A thirdshift-plugin-* temp dir can outlive its Run #24

Activity

1. JacobStephens2

   added

   needs-triageMaintainer needs to evaluate this issueMaintainer needs to evaluate this issue

   on Sep 25, 2026
2. JacobStephens2 commented on Sep 25, 2026

   

   JacobStephens2

   on Sep 25, 2026

   OwnerAuthor

   More actions

   This was generated by AI during triage.

   Agent Brief

   Category: bug
   Summary: Treat SIGHUP like SIGINT and SIGTERM, so a Run whose terminal closes goes through the Failed run path and cleans up

   Current behavior:
   SIGINT and SIGTERM are recorded, not fatal: the Run stops its agent session, goes through the Failed run path, and removes the plugin temp directory. Verified during triage: with an extra check added to the existing SIGINT/SIGTERM tests, the scenario's $TMPDIR is empty afterwards.

   SIGHUP (such as a closed terminal) isn't handled, so thirdshift dies on the spot. Verified during triage with a throwaway test that sends HUP while the fake agent runs:

   • the thirdshift-plugin-* directory is left in $TMPDIR
   • the agent session keeps running after thirdshift is gone, because it runs in its own process group and nothing stops it (the fake agent finished its full sleep 30)
   • there is no Failed run: no failure commit and no push, so the work in the worktree isn't saved

   For context: the leak that prompted this issue came from a real Run for #7 (started 11:55:33, and its session log ends 12s later with [Request interrupted by user]), not from a test. Tests set TMPDIR to their own scenario directory, so they can't leak into /tmp. Which signal killed that Run couldn't be pinned down, but SIGHUP is the confirmed path the current code doesn't handle.

   Desired behavior:
   SIGHUP behaves the same as SIGINT and SIGTERM:

   • the Run is marked interrupted
   • the running agent session is stopped (its whole process group)
   • the Run goes through the Failed run path: work is pushed and the failure is reported as interrupted
   • the plugin temp directory is removed

   Key interfaces:

   • The interrupt module's install(): it registers the signals that set the interrupt flag, and should include SIGHUP. Its module doc names the handled signals and should be updated to match.
   • interrupt::requested(): unchanged; SIGHUP just sets the same flag.
   • The interrupt tests in the Failed run test suite (assert_interrupt_fails_the_run(signal) and the per-signal tests that call it).

   Acceptance criteria:

   • Sending SIGHUP to thirdshift while an agent session runs stops the session within the existing stop grace period, and the Run finishes well before the fake agent's own sleep would end
   • After SIGHUP, the Issue branch on origin has the thirdshift: failed run (interrupted) commit containing the agent's half-done work, the same as the SIGINT and SIGTERM tests check
   • A SIGHUP test exists alongside the SIGINT and SIGTERM tests
   • The shared interrupt assertion checks that the scenario's tmp/ directory is empty afterwards, for all three signals
   • cargo test passes

   Out of scope:

   • Sweeping stale thirdshift-plugin-* directories at startup. Runs can overlap, so a blind sweep would delete a live Run's plugin; a safe sweep needs a liveness design, which is a separate decision
   • SIGKILL, which can't be handled
   • Changing the interrupt semantics or the Failed run path beyond adding SIGHUP
3. JacobStephens2

   added

   bugSomething isn't workingSomething isn't working

   ready-for-agentFully specified, ready for an AFK agentFully specified, ready for an AFK agent

   and removed

   needs-triageMaintainer needs to evaluate this issueMaintainer needs to evaluate this issue

   on Sep 25, 2026
4. JacobStephens2

   closed this as completedin #25on Sep 25, 2026
5. JacobStephens2

   added a commit that references this issue on Sep 26, 2026

   Merge pull request #76 from JacobStephens2/issue-64

   ...

   ed1fff7