[Alexandru Dima (alexdima)]

Follow-up to #342698 and #343092.

Problem

An audit of origin/main at eeedb17 found waits for state in four agent host test files that local/code-no-timeout-polling-loop does not report. The rule reports loops whose body only waits on a timer; these loops wait on setImmediate, a setInterval callback or a microtask, or do other work in the same body (a try/catch re-read, answering a request). npx eslint --rule '{"local/code-no-timeout-polling-loop": ["warn", {"exclude": []}]}' reports 0 problems on all four original files. Each wait stood in for a specific piece of work:

• Claude bootstrap model refresh (claudeAgent.test.ts, 6 loops). ClaudeAgent queues its first model refresh with queueMicrotask in the constructor and keeps it only in the private _modelRefreshInFlight. Five tests ran up to 100 setImmediate ticks until the fake SDK had counted a supportedModels() or accountInfo() call, then one more tick for the refresh to publish. A sixth, in the SDK setup channel suite, ticked until the bootstrap reached enumeration, so that a gate it raises next blocks only the download's enumeration. The comment on one of them said _fetchNativeModels awaits a real mkdtemp; it no longer does (buildModelEnumerationOptions() uses tmpdir()).
• Published models (claudeAgent.test.ts, 2 loops) and a download enumeration (1 loop). Two loops ticked until agent.models was not empty, one until the fake SDK had counted the download's supportedModels() call.
• Stub flags (claudeAgent.test.ts, 2 loops; claudeProxyService.test.ts, 1 waiter). Two while (!persistEntered) loops spun on setImmediate until the test's stubbed setMetadata had been entered. The proxy test polled a flag set by its stubbed messages every 10 ms with setInterval.
• Root config write (agentService.test.ts, 1 loop). The test re-read the root config file up to 20 times, every 5 ms, until it parsed and matched, and continued with a failure flag when it ran out. It waited for the serialized write that AgentConfigurationService.persistRootConfig() chains onto _rootConfigWrite, which whenIdle() already exposes.
• Import chunk staging (agentHostProtocolClient.test.ts, 1 loop). After acknowledging each vscode/stageCreateSessionImportChunk request, the test waited setTimeout(0) for the client to send its next message (agentHost: support turn-by-turn conversation import for remote sessions #343312).
• Request helpers (agentHostProtocolClient.test.ts, 2 helpers). waitForRequest and waitForRequestAt spun on await Promise.resolve() until the transport's outbox had the request. They have 99 call sites in the soft reconnect suite; a probe (P1 below) counted 123 calls in that suite, 79 of which had to wait for a send. A microtask spin never yields to timers, so it hangs if the send ever moves behind one.

The tick and timer budgets depended on machine load, and the bounded loops continued with stale state when they ran out.

Change

Wait	Waited for	Now awaits
claudeAgent.test.ts 1727, 1755, 1775, 1796, 6398, 6981	the constructor's model refresh to enumerate and publish	bootstrapModelRefresh(agent, sdk): sdk.waitForCalls(() => sdk.accountInfoCallCount > 0), then the agent's _modelRefreshInFlight
claudeAgent.test.ts 1688, 6421	agent.models not empty	waitForState(agent.models, models => models.length > 0)
claudeAgent.test.ts 6415	the download's supportedModels() call	sdk.waitForCalls(() => sdk.supportedModelsCallCount > enumerationsBefore)
claudeAgent.test.ts 4501, 5878	the stubbed setMetadata entered	a DeferredPromise completed by the stub
claudeProxyService.test.ts 1360	the stubbed upstream messages called	a DeferredPromise completed by the stub
agentService.test.ts 8106	the serialized root config write	await svc.dispatchAction(...), then getConfigurationService(svc).whenIdle(), then one read
agentHostProtocolClient.test.ts 2154	the client's next send after a chunk acknowledgement	Event.toPromise(transport.onDidSend), subscribed before the acknowledgement
agentHostProtocolClient.test.ts 4869, 4879	the request in the outbox	the same check, re-run after each transport.onDidSend

Line numbers are from eeedb17; none of these files changed between it and the base of this PR.

Details worth a look:

• FakeClaudeAgentSdkService gets waitForCalls(condition), in the shape of TestCopilotClient.waitForCalls from agentHost: await real signals instead of polling timers in copilotAgent tests #343092: it checks the condition and otherwise awaits Event.toPromise(Event.filter(...)) on an emitter that FakeQuery fires after it counts a supportedModels() or accountInfo() call.
• bootstrapModelRefresh reads the private _modelRefreshInFlight through element access in a single helper, as proxyRefresh(agent) does in copilotAgent.test.ts. The field is only set once the constructor's microtask has run, so the helper first waits for the refresh's accountInfo() call; after that the field holds the bootstrap refresh, or is already cleared because the refresh has published. Nothing else starts a refresh in these tests. Calling agent.refreshModels() instead would coalesce or start a second enumeration depending on timing, and would hide a missing bootstrap.
• Negative check: "an SDK account report of "nothing configured" publishes an empty catalog" asserts [], which is also the initial value, so it is only meaningful after the refresh has published. M3 and M3b below show that awaiting the in-flight refresh is what makes it fail when the product publishes the static list.
• The SDK setup channel test now waits for the whole bootstrap refresh rather than only its enumeration call. The old setImmediate ticks gave the same result, and the test relies on it: the refreshModels() that follows must start a new refresh instead of coalescing onto the bootstrap. Its comment says so.
• dispatchAction returns the queued dispatch promise when it defers a client action and undefined when it applies it at once; awaiting it covers both. persistRootConfig() snapshots the values when it is called, so the write that whenIdle() awaits already contains the change. The existing whenIdle() before cleanup is kept.
• The protocol client helpers keep their signatures and call sites. Each iteration of their loops awaits the next send; the new check is the old one.
• A condition that never holds now fails at Mocha's timeout instead of continuing with stale state.

There are no product-code changes. All scenarios and assertions are kept.

Not changed:

• The settle() helper in the "ClaudeAgent — agent SDK setup channel" suite runs a fixed 20 setImmediate ticks and has 21 call sites, including negative checks. Each site needs its own signal, probably an event on the root transient values that the setup channel publishes, so it is left for a separate change. "a download that lands stays downloading until the catalog does", changed here, still calls it twice, once right after the waitForState this PR adds (M5 below).
• Standalone single waits that are not loops: await tick() calls, await new Promise(resolve => setImmediate(resolve)) in the shutdown test at claudeAgent.test.ts 4401, setTimeout(resolve, 0) waits in agentHostProtocolClient.test.ts 4230 to 4395 and agentService.test.ts 41652, and the 2 s Promise.race in the streaming abort test in claudeProxyService.test.ts.

Validation

The baseline ran on eeedb17. The change was then rebased onto 37b0141, the base of this PR; the rebase was clean and no upstream commit in between changed the four test files or claudeAgent.ts, claudeProxyService.ts, agentConfigurationService.ts, agentService.ts or agentHostProtocolClient.ts. The mutation checks ran before the rebase, the 10 runs, CPU load runs and final checks after it. The machine was shared, so load averages are given with each timed run.

• Baseline before the change, ./scripts/test.sh --run src/vs/platform/agentHost/test/node/claudeAgent.test.ts --run src/vs/platform/agentHost/test/node/claudeProxyService.test.ts --run src/vs/platform/agentHost/test/node/agentService.test.ts --run src/vs/platform/agentHost/test/electron-browser/agentHostProtocolClient.test.ts: 2027 passing, 16 pending (1 min Mocha time, 263 s wall, load average 20 to 42).
• git grep -n -E 'for \(let i = 0; i < 100 && .*; i\+\+\) \{|while \(!persistEntered\)|const i = setInterval\(|attempt < 20; attempt\+\+|await new Promise\(resolve => setTimeout\(resolve, 0\)\);|while \(true\) \{|^\t\t\t\tawait Promise\.resolve\(\);$' <ref> -- <the 4 files>: on eeedb17 it lists the 9 tick loops and 2 persistEntered loops in claudeAgent.test.ts, the setInterval waiter, the 20-attempt loop, the chunk loop's setTimeout(0) and the two helper loops. After the change none of them remain; the remaining matches are unrelated: the event-driven while (true) loops in waitForPendingWorktreeRemovals and waitForCondition (agentHost: await real signals instead of polling timers in the remaining agentService tests #343603), the chunk loop itself (each iteration now awaits a send), the fake query's prompt drain, and the standalone setTimeout(0) waits listed above.
• npx eslint --rule '{"local/code-no-timeout-polling-loop": ["warn", {"exclude": []}]}' <the 4 files>: 0 problems, before (run on the eeedb17 contents through --stdin --stdin-filename) and after.
• 10 runs, for i in $(seq 10); do ./scripts/test.sh --run <the 4 files> 2>&1 | grep -E 'passing|failing'; done: 10/10 with 2027 passing and 16 pending. Mocha time 33 s to 1 min and wall time 35 to 70 s, at a 1-minute load average of 6.7 to 26.
• Under CPU load with 18 yes processes, lockf -t 3600 /tmp/vscode-no-timeout-polling-stress.lock bash -c 'pids=(); for i in $(seq 18); do yes > /dev/null & pids+=($!); done; ./scripts/test.sh --run <the 4 files> 2>&1 | grep -E "passing|failing"; kill "${pids[@]}"', alternating with the original files:
  • New: 2027 passing, 16 pending (45 s Mocha time, 68 s wall, load average 8 to 27); 2027 passing, 16 pending (42 s, 104 s wall, load 32 to 33).
  • Original: 2027 passing, 16 pending (38 s, 81 s wall, load 27 to 32); 2027 passing, 16 pending (44 s, 91 s wall, load 33 to 34).
  • The 18 yes processes raised the 1-minute load average to about 27 to 34 on this machine. Neither version failed, so this load does not show a difference between them.
• Mutation checks, each applied by a script, run with --grep, and restored byte for byte (verified with filecmp):
  • M1, in-flight refresh: without await agent['_modelRefreshInFlight'] in bootstrapModelRefresh, three of the four bootstrap tests fail with an empty catalog ("signed out with an SDK-reported account", "the SDK default alias row is dropped", "signed out, the native catalog bootstraps"). "native model enumeration closes the throwaway query" still passes: the query closes before the test resumes, earlier in the same refresh.
  • M2, fake SDK signal: with fireDidRecordCall() not firing, "the SDK default alias row is dropped", "nothing configured" and "a download that lands stays downloading" hit the Mocha timeout.
  • M3 and M3b, negative check: with the product publishing the SDK's static list without an account, "nothing configured" fails and lists the model. With the same mutation and without the in-flight await, it passes.
  • M4, models observable: without the waitForState, "first sign-in keeps the native catalog published" fails with bootstrap: [].
  • M5 and M6, setup channel test: without the waitForState after releasing the enumeration, the test still passes, because the settle() that follows also covers it. Without the waitForCalls for the download's enumeration, it also passes: the download publishes downloading synchronously and the catalog is empty until the gate opens, so the snapshot is the same. The wait sets up the state the test names but does not decide its result, as with the old loop.
  • M7 and M8, stub deferreds: with the setMetadata stubs never completing persistEntered, both race tests hit the Mocha timeout. Without awaiting it, "agent.dispose() during a racing first sendMessage" fails (rejectedIsCancellation: false, warmQueryDisposed: false); "teardown racing _writeCustomizationDirectory" still passes.
  • M9, stub deferred: without awaiting upstreamCalled, "dispose() with in-flight non-streaming aborts the upstream call" fails with "expected abort to fire on dispose".
  • M10 and M11, root config write: without the whenIdle() before the read, and separately with the product's whenIdle() not awaiting _rootConfigWrite, "applies and persists root config changes from clients" fails with ENOENT for the config file.
  • M12 and M13, transport sends: with the helpers' onDidSend wait never resolving, the first two soft reconnect tests that call them hit the Mocha timeout, and the failing afterEach stops the suite. With the chunk loop's next send never resolving, "stages large imported conversations in bounded chunks" hits the Mocha timeout.
  • P1, probe: counting in the helpers over the soft reconnect suite gave 123 calls, 79 of which had to wait for a send.
• lockf -t 3600 /tmp/vscode-no-timeout-polling-typecheck.lock npm run typecheck-client: no errors in the changed files, before the rebase and on 7e76185. The command still exits with errors in files this PR does not touch, 19 before the rebase and 20 after it: fusionTier in the Copilot agent files and their tests, and after the rebase setParentWindow in terminalProcess.ts from Parent the ConPTY console window to the VS Code window #342784. Both come from the shared node_modules, whose @github/copilot-sdk (1.0.24-preview.0) and node-pty (1.2.0-beta.15) are behind this tree's pins (1.0.26 and ^1.2.0-beta.18).
• npm run eslint -- <the 4 files>: no problems.
• Self-review with the code-review skill: no findings. Earlier, the first test run after the change found a test that used agent without destructuring it, and the first type check found that the type guards in a first version of waitForRequest and waitForRequestAt did not narrow; both were fixed before the runs above.

[Copilot]

🟢 Approval recommended

The test-only changes use the relevant completion signals, preserve assertions, and have no unresolved blocking findings.

No open issues 🎉

What changed in this PR

This PR continues the Agent Host test synchronization cleanup by replacing hidden polling waits with explicit completion signals. It changes only tests, preserving existing scenarios and assertions.

Changes:

• Await fake SDK events, model publication, and deferred stub signals in Claude tests.
• Await root configuration persistence before reading the saved file.
• Replace protocol request microtask loops and chunk-acknowledgement timer waits with transport send events.

No actionable findings remain.

File	Description
src/​vs/​platform/​agentHost/​test/​node/​claudeProxyService.test.ts	Waits for an explicit upstream-call signal before testing disposal.
src/​vs/​platform/​agentHost/​test/​node/​claudeAgent.test.ts	Uses SDK events, refresh completion, observable state, and deferred signals instead of polling.
src/​vs/​platform/​agentHost/​test/​node/​agentService.test.ts	Awaits dispatch and configuration persistence before asserting saved values.
src/​vs/​platform/​agentHost/​test/​electron-browser/​agentHostProtocolClient.test.ts	Waits for transport sends during chunk staging and request lookup.

_{Review effort: Balanced}