Codex review: needs real behavior proof before merge.

Summary
Review failed before ClawSweeper could summarize the requested change.

Reproducibility: unclear. The review failed before ClawSweeper could establish a reproduction path.

Real behavior proof
Not applicable: Real behavior proof was not assessed because the Codex review failed.

Next step before merge
Review did not complete, so no work-lane recommendation was made.

Heads up — this isn't actually test-only. The PR's runtime change introduces a buffer-loss bug that the new test doesn't surface.

The bug. In src/agents/openai-ws-stream.ts, the new "ignore stale terminal event" branches at :1200-1202 (response.completed) and :1236-1238 (response.failed) call:

outputItemPhaseById.clear()
outputTextByPart.clear()
emittedTextByPart.clear()

…then return. Those three maps (declared at :1018-1020 inside the per-request closure) are written by the active turn's deltas (:1165-1166, :1072) and read by the normal completion path (:1205-1207) to assemble the final assistant message before the done event at :1224. Every other clear site in the file is terminal (followed by cleanup() + resolve/reject): abort :1086-1088, WS close :1099-1101, real completed :1205-1207, real failed :1241-1243, error :1255-1257. Only the new stale branches clear AND return — the subscription stays live but the buffer is gone.

So if turn B has accumulated any deltas before a stale response.completed arrives carrying turn-A metadata, the stale-ignore branch destroys them. When turn B's real completion lands, the assistant message is missing every delta that landed before the stale event. State-wise this is strictly worse than no-op — "ignore stale" in the sense of "leave state alone" was the right intent, but the implementation does the opposite.

The added test (openai-ws-stream.test.ts:3075-3158) doesn't surface this because it emits zero deltas between the turn-B send and the stale response.completed.

Smallest fix — six lines deleted, no other changes:

Delete the three .clear() calls in each stale branch (:1200-1202 and :1236-1238). No leak risk: the maps are per-request closure scope (new Map() inside the inner streamFn), so they GC when the request resolves; turn B's items have fresh server-assigned IDs that won't collide with leftover turn-A keys; the phase-replay loop at :1149 filters by ${event.item.id}: prefix, so leftover entries are skipped anyway.

Add the missing interleaved-delta regression test:

emit response.created (turn B)
emit response.output_item.added { item.id: "msg_b", phase: "answer" }
emit response.output_text.delta { item_id: "msg_b", delta: "Hello " }
emit response.output_text.delta { item_id: "msg_b", delta: "world" }
emit response.completed { id: "resp_turn_a", metadata.openclaw_turn_id: "turn_a" }   // STALE
emit response.output_text.done { item_id: "msg_b", text: "Hello world" }
emit response.completed { id: "resp_turn_b", metadata.openclaw_turn_id: "turn_b" }   // REAL
assert finalAssistantMessage.text === "Hello world"

Today this test fails on this PR head; after the six-line removal it passes.

Plus: Real behavior proof needs a captured trace, and check-test-types (= pnpm tsgo:test) needs to clear.

Alternative shape if you'd rather have cleaner history: split into a runtime-fix PR (clear-removal + interleaved-delta test) and keep this one as test-only against the post-fix tree.

Thanks @100yenadmin. Closing this as superseded by #79726.

This PR was plausible while OpenClaw owned the OpenAI Responses WebSocket event loop, but the better fix is to delete that owner boundary instead of hardening it. #79726 removes the custom src/agents/openai-ws-* transport, the stale terminal-event test surface, the custom WS session cleanup path, and the openaiWsWarmup config/docs surface. Explicit openai-codex/* Responses runs now go through PI native Codex Responses streaming, with OpenClaw only wrapping the stream for auth injection, abort signals, session id propagation, and prompt-cache boundary stripping.

That means the stale response.completed / response.failed metadata validation added here no longer has a runtime call site in OpenClaw. PI now owns the WebSocket/SSE transport behavior; OpenClaw should not keep parallel stale-event handling for a deleted transport.

Proof in #79726:

• deleted src/agents/openai-ws-connection.ts, src/agents/openai-ws-stream.ts, src/agents/openai-ws-message-conversion.ts, and related tests
• removed createOpenAIWebSocketStreamFn, releaseWsSession, shouldUseOpenAIWebSocketTransport, and openaiWsWarmup references from the OpenClaw-owned paths
• rg over src/agents, extensions/openai, docs, src/proxy-capture, and test/helpers finds no remaining custom WS symbols/config/docs
• targeted PI/OpenAI tests passed, and Testbox pnpm check:changed passed

#78055 is still referenced from #79726, but I am not treating this PR as the canonical issue fix because that issue also contains non-WS subagent announce/session-history contamination evidence.