Repository navigation
fix(examples): close multi-agent WebSockets on terminal responses - #2533
Merged
HAYDEN-OAI merged 1 commit intoSep 3, 2026
Merged
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Contributor
Contributor
Castiron custom code✅ No new custom-code files detected. 32 mixed files remain; 0 existing customizations changed. Compared 32 existing customizations unchanged
A changed generated baseline means this report cannot reliably identify which handwritten lines changed. Inspect the custom-code diffDownload the exact patch produced by this run (requires repository access): gh run download 33390829019 --repo openai/openai-node \
--name castiron-custom-code-33390829019-1 --dir /tmp/castiron-custom-code-33390829019-1
git apply --stat /tmp/castiron-custom-code-33390829019-1/custom-code.patch
cat /tmp/castiron-custom-code-33390829019-1/custom-code.patchOr reproduce it from an SDK checkout containing the vendored reporter: git fetch --no-tags origin eea2292a4a523da9405161dde0a79ac5dc2ecb2a 568091772e29ab8b15499cb1a9c143883cd120c2
python3 scripts/castiron/custom_code_report.py report \
--base eea2292a4a523da9405161dde0a79ac5dc2ecb2a \
--head 568091772e29ab8b15499cb1a9c143883cd120c2 --fetch --require-head-hash --public \
--out /tmp/castiron-custom-code-568091772e29
cat /tmp/castiron-custom-code-568091772e29/custom-code.patchThis is the current full custom patch for mixed files, not an attribution of only the handwritten lines changed by this PR. |
jbeckwith-oai
approved these changes
Aug 31, 2026
HAYDEN-OAI
deleted the
dev/hayden/autoimprove-multi-agent-websocket-cleanup
branch
September 3, 2026 20:32
Merged
gh-actions-shared Bot
pushed a commit
to xf-qubit/openai-node
that referenced
this pull request
Sep 3, 2026
## Summary Reject an unfinished response when the standard Responses WebSocket example receives a socket close event. `runResponse()` listens for response events and errors, but not connection closure. A normal close frame or abrupt TCP disconnect can therefore leave its promise pending. Because a pending promise does not keep Node running, the actual CLI exits **0 with no error**, even though it never received the response or completed the remaining demo turns. The fix is six added lines in the handwritten example: register a close listener that uses the existing failure path, and remove that listener alongside the existing event/error listeners on every settled response. Successful completion and API errors continue to win before a subsequent socket close. The new diagnostic is fixed text and does not expose the peer's close reason. No SDK implementation, generated file, public declaration, dependency, policy, or custom-code budget changes are included. Both changed files are absent from the verified generated snapshot, and no open PR covers this standard example. This is independent of the separate multi-agent example in openai#2533. ## Regression coverage The tests run the actual TypeScript example in a child process using its documented `ts-node`/`tsconfig-paths` runtime and source SDK entrypoints. An ephemeral localhost WebSocket server provides synthetic responses. The child receives an explicitly isolated environment with a synthetic API key and local base URL; no live API access is used. Before the fix, the clean-close and abrupt-disconnect tests failed because the CLI returned exit code 0. The complete-conversation and API-error controls already passed. All four pass after the fix: - Clean close after partial response text: exit 1, fixed diagnostic, no completed assistant output or second turn. - Abrupt disconnect before response content: the same failure behavior. - All three demo turns, including tool execution and six response requests: exit 0, unchanged output and `previous_response_id` chain, normal final cleanup. - API error followed immediately by socket closure: preserve the original API diagnostic instead of replacing it with the new close error. Every case also checks that neither the synthetic credential nor the peer's synthetic private close reason appears in output. ## Validation - Focused CLI regressions and existing stable/beta WebSocket suite: **59 passed**. - `CI=true ./scripts/test`: **7,013 handwritten + 556 generated tests passed**; the existing two generated test skips remain and all 12 snapshots passed. - `pnpm lint`, `pnpm exec tsc --noEmit` (TypeScript 6.0.3), and `pnpm build`: passed. - Actual CLI execution compiled as CJS and native ESM, using built public SDK entrypoints on Node **22.0.0, 24.20.0, and 26.7.0**: **24 lifecycle scenarios passed**. - Additional legacy-compiler comparison: TypeScript **4.5.5 and 4.9.5** both report the same pre-existing `TS7022` for the unrelated `turnResult` initializer on main and this branch (line 510 on main, 516 after these six additions). Those legacy example checks do not pass on either revision; this PR adds no new diagnostic and does not broaden scope to that typing issue. ## Adversarial self-review Reviewed and rechecked listener ownership, closure before/after terminal response events, consecutive tool-call turns, error precedence, and peer-controlled diagnostic data. The listener is registered before sending each request and removed by the existing shared cleanup path on both success and failure. No new state, parser, retry, timeout, or transport behavior is added to the SDK.
HAYDEN-OAI
added a commit
to 1fanwang/openai-node
that referenced
this pull request
Sep 4, 2026
## Summary Make the multi-agent SSE example report unsuccessful coordinator responses instead of silently exiting successfully. The example currently prints message/text events but ignores `response.failed` and `response.incomplete`. Both are valid beta Responses stream events. Unlike named SSE `error` frames, the SDK correctly yields these ordinary response events to the caller. When the stream ends, the example prints a newline and exits 0—even if the coordinator failed before producing any text or after partial output. ## Change Add a five-line branch in `examples/responses/multi-agent-streaming.ts` that throws a concise diagnostic for failed/incomplete coordinator responses. - Treat omitted/null ownership and explicit `/root` ownership as the coordinator. - Keep child-agent terminal events non-fatal so the coordinator can recover and continue. - Preserve successful output, agent labels, request parameters, and the existing SDK handling of named SSE errors. - Report only the terminal event type, not response/error payload details. This is separate from openai#2533 and openai#2544, which address WebSocket examples. The SSE example needs no WebSocket-style close handling: its existing iterator closes normally on completion and performs its existing abort cleanup when the loop throws. The shared ecosystem CLI fixtures also persist offline npm settings in each temporary project’s `.npmrc`. pnpm strips `npm_config_*` environment variables when starting workers, which allowed Node 22’s nested npm installs to make audit requests and time out. The dependency-free wrapper also disables pnpm’s automatic root installation, preventing parallel workers from racing on shared pnpm state. The fixtures retain real npm package installations and the existing 15-second timeout; failures now include subprocess output. Only this handwritten example and two handwritten test files change. No generated code, SDK APIs, dependencies, or budget rules change. ## Validation The ecosystem fixture correction passes all three install modes on Node 22.23.2 and all 31 ecosystem CLI/SSE tests on Node 24.20.0. Lint, TypeScript, and the CJS/ESM build pass. Validation of the streaming change: - Four executable regressions fail on main because failed/incomplete root responses exit 0; all eight final cases pass with the fix. - Tests run the actual documented `ts-node` entrypoint and public SDK against a local synthetic HTTP/SSE server. They cover failures before/after partial text, omitted/null/explicit-root ownership, all three child terminal statuses followed by coordinator success, and named SSE error handling. They also check the actual request path/body, exact output, exit status, and absence of credentials/private failure details in new diagnostics. - Complete `CI=true ./scripts/test`: 7,017 handwritten tests across 164 files, 556 generated tests across 82 suites, and 12 snapshots pass. Existing generated skips are unchanged. - `pnpm lint`, `pnpm exec tsc --noEmit`, and `pnpm build` pass. - The changed example compiles against the built SDK with strict TypeScript 4.5.5 and 4.9.5 checks (`skipLibCheck: false`). - An additional 138-case runtime matrix passes across Node 22.0.0, 24.20.0, and 26.7.0, loading actual built CJS/ESM SDK classes with the transpiled example. It covers root/child terminal events, partial output, SSE errors, original stdout-error identity, and request-signal abort on failure. The same matrix confirms the corresponding baseline failures. Repeated adversarial self-review and an independent bounded review found no remaining defect. Fixtures are synthetic and require no live API calls or credentials.
gh-actions-shared Bot
pushed a commit
to xf-qubit/openai-node
that referenced
this pull request
Sep 8, 2026
Automated Release PR --- ## [7.11.0](openai/openai-node@v7.10.0...v7.11.0) (2026-09-08) ### Features * **api:** Add prompt cache diagnostics ([openai#2608](openai#2608)) ([f83de8f](openai@f83de8f)) * **api:** add service-account API key expiration fields ([openai#2618](openai#2618)) ([eecbebe](openai@eecbebe)) * **api:** correct function argument completion event fields (openapi-545) ([openai#2611](openai#2611)) ([31bf7e7](openai@31bf7e7)) * **api:** recognize incomplete web search call status ([openai#2576](openai#2576)) ([c093c44](openai@c093c44)) ### Bug Fixes * **assistants:** retain terminal message snapshots ([openai#2650](openai#2650)) ([414677e](openai@414677e)) * **audio:** only signal started recording processes ([openai#2665](openai#2665)) ([cebdf7c](openai@cebdf7c)) * **audio:** prefer readable inputs over body metadata ([openai#2548](openai#2548)) ([86ff7d5](openai@86ff7d5)) * **audio:** use ALSA card identifiers when recording ([openai#2564](openai#2564)) ([67da19c](openai@67da19c)) * **audio:** wait for input cleanup before reporting success ([openai#2624](openai#2624)) ([5ed87fe](openai@5ed87fe)) * **auth:** account for token exchange elapsed time ([openai#2609](openai#2609)) ([b0f6f20](openai@b0f6f20)) * **auth:** validate Azure metadata access token values ([openai#2523](openai#2523)) ([32b4d35](openai@32b4d35)) * **azure:** carry declared options through withOptions and fix deployment paths ([openai#2514](openai#2514)) ([f8f29fb](openai@f8f29fb)) * **bedrock:** cancel requests awaiting AWS credentials ([openai#2610](openai#2610)) ([97c94e4](openai@97c94e4)) * bound partial number recovery to its token ([openai#2692](openai#2692)) ([0d306ca](openai@0d306ca)) * **build:** fail on reported transpile errors ([openai#2638](openai#2638)) ([833e3e8](openai@833e3e8)) * **build:** fail when a compiler worker is signaled ([openai#2633](openai#2633)) ([cd79cd1](openai@cd79cd1)) * **build:** honor configured compiler worker limits ([openai#2635](openai#2635)) ([aa08b44](openai@aa08b44)) * **build:** narrow Git-install staging detection ([openai#2621](openai#2621)) ([a6170d3](openai@a6170d3)) * **build:** postprocess the selected package manifest ([openai#2620](openai#2620)) ([1850d58](openai@1850d58)) * **build:** preserve CommonJS source-map positions ([openai#2623](openai#2623)) ([df0034d](openai@df0034d)) * **build:** preserve config-relative compiler project paths ([openai#2637](openai#2637)) ([3acb064](openai@3acb064)) * **build:** preserve declaration file import specifiers ([openai#2659](openai#2659)) ([0098c2e](openai@0098c2e)) * **ci:** propagate Node matrix check failures ([openai#2569](openai#2569)) ([bff92c9](openai@bff92c9)) * **ci:** serialize release-please branch updates ([openai#2631](openai#2631)) ([bb0e400](openai@bb0e400)) * clean up failed mock-server startups ([openai#2558](openai#2558)) ([83c92f0](openai@83c92f0)) * **deps:** bump qs to 6.16.0 for security fixes ([openai#2588](openai#2588)) ([4e5c6e3](openai@4e5c6e3)) * **docs:** await headers in Next.js webhook examples ([openai#2595](openai#2595)) ([7e48f05](openai@7e48f05)) * **docs:** correct the chat streaming helper contract ([openai#2596](openai#2596)) ([135ac7e](openai@135ac7e)) * **docs:** include image detail in the Vision example ([openai#2598](openai#2598)) ([e76bb16](openai@e76bb16)) * **docs:** omit absent Bedrock provider session tokens ([openai#2597](openai#2597)) ([41f85bd](openai@41f85bd)) * **ecosystem:** preserve retry delays in parallel workers ([openai#2673](openai#2673)) ([6164745](openai@6164745)) * **ecosystem:** reject invalid retry counts ([openai#2642](openai#2642)) ([d0e2bcf](openai@d0e2bcf)) * **ecosystem:** restore explicit web type-test module markers ([openai#2584](openai#2584)) ([3c189f4](openai@3c189f4)) * **embeddings:** preserve numeric output with body overrides ([openai#2549](openai#2549)) ([97d1ae7](openai@97d1ae7)) * **embeddings:** preserve numeric responses after request hooks ([openai#2590](openai#2590)) ([5202b13](openai@5202b13)) * **embeddings:** reject incomplete float32 vectors ([openai#2615](openai#2615)) ([3940a20](openai@3940a20)) * **embeddings:** validate response collection before decoding ([openai#2686](openai#2686)) ([888811f](openai@888811f)) * **events:** preserve unbound listener receivers ([openai#2644](openai#2644)) ([c8238e7](openai@c8238e7)) * **examples:** annotate Responses WebSocket turn result ([openai#2585](openai#2585)) ([dde19c5](openai@dde19c5)) * **examples:** avoid extra requests for empty tool calls ([openai#2672](openai#2672)) ([be292d3](openai@be292d3)) * **examples:** avoid unhandled WebSocket rejections ([openai#2669](openai#2669)) ([744b0ad](openai@744b0ad)) * **examples:** cancel Edge streams when requests abort ([openai#2530](openai#2530)) ([18100a1](openai@18100a1)) * **examples:** close failed streaming responses ([openai#2531](openai#2531)) ([eb9b468](openai@eb9b468)) * **examples:** close multi-agent WebSockets on terminal responses ([openai#2533](openai#2533)) ([51ec453](openai@51ec453)) * **examples:** drain manual conversation error output ([openai#2629](openai#2629)) ([e488b53](openai@e488b53)) * **examples:** find structured-output tool calls ([openai#2572](openai#2572)) ([84870a8](openai@84870a8)) * **examples:** keep tracking fine-tuning file validation ([openai#2528](openai#2528)) ([63312f2](openai@63312f2)) * **examples:** preserve background stream resume cursors ([openai#2671](openai#2671)) ([b7398e3](openai@b7398e3)) * **examples:** reject early multi-agent WebSocket closure ([openai#2599](openai#2599)) ([c6320f0](openai@c6320f0)) * **examples:** reject early Realtime socket closure ([openai#2639](openai#2639)) ([7db48ec](openai@7db48ec)) * **examples:** reject incomplete image streams ([openai#2617](openai#2617)) ([ef39397](openai@ef39397)) * **examples:** reject incomplete multi-agent streams ([openai#2666](openai#2666)) ([152b98d](openai@152b98d)) * **examples:** reject unsuccessful background responses ([openai#2600](openai#2600)) ([902a931](openai@902a931)) * **examples:** reject unsuccessful Responses streams ([openai#2645](openai#2645)) ([86acb95](openai@86acb95)) * **examples:** reject unsuccessful structured-tool responses ([openai#2668](openai#2668)) ([99aa303](openai@99aa303)) * **examples:** report audio demo failures ([openai#2570](openai#2570)) ([29c0572](openai@29c0572)) * **examples:** report image-streaming failures ([openai#2563](openai#2563)) ([947c555](openai@947c555)) * **examples:** report multi-agent streaming failures ([openai#2550](openai#2550)) ([19d9151](openai@19d9151)) * **examples:** report nonstreaming request failures ([openai#2603](openai#2603)) ([384dc26](openai@384dc26)) * **examples:** report premature WebSocket closure ([openai#2544](openai#2544)) ([7eb5f7e](openai@7eb5f7e)) * **examples:** report unsuccessful Azure Realtime responses ([openai#2612](openai#2612)) ([2b62209](openai@2b62209)) * **examples:** report unsuccessful fine-tuning jobs ([openai#2616](openai#2616)) ([e761c16](openai@e761c16)) * **examples:** report unsuccessful Realtime responses ([openai#2559](openai#2559)) ([5d3e584](openai@5d3e584)) * **examples:** respect raw streaming proxy backpressure ([openai#2663](openai#2663)) ([a11bb1b](openai@a11bb1b)) * **examples:** reuse completed background responses ([openai#2557](openai#2557)) ([8e1977b](openai@8e1977b)) * **examples:** serialize missing book lookup results ([openai#2573](openai#2573)) ([db7e371](openai@db7e371)) * **examples:** stop polling failed training files ([openai#2592](openai#2592)) ([1d67945](openai@1d67945)) * **examples:** support redirected streaming output ([openai#2546](openai#2546)) ([8df2e3a](openai@8df2e3a)) * **examples:** use the default audio recording input ([openai#2593](openai#2593)) ([6153484](openai@6153484)) * **format:** ignore missing literal file-list paths ([openai#2670](openai#2670)) ([516a849](openai@516a849)) * **format:** preserve file-list scope ([openai#2568](openai#2568)) ([7faff0e](openai@7faff0e)) * identify modern Edge browser runtime headers ([openai#2628](openai#2628)) ([9cdf6f0](openai@9cdf6f0)) * isolate monthly node review publication ([openai#2691](openai#2691)) ([3970dae](openai@3970dae)) * isolate realtime connection credentials ([openai#2690](openai#2690)) ([e9db491](openai@e9db491)) * keep formatter inputs on LF across Git checkouts ([openai#2515](openai#2515)) ([074618f](openai@074618f)) * **lint:** preserve leading-dash generated file paths ([openai#2640](openai#2640)) ([3b91a83](openai@3b91a83)) * **pagination:** follow next cursors through empty pages ([openai#2561](openai#2561)) ([9ea39ce](openai@9ea39ce)) * parse JSON media types case-insensitively ([openai#2551](openai#2551)) ([7ea7926](openai@7ea7926)) * **polling:** honor zero custom polling intervals ([openai#2518](openai#2518)) ([0cd94bb](openai@0cd94bb)) * **polling:** treat a cancelled vector store file as terminal ([openai#2506](openai#2506)) ([43c1007](openai@43c1007)) * preserve alternate Headers inputs ([openai#2560](openai#2560)) ([5d36ff4](openai@5d36ff4)) * preserve CommonJS constructor inheritance ([openai#2547](openai#2547)) ([aafb64e](openai@aafb64e)) * preserve cross-realm error causes ([openai#2555](openai#2555)) ([bfec504](openai@bfec504)) * preserve own Zod schema definition names ([openai#2695](openai#2695)) ([f65d190](openai@f65d190)) * preserve package selection in parallel ecosystem tests ([openai#2516](openai#2516)) ([fa0ce41](openai@fa0ce41)) * preserve streamed API error classification ([openai#2694](openai#2694)) ([8e2c6ee](openai@8e2c6ee)) * **realtime:** preserve base URL routing components ([openai#2567](openai#2567)) ([0f2e22d](openai@0f2e22d)) * **responses:** omit parsed values from replayed history ([openai#2634](openai#2634)) ([124b652](openai@124b652)) * **responses:** parse namespaced function tools ([openai#2574](openai#2574)) ([4d86684](openai@4d86684)) * **responses:** validate indexed response collections ([openai#2687](openai#2687)) ([1ecbf52](openai@1ecbf52)) * **runners:** observe cancellation after completion callbacks ([openai#2591](openai#2591)) ([27f5c6a](openai@27f5c6a)) * **runners:** preserve frozen input history ([openai#2525](openai#2525)) ([1211038](openai@1211038)) * **runners:** reuse caller abort listeners across tool turns ([openai#2594](openai#2594)) ([57d0da4](openai@57d0da4)) * **runners:** suppress content events for input history ([openai#2520](openai#2520)) ([37d6fc2](openai@37d6fc2)) * serialize tool-call IDs once per completion ([openai#2688](openai#2688)) ([8afb23e](openai@8afb23e)) * **streaming:** abort before early-exit cleanup ([openai#2613](openai#2613)) ([03acf94](openai@03acf94)) * **streaming:** avoid duplicate terminal tool argument events ([openai#2658](openai#2658)) ([12e5db6](openai@12e5db6)) * **streaming:** interrupt pending reads on cancellation ([openai#2556](openai#2556)) ([2d7a767](openai@2d7a767)) * **streaming:** surface server error frames and settle streams past throwing listeners ([openai#2502](openai#2502)) ([17a661e](openai@17a661e)) * support Edge Runtime without Node process ([openai#2552](openai#2552)) ([5dc4deb](openai@5dc4deb)) * **test:** cancel Steady commands waiting for cache locks ([openai#2636](openai#2636)) ([8896395](openai@8896395)) * **test:** preserve unrelated processes on mock startup failure ([openai#2622](openai#2622)) ([6896e66](openai@6896e66)) * **test:** reap Steady commands when lease creation fails ([openai#2664](openai#2664)) ([518756a](openai@518756a)) * **test:** reject HTTP errors in mock health checks ([openai#2667](openai#2667)) ([7b9e0b0](openai@7b9e0b0)) * **test:** reject unknown ecosystem project names ([openai#2626](openai#2626)) ([fcd12bb](openai@fcd12bb)) * **test:** run Bun validation with --from-npm ([openai#2625](openai#2625)) ([74f58a9](openai@74f58a9)) * **test:** use a monotonic Steady lock timeout ([openai#2661](openai#2661)) ([7b5d120](openai@7b5d120)) * **types:** account for optional event arguments ([openai#2527](openai#2527)) ([548a651](openai@548a651)) * **types:** include missing parsed response stream events ([openai#2521](openai#2521)) ([6a9db07](openai@6a9db07)) * **types:** preserve custom tool factory metadata ([openai#2526](openai#2526)) ([13dca7b](openai@13dca7b)) * **types:** preserve runnable function tuple positions ([openai#2524](openai#2524)) ([e345d0f](openai@e345d0f)) * **uploads:** accept cross-realm ArrayBuffers ([openai#2529](openai#2529)) ([69d7bda](openai@69d7bda)) * **uploads:** preserve explicit empty filenames ([openai#2532](openai#2532)) ([eabb487](openai@eabb487)) * **uploads:** reject zero batch upload concurrency ([openai#2519](openai#2519)) ([8b9405a](openai@8b9405a)) * validate partial image indices in streaming example ([openai#2693](openai#2693)) ([a9a4440](openai@a9a4440)) * **zod:** preserve strict schema compatibility with Zod 4.5.4 ([openai#2677](openai#2677)) ([c11870e](openai@c11870e)) * **zod:** retain the original tool parameter schema ([openai#2662](openai#2662)) ([fdb038e](openai@fdb038e)) ### Performance Improvements * **bedrock:** avoid copying repeated query values ([openai#2545](openai#2545)) ([80ab919](openai@80ab919)) ### Chores * **api:** document throttling and model overload responses — SDK-235 ([openai#2632](openai#2632)) ([9dd0524](openai@9dd0524)) * clarify Response types and remove unused handwritten code ([openai#2660](openai#2660)) ([86152ab](openai@86152ab)) * **deps-dev:** bump @aws-sdk/credential-provider-node from 3.972.78 to 3.972.81 ([openai#2681](openai#2681)) ([48629a3](openai@48629a3)) * **deps-dev:** bump @cloudflare/workers-types from 5.20260816.1 to 5.20260823.1 in /ecosystem-tests/cloudflare-worker ([openai#2535](openai#2535)) ([dc14bf5](openai@dc14bf5)) * **deps-dev:** bump @cloudflare/workers-types from 5.20260823.1 to 5.20260830.1 in /ecosystem-tests/cloudflare-worker ([openai#2676](openai#2676)) ([ae04066](openai@ae04066)) * **deps-dev:** bump @smithy/signature-v4 from 5.7.2 to 5.7.3 ([openai#2679](openai#2679)) ([41ca20d](openai@41ca20d)) * **deps-dev:** bump @swc/core from 1.15.47 to 1.16.1 ([openai#2538](openai#2538)) ([5c04dc4](openai@5c04dc4)) * **deps-dev:** bump fast-uri from 3.1.5 to 3.1.7 in /ecosystem-tests/ts-browser-webpack ([openai#2579](openai#2579)) ([34380bf](openai@34380bf)) * **deps-dev:** bump fast-uri from 3.1.5 to 3.1.7 in /ecosystem-tests/vercel-edge ([openai#2577](openai#2577)) ([511d1e4](openai@511d1e4)) * **deps-dev:** bump oxlint from 1.76.0 to 1.79.0 ([openai#2542](openai#2542)) ([2683c3a](openai@2683c3a)) * **deps-dev:** bump oxlint from 1.79.0 to 1.80.0 ([openai#2683](openai#2683)) ([02d5fb0](openai@02d5fb0)) * **deps-dev:** bump puppeteer from 25.7.0 to 25.8.0 in /ecosystem-tests/browser-direct-import ([openai#2534](openai#2534)) ([f039c09](openai@f039c09)) * **deps-dev:** bump puppeteer from 25.7.0 to 25.8.0 in /ecosystem-tests/ts-browser-webpack ([openai#2539](openai#2539)) ([5828818](openai@5828818)) * **deps-dev:** bump puppeteer from 25.8.0 to 25.9.0 in /ecosystem-tests/browser-direct-import ([openai#2675](openai#2675)) ([bebd72f](openai@bebd72f)) * **deps-dev:** bump puppeteer from 25.8.0 to 25.9.0 in /ecosystem-tests/ts-browser-webpack ([openai#2680](openai#2680)) ([fe3770f](openai@fe3770f)) * **deps-dev:** bump ultracite from 7.10.0 to 7.10.5 ([openai#2537](openai#2537)) ([6d37b30](openai@6d37b30)) * **deps-dev:** bump vitest from 4.1.10 to 4.1.11 ([openai#2541](openai#2541)) ([6ea24b3](openai@6ea24b3)) * **deps-dev:** bump webpack from 5.109.2 to 5.110.2 in /ecosystem-tests/ts-browser-webpack ([openai#2682](openai#2682)) ([73405d1](openai@73405d1)) * **deps-dev:** bump wrangler from 4.125.0 to 4.127.1 in /ecosystem-tests/cloudflare-worker ([openai#2678](openai#2678)) ([0d92742](openai@0d92742)) * **deps:** bump @azure/identity from 4.13.1 to 4.13.2 ([openai#2540](openai#2540)) ([0318287](openai@0318287)) * **deps:** bump express and @types/express ([openai#2684](openai#2684)) ([c581c17](openai@c581c17)) * **ecosystem:** remove unused proxy stub ([openai#2641](openai#2641)) ([ce1713a](openai@ce1713a)) * migrate to forked steady ([openai#2587](openai#2587)) ([67ce45d](openai@67ce45d)) ### Documentation * correct Bun ecosystem test instructions ([openai#2627](openai#2627)) ([3434b42](openai@3434b42)) * correct error handling diagnostics ([openai#2630](openai#2630)) ([d334d8c](openai@d334d8c)) * **helpers:** remove undefined main call from parsing example ([openai#2648](openai#2648)) ([d8f6be9](openai@d8f6be9)) * **realtime:** report unsuccessful responses in the guide ([openai#2647](openai#2647)) ([f58b1fd](openai@f58b1fd)) * **streaming:** reject unsuccessful background responses ([openai#2614](openai#2614)) ([1025f7a](openai@1025f7a)) * **tools:** support TypeScript 4.9 in the function loop ([openai#2649](openai#2649)) ([42e852b](openai@42e852b)) --- This PR was generated with [Release Please](https://github.com/googleapis/release-please). See [documentation](https://github.com/googleapis/release-please#release-please). Co-authored-by: openai-sdks[bot] <284451331+openai-sdks[bot]@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Make the multi-agent Responses WebSocket example finish on the coordinator's terminal response and close its connection on every exit path.
The existing loop closes only for
response.completed. A rootresponse.failedorresponse.incompleteleaves it waiting indefinitely on the still-open socket. API errors and exceptions while writing output leave the iterator without closing the connection; the public WebSocket iterator explicitly does not own socket closure. A child agent's completion also currently closes the example before the coordinator can finish.This change:
try/finally, with onews.close()cleanup site./rootconvention.Only the handwritten example and its regression tests change. No generated SDK files, public declarations, dependencies, policies, or custom-code budget changes are involved. Neither file appears in the verified generated snapshot, and no other open PR covers this example.
Regression coverage
The tests execute the actual example with the public
OpenAIand betaResponsesWSimplementations against an ephemeral localhost WebSocket server. Only client configuration and stdout are injected; socket transport, protocol dispatch, iteration, and the close handshake are real. All credentials and events are synthetic.The first six regressions were run before the fix: the completed-response control passed and all five failure/cleanup cases failed because the connection stayed open. The adversarial review then added child-agent termination cases, which exposed premature completion and prevented the initial failure handler from terminating the coordinator on a child's failure.
The final 11 tests cover success, failed/incomplete coordinator responses before and after partial text, each child terminal status followed by coordinator output, both API error shapes, and output-write failures. They verify a clean close handshake, unchanged request settings, preserved output formatting, and propagation of existing errors.
Validation
CI=true ./scripts/test: 7,020 handwritten + 556 generated tests passed, with the existing two generated test skips; all 12 snapshots passed.pnpm lint,pnpm exec tsc --noEmit, andpnpm build: passed.skipLibCheck: false.Adversarial self-review
Repeated the review after adding the child-agent safeguard. The final change has a single connection owner and cleanup path, does not add retries or parser state, preserves streamed output and existing error objects, and adds no response-body data to failure diagnostics. The request URL and beta header were checked in the compatibility smoke tests. API behavior and runtime requirements outside this example are unchanged.