Вызываете sessions_spawn из main Agent — получаете ACP_TURN_FAILED, invalid handshake 1008 или queue owner unavailable, при этом прямой чат на том же Gateway жив. Статья закрывает: ① когда runtime=acp vs runtime=subagent；② почему streamTo/resumeSessionId legal только на acp и как triage'ить misconfig на subagent；③ subagent fallback при ACP handshake fail + Windows OPENCLAW_ACPX_RUNTIME_STARTUP_PROBE；④ autofill streamTo under Completions vs Responses. Дополняет upgrade escort ACP triage и Docker subagent 1008 pairing — фокус только на runtime selection для multi-Agent spawn.
runtime=acp без ACP bridge: main channel OK, spawn → ACP_TURN_FAILED — acpx не зарегистрирован или queue owner offline, не quota модели.streamTo/resumeSessionId в runtime=subagent: поля только для ACP session resume; subagent = embedded Gateway RPC — invalid params или silent drop, симптом «sub-agent без контекста».OPENCLAW_ACPX_RUNTIME_STARTUP_PROBE.streamTo на Responses, не на Completions — кажется «spawn сломан», на деле API shape + runtime mismatch.В 2026 upstream sessions_spawn — два mutually exclusive runtime: runtime=acp через ACP bridge к acpx, streamTo для backflow в main UI; runtime=subagent — lightweight sub-agent inside Gateway, lower latency, no acpx, без ACP resume fields. Wrong runtime + wrong fields = самый дорогой on-call «fake complexity» — выровняйте path, потом model/allowlist (tools.profile triage).
Каждый spawn — scheduling contract в ticket: runtime, нужен ли UI backflow, Completions vs Responses на main session, fallback (subagent/rollback). Без контракта «вчера spawn работал» не объяснить — diff обычно в template fields/API shape, не в Gateway PID.
| Уже на сайте | Эта статья | Не дублируем |
|---|---|---|
| Upgrade escort ACP triage | acp vs subagent при spawn + fallback | backup create, full probe ladder |
| Docker subagent 1008 | Граница subagent vs pairing 1008 | Compose trustedProxies step-by-step |
| tools.profile triage | «Spawn OK, sub-agent без tools» | allowlist layering runbook |
| SSH dedicated Gateway | acpx + subagent на remote Mac | port-forward, launchd детали |
runtime=acp vs runtime=subagent: какой stack когдаRule of thumb: UI backflow, resume ACP session, align с Cursor/IDE acpx → acp；Gateway closed loop, low deps, acp down → subagent. Четыре prod shapes (~80% tickets):
| Task shape | Runtime | Key params | Избегать |
|---|---|---|---|
| Sub-agent output stream в main chat | acp |
streamTo → main session; opt resumeSessionId |
subagent + streamTo |
| Background batch, no UI | subagent |
task + timeout; no streamTo |
Force acp → bridge blast radius |
| queue owner unavailable | subagent | Ticket fallback; fix acpx reg parallel | Retry acp → MTTR |
| Multi-container Docker, RPC OK, spawn 1008 | Fix pairing/net, then subagent |
trustedProxies; Docker runbook | Runtime switch до fix bind |
streamTo / resumeSessionId: acp-only misconfig triage«Params выглядят OK, sub-agent пустой» — почти всегда field/runtime cross-contamination. Gateway на subagent path strip/reject ACP fields; template с Completions streamTo:"main" + runtime=subagent → generic RPC error, не «invalid field» — смотрите call JSON.
На runtime=acp: resumeSessionId resume acpx session; streamTo направляет token stream в Control UI/channel. 2026 Responses router autofill streamTo при bound Responses session; Completions — нет. Migrate Responses→Completions без template change → «раньше backflow, теперь sub-agent тихо в фоне» — не upgrade bug, API+runtime combo.
Field-strip experiment: copy fail JSON, drop streamTo/resumeSessionId, runtime=subagent. Instant OK → ACP path/misconfig, не task/model. Else pairing/Token/tools surface. Step 4 runbook — не крутите acp logs часами.
// ✅ acp: UI backflow + optional resume
{
"tool": "sessions_spawn",
"runtime": "acp",
"task": "Research competitor pricing, output table",
"streamTo": "main",
"resumeSessionId": "acp-sess-abc123"
}
// ✅ subagent: Gateway loop, no streamTo
{
"tool": "sessions_spawn",
"runtime": "subagent",
"task": "Batch rename files in logs/"
}
// ❌ classic misconfig
{
"runtime": "subagent",
"streamTo": "main"
}
ACP_TURN_FAILED, 1008, queue ownerruntime=acp fail, main chat OK — route по log fingerprint. Не смешивать с «Gateway dead» (channel/model runbooks).
| Log / symptom | Suspect first | Action |
|---|---|---|
ACP_TURN_FAILED |
acpx not ready; turn timeout | Windows OPENCLAW_ACPX_RUNTIME_STARTUP_PROBE=1; bridge reg |
| invalid handshake / WS 1008 | CLI/Gateway/acpx version split | Pin versions; single reload; pin matrix |
| queue owner unavailable | ACP bridge reg lost (2026.3.x window) | host acpx; temp runtime=subagent |
| subagent тоже 1008 | pairing/token/net (not ACP-only) | Docker 1008 runbook |
| spawn OK, sub-agent no tools | tools.profile / agent override | tools.profile triage |
Fallback policy: acp fail два тура подряд (post-reload retest), subagent minimal probe pass → ticket «temp runtime=subagent», restrict UI-backflow tasks до fix acpx. Orthogonal к digest rollback: fallback = SLA, rollback = known regression.
OPENCLAW_ACPX_RUNTIME_STARTUP_PROBEProvider extensions + Defender на Windows растягивают acpx cold start; spawn до ready bridge → ACP_TURN_FAILED или invalid handshake. OPENCLAW_ACPX_RUNTIME_STARTUP_PROBE=1 (semantics — docs вашей версии) — extra health wait перед first spawn. Still fail → same-box subagent minimal task validates scheduling stack, isolate acpx install/ACL — не путать slow Windows boot с mandatory OpenClaw downgrade.
Upgrade escort = post-migration probe/ACP regression; эта статья = stable version, spawn layer mis-pick или flaky handshake. Ladder pass first — иначе split-brain маскируется под streamTo misconfig.
# Windows: extend acpx startup probe $env:OPENCLAW_ACPX_RUNTIME_STARTUP_PROBE = "1" openclaw gateway status openclaw gateway probe # Minimal subagent probe (no streamTo) # Control UI or CLI: equivalent sessions_spawn, runtime=subagent
| Blast radius | Keep fixing acp | Temp subagent | Pin/rollback |
|---|---|---|---|
| Only spawn/acp red; main + subagent probe OK | acpx reg, startup probe | Default: no-UI tasks → subagent | If known regression window |
| acp + subagent fail; probe red | Only after backup restore | Not first choice | First backup or digest rollback |
| Must UI backflow; acp dead | Business hurt before bridge fix | Can't replace streamTo | Rollback or remote Mac dedicated Gateway |
runtime=subagent — prod target 0 (lint/templates).Notebook Gateway + acpx + sleep + multi-provider plugins → «OpenClaw unstable». Better: authoritative Gateway + acpx на always-on remote Mac, local только SSH forward Control UI — spawn/probe same node, aligned log timeline.
Крутить prompt без проверки runtime/streamTo match — раздувает ACP handshake до «multi-Agent down». Decision matrix + misconfig triage + subagent fallback в runbook сжимают on-call с ночи blind tries до probed, fallback'd, metered minute incidents.
Hard-push acp bridge на Windows laptop или split Docker topology — три hidden costs: race fake 1008, Completions/Responses streamTo drift, overlapping failure domains. Prod Gateway 7×24, ticketed spawn, acp/subagent switch — MACCOME Mac mini M4/M4 Pro + six-region lease обычно дешевле total cost, чем queue owner на clamshell. Regional guide + SSH runbook для topology.
FAQ
streamTo и resumeSessionId при runtime=subagent?
Нет — только runtime=acp. subagent: task и embedded fields. Template review: цены аренды.
ACP 1008 или queue owner unavailable — обязателен rollback?
Не обязательно — symptom table; temp runtime=subagent + Windows probe. Dual-path fail → digest rollback; access → центр помощи.