test: isolate root-suite shared state

[desplega-bot]

Summary

• Restore absent environment variables by deleting them in artifact-sdk.test.ts.
• Reset the production Slack handler caches before and after the test suite that populates them.
• Keep the affected assertions, file order, and retry behavior unchanged.

Serial contamination

Artifact SDK → Codex OAuth

artifact-sdk.test.ts saved optional environment variables and restored them with assignment. When MCP_BASE_URL was originally absent, assigning undefined to process.env.MCP_BASE_URL stored the literal string "undefined". The later Codex OAuth sweep inherited that value and failed while constructing its lock URL.

Minimal reproducer before the fix:

env -u MCP_BASE_URL bun test src/tests/artifact-sdk.test.ts src/tests/codex-oauth-keep-warm.test.ts
41 pass / 2 fail

After the fix: 43 pass / 0 fail.

The fix is in the contaminating test and restores the original environment state exactly: absent keys are deleted, while present values are reassigned.

Slack production handler test → thread buffer

slack-assistant-comention-production.test.ts populated cachedBotUserId, cachedBotId, and swarmThreadRootCache in slack/handlers.ts. Those module-level caches survived into slack-thread-buffer.test.ts.

Minimal reproducer before the fix:

bun test src/tests/slack-assistant-comention-production.test.ts src/tests/slack-thread-buffer.test.ts
30 pass / 1 fail

After the fix: 31 pass / 0 fail.

The production test now explicitly owns cache setup and teardown through a narrow testing reset. The affected thread-buffer test does not depend on ambient state.

Full serial result

Bun 1.4.0, 506 files:

• Before: 7991 pass / 7 skip / 3 fail
• After: 7994 pass / 7 skip / 0 fail

Both affected files also remain green in isolation: Codex OAuth 6/6; Slack thread buffer 25/25.

Parallel verdict: upstream Bun 1.4.0

This half is upstream, not a repository port/tmpdir collision.

Taras's consecutive four-worker runs produced different failure sets (41 fail / 2 errors, then 99 fail / 3 errors) with Bun child-process panic output. On this SHA, the same command completed 7994 pass / 7 skip / 0 fail in 95.89s. The divergent failure sets plus a green repetition are nondeterministic.

The failing surface is spawn-heavy, but the repository audit found no literal hard-coded .listen(...) or Bun.serve({ port: ... }) calls in tests; network tests use the shared allocation helpers, and subprocess temp directories are uniquely named.

Bun 1.4.0 has matching upstream reports:

• #39783: test scanner leaks directory file descriptors
• #39852: mock.module followed by enough Bun.spawn calls reaches EBADF and panics
• #40016: scanner FD-leak fix merged after Bun 1.4.0

No repository workaround is included.

Verification

• bun install --frozen-lockfile
• bun run lint
• bun run tsc:check
• bun run test:root — 7994 pass / 7 skip / 0 fail
• bun run test:root -- --parallel=4 — 7994 pass / 7 skip / 0 fail
• bun run check:bun-version
• bash scripts/check-db-boundary.sh
• bash scripts/check-test-spawn-sync.sh
• bash scripts/check-audit-columns.sh
• bun run check:dep-graph
• Pre-push hooks passed

CI's sharded test:root is green on ba6408ee: both shards passed on attempt 2. Attempt 1's shard 1 hit an unrelated existing migration-117 test timeout at 10 seconds; that file passed locally in 651 ms and the failed shard passed unchanged on its single rerun.

Fixes #1266

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
agent-swarm	Ready	Preview	Aug 29, 2026 8:05pm
agent-swarm-docs	Ready	Preview	Aug 29, 2026 8:05pm
agent-swarm-templates	Ready	Preview	Aug 29, 2026 8:05pm

[desplega-bot]

CI timings

Action	Shards	Wall clock	vs main	Compute	Status
test:root	1/2	52s	-39.5% ▼	52s	pass

Sum of wall clock: 52s · sum of compute: 52s

Shard breakdown

• test:root — 1/2: 52s

_{updated 2026-08-29T20:09:59Z · ba6408e · run 33272530183 · swarm script ci-timings}

[desplega-bot]

📊 Docker image sizes (uncompressed, amd64)

fix-issue-1266-test-order @ ba6408e vs main @ a7b2317

🔴 1 regression

Metric	main	this PR	Δ
image.api	310.4 MiB	310.4 MiB	0	⚪
image.worker	3.78 GiB	—	—	⏳
image.worker-slim	1.93 GiB	1.95 GiB	+12.7 MiB (+0.64%)	🔴

_{updated 2026-08-29T20:08:20Z · run · swarm script ci-metrics}

[desplega-bot]

Lead review — no changes requested.

Scope is minimal and matches issue #1266: two test files fixed for state leakage, plus one exported resetSlackHandlerCachesForTesting() in src/slack/handlers.ts. That helper is test-only, side-effect free, and follows the existing "Exported for testing" convention in the same file; no production behaviour changes. No src/be/db/bun:sqlite imports on the worker side.

The env-restore fix is the right root cause: assigning process.env.KEY = undefined writes the literal string "undefined", so an originally-absent key must be deleted, not assigned.

CI green on ba6408e (both sharded Run Tests jobs, Lint/Type Check, all three Docker builds, Merge Gate). Serial root suite goes 3 fail → 0 fail; both files stay green in isolation.

The parallel-mode panics are correctly left alone and attributed upstream (Bun 1.4.0, oven-sh/bun#39783 / #39852, scanner FD-leak fix landed post-1.4.0 in #40016) rather than papered over with a repo workaround. That means --parallel flakes will persist until the Bun bump — worth tracking separately.

Not merging: swarm policy requires explicit authorization. Ready for @tarasyarema.