[szweibel]

Summary

The local Worker browser acceptance failed three times in CI on 2026-09-10 while passing locally and on reruns. Two independent faults, both reproduced and fixed at their real boundary. One of the fixes is a dependency patch and needs Stephen's explicit approval per the fleet AGENTS.md rule on patching upstream internals; see the last section.

1. Port collision between the smoke step and the acceptance

The smoke step kills the bun run dev wrapper and returns as soon as it exits; Wrangler releases port 8787 a moment later. The acceptance then spawned its own Worker with output discarded and never checked it was alive. Locally reproduced: a second wrangler dev on a bound port dies with "Address already in use", so the acceptance's health poll passed against the smoke step's Worker while it shut down. Depending on when that stranger died, the journey failed with net::ERR_CONNECTION_REFUSED, "Failed to fetch" (CI attempt 2), or a missing Ready status.

• scripts/agent-studio-browser-acceptance.ts: refuses to start if anything answers on the port ("Port 8787 is already in use; the browser acceptance must start its own local Worker"), tracks the Worker's exit, fails the health wait or the running journey with the Worker's exit reason, and keeps only the bind-failure signal from Wrangler's stderr.
• .github/workflows/ci.yml: after the smoke, the step waits until nothing answers on 8787 (bounded at 30 s) before ending.

2. React #185 in the throttled resume scenario (real client defect)

The resumed-history scenario throttles its page 8x on purpose. Under load it ended in "Something went wrong". With an instrumented react-dom dev build the chain is: each replay chunk causes a synchronous store commit; after every commit, useAgentChat's stale client-tool-result cleanup effect dispatches setClientToolResults(prev => …) even when it prunes nothing; React cannot bail that update out eagerly while the fiber has pending work, so each commit ends with a Default-lane update pending and the nested-update counter climbs (no reset) until the 51st chunk throws Maximum update depth exceeded. The AI SDK catches it into chat error. Reproduced at 16x and 32x on this laptop; at 8x the throttle batches chunks so the counter never gets there, which is why CI sat on the edge.

Upstream tracks exactly this as cloudflare/agents#2217 (open since 2026-09-04, same diagnosis and fix); agents@0.22.0 and current main still carry the effect, so there is no version to upgrade to. The patch mirrors the map in a ref and dispatches only when a stale entry exists; the updater still recomputes from prev. With the patch the replay reaches Ready at 16x, 24x, and 32x.

3. Measured bound for the throttled scenario

Ready after the 100-chunk replay on the production build with the patch: 3.7 s at 16x, 10 s at 24x, 16 s at 32x. CI runners are slower than this laptop, so 8x there lands in that range, past Playwright's default 5 s. The scenario's turns now wait up to 60 s with a comment recording the measurements. The assertions are unchanged: Ready, the exact full replayed and continued text, no page errors, one acknowledgement.

Verification

• bun install --frozen-lockfile applies the patch; bun run check passes (469 backend, 287 frontend, 22 tool tests, build, staging dry run).
• bun run test:browser -- --no-build passes at the scenario's 8x, and an instrumented copy passes at 24x (Ready 10 s after replay).
• Port guard: with a stranger on 8787 the script fails immediately with its fixed message. The workflow's release wait was exercised locally (port released after the wrapper exited).
• CI runs on this branch: see the checks; I am rerunning the validate job several times to show consecutive passes.

Needs Stephen's approval

patches/agents@0.20.1.patch changes upstream hook internals. Behavior changed: the cleanup effect no longer schedules a React update when there is nothing to prune. Why supported extension points cannot do it: the effect lives inside useAgentChat and runs on every messages change; no option disables or replaces it, and experimental_throttle only lowers the odds (upstream's own advice on #1732), it does not remove the accumulator. Impact: none on tool-result rendering (the updater is unchanged when it runs); upstream's own regression test noted the replayed prefix paints a few tens of milliseconds later without the accidental extra render. Upgrade path: when bumping agents, check whether #2217 has landed and drop the patch. AGENTS.md now records both patches.

🤖 Generated with Claude Code

[szweibel]

CI on this branch: run 34486682158 passed on four consecutive attempts (14:06, 14:09, 14:13, and the fourth), each with the browser acceptance step green in 65–80 s. On main earlier today the same step failed on two of three attempts of run 34478915492 and on the #75 branch run.