[freeqaz]

Problem

PostgreSQL tests open a pool and close it in an after hook with the same two lines: const pool = new pg.Pool({...}) and then t.after(() => pool.end()). A survey of tests/ on main found 65 such pairs in 26 files.

pool.end() resolves once the pool's idle clients are removed, before their sessions close. When the test's database is dropped right after (DROP DATABASE ... WITH (FORCE)), the drop can end a session that is still closing. The server's 57P01 then reaches the pool, which has no error listener, and the test fails in its teardown. That is finding 2379, fixed for the revision fixture in #3547. withPool in tests/helpers/postgres-database.mjs already avoids it by waiting for each client's own end. The hand-written after hooks do not.

Change

• tests/helpers/postgres-database.mjs:
  • withPool's close (end the pool, then wait for each client's end) moves into a private closer(pool). withPool uses it unchanged.
  • New testPool(context, config) returns new pg.Pool(config) and registers closer(pool) with context.after at once, where the old after hook was registered. context is a test context or anything with after (the repository credential resource scope).
• 19 pairs in 13 files now call testPool, with the same pool configuration. postgres-presets' two-pool hook (Promise.all of two ends) becomes two testPool calls. Imports of pg that became unused are removed.
• The helper adds no error listener: an idle client error still fails the test as before.
• Left alone:
  • feat(occdev): wire the OpenShell Git route into occ dev up with an installed acceptance #3045's files. That includes postgres-platform-state-kubernetes.test.mjs, whose four case pools end right before their case database is dropped in the same process (the 2379 shape).
  • native-admin-session-postgres.test.mjs (feat: add opt-in host-only Agent browser sessions #3255 rewrites its imports).
  • Hooks that release a held client before pool.end(), or that guard against a second end.

Test-only. No product code or assertions change.

Verification

• On a remote test VM, the lanes that run the touched files passed: postgres-platform 569/569 (0 skipped), postgres-application 305/305, postgres-auth 159/159.
• Locally (Node 24, CPUs 20-27): secret-driver-startup 2/2, ssh-compute-startup 3/3.
• Mutant: testPool ignoring its config (connecting to a closed port) fails all nine postgres-platform caller files (openshell-git-routes, openshell-standing-git-transfer, prepared-standing-removal, presets, repository-broker-receipts, repository-sessions, repository-transactions, service-account-issuance-recovery, standing-approval-api); the two startup files never connect, so they survive.
• Survivor: the close without the wait for each client's end passes postgres-application and postgres-platform. Those files' databases are dropped only after their process exits, so nothing there can reach the race; the wait is withPool's existing behaviour, now shared.
• tsc --build tsconfig.json tests/tsconfig.json, ESLint and Prettier are clean.
• git merge-tree: feat: assign configured OpenClaw roles to Agent users #906, fix(controller): close refused native-admin sockets before audit settlement #1707, feat: add Gateway managed mode to Agent deployments #2982 and feat: add opt-in host-only Agent browser sessions #3255 (the open PRs on these files) get no new conflicted file or hunk. This PR merges cleanly with test: assert API error responses through a shared helper #3558.

Risks

Low. Teardown now also waits until the pool's sessions have closed, as withPool already does. A pool whose client never closes would hold its after hook, but pool.end() already waits for checked-out clients before it resolves.

[clawsweeper]

🦞👀
ClawSweeper picked this up.

Pull request received. I will update this pull request when review starts.