[mchwang]

Problem

blocks a partially received merge request when shutdown starts failed about 1 in 5 runs on main. The errors were connect ECONNRESET 127.0.0.1:<port>, then Server is not running.

Cause

The test was wrong; the server is correct. requestAdmitted() resolved on the server's first 'request' event from any client. When another client reached the port first, the test called close() before the server had accepted its own partial merge request. Examples of another client: a polling tab, or a pooled connection to a reused ephemeral port.

• The listener closed while the test's connection was still in the accept queue. The client saw connect ECONNRESET (the connect syscall shows that the server never accepted the connection).
• Server is not running is a side effect. The test threw before its last line restarts the server, so afterEach closed an already-closed server.

close() in web/server.ts works as designed: it blocks new work, bounds the drain, and aborts active requests.

Fix

This PR changes only test/browser/review.spec.ts.

• requestAdmitted() now returns a unique x-codeboost-test-request header. It resolves only when the server parses the headers of the request that carries that header. The three partial-request shutdown tests use it.
• Regression: before its partial request, the merge test sends a complete request from another client (GET /api/settings).

Evidence

Results on head a2328a0:

Check	Result
Regression with the old helper	Fails every run with connect ECONNREFUSED (or ECONNRESET), then Server is not running
Regression with the new helper	Passes
Mutation: disable the shutdown merge guard at web/server.ts:58	The test fails (expected 503, received 409), so the guard is still covered
Flaky test, --repeat-each 20	20/20 pass
Full Playwright suite	58/58 pass
npm run typecheck	Clean
Vitest	3 failures, all in test/agent-supervisor.test.ts: Docker container tests that time out at 60 s on the local host. Vitest includes only test/*.test.ts, so it does not load the changed file.

I could not reproduce the original flake naturally. It did not occur in 40 Playwright runs under CPU load or in 320 standalone harness runs. The deterministic regression reproduces the same failure signature.

Review-lesson audit

Covered by the existing rule in AGENTS.md "Required race regressions": A request is partially received, then shutdown starts, then the request completes. This PR adds no new rule.

Review rounds

• Round 1 (Copilot, head a2328a0): no findings; Copilot recommended approval. Nothing changed and nothing was declined.

🤖 Generated with Claude Code

[Copilot]

Copilot review overview

🟢 Approval recommended

No unresolved review issues remain.

Review effort: Lite
Findings: None

What changed in this PR

Fixes flaky shutdown tests by ensuring they wait for admission of their own marked request.

Changes:

• Adds unique request markers to requestAdmitted().
• Applies markers to partial-request shutdown tests.
• Adds a competing request regression scenario.

File	Description
test/​browser/​review.spec.ts	Makes shutdown request-admission tests deterministic.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.