[ZacxDev]

Flake

TestAPoolWorkerIsVisibleWhereAgentsAreListed failed once in CI on #387 (run 41274534531, attempt 1, job "test (postgres)"):

worker_pool_form_test.go:62: /agents/spry-newt carries 1 pool chips, want 0:

Root cause (test fixture, not production)

internal/api/kickoff_preflight_test.go:123 (origin/main): preflightStore.NameExists returned false unconditionally, so agents.BuildUniqueAgentName (internal/agents/names.go:207) never redrew. The pool is 16 adjectives x 16 nouns = 256 names, so the test's two dispatches drew the same name about once in 256 runs. pageablePreflight.GetByName (worker_pool_form_test.go) returns the first row with that name, so the pool-less control's page (plain, line 54) rendered the pooled agent with its chip, and the assertion at line 62 failed.

Production is fine: PGStore.NameExists checks live rows plus the retired-names ledger, so a real deployment redraws.

Fix

preflightStore.NameExists now answers from the rows the fixture has created, the way PGStore does. No retry, no sleep. nameExistsNever in officers_test.go stays as it is on purpose: it samples the generator and creates no rows.

Evidence

go test ./internal/api -run '^TestAPoolWorkerIsVisibleWhereAgentsAreListed$', with counts taken from --- PASS/--- FAIL lines:

tree	mode	FAIL	PASS
origin/main 0a2ef0e	natural, -count=3000	9	2991
HEAD	natural, -count=3000	0	3000
origin/main fixture + generateName pinned to one name (uncommitted)	forced collision, -count=200	200	0
HEAD fixture + same pin	forced collision, -count=200	0	200
HEAD	-race -cpu 1,4 -count=500	0	1000

The natural rate of 9/3000 matches the expected 3000/256 ≈ 11.7.

Full suite, run as CI does (go test -race -cover -timeout 20m -json ./... piped to tests/verdict.py with the three --requires, against a local Postgres 17): internal/api ok, 0 skipped, 3/3 required tests passed. go vet ./... is clean. My local run also showed failures in internal/connectors and internal/db. Both come from the local environment, and this diff doesn't touch either package (it changes one internal/api _test.go file):

• connectors (a wrong password connected): my pg_hba first used trust. With scram it passes.
• db: DROP DATABASE … WITH (FORCE) hit statement timeouts under full parallel -race load. The same tests pass when run alone.

🤖 Generated with Claude Code

[ZacxDev]

Audit ladder: round 0 = proceed (0 deletion candidates); round 1 = no 🔴/🟡, one 🟢 (comment wording) fixed in 1fdc696 without re-audit. Red/green independently reproduced: base 5/1500 and 8/2000 FAIL, HEAD 0/1500 and 0/2000; forced collision 50/50 FAIL at base, 0/50 at HEAD.

1. preflightStore.NameExists answers from the fixture's created rows instead of false unconditionally, so BuildUniqueAgentName redraws on a taken name
2. TestAPoolWorkerIsVisibleWhereAgentsAreListed no longer fails about 1 run in 256 when two dispatches draw the same name