readers kept seeing an unavailable state. The defect is marked fixed in tests on 30 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Forking a running session could put the child's credential in the parent

fixed-in-test

With signed capabilities enabled, forking a running session wrote the new child's signed capability into the parent session's machine, and the child kept the parent's stale one. The source records that this mode was not enabled in production at the time. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Resume and kill could signal an unrelated process reusing a stored id

fixed-in-test

Resume and kill checked only that a stored process id was alive before signalling it, so after a host restart a reused id could lead to an unrelated process being killed. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Regression guards relied on a copied signing key and hand-built credentials

fixed-in-test

Three regression guards started a test engine with a key setting it ignored, minted credentials with hand-written code instead of the product's own, and handled a copy of a signing key unsafely on the test machine. The defect is marked fixed in tests on 30 September 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Disk space held by parked sessions was never reclaimed

fixed-in-production

No scheduled job reclaimed the rebuildable disk space held by parked sessions, so that space was never freed. A scheduled cleanup with its own monitor was added. The defect is marked fixed in production on 7 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-07

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Two engine checks left network resources behind after their runs

open

Two engine checks did not force-delete their sessions or remove fork network resources, leaving them on the test machine where they later made another check's preflight fail. Amendments were drafted but not applied. The issue remains open.

Result: Recorded status: open.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

An engine gateway check had three assertions that always passed

open

Three assertions in the engine gateway check passed whichever outcome they observed, so they gave no assurance. An amendment making each able to fail was drafted but not applied. The issue remains open.

Result: Recorded status: open.

Date: 2026-09-30

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

Checkpoint soak sabotage run could report green on no evidence

closed

The sabotage run of the long checkpoint soak check fell back to a missing record when its own short soak wrote nothing, and could then report a green result on no evidence. The finding is closed on 12 October 2026: the fallback was removed, and a later sabotage run judged its own record and caught the sabotage as intended.

Result: Recorded status: closed.

Date: 2026-10-12

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Two regression tests were not run by any automated check

closed

Two named regression tests, one for the verifier's refusal handling and one for the image builder's refusals, were run by nothing, so they gave no protection. The finding is closed on 1 October 2026: both now run in continuous integration and each was shown able to fail; it concerned only the test process, with no product exposure.

Result: Recorded status: closed.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

An interrupted image build left a partial file that blocked bench checks

fixed-in-test

When copying a published machine image failed partway, the image build left a partial file among the preserved images, and every check whose preflight ran the health check on the test machine then stopped before its first assertion. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Checkpoints could not be restored on an engine without the base image

fixed-in-test

After a change that compares the saved software stack at restore, a checkpoint could no longer be revived on a second engine holding only its stored content and not the base image, so a portable restore was refused. The restore now proceeds and reports compatibility as unknown. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Restore check sampled its baseline on a fixed timer

open

One part of the portable restore check read its baseline a fixed 3 seconds after boot, so a slow first line from the test machine produced a failure on missing evidence instead of an abort. An amendment that waits for the line was drafted but not applied. The issue remains open.

Result: Recorded status: open.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

Deleting a session did not delete its checkpoint content

fixed-in-production

Deleting a session removed its records, but the content of its checkpoints stayed readable with the account's key until the whole account was deleted, although the retention policy said checkpoint content was deletable per session. The source records that sessions saved in the older format still await conversion. The defect is marked fixed in production on 6 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-06

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Account deletion did not purge content from the optional resume cache

fixed-in-test

With an optional resume cache turned on, deleting an account did not purge that cache, so decompressed content from the account's checkpoints could remain on the host. The cache was off by default. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Account data keys were kept separately on each host

fixed-in-production

Each host kept its own copy of an account's data key while failover shared the store, so a second host could not read an encrypted session or could create a different key, and deleting an account on one host could leave a key on another. The second host was not running at the time. The defect is marked fixed in production on 6 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-06

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Encrypted checkpoints had no long-running soak check

fixed-in-test

The long-running checkpoint checks exercised only unencrypted sessions, so the encrypted configuration used in production had a round-trip record but no soak. An encrypted soak check was added and passed a three-hour run. The defect is marked fixed in tests on 5 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-05

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Regression harness exit code ignored failures in later sections

closed

The regression harness decided its exit code before its later sections ran, so a failure in any of those sections was printed but the run still reported success. The finding is closed on 1 October 2026: it concerned only the test harness, with no product exposure, and the exit code now reflects every section.

Result: Recorded status: closed.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Agent commands on a forked session reached the parent session

fixed-in-test

Starting, reading or stopping an agent on a forked session sent the commands to the parent session's machine instead of the child's. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A fixed timing bound made a resume unit test fail under load

fixed-in-test

A unit test used a fixed 150 millisecond bound that a busy disk could exceed, so it failed intermittently on a loaded machine. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Payment events could grant credit without a verified paid order

fixed-in-test

The payment webhook would have granted credit on a checkout update without checking that it was paid, and could grant a paid checkout twice, while a refund-created event netted nothing; checkout update events were not subscribed at the time. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Resuming through the newer API bypassed the free-tier session limit

fixed-in-test

A free-tier account could run two sessions at once by resuming a second one through the newer API, which lacked the one-session limit the older route applied. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Console sessions defaulted to the free-tier budget for paid accounts

fixed-in-test

Sessions started from the console defaulted to a 500,000 token budget for every account, although paid tiers default to 2,000,000 tokens. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Billing reconciliation monitor could alert on normal settlement lag

fixed-in-test

After a billing change, the internal reconciliation monitor could flag an account whose usage briefly exceeded its balance before the funds pause took effect, which is expected settlement lag. The monitor now tolerates that lag within set bounds. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Free-tier session limit was skipped when the engine could not be read

fixed-in-test

When the engine's list of running sessions or its session count could not be read, the free tier's one-session limit and capacity cap were not applied, so a start could proceed unchecked. Such starts are now refused with a retryable error. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Project members saw the gateway usage figure as unreadable

fixed-in-test

In the usage view, a member of the project that holds a session saw the gateway meter as unreadable, because it was read under the member's account instead of the session's own account. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Permission request details were readable across accounts

fixed-in-test

Any holder of service credentials, including code inside a session of another account, could read any permission request by its id, including its arguments and decision. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Code in one session could act on another session's tasks

fixed-in-test

Code inside one session could act for another session's agent in the same account: finishing or acknowledging its tasks, posting its results, answering its consultations and spending its reserved budget. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Permission requests could be raised for another session or agent

fixed-in-test

Code inside a session could raise a permission request in another session of the same account, or attributed to an agent that never asked. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Pending tasks lacked structured fields naming what held them

fixed-in-test

The pending task view described why a held task was waiting only in a text reason, without the hold and blocking task identifiers the task view already provided. Both fields were added. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Failed account deletion could leave sessions behind with no retry

fixed-in-test

When a customer deleted their account and the engine's erase step failed, the account was already closed and its key destroyed, so its sessions and per-account engine key could be left behind and the customer could not retry the deletion. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Release check cleanup removed test data whatever the deletion result

open

A release check's cleanup removed its bystander test account's data directly from the database even when the service's deletion route had not confirmed the deletion, which could hide a failed deletion. As of 1 October 2026 a correction has been drafted but not applied, and the issue remains open.

Result: Recorded status: open.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

Admin overview could wait up to a minute on a hanging engine

fixed-in-test

The admin console's overview asked the engine for its health on every load and could wait up to a minute when the engine hung, and its fleet tile could disagree with the public status feed. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Attach on Windows did not switch the terminal to raw mode

fixed-in-test

On Windows, ks attach relied on a Unix terminal tool to switch the terminal into raw mode, so an attached session ran in line-buffered (cooked) mode there. The defect is marked fixed in tests on 1 October 2026, in an unreleased client build that has not yet run on a real Windows console.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Windows client not yet run on a real Windows machine

open

The Windows build of the ks client has been checked only by an automated build job that runs a subset of its tests; no part of it has run on a real Windows machine, so its behavior there is not established. As of 1 October 2026 the issue remains open.

Result: Recorded status: open.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

Test harness could act on the hosted service when signed in

closed

Some test-harness checks that run ks commands used whatever account the client on the test machine was signed in to, so with a hosted login present they could have started, stopped or metered sessions on the hosted service. The finding is closed on 1 October 2026: it concerned only the test harness, which now refuses to run while signed in.

Result: Recorded status: closed.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Launch rehearsal could leave the host on the rehearsal's engine process

closed

The pre-launch rehearsal did not stop one engine process it had started, so afterwards the host could keep serving from the rehearsal's process instead of the production service. The finding is closed on 12 October 2026: the rehearsal was replaced by one that runs beside production with its own separate processes.

Result: Recorded status: closed.

Date: 2026-10-12

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Published-claims check did not run on every change

closed

The check that every published claim cites an existing, sealed test result ran only when a release check called it, so a claim citing a missing result could go unnoticed between runs. The finding is closed on 1 October 2026: it concerned only the build checks, which now run this check on every change.

Result: Recorded status: closed.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Many release checks do not declare the kind of claim they prove

open

A rule requires every release check to state whether it proves a point-in-time or an ongoing property, but about half of the check scripts declare neither. A lint now stops new undeclared checks; the existing declarations are still owed, and as of 12 October 2026 the issue remains open.

Result: Recorded status: open.

Date: 2026-10-12

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

Engine kill-cycle check could pass with no model call

open

A release check of the engine's kill-and-resume cycle counted an agent cycle as passing without any model call being made, and whether it spent money depended on the test machine's gateway. As of 12 October 2026 a correction has been drafted but not applied, and the issue remains open.

Result: Recorded status: open.

Date: 2026-10-12

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

A Cruise check overstated request counts in failure messages

open

A Cruise release check kept counting requests after it reset its test stub, so its failure messages reported more requests than were made. As of 1 October 2026 a correction has been drafted but not applied, and the issue remains open.

Result: Recorded status: open.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

A release check reported setup failures as failed assertions

open

A release check counted any failure of its scenario, including a failure to set the scenario up, as a failed assertion, so a broken setup read as a product failure instead of an aborted run. As of 12 October 2026 a correction has been drafted but not applied, and the issue remains open.

Result: Recorded status: open.

Date: 2026-10-12

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

Engine harness checks left network devices behind on the test machine

closed

Three engine test-harness checks stopped their virtual machines without deleting their sessions, leaving network devices and data behind on the test machine. The finding is closed on 1 October 2026: it concerned only the test harness, whose checks now delete what they create and confirm nothing is left.

Result: Recorded status: closed.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

A failed checkpoint could leave a session paused but shown as running

fixed-in-test

A checkpoint whose pause step timed out could end without taking the snapshot and leave the session's virtual machine paused while its record still read running. The engine now resumes and confirms the machine after any failed checkpoint; the defect is marked fixed in tests on 6 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-06

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Engine could delete leftover database files that may hold usage records

fixed-in-test

When the engine's database file was removed while it ran, the engine deleted leftover journal files that could hold gateway usage records used for billing, instead of setting them aside, and several requests could reopen the database at once. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Cruise job build record did not cover all worker code

fixed-in-test

The worker build a Cruise job records covered only one file of the worker, so changes to the other code the worker runs left the recorded build unchanged. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Resumed Cruise job could restart from its first step

fixed-in-test

In production, a resumed Cruise job could start again from its first step with nothing counted as already spent when it resumed on a worker host without its earlier local record. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Price-change check would ignore changes to per-model rates

fixed-in-test

The check that marks money-related release checks stale after a price change did not watch the per-model rate book, so a future change to ratified rates would not have required those checks to re-run. It was dormant while the rates are unratified; the defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Dollar budget rule would round small per-token rates up

fixed-in-test

The rule for reserving a dollar budget before a call could not express a rate below one micro-dollar per token, so ratified per-million rates would have been rounded up and a dollar limit could refuse calls it had room for. It is dormant while dollar admission is unavailable; the defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Internal record misdescribed how Cruise attempts are metered

closed

An internal finding record said each Cruise attempt is metered from the parent session's usage, when each attempt is metered on its own fork's gateway usage. The finding is closed on 1 October 2026: the record was corrected and no product behavior changed.

Result: Recorded status: closed.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Cruise check passed only after a manual re-signing step

open

A Cruise release check verified a committed test fixture signed with a key no run holds, so it could pass only after a manual re-signing that no run can do. As of 1 October 2026 a correction has been drafted but not applied, and the issue remains open.

Result: Recorded status: open.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

Adapter supervisor dropped settings meant for the agent adapter

fixed-in-test

When run under its supervisor, the agent adapter received only some of its settings: 16 of the 30 it acts on, such as log format and approval timing, were dropped. The engine passed none of the dropped settings, so the effect was latent; the defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Limits page misdescribed unanswered approval requests

fixed-in-test

The published Limits page said an unanswered approval request stays pending, but since 22 September 2026 an unanswered request is closed when its window ends and a later decision is refused. The page source was corrected, and the defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Certification harnesses wrote where the agent runner has no access

closed

Certification test harnesses had the agent runner write files under the administrator's home folder, which fails now that the runner runs as a separate, unprivileged user. The finding is closed on 1 October 2026: it concerned only the test harness, which now writes under the runner's own home.

Result: Recorded status: closed.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Interrupted image publish could leave partial files behind

fixed-in-test

If publishing a new session image was killed or the host crashed partway, a partial copy or an image without its marker could be left in the preserved image store. The defect is marked fixed in tests on 1 October 2026; one narrow interruption point is completed by the next build.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Commit checks on side branches ran the main branch's version

closed

Developer commit checks in every working copy ran the main branch's version of the check, so a check changed on main could refuse commits on branches that did not yet carry its scripts. The finding is closed on 1 October 2026: it concerned only developer tooling.

Result: Recorded status: closed.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Forked session traffic was misaddressed for up to 40 seconds

fixed-in-test

After a fork, the child session's own outbound traffic, gateway calls included, went to the wrong network address for about 15 to 40 seconds until it re-learned the address, because the host side of each session's network link got a random hardware address. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Cancelled Cruise job could leak its parked parent session

fixed-in-test

A Cruise job's parked parent session was known only to the worker host that parked it, so a resume on another host built a new parent and a cancel there left the old one stored and unused. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Cruise test stand-in handled cancellations unlike production

fixed-in-test

The test stand-in for the service handed a cancelled Cruise job back to the worker, while production lists it for teardown, so a Cruise release check proved a path production does not take. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Cruise job build record did not cover model and rate tables

fixed-in-test

The worker build a Cruise job records covered the worker's code but not the model catalogue and rate book it reads, so a change to either left the recorded build unchanged. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Approval unit test failed intermittently under load

fixed-in-test

A unit test of approval requests whose hook had gone failed intermittently under load, because it closed the hook before the adapter knew the request's id. The test was corrected and the defect is marked fixed in tests on 1 October 2026; the underlying timing case in the product was proposed as a separate finding.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Admin account view always showed no device tokens

fixed-in-test

The admin console's account view always showed an empty device-token list because its query named a column that does not exist, and a failed list read was shown as empty. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Monitor incidents could sort slightly out of order on the status feed

fixed-in-test

The billing reconciliation monitor stamped the incidents it opened in whole seconds, so on the public status feed and admin incident list they could sort a fraction of a second out of order against the service's other incidents. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Encrypted sessions never resumed warm and could leave plain copies

fixed-in-test

For sessions with stored-data encryption on, the resume cache could never serve an entry, so they always resumed cold, and with optional cache population turned on it wrote decrypted copies of their data to the host's disk. Encrypted stores now bypass the cache; the defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Engine and gateway could set aside each other's live database files

fixed-in-test

When the shared session database was missing, the engine and the gateway could each set aside its leftover files at the same moment, so one could move the other's live files and session or usage records could become invisible. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Agent image publishing was not safe against an interrupted run

fixed-in-test

If the step that publishes the agent environment image was interrupted, the service could be left with a partial image copy or an image without its integrity marker, and a later run would not repair it. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

On Windows the client showed a pasted key while adding it

fixed-in-test

On Windows, adding a provider key with the client could not hide the pasted key on screen, and the client also could not save terminal settings or read the window width there. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Attaching from a Windows console could print raw escape codes

fixed-in-test

On some Windows consoles, attaching to a session did not turn on terminal escape processing, so the session's formatting codes could print as raw characters; this was not confirmed on a real Windows console. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A member's agent service could act in another account's session

fixed-in-test

Within a shared project, an agent service acting for one account could raise an approval question or reserve spending in a session owned by another account, with the record kept under the wrong account. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Another account could close or read an adviser consultation

fixed-in-test

A caller acting for another account could close a pending or late adviser consultation, or read back an answered one, because the account check ran after those steps. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Billing monitor would flag a false stall when usage starts after idle

fixed-in-test

With a pending usage accounting change, the monitor that checks usage is settled would have read any account that starts using the service after a quiet period as stalled and raised a major incident until the next settlement cycle ran. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A partial refund removed the whole purchased credit

fixed-in-test

A partial refund of a credit purchase removed all the credit the order had granted, and a refund that later failed did not give the credit back. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Account deletion page said sessions were kept when they were erased

fixed-in-test

After deleting an account, the confirmation page said running sessions had been parked at a checkpoint, when in fact every session was erased and its checkpoints made unreadable, so the page suggested work had been kept. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A service test failed rarely on random identifiers

closed

An automated test of the service failed in about 0.03% of runs when a randomly generated identifier happened to contain a number the test treated as a leak; there was no product impact. The finding is closed on 1 October 2026 because it concerned only the test, which was corrected.

Result: Recorded status: closed.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Older session creation used fixed budget figures, not the price list

fixed-in-test

Creating a session through the older interface used built-in default budget figures instead of reading the published price list; the figures matched at the time, but a price list change would not have reached these sessions. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

An approval could stay open after the agent stopped waiting for it

fixed-in-test

If the agent's permission request went away while its approval question was being raised, the question stayed open and could be approved for an action that would never run, until it expired. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Test bench cleanup checks missed leftover processes

closed

The test bench's cleanup checks did not list leftover processes still holding the product's isolated networks, so a run could report a clean teardown when something was left behind. The finding is closed on 1 October 2026 because it concerned only the test harness.

Result: Recorded status: closed.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Bench test drivers looked for the wrong file names

closed

Two test bench drivers looked for files under names the scripted test service never wrote, so one check of a denied action could pass on missing evidence. The finding is closed on 1 October 2026 because it concerned only the test harness.

Result: Recorded status: closed.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Storage held by a failed Cruise job was never released

fixed-in-test

When a Cruise job failed, the parked session it kept for resuming was never released, and the promised seven-day expiry was not implemented, so that storage could be held indefinitely. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Cruise jobs on several engines could leave parked sessions behind

fixed-in-test

When the service runs more than one engine, a Cruise job's parked session on one engine could neither be reused nor cleaned up by a worker on another engine, and one engine's cleanup could end cleanup on the others. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Simulated Cruise service recorded events production never records

fixed-in-production

A simulated control service used in Cruise tests recorded events the real worker never sends and accepted verdicts the real service refuses, so tests run on it could read a job history that differs from production. The defect is marked fixed in production on 12 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-12

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Two Cruise test modules failed when run in one process

closed

One Cruise test left its temporary settings in place after finishing, so another test failed when both ran in the same process; automated checks were not affected and there was no product impact. The finding is closed on 1 October 2026 because it concerned only the tests.

Result: Recorded status: closed.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

A Cruise acceptance check could not pass with current clients

closed

Since client version 0.1.13, one Cruise acceptance check could not pass with any client, because its simulated service did not answer the model catalog and key list requests the approve step now makes. The finding is closed on 4 October 2026 after the check was amended and passed.

Result: Recorded status: closed.

Date: 2026-10-04

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

A billing safety check no longer matched the duplicate charge rule

fixed-in-test

A billing safety check and the usage meter report counted duplicates by request rather than by attempt, which no longer matched the rule the database enforces, so the check failed and could not tell a real double charge from a legitimate repeat. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Admin and activation test scenarios could reach a previous server

open

Several admin console and automatic activation test scenarios restarted the control service on the same port while the previous one was still shutting down, so a run could reach the old service and its result could not be trusted. The issue remains open.

Result: Recorded status: open.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

Two release checks leave a network device behind on each run

open

Two release checks leave one virtual network device behind on the test bench each time they run, because their cleanup does not remove what they create. The issue remains open.

Result: Recorded status: open.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

A retried call after resume could be run and billed twice

fixed-in-production

From 20 September to 6 October 2026, if a session was paused while a model call was in flight and its answer had already been sent, the retry after resume was run again and billed again, so a customer could pay the provider twice on their own key and be metered twice; the source states this applied to the production service. The defect is marked fixed in production on 13 October 2026: the fix has been in the production gateway since 6 October 2026, and the paid gateway test re-ran green on 13 October 2026.

Result: Recorded status: fixed-in-production. Earlier recorded status: fixed-in-test on 13 October 2026.

Date: 2026-10-13

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: critical. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Repeated identical requests could be answered from storage unbilled

fixed-in-test

After one lost reply was replayed, every later identical request in that session was answered from the stored reply instead of being run, so new work was neither performed nor billed and could return an old answer. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A delivery failure record was saved where the retry check never looked

fixed-in-test

When sending an answer failed partway through, the record of that failure was saved under a different key from the one the retry decision reads, so the retry decision never saw it. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A session start test failed when a random id matched a pattern

closed

An automated test occasionally failed when a randomly generated session identifier happened to contain a number the test treated as a leak; there was no product impact. The finding is closed on 1 October 2026 because it concerned only the test, which was corrected.

Result: Recorded status: closed.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

A resumed session's machine once failed to connect its network

closed

In one test run, a resumed session's virtual machine never connected to its network device, and the cause was not established. The finding is closed on 5 October 2026 because it was not reproduced in more than 450 further resumes, and a permanent check was added.

Result: Recorded status: closed.

Date: 2026-10-05

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

A forked session's network address could change after setup

fixed-in-test

For a forked session, the host could replace the hardware address of its network link after the session's network had already learned it, which could disrupt the forked session's connectivity; this was not observed. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

The test bench clock was adjusted by two competing time services

closed

The test machine's clock ran fast and was stepped back every ten seconds because two time services adjusted it, making timing on the bench unreliable; there was no product exposure. The finding is closed on 1 October 2026 because it concerned only the test bench configuration, which was corrected.

Result: Recorded status: closed.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. Closure applies to this recorded case, not every workload.

Workaround: No verified customer workaround is recorded.

Forked sessions could be corrupted when host memory ran low

fixed-in-production

When the host ran short of memory, restoring several forked sessions in turn could corrupt the memory of children already running, causing them to hang or crash. The defect is marked fixed in production on 6 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-06

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Cruise and setup screen budgets used fixed figures, not the price list

fixed-in-test

Cruise job sessions used built-in budget figures instead of the published price list, and the onboarding setup screen showed every account the paid tier budget, so a free account was shown four times the budget its session received. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

The price consistency check treated test files as prices

fixed-in-test

An internal check that keeps prices in one source flagged durations in test files as prices, so a pre-release check failed while automated checks passed; there was no product impact. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

The link on the account deletion page did not work

fixed-in-test

After deleting an account, the link on the confirmation page pointed to a broken address on the service instead of the KeepState website. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A fleet test checks a fixed free budget, not the price list

open

One fleet acceptance test checks a fixed free tier budget figure instead of the figure in the price list, so it would fail if the free budget were lowered, even with correct service behavior. The issue remains open.

Result: Recorded status: open.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

An interrupted deploy could leave an agent image the engine cannot use

fixed-in-test

If a deploy was interrupted while publishing an agent image, the image could be left without its description file, and later deploys reported it as already built without repairing it, so the engine could not start agents from it. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Cruise sessions could run on a guessed token budget when none was given

fixed-in-production

When a Cruise job's claim carried no token budget, its sessions ran on a fixed fallback budget instead of the price book figure for the account's tier; such a claim is now refused before any session starts. The defect is marked fixed in production on 12 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-12

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

A permission request could stay open after its reply was lost

fixed-in-test

If an agent's permission request reached the service but the reply was lost in transit, the tool call was refused yet the request stayed open and decidable until its time limit, with nothing able to close it. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

No notification was sent when a Cruise job moved to review

fixed-in-production

A published claim said a Cruise job whose attempts were exhausted reached the review queue with a notification, but no part of the service sent or recorded one. The service now emails the account and records the notification, or records why it was not sent. The defect is marked fixed in production on 12 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-12

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

A failed Cruise job did not say why it failed

fixed-in-test

When a Cruise job failed, the reason was not passed to the service, so the failure record a customer reads carried no error, contrary to the documented behavior. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Cruise job status showed an empty reason for jobs in review

fixed-in-test

The status of a Cruise job waiting for review always reported an empty review reason, because it read a field the service never records. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A Cruise test stand-in accepted attempt changes the service refuses

fixed-in-production

The simulated control plane used in Cruise tests kept and accepted attempt data the real service does not, including changes after a job had stopped, so one test check read a value the service never stores. The defect is marked fixed in production on 12 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-12

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Evidence sessions kept by failed Cruise jobs were never released

fixed-in-test

Sessions that a failed Cruise job kept parked as evidence were never released, so they stayed parked after the job had ended. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

The Cruise runner recorded a parent released when its deletion failed

fixed-in-test

When deleting a Cruise job's parked parent session failed at acceptance, the runner's own record still said the parent was released while the session stayed parked; the service's release listing was not affected. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A production check's cleanup could leave a job's parent session behind

open

The cleanup step of one production Cruise check deleted a job's records directly before the normal teardown ran, which could leave the job's parent session running and some records orphaned. A correction has been drafted but not applied. The issue remains open.

Result: Recorded status: open.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

A Cruise test stand-in used a fixed budget instead of the price book

fixed-in-production

The simulated control plane used in Cruise tests gave every claim a fixed token budget instead of reading the price book figure that the service uses. The defect is marked fixed in production on 12 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-12

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

An interrupted agent image build could leave an unusable image

fixed-in-test

If building and publishing an agent image was interrupted, the image or its description file could be left incomplete, and a rebuild reported it as already published without checking, so the engine could not use it. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A lost withdrawal could leave a permission request decidable

fixed-in-test

When an agent withdrew a permission request and that withdrawal was lost in transit, it was not sent again, so the request stayed open and decidable until its time limit. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

The Cruise contract named the wrong stop reason for a refused attempt

fixed-in-test

The Cruise contract said an attempt stopped for lack of a usable inference key would record one stop reason, while the service records a different one; the contract text was corrected to match and no code changed. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

The administration overview could stall on a hanging engine

fixed-in-test

Loading the service's administration overview asked the engine for its live session count every time, so a hanging engine could hold the page for up to a minute; the count now comes from the regular health probe. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Administration pages showed lists that could not be read as empty

fixed-in-test

Several administration pages showed a list that could not be read as empty or left it out, and the billing reconciliation view could skip unreadable statements and report no drift. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

The public status feed read operational when incidents were unreadable

fixed-in-test

If the service could not read its incident records, the public status feed reported the service as operational instead of unavailable. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A test scenario stamped its incident in local time marked as UTC

open

One test scenario recorded the time of the incident it creates in whole seconds and in local time while labelling it as UTC. A correction has been drafted but not applied. The issue remains open.

Result: Recorded status: open.

Date: 2026-10-01

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: low. The issue remains open in this dated source; current exploitability or impact was not independently tested.

Workaround: No verified customer workaround is recorded.

The website status footer showed a critical incident as paused

fixed-in-test

When the status feed reported a major outage, the website's status footer had no display for it and showed live measurement as paused instead of an incident in progress. The change had not been merged or deployed at the time of this record. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

The administration accounts list showed zero when a read failed

fixed-in-test

The administration accounts list showed figures that could not be read as zero and silently dropped accounts whose rows could not be read, and its export could be written short. The defect is marked fixed in tests on 1 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-01

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A cancel during an attempt report was logged as a runner crash

fixed-in-production

If a customer cancelled a Cruise job while an attempt was still being reported, the job was torn down correctly but the worker recorded a failure and logged a crash. The defect is marked fixed in production on 12 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-12

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

A cancel just after a Cruise job was claimed was logged as a crash

fixed-in-production

If a customer cancelled a Cruise job between its claim and the start of its work, the worker recorded a failure and logged a crash instead of a clean teardown. The defect is marked fixed in production on 12 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-12

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

A stopped Cruise job could still gain new attempt records

fixed-in-production

The service accepted new attempt records for a Cruise job that had already stopped (in review, accepted, failed or cancelled), so a worker still running after a cancel could add rows to a stopped job's record. The defect is marked fixed in production on 12 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-12

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Resumed sessions could stop answering shortly after a resume

fixed-in-test

A session that answered after being resumed from a checkpoint could stop responding 10 to 60 seconds later and never recover, so a resumed session could be silently lost; the cause was identified in the engine and a fix made there. The defect is marked fixed in tests on 6 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-06

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A test stand-in made one engine test fail intermittently

fixed-in-test

A stand-in program used by one engine test never read its input, so the test failed intermittently with a broken pipe and turned one build red; the product was not affected. The defect is marked fixed in tests on 12 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-12

Scope: Test or evidence-quality finding; customer production behavior is not established by this record.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

A resume whose session died could hold the engine for about 20 minutes

fixed-in-test

If a session's virtual machine died or stalled during a resume's clock step, the resume could take about 20 minutes, blocking every checkpoint, resume, fork and kill on that engine meanwhile, and could then report success for a session that was gone. The defect is marked fixed in tests on 6 October 2026.

Result: Recorded status: fixed-in-test.

Date: 2026-10-06

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. A test fix does not establish a production fix.

Workaround: No verified customer workaround is recorded.

Older sessions could not be erased on their own

fixed-in-production

Sessions saved before a 6 October 2026 change could not be erased individually, and content from their earlier checkpoints stayed readable for as long as the account's key lived. Existing accounts were re-sealed to remove this. The defect is marked fixed in production on 10 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-10

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: high. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Client app help, confirmations and typed commands behaved wrongly

fixed-in-production

In the ks client's app, help on a command group did not list its commands, a typed ks command was sent to the agent as an instruction (a paid model call), confirmation steps blamed a flag nobody had typed, and kill printed its plan as raw JSON. The defect is marked fixed in production on 11 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-11

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Deletion plans said erasure was incomplete for every session

fixed-in-production

A session's deletion plan always reported that erasure would not be complete, even for a session whose deletion erases everything, because it only repeated the general retention statement. The defect is marked fixed in production on 11 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-11

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Console actions relied on a cookie setting alone against forged requests

fixed-in-production

The console's signed-in actions, including approving a device login, were protected only by the session cookie's same-site setting, so a page on another subdomain of the service's domain could have submitted them for a signed-in customer; every such subdomain belonged to KeepState, so this was defence in depth. The defect is marked fixed in production on 12 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-12

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

An open Agents page stream delayed restarts and was cut without notice

fixed-in-production

An open event stream on the console's Agents page held a service restart for the full 30 second shutdown limit and was then cut without an end message, so the page reported a lost connection instead of reconnecting at once. The defect is marked fixed in production on 12 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-12

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Asynchronous Agents page operations ran without the signed-in account

fixed-in-production

An operation that the console's Agents page ran asynchronously lost the signed-in account and was refused as unauthenticated; no page used this path before the new agent flow, so nothing in production reached it. The defect is marked fixed in production on 12 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-12

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: medium. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

The client accepted agent names that the service then refused

fixed-in-production

The ks client accepted some agent and session names, such as ones starting with a digit, containing a dot or longer than 48 characters, that the service refuses, so the request failed only after it was sent. The defect is marked fixed in production on 12 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-12

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.

Session journals showed customers internal provisioning detail

fixed-in-production

A session's event journal showed its readers internal provisioning detail, such as internal identifiers and engine configuration errors, that the provisioning answer deliberately withheld; nothing crossed between accounts. Journal entries written before the fix are kept as they were. The defect is marked fixed in production on 12 October 2026.

Result: Recorded status: fixed-in-production.

Date: 2026-10-12

Scope: Dated product finding; current deployment has not been independently verified.

Limit: Recorded severity: low. The register reports a production fix; this publication did not independently retest it.

Workaround: No verified customer workaround is recorded.