You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Gap e2e: witnesses owed for session-directory removal (review of #164: F2, F4, F5, F6) #168
Follow-ups from the review of #164 (#150's fix). Findings F4, F5 and F6 were not fixed in #164, and neither was a witness for F2(a). Each is a gap in what the gap report's end-to-end oracle judges. None is a known defect.
F2(a) and F4: no history makes a session directory's removal fail.
removeSession (packages/oracles/oracle-gaps/paths.mjs) catches the failure and leaves the directory for the next run. The catch keeps one done line: without it, a throw after the done line would reach uncaughtException and end the run a second time.
A mutant that removes the catch survives every history.
Witness owed: a test-only fault that makes the directory unremovable (under G43/R57). The removal then fails, the run ends with one done line and its own exit code, and the next run removes the directory.
The README does not say that a directory which cannot be removed is left. The witness should land with that sentence.
F2(b): the pid-reuse guard has no witness.
removeSession keeps the directory of a pid that a live process runs under. No history reuses a pid.
F5: each removal point is judged on only some paths.
The supervisor's removal is judged only by abort histories. A SIGINT or SIGTERM stop ends the child through process.exit, and the session's exit listener removes the directory first.
The exit listener is judged only on win32 (its hookStops histories), whose e2e jobs are non-blocking. POSIX's one unsupervised history stops at its note, before a session opens.
Witness owed: a POSIX unsupervised history that opens a session and is stopped. Today, a stop while the witness spins leaves the test's own child on unsupervised POSIX (C 2026-11-13).
F6: whether a stopped run was killed is read from the platform, not observed.
The killed field of the stopped observation (gap-report-oracle.property.test.mjs) treats every win32 abort as killed.
More follow-ups from the delta review of #164, which found #164 READY at ef875d8:
README sentence. When a run that completed, or a stopped run, cannot remove its session directory (removeSession catches the failure, for example a file another process holds open on win32), the directory is left for the next run. The README says only "which it removes when it ends". The sentence that says what happens when the removal fails should land with the F2(a)/F4 witness.
Stale-claim scenarios. The stale-claim scenarios (claimScenarioRuns in gap-report-oracle.property.test.mjs) use the synthetic pid 999999. On Linux that pid is below the largest pid_max (2^22), so a live process can have it, the same exposure as F7. If one does, a "stale" claim reads as live, and the refusal names another cause. The fix is the one F7 got: a pid that no process can have, such as 2^30 + 1.
D1: the order of the kill. The kill stop (stoppedRun) SIGKILLs the check's process group first and the CLI second. Between the two signals, the supervisor can see its child end and start removing its session directory, which is the path a killed CLI must not take. Kill the supervisor first, then the group, so that no code of the run runs after the kill. The child's watchdog would end the group anyway, and the driver's group kill makes that deterministic.
More follow-ups from the delta review of #164, which found #164 READY at ef875d8:
README sentence. When a run that completed, or a stopped run, cannot remove its session directory (removeSession catches the failure, for example a file another process holds open on win32), the directory is left for the next run. The README says only "which it removes when it ends". The sentence that says what happens when the removal fails should land with the F2(a)/F4 witness.
Stale-claim scenarios. The stale-claim scenarios (claimScenarioRuns in gap-report-oracle.property.test.mjs) use the synthetic pid 999999. On Linux that pid is below the largest pid_max (2^22), so a live process can have it, the same exposure as F7. If one does, a "stale" claim reads as live, and the refusal names another cause. The fix is the one F7 got: a pid that no process can have, such as 2^30 + 1.
D1: the order of the kill. The kill stop (stoppedRun) SIGKILLs the check's process group first and the CLI second. Between the two signals, the supervisor can see its child end and start removing its session directory, which is the path a killed CLI must not take. Kill the supervisor first, then the group, so that no code of the run runs after the kill. The child's watchdog would end the group anyway, and the driver's group kill makes that deterministic.
Activity
docs(oracle-gaps): cite the register entry, fix a stale changeset, name
KyleAMathews commentedon Nov 19, 2026
More follow-ups from the delta review of #164, which found #164 READY at ef875d8:
removeSessioncatches the failure, for example a file another process holds open on win32), the directory is left for the next run. The README says only "which it removes when it ends". The sentence that says what happens when the removal fails should land with the F2(a)/F4 witness.claimScenarioRunsingap-report-oracle.property.test.mjs) use the synthetic pid999999. On Linux that pid is below the largestpid_max(2^22), so a live process can have it, the same exposure as F7. If one does, a "stale" claim reads as live, and the refusal names another cause. The fix is the one F7 got: a pid that no process can have, such as 2^30 + 1.killstop (stoppedRun) SIGKILLs the check's process group first and the CLI second. Between the two signals, the supervisor can see its child end and start removing its session directory, which is the path a killed CLI must not take. Kill the supervisor first, then the group, so that no code of the run runs after the kill. The child's watchdog would end the group anyway, and the driver's group kill makes that deterministic.Owner: Isaac, as the follow-up of #164.
This comment was written by Isaac.
KyleAMathews commentedon Nov 19, 2026
More follow-ups from the delta review of #164, which found #164 READY at ef875d8:
removeSessioncatches the failure, for example a file another process holds open on win32), the directory is left for the next run. The README says only "which it removes when it ends". The sentence that says what happens when the removal fails should land with the F2(a)/F4 witness.claimScenarioRunsingap-report-oracle.property.test.mjs) use the synthetic pid999999. On Linux that pid is below the largestpid_max(2^22), so a live process can have it, the same exposure as F7. If one does, a "stale" claim reads as live, and the refusal names another cause. The fix is the one F7 got: a pid that no process can have, such as 2^30 + 1.killstop (stoppedRun) SIGKILLs the check's process group first and the CLI second. Between the two signals, the supervisor can see its child end and start removing its session directory, which is the path a killed CLI must not take. Kill the supervisor first, then the group, so that no code of the run runs after the kill. The child's watchdog would end the group anyway, and the driver's group kill makes that deterministic.Owner: Isaac, as the follow-up of #164.
This comment was written by Isaac.
fix(oracle-gaps): a stopped run removes its session directory (#150) (#…