xAI grok-4.6 completions streams stall at the 120s idle watchdog and subagent reviews are marked failed despite completed thinking

[nerdy2quant]

Summary

xai/grok-4.6 subagent reviews (architect / critic) repeatedly die with:

OpenAI completions stream stalled while waiting for the next event

The reviewer has already finished the review in thinking. The official assistant text never arrives, so the parent session receives only failed and cannot apply the review.

This is the xAI / openai-completions counterpart of #4096 (Anthropic max-reasoning vs the shared 120s idle watchdog). #4096 is closed and Anthropic-specific. The same watchdog still kills Grok xhigh after first-token / first-thinking progress.

Environment

• Host: macOS darwin 25.5.0, arm64 (Apple M4 Pro)
• GJC binary: /opt/homebrew/bin/gjc (local build dated 2026-08-20)
• Parent + reviewers: xai/grok-4.6, api: openai-completions, thinking xhigh
• Workflow: superpowers-brainstorming independent spec review, then superpowers-writing-plans critic
• Repo: private product session, not a GJC source checkout
• PI_STREAM_IDLE_TIMEOUT_MS, PI_OPENAI_STREAM_IDLE_TIMEOUT_MS, and GJC_OPENAI_STREAM_IDLE_TIMEOUT_MS unset

Reproduction

1. Run superpowers-brainstorming on a non-trivial spec with default grok-4.6 / xhigh.
2. Dispatch read-only task(agent="architect") for the spec review.
3. The reviewer reads files and starts writing ## Spec Review in thinking.
4. After ~120s of no further SSE events, the turn ends as stopReason: error.
5. Retrying the same review dies the same way. A later critic plan review dies the same way.

Evidence

Three consecutive reviewer turns in one session:

Subagent	Role	TTFT	Duration	Result
0-GrokSpecReview	architect	1683ms	141813ms	error, usage all zeros
1-GrokSpecReview2	architect	1173ms	132381ms	error, usage all zeros
2-GrokPlanReview	critic	1435ms	121933ms	error, usage all zeros

Canonical error on all three:

stopReason: error
errorMessage: OpenAI completions stream stalled while waiting for the next event
api: openai-completions
provider: xai
model: grok-4.6
usage: input=0 output=0 cacheRead=0 cacheWrite=0 totalTokens=0

The first two turns already contained a completed review in thinking, including ## Spec Review, Status: Issues Found, and a concrete finding about --model default vs test_main_cli_knobs.py. There was no official text part, so the parent inspect/await path reported:

0-GrokSpecReview — failed
1-GrokSpecReview2 — failed

The parent then cancelled the still-running second reviewer and proceeded without an independent review. The review content was recoverable only by manually reading the child jsonl after the user asked why it failed.

GPT-family reviewers on openai-codex-responses do not reproduce this in the same workflow, because they keep emitting reasoning/stream events during thinking. Grok xhigh goes silent after a fast TTFT.

Why this is not just #4096

#4096 already established:

• shared default is DEFAULT_STREAM_IDLE_TIMEOUT_MS = 120_000
• Anthropic can open a thinking block, then stay silent long enough to trip the watchdog
• workaround: export PI_STREAM_IDLE_TIMEOUT_MS=300000
• expected fix: Anthropic-specific 300s default

This report adds a still-open provider path:

• openai-completions + xai + grok-4.6 xhigh
• first token / thinking arrives in ~1–2s
• then 120s+ of no events during silent reasoning
• mid-stream stall is non-retryable once firstTokenTime is set (team: worker turns die on non-retryable mid-stream idle stall and are never auto-continued #2580)
• subagent result promotion drops the completed thinking and returns only failed

So even if the user knows the env workaround, a stalled reviewer still wastes the completed review.

Expected behavior

1. openai-completions / xAI should not use the 120s Anthropic-era default for xhigh reasoning gaps. A provider-specific default (300s, matching Anthropic max-reasoning streams repeatedly abort at the 120s idle watchdog #4096) or a documented GJC_OPENAI_STREAM_IDLE_TIMEOUT_MS default would avoid the false stall.
2. If a subagent turn still stalls after thinking already contains a structured review (## Spec Review, status, findings), promote that thinking to the parent as a successful/partial result instead of failed with no payload.
3. Parent subagent inspect/await should include errorMessage and the recovered thinking excerpt, not just failed.

Workaround

• Send only the reviewer (architect / critic) to a GPT-family model.
• Or lower Grok reasoning below xhigh.
• Or set PI_STREAM_IDLE_TIMEOUT_MS=300000 / GJC_OPENAI_STREAM_IDLE_TIMEOUT_MS=300000 if those still apply to openai-completions.

The first workaround is what we will use locally. The second/third do not fix the lost-review promotion bug.

[Yeachan-Heo]

OWNER_CONFIRMATION_REQUIRED — implementation evidence is terminal on PR #4748 at exact head ad25c508aa3469f2e181aa2099f3d4de71e2d8f9: Dev CI 32328081510 is green, the exact three-dot SHA-256 is faadc5d118f1fe3a4acbc8cf7485b35fe97ee323a2c201089f96d60af4e7ae4e, and git verify-commit verifies the head SSH signature. The sole blocker is eligible independent exact-head approval: @probepark was requested and has write permission, but has not approved. The canonical PR verdict remains needs-human; no merge or close is authorized.
—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[Yeachan-Heo]

RESOLVED — PR #4748 merged to dev as dadaf198804b13d5cff3a19767b5418e43dbf649. Exact-head review 4979337021, current-head Dev CI 32339420717, exact-head PR contract 32339420913, and fresh merged-dev bun run build all passed. The merged exact head was SSH-signed and bound to digest 20f9f5e3470812c5e756ce614d0080add22d687cc004c065d468ebea005496d9. No release, tag, or publish action occurred.
—
[repo owner's gaebal-gajae (clawdbot) 🦞]