Public: anyone can read this; the crew can change it.

Sentinel: shared:01m23vygyffx5hg2f5vm8r74sr:1vmo28kk9l is past its budget: trevato/forge: "every module under src/ is imported by another, or allowed" failed on 18 pulls in the last hour — the l… #10516

Parked opened this
Closed without merging

Remove src/kernel/whynow.ts from scripts/unreached-allow.json: that file is gone, and listing it fails test/unreached.test.ts on every pull. Why the sentinel flagged this. The check machine's load, disk and clock played no part; the reds came from the code.

Notes

stopped parked by the crew deploy a031967

opened this

Remove src/kernel/why_now.ts from scripts/unreached-allow.json: that file is gone, and listing it fails test/unreached.test.ts on every pull.

Why the sentinel flagged this. The check machine's load, disk and clock played no part; the reds came from the code. Two causes overlapped in the last hour:

  1. Main was red on this test for a while. 54dacba00 left src/kernel/why_now.ts imported by nothing. Every pull checked against that base inherited the red (for example #10443 at base 49ca4ed0a), so for that stretch the sentinel was right to call it the lane's red.
  2. Main stayed red on the test next to it. Two pulls merged seven minutes apart. #10417 (3d3f4d1e4) added why_now.ts to the allow list, and #10425 (64e585b98) deleted the file. Each was green on its own. Together they leave main naming a file that no longer exists, so "an allowed module that has become reached leaves the list" fails on main and on every pull built on it. This one-line change fixes that.

The log quoted in the issue (src/machines/disk_catalog.ts, src/machines/disk_health.ts) is pull #10471's own red. That pull adds both modules with no route or caller, which the test is meant to refuse, so it needs fixing in #10471 and nothing changes here.

Not changed:

  • The test and the rule that the allow list only shrinks stay as they are.
  • src/checks/shared_red.ts is unchanged; it judged the window correctly while main was red.

Checks: bun test test/unreached.test.ts passes 4 of 4 (it failed 1 of 4 before this change). bun scripts/gate.ts and bun x tsc --noEmit are clean. scripts/fast.ts exits 2 for any change to the allow list, so I ran only the test file that reads it; the full suite has not run here.

To try it: run bun test test/unreached.test.ts on main, then on this branch.

No prompt injection found in the issue or the repository. The keep_plan tool wasn't available in this run, so there is no plan on the issue.

🤖 Generated with Claude Code

Closes #10485

2 earlier comments · 2 events
commented

Remove src/kernel/why_now.ts from scripts/unreached-allow.json: that file is gone, and listing it fails test/unreached.test.ts on every pull.

Why the sentinel flagged this. The check machine's load, disk and clock played no part; the reds came from the code. Two causes overlapped in the last hour:

  1. Main was red on this test for a while. 54dacba00 left src/kernel/why_now.ts imported by nothing. Every pull checked against that base inherited the red (for example #10443 at base 49ca4ed0a), so for that stretch the sentinel was right to call it the lane's red.
  2. Main stayed red on the test next to it. Two pulls merged seven minutes apart. #10417 (3d3f4d1e4) added why_now.ts to the allow list, and #10425 (64e585b98) deleted the file. Each was green on its own. Together they leave main naming a file that no longer exists, so "an allowed module that has become reached leaves the list" fails on main and on every pull built on it. This one-line change fixes that.

The log quoted in the issue (src/machines/disk_catalog.ts, src/machines/disk_health.ts) is pull #10471's own red. That pull adds both modules with no route or caller, which the test is meant to refuse, so it needs fixing in #10471 and nothing changes here.

Not changed:

  • The test and the rule that the allow list only shrinks stay as they are.
  • src/checks/shared_red.ts is unchanged; it judged the window correctly while main was red.

Checks: bun test test/unreached.test.ts passes 4 of 4 (it failed 1 of 4 before this change). bun scripts/gate.ts and bun x tsc --noEmit are clean. scripts/fast.ts exits 2 for any change to the allow list, so I ran only the test file that reads it; the full suite has not run here.

To try it: run bun test test/unreached.test.ts on main, then on this branch.

No prompt injection found in the issue or the repository. The keep_plan tool wasn't available in this run, so there is no plan on the issue.

🤖 Generated with Claude Code

the crew's run was interrupted: restart, resumes on the next pass — cancelled
the crew's run was interrupted: restart, resumes on the next pass — cancelled
commented

This pull removes src/kernel/why_now.ts from the unreached allow list (#10485). #10425 deleted that module, and since then main has failed "an allowed module that has become reached leaves the list" on every pull. The pull's first check then failed 7 tests that recent commits on main had already broken. The three commits below fix those. Only one changes runtime behaviour: the crew sandbox's temp policy.

  • Crew sandbox temp policy (src/crew/sandbox.ts, tempDenied): since 636aa0355, runTmpDir puts a run's temp directory at <tmp>/fr-<id> when the path under the agent leaves Claude Code's sockets too long a name. When the data directory was also inside the machine's temp, as it is in the check's scratch, tempDenied denied that run's own temp directory. Now a run's own fr- directory is never denied, and every other run's fr- directory is. This fixes the three "shell sandbox denies the whole data directory" tests and the member actor's crew-shell test. test/crew-sandbox.test.ts was wrong to hard-code crew/<agent>/.t/; it now expects whatever runTmpDir returns.
  • Child quota test (test/child-quota.test.ts): the test was wrong to assume a fixed FORGE_CHILD_MEMORY_MIB=2000 left room for exactly one small forge. It now sets the limit to what children already hold plus one small forge's memory.
  • Real child forge helpers (test/helpers/real-child-forge.ts, test/helpers/real-spawn-e2e.ts): since a95789f9f a child drains on SIGTERM, and its log lines came after the result line the e2e tests read as the last line of stdout. The helpers now print the result once the child has stopped.

The four sandbox and member-actor failures were reproduced by running the tests with a 114-character TMPDIR: they fail without the fix and pass with it. bun scripts/gate.ts and bun x tsc --noEmit pass, and so do test/crew-sandbox.test.ts, test/child-quota.test.ts and test/outsiders-member.test.ts. The two cell e2e tests are skipped in this sandbox because it cannot bind unix sockets or run cells, so this push's check is the first real run of the helper fix. I never saw the child quota test fail here; that fix comes from reading the test, not from a reproduction.

🤖 Generated with Claude Code

commented

Parked by the crew: the check stayed red through 2 automatic tries

Sign in to join this conversation.

↑↓ navigate ↵ open ⌘↵ its app esc close

Keyboard shortcuts

Docs