[cjhyy]

Problem

packages/server/src/panels/project-packages.test.ts ("real Hub HTTP serves and executes each project's retained package…") failed intermittently in tests (rest). The task job failed with one of three errors:

• process is unavailable or belongs to another Panel App
• process stdin is unavailable
• Tool did not return a completed result

Root cause (runner evidence)

• Alone: 50/50 passes on the runner (2 jobs × 25 runs).
• In the full rest shard: failed in 3/8 runs (diag run 39713091209) and 1/10 runs (diag run 39714111699).
• Instrumentation: I added logging to the executor and process service. In each failure, the fixture tool's output and exit arrived before the Host had ended stdin. In some runs the child error event was EBADF: bad file descriptor, epoll_ctl, with exit(code=null) and the parent's stdin closed, while the real child kept running and printed its result later.
• strace (close, epoll_ctl, pidfd_open, socketpair, vfork): late in the shard the single Bun process holds about 1400–4500 descriptors. Inside Bun's synchronous spawn path, the main thread does this:

  pidfd_open(8734) = 4365<pidfd>
  close(4362<socket>) close(4364<socket>)   # child ends, normal
  close(4365<pidfd>) = 0                    # closes the pidfd it just opened
  close(4365) = -1 EBADF  (x2)
  epoll_ctl(ADD, 4365) = -1 EBADF           # Bun reports the spawn as failed
  

  A spawn that succeeded a few seconds earlier, with the same fd numbers, shows no such close. In the other variant, a parent stdio socket is closed under the child: the fixture reads EOF, answers, and exits before processes.end().
• Not specific to this suite: the same EBADF ... epoll_ctl also failed an unrelated Bun.spawn test in the shard (hub-automations.test.ts lease test).

The product code is not involved. Hub and Desktop run on Node, and the executor and process service behave correctly when their descriptors are intact. Recent Bun PRs fix similar double-close and fd-confusion bugs in spawn and IO (oven-sh/bun#38354, #43626, #43814). Upgrading Bun is a separate, repo-wide decision.

Fix

• The rest shard now ignores **/server/src/panels/project-packages.test.ts.
• A new rest step runs that suite in its own Bun process with the same 30s timeout. The shard already uses this pattern for other suites that a long shared Bun process destabilizes.

Verification

• bun run lint:workflow-test-paths passes.
• Running the panels directory with the ignore pattern excludes the suite.
• Runner verification (diag run 39719250655): 10 × (rest shard with the ignore + this suite run 5 times alone). Results are in the PR comments.

Follow-ups (not in this PR)

• The rest-shard Bun process climbs to about 4500 open descriptors (pidfd numbers reach 4550). Something in the shard leaks fds.
• The same Bun spawn defect can still hit other spawn-heavy tests in the shard.

🤖 Generated with Claude Code

[cjhyy]

Runner verification (workflow_dispatch run 39719250655, diag branch): 10 parallel jobs, each running the full rest shard with the new ignore and then this suite alone 5 times. All 10 shards: 5625 tests, 0 failures. Isolated suite: 50/50 passes. Before the fix, the same rest shard failed this suite in 3/8 (run 39713091209) and 1/10 (run 39714111699) runs.