test/apps-api.test.ts runs in under 8 s #11060

test/apps-api.test.ts used to take 19–23 s alone; it now takes about 6.2–7.3 s on this shared machine (load average 40–70). All 12 cases are kept. It passed five times in a row with --randomize, and once in file order. scripts/gate.ts is untouched; the gate and bun x tsc --noEmit are green.

Where the time went: 13.7 s of it was one case waiting for agent runs. The crew only checks its queue every 5 s, and the runs queued one behind another.

What changed (only this test file):

• Faster crew tick: the crew is started with tickMs: 50, so a queued run starts right away.
• Waiting on events: untilRun(repo, where) now wakes when the app's event bus reports a finished run (run.finished), instead of sleeping 100 ms between polls. The listener is attached before each read, so a run that finishes in between still wakes it. The 30 s limit only catches a run that never finishes.
• One app per case that changes one: before, the cases shared hello and depended on running in order, since later cases stopped it, made it private, deleted it and retired bot. Now opened, envy, hidden, crewed and checked each belong to one case. They are all created at the same time while the forge boots, so their first deploys overlap. The cases that only read share a hello app, made once by whichever case asks first. bot is created in beforeAll. The "create" case makes its own maker agent and the "retire" case its own retiree.
• Pick case: its four apps are created at the same time.
• Disable case: the sleep loop is gone. The proxy reads the app's database row on every request, and the DELETE removes that row before it answers, so the 404 is checked once, directly.

Assertions widened because other cases now share the forge:

• The agent list must contain maker, where it used to equal ["bot"].
• The public app list must contain hello, where it used to equal ["hello"].
• After hidden is made private, the public list must not contain it, where the list used to have to be empty.
• The leftover-credential check looks for this run's own token, not every run:% token.
• firstPreviewMs clears serving_at for hello's preview runs only, not every repository's.

Not done:

• Each deploy still waits on the app process's 250 ms health check in src/apps/process.ts. That code is outside this issue's files.
• About 1 s of each run is importing the code, and the rest is real deploys. Under heavy load a run can still come close to 8 s; I saw 8.0–9.4 s on the heavier runs before the last two speed-ups.
• The issue asks for a keep_plan tool; it isn't available in this session, so no plan was posted.

Try it: bun test test/apps-api.test.ts --randomize (three times), or bun test test/apps-api.test.ts.

🤖 Generated with Claude Code

Closes #11054

test/apps-api.test.ts used to take 19–23 s alone; it now takes about 6.2–7.3 s on this shared machine (load average 40–70). All 12 cases are kept. It passed five times in a row with --randomize, and once in file order. scripts/gate.ts is untouched; the gate and bun x tsc --noEmit are green.

Where the time went: 13.7 s of it was one case waiting for agent runs. The crew only checks its queue every 5 s, and the runs queued one behind another.

What changed (only this test file):

• Faster crew tick: the crew is started with tickMs: 50, so a queued run starts right away.
• Waiting on events: untilRun(repo, where) now wakes when the app's event bus reports a finished run (run.finished), instead of sleeping 100 ms between polls. The listener is attached before each read, so a run that finishes in between still wakes it. The 30 s limit only catches a run that never finishes.
• One app per case that changes one: before, the cases shared hello and depended on running in order, since later cases stopped it, made it private, deleted it and retired bot. Now opened, envy, hidden, crewed and checked each belong to one case. They are all created at the same time while the forge boots, so their first deploys overlap. The cases that only read share a hello app, made once by whichever case asks first. bot is created in beforeAll. The "create" case makes its own maker agent and the "retire" case its own retiree.
• Pick case: its four apps are created at the same time.
• Disable case: the sleep loop is gone. The proxy reads the app's database row on every request, and the DELETE removes that row before it answers, so the 404 is checked once, directly.

Assertions widened because other cases now share the forge:

• The agent list must contain maker, where it used to equal ["bot"].
• The public app list must contain hello, where it used to equal ["hello"].
• After hidden is made private, the public list must not contain it, where the list used to have to be empty.
• The leftover-credential check looks for this run's own token, not every run:% token.
• firstPreviewMs clears serving_at for hello's preview runs only, not every repository's.

Not done:

• Each deploy still waits on the app process's 250 ms health check in src/apps/process.ts. That code is outside this issue's files.
• About 1 s of each run is importing the code, and the rest is real deploys. Under heavy load a run can still come close to 8 s; I saw 8.0–9.4 s on the heavier runs before the last two speed-ups.
• The issue asks for a keep_plan tool; it isn't available in this session, so no plan was posted.

Try it: bun test test/apps-api.test.ts --randomize (three times), or bun test test/apps-api.test.ts.

🤖 Generated with Claude Code

Left #11054 open: 1 unticked item remain (The file runs in under 8 s alone, with every case kept. Share one booted forge across cases where they do not mutate the same app, and wait on events, not sleeps.).