fix(server): Grok replies finish when Grok finishes them, not when background work does

[juliusmarminge]

When Grok ends its prompt while a background subagent or monitor keeps running, the ACP adapter keeps the run open for that work (#13594, #13609). But it also left the last assistant message streaming, because it closes a text stream only when the next item starts or the run ends. In the maintainer's live thread, Grok's reply ROOT_DONE showed as streaming from 22:57:10 until 22:57:33, and completed only after the subagent finished at 22:57:31. In the monitor run it streamed from 22:58:37 until 22:58:44, when the first monitor tick arrived.

Fix

When the prompt ends and the adapter holds the run for background work (deferFinalizeForBackgroundWork, Grok only), it now closes the turn's text streams right away. Grok ends the reply with its prompt: session/prompt resolves, or the turn_completed / prompt_complete notifications for that prompt id settle it through makeXAiPromptCompletionRuntime. So the reply completes then, not when the background work does. The run stays open exactly as before. Anything Grok sends later in the run opens a new segment, as it already did after a tool call.

Proof

grok_background_subagent (recorded live) now asserts that the root run's ROOT_DONE assistant item completes before the subagent completes.

• Before the fix: AssertionError: ROOT_DONE streamed until the subagent finished: expected 160 to be below 157.
• After the fix: passes.

I added no such assertion to grok_monitor, because it would pass without the fix. The replay feeds recorded frames back to back, so the first monitor tick (a new tool update) already closes the reply before the monitor completes. Live, that tick came 7 s later (22:58:44), which is the gap this fix removes. Only the subagent recording can tell the two behaviors apart, since its later frames are on the child session and never close root text.

Verification

• vp test run src/orchestration-v2/testkit/OrchestratorReplayFixtures.integration.test.ts -t "grok|acp": 21 passed
• vp test run src/orchestration-v2/Adapters/AcpAdapterV2.test.ts src/orchestration-v2/Adapters/GrokAdapterV2.test.ts src/orchestration-v2/testkit/OrchestratorReplayFixtures.contract.test.ts: 133 passed
• vp exec tsc --noEmit -p . (apps/server): no error TS / warning TS
• vp run knip:check: clean
• vp lint on the touched files: only the existing unused NodePath import warning in AcpAdapterV2.ts
• Not run: repo-wide checks, or a new live Grok recording (the existing recordings already show the behavior)

Model: Claude Opus 5.5 (Claude Code)

🤖 Generated with Claude Code

---

[macroscopeapp]

Approvability

Verdict: Approved at cc8115c

Macroscope's review found this PR approvable — This focused Grok server fix closes the assistant message when the prompt completes while preserving the run for background work, with a targeted regression assertion covering the timing. The remaining change is test-only, and no defaults, schemas, infrastructure, or static-analysis settings are modified.

^{You can add or adjust custom eligibility rules. Learn more.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ No successful main baseline artifact is available yet. This run establishes the initial measurement.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	—	4.9 KiB	—	6.8 KiB	✅
Codex	Thread snapshot wire	—	3.7 KiB	—	4.9 KiB	✅
Codex	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Codex	Live turn WebSocket decoded	—	20.4 KiB	—	29.3 KiB	✅
Codex	Live turn messages	—	2	—	8	✅
Claude	Total thread wire	—	4.9 KiB	—	6.8 KiB	✅
Claude	Thread snapshot wire	—	3.7 KiB	—	4.9 KiB	✅
Claude	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Claude	Live turn WebSocket decoded	—	20.7 KiB	—	29.3 KiB	✅
Claude	Live turn messages	—	1	—	8	✅

Baseline: unavailable · PR result: cc8115c · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 106.1 KiB
• Claude decoded thread snapshot: 106.4 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.