[danshapiro]

Fixes intermittent failures in the agent crates and five cloud e2e tests. Each was reproduced first (or, where it would not reproduce, the race was shown by forcing the timing), then fixed at the cause. No test was skipped or loosened, and no fixed sleep was made longer. "Under load" below means the host load average was about 100–900 on 96 cores (busy-loop burners plus other agents' work).

Rust tests

1. session_watcher metadata_late_subagent_named_mkdir_arms_via_deferred_recheck_without_a_discover (freshell-sessions)

   • Cause (test): the watcher marks the recheck as armed before it applies the index entry, and a background refresh publishes the new row. The test read the index once, right after it saw the arm.
   • Fix: wait for the row within the same 5 s budget, then check that no full discover ran.
   • Before: 2/20 failed quiet, 13/30 under load. After: 140/140 under load.

2. session_handoff opencode_handoff_stop_waits_for_a_mid_dispatch_rest_drive_under_the_gate (freshell-freshagent)

   • Cause (test): the release used notify_waiters(), which is lost if the drive has not reached its pause yet. The drive then holds the gate and the handoff times out. The "prompt POST" check also matched the first drive's POST, so it proved nothing about the second.
   • Fix: a test-only handshake (stored permits) in lib.rs / opencode_ws.rs. Assertions now use the second drive's POST.
   • Did not fail naturally in about 90 runs. A 400 ms delay injected before the pause failed it every time; with the fix it passes. 90/90 under load. Removing the product gate now fails the test at once.

3. session_handoff an_opencode_placeholder_handoff_resolves_the_durable_owner_and_completes (freshell-freshagent)

   • Cause (test): a claude_snapshot test removed HOME for the whole process while this test was reading it.
   • Fix: that test no longer removes HOME (its own result never depended on it).
   • Before: 20/30 failed when run with that test. After: 40/40 for the pair, 30/30 with all 67 claude_snapshot tests under load.

4. claude a_sidecar_exit_after_the_live_verdict_never_gets_its_stamp_restored and seven sibling tests (freshell-freshagent, test-only changes in claude.rs)

   • Cause (test): a new session first publishes its alias, then writes its identity record, then records ownership together with the release stamp. These tests waited for the alias or the identity record and then read or took the stamp, which might not exist yet. One test also set a pause that the first session could get caught in.
   • Fix: shared helpers wait for the stamp itself, which is the real condition.
   • A 60 ms delay injected before the ownership step failed seven of the eight old tests. With it injected, all 201 claude unit tests pass. Without it, all eight passed 40/40 alongside 200 busy loops.

5. freshagent_session_lease expired_lease_kill_sweeps_the_sidecar_tree_before_release (freshell-ws)

   • Product fix: the expired-holder kill reads /proc files of processes it is killing. A process reaped between open and read returns ESRCH, which was treated as "can't tell", so the kill counted as unconfirmed and the second client was refused. ESRCH now means "process gone". A regression test reaps a real child mid-read and fails under the old handling.
   • Test fixes:
     • The test's decoy sleep 300 is now killed on cleanup. Before, it was left running after every run.
     • The second client's sidecar no longer inherits the decoy setting or the first client's 5 s create delay.
     • The test fails at once on create.failed.
     • The decoy now ignores SIGTERM. Before, the claude lane's own SIGTERM of its tagged processes also killed the decoy, so the test kept passing with the lease sweep removed. With the sweep removed, the test now fails.
   • The created timeout did not reproduce in about 110 sandbox runs. After the fix: 30/30 full-file runs under load, with no processes left behind.

E2E tests (cloud)

6. unified-agent-names-restart "[freshopencode] pending rename survives a zero-turn restart"

   • Cause (product, found from two retained cloud traces of failures): after a server restart, the browser's reconnect attach names the pane's not-yet-created freshopencode-* placeholder. The server tried to resume it. It took the ownership claim for that id and held it while it started the shared opencode daemon, only to get "not found". The user's first message landed inside that window. Its claim on the same id was refused with SESSION_RESERVED. The browser then retried the attach, not the message, so the message was silently lost. The traces show this: the message was echoed, status went "running" then "idle" within 300 ms, and there was no error banner and no new session.
   • Fix: resuming a placeholder id now answers "not found" at once, with no claim and no daemon start. A unit test that holds the daemon start open reproduces the exact failure on the old code.
   • Before: about 1 failure in 30–90 runs. After: 100/100 on cloud (3 workers), and the unit test passed 40/40 under load.

7. handoff-two-device-rust "codex: reopen-as-CLI on desktop converges the phone pane"

   • Cause (product, browser client): the handoff's HTTP answer can switch the pane to the new terminal before the WebSocket owner update arrives. The attach is refused while the old owner is still recorded, and nothing retried it.
   • Fix: TerminalView retries the attach once the owner record names this pane's terminal. A unit test fails before the fix.
   • Before: 8/20 failed. After: 90/90 across three cloud runs.

8. fresh-agent-rollback-rust "codex: undo-to-here … /redo is refused with the codex copy"

   • Cause (test): the pane's own undo confirmation could arrive after /redo was typed. It refilled the composer and replaced the redo notice.
   • Fix: wait for the refill before typing.
   • Before: 4/20 failed. After: 177/180 across five cloud runs. The 3 failures came back-to-back inside one 40-second window of a single cloud task, in a run without retries, so no trace was kept. The 160 runs after that (with retries on, so a failure would have kept its trace) all passed.

9. layout-sync-authoritative "REST snapshot exactly mirrors a visible-UI-driven layout"

   • Cause (test): the drag-and-drop library ignores clicks for 50 ms after a drag ends. A click right after the drop was lost.
   • Fix: wait on a page timer that fires after the library's timer.
   • After: 90/90 on cloud.

10. unified-agent-names-cross-mode "tab source survives focus swap move and close"

    • Cause (test): the rename helper resubmitted while the first rename was still in flight, which closed the editor under the second Enter.
    • Fix: resubmit only when the rename error is shown, as the pane-header helper already does.
    • After: 90/90 on cloud.

Not fixed here (noted for follow-up)

• In the browser, a send that the server refuses with a session-wide SESSION_RESERVED is never resent. The message stays on screen as sent. Item 6 removes the cause seen in these tests, but any other short-lived claim would lose a message the same way.
• The OpenCode test server picks a free port, releases it, and then uses it. With 30 test processes running at once, one test's health check was answered by another test's server.
• Several OpenCode test processes using /tmp as their working directory compete for the /tmp/.opencode lock.

All changed crates pass cargo fmt --check and clippy -D warnings, and the branch passed the pre-push hook. The branch merges the current origin/main.

🤖 Generated with Claude Code