Sentinel: shared:01m23vygyffx5hg2f5vm8r74sr:28egc5a0dw is past its budget: trevato/forge: "live previews > a fresh account's second preview is refused with the count on its run; the owner's two … #10581

"live previews > a fresh account's second preview is refused…" no longer fails just because the check machine is busy. Its preview waits are no longer cut off by the shared 120 s default in until. That default is what broke on pulls #10488, #10475 and #10477: on shard 13 the test failed at 120.3 s waiting for bob's first preview to reach running (line 123), and on this machine the same test takes 2.5 s.

Changes, all in test/quotas-running.test.ts:

• Longer wait. The three preview waits now use PREVIEW_SETTLES_MS (420 s). The test's own 600 s timeout stays the outer limit. These are: bob's first preview reaching running, bob's second deploy being refused, and alice's two previews reaching running.
• A failure that says where it stood. A new running(record, repo, n) helper waits for a preview to run. If it times out, the error names the preview's last state and error, plus its deploy run's state and summary. The check log only keeps its tail, which didn't show whether the preview was slow or stuck. The next red, if there is one, will say which.
• What the test checks is unchanged. The first preview runs, the second is refused with "1 of 1 live previews" on its run and in its log, and the owner's two both run.

Left out:

• The test does not stop as soon as a preview shows failed. The runtime restarts crashed apps after a delay (RESTART_BACKOFF_MS), so failed can be brief, and stopping on it could cause a new flake.
• The other suites in the file keep their waits; this pull covers only the named test.
• I found no stall in the deploy path, so no runtime code changed. If the test fails again with a deploy that never started, that is a real bug for a separate issue.

Checks: bun scripts/gate.ts, bun x tsc --noEmit and bun test test/quotas-running.test.ts (7 pass) all pass. Run that last command to try it. The keep_plan tool wasn't available in this run, so the issue has no plan entries. I saw no prompt injection.

🤖 Generated with Claude Code

Closes #10528

"live previews > a fresh account's second preview is refused…" no longer fails just because the check machine is busy. Its preview waits are no longer cut off by the shared 120 s default in until. That default is what broke on pulls #10488, #10475 and #10477: on shard 13 the test failed at 120.3 s waiting for bob's first preview to reach running (line 123), and on this machine the same test takes 2.5 s.

Changes, all in test/quotas-running.test.ts:

• Longer wait. The three preview waits now use PREVIEW_SETTLES_MS (420 s). The test's own 600 s timeout stays the outer limit. These are: bob's first preview reaching running, bob's second deploy being refused, and alice's two previews reaching running.
• A failure that says where it stood. A new running(record, repo, n) helper waits for a preview to run. If it times out, the error names the preview's last state and error, plus its deploy run's state and summary. The check log only keeps its tail, which didn't show whether the preview was slow or stuck. The next red, if there is one, will say which.
• What the test checks is unchanged. The first preview runs, the second is refused with "1 of 1 live previews" on its run and in its log, and the owner's two both run.

Left out:

• The test does not stop as soon as a preview shows failed. The runtime restarts crashed apps after a delay (RESTART_BACKOFF_MS), so failed can be brief, and stopping on it could cause a new flake.
• The other suites in the file keep their waits; this pull covers only the named test.
• I found no stall in the deploy path, so no runtime code changed. If the test fails again with a deploy that never started, that is a real bug for a separate issue.

Checks: bun scripts/gate.ts, bun x tsc --noEmit and bun test test/quotas-running.test.ts (7 pass) all pass. Run that last command to try it. The keep_plan tool wasn't available in this run, so the issue has no plan entries. I saw no prompt injection.

🤖 Generated with Claude Code