[bborbe]

Problem

scripts/session-self-exit-test.sh hardcoded PORT=18099 for its stub attention store. On a host where anything else already listens on that port, the stub server fails to bind silently — and the test then reads whatever is on that port. A 404 on /api/1.0/attention is treated by session-self-exit.py as "store unreadable, not a clear gate", so the five cases that expect a clear gate fail.

Observed 2026-10-20: two unrelated processes held 18099 on the operator's Mac — a coding/bench node fixture up two days, and an orphaned /private/tmp/sc2-stub server. make test was red on the host and green in CI, where nothing squats the port.

❌ completed + phase done + zero boxes exits: want exit 0, got 10 — SELF-EXIT: stay — attention store unreadable (HTTP Error 404: Not Found)
❌ aborted exits even with open boxes: want exit 0, got 10 — …
session-self-exit: 13 passed, 5 failed

Fix

• Bind an ephemeral port instead of a fixed one, so the test no longer depends on what else runs on the host.
• Fail loudly when the stub server never answers, instead of silently reading another server's directory.

Verification

• bash scripts/session-self-exit-test.sh → session-self-exit: 18 passed (was 13 passed, 5 failed)
• make precommit → exit 0

[ben-s-pull-request-reviewer]

Review complete. This is a small, targeted bugfix. Here's the assessment:

Diff summary: 2 files — CHANGELOG.md (+1 unreleased bullet) and scripts/session-self-exit-test.sh (+20/−2: ephemeral port function + server-up guard).

Correctness:

• free_port() correctly binds port 0 and reads back the assigned ephemeral port
• The server_up guard (lines 158–169) properly fails with exit 2 if the stub never responds, solving the original silent failure mode
• trap cleanup at line 31 correctly kills the server on exit
• Port bound to 127.0.0.1 (not localhost) — no DNS dependency

No issues found.

{
  "verdict": "approve",
  "summary": "Targeted bugfix that replaces a hardcoded port (18099) with an ephemeral port and adds a loud failure if the stub server doesn't respond — directly addresses the observed CI failure where port collisions were silent.",
  "comments": [],
  "concerns_addressed": [
    {
      "concern": "correctness: test script port binding may fail if ephemeral port selection logic is incorrect",
      "disposition": "addressed",
      "detail": "free_port() uses Python socket.bind(('127.0.0.1', 0)) which delegates port selection to the OS — correct and standard pattern. The server-up guard (lines 158-169) additionally ensures the stub actually responds before tests run."
    }
  ]
}