test/crew-interrupted.test.ts runs in under 10 s #11062

test/crew-interrupted.test.ts now runs alone in about 8.4–9.0 s, down from about 65 s, with the same 14 cases and the same 102 expect() calls. Only this test file changed. Nothing in src/ changed and scripts/gate.ts is untouched.

Speed. Almost all of the time was cases waiting for the crew's default 5 s tick. startCrew already accepts tickMs, so every crew in the file now gets tickMs: 50 and networkRetryScale: 0. Smaller cuts:

• Wait loops poll every 20 ms instead of 100–200 ms.
• The #7799 drain case uses drainMs: 250 instead of 1 s. It still asserts the stop finishes well inside 15 s.
• The drain-and-land stand-in sleeps 0.5 s instead of 2 s, which is still long enough for the run to be in flight when the drain starts.

What's left is mostly real runner work: each kept-clone case does two real runs (clone, stand-in, push, open pull), about 0.5–0.9 s per case. Loading the imports takes about 1 s.

Order (--randomize). On main the file only passed in written order:

• The lease-lapse case reused the garden repo and the idle crew that the first case left behind.
• Runs left queued by earlier cases took the single slot that later cases needed.

Now every case in the first describe starts with atRest(), which stops any crew left up and cancels runs left queued. garden() gets or creates the shared repository, and idleCrew() starts the concurrency-0 crew. The lease-lapse case starts its own idle crew before the lease lapses, so it still tests what happens while a crew is up. The two cases in the second describe were already independent.

Checked here: two seeds that failed before the order fix now pass, as do three new --randomize runs and the default order (8.4–9.0 s each). bun scripts/gate.ts and bun x tsc --noEmit pass. The keep_plan tool the issue asks for isn't available in this session, so the issue has no plan on it.

To try it: bun test --randomize test/crew-interrupted.test.ts

🤖 Generated with Claude Code

Closes #11053

test/crew-interrupted.test.ts now runs alone in about 8.4–9.0 s, down from about 65 s, with the same 14 cases and the same 102 expect() calls. Only this test file changed. Nothing in src/ changed and scripts/gate.ts is untouched.

Speed. Almost all of the time was cases waiting for the crew's default 5 s tick. startCrew already accepts tickMs, so every crew in the file now gets tickMs: 50 and networkRetryScale: 0. Smaller cuts:

• Wait loops poll every 20 ms instead of 100–200 ms.
• The #7799 drain case uses drainMs: 250 instead of 1 s. It still asserts the stop finishes well inside 15 s.
• The drain-and-land stand-in sleeps 0.5 s instead of 2 s, which is still long enough for the run to be in flight when the drain starts.

What's left is mostly real runner work: each kept-clone case does two real runs (clone, stand-in, push, open pull), about 0.5–0.9 s per case. Loading the imports takes about 1 s.

Order (--randomize). On main the file only passed in written order:

• The lease-lapse case reused the garden repo and the idle crew that the first case left behind.
• Runs left queued by earlier cases took the single slot that later cases needed.

Now every case in the first describe starts with atRest(), which stops any crew left up and cancels runs left queued. garden() gets or creates the shared repository, and idleCrew() starts the concurrency-0 crew. The lease-lapse case starts its own idle crew before the lease lapses, so it still tests what happens while a crew is up. The two cases in the second describe were already independent.

Checked here: two seeds that failed before the order fix now pass, as do three new --randomize runs and the default order (8.4–9.0 s each). bun scripts/gate.ts and bun x tsc --noEmit pass. The keep_plan tool the issue asks for isn't available in this session, so the issue has no plan on it.

To try it: bun test --randomize test/crew-interrupted.test.ts

🤖 Generated with Claude Code

test/crew-interrupted.test.ts now runs in about 9 s and passes in any order (#11053). It keeps all 14 cases. src/ and scripts/gate.ts are not touched.

Speedup (first commit): every crew in the file starts with the tickMs option startCrew already accepts, set to 50 ms instead of the 5 s default, and with networkRetryScale: 0. Waits check every 20 ms, and the drain-budget case drains for 250 ms.

Order independence (second commit, from review):

• No shared repository: the two lease cases used to share a lazily created garden repository. Each now makes its own (garden, allotment) through a crewedRepo(name) helper. That adds one addToCrew expect, so the file now makes 103 expect calls.
• No leftover runs: atRest() now cancels runs left running as well as queued. With no crew up, a running run can only be left over from an earlier case. A later crew would count it against its single slot, or requeue it once its lease lapsed.
• Drain-and-land waits for the end: after stop({ drain: true }), the case polls the run until it is ok, failed or cancelled (up to 20 s). It no longer reads the state the moment stop() returns. The same settled() helper replaces the local wait in keptClone.
• Boot case timeout: the boot case gets a 30 s timeout like the cases around it. It was on bun's 5 s default, but it first stops the previous crew, and that stop can wait up to 10 s for the crew's background sweeps.

The seed failures from the review did not reproduce here, before or after the change. On the original commit, all 40 random seeds passed, and so did the three named seeds (3068874285, 518429626, 1427295368), both at normal speed and with six copies pinned to one core (about 77 s each). After the change, all 36 runs passed 14/14: those three seeds plus 30 random ones, and the three named seeds again pinned to one core (about 42 s). Seed-to-order mapping may differ between bun builds (here: 1.3.14), so your seeds may not give the order you saw. The changes remove the shared state behind each failure you described, but they are not confirmed against your environment.

Under random order, a drain-and-land run still running after stop() can only belong to a different crew or be a leftover. Draining waits up to 30 s, so the 0.5 s stand-in and the 50 ms tick were not the cause. The case now waits on the run's terminal state as asked.

To try it: bun test --randomize --seed=3068874285 test/crew-interrupted.test.ts (and the other seeds). bun scripts/gate.ts and bun x tsc --noEmit are clean.

🤖 Generated with Claude Code

Left #11053 open: 2 unticked items remain (The file runs in under 10 s alone, with the same 14 cases still asserting the same things. Use an injected clock or shorter injected timeouts, not fewer cases.; It passes three times with --randomize.).