Sub-agent concurrency, recovery, and stopping

Concurrency

Each spawning session has its own in-process queue for ordinary sub-agents. The setting agents.defaults.subagents.maxConcurrent limits concurrent child runs for that session (default 8). Independent sessions have independent budgets, so one busy session does not consume another session's sub-agent slots. A nested orchestrator's children use that orchestrator's budget, not its parent's.

The budget belongs to the child's immediate spawning/controller session. Changing where a child's completion is delivered does not move its execution to another session's budget. Accepted runs above the execution limit queue until a slot is available.

Swarm collector children (collect: true) instead use a dedicated subagent:swarm:<schedulerGroupKey> lane. Its cap is the group's resolved tools.swarm.maxConcurrent (default 32), independent of the parent's ordinary subagent:<immediate session> lane. An ordinary spawn from the same parent can start while its Swarm lane is full. Ordinary children spawned by a collector use that collector's own session lane. Each running collector costs one model stream and one Code Mode worker isolate; tune the Swarm cap for the Gateway's resources.

maxChildrenPerAgent is a separate admission limit on active children per session (default 5); increasing execution concurrency does not raise that limit. Collector admission instead uses Swarm's maxChildrenPerGroup (default 50) and maxTotalPerGroup (default 200); raising execution concurrency does not raise either group limit. Codex-native subagents use Codex's own scheduler and limits independently of these OpenClaw queues.

Suspended completion deliveries do not block new work. Native subagents, ACP sessions, and visible sessions retain their normal active-run limits and authorization checks independently of the delivery backlog.

OpenClaw warns when the delivery backlog reaches 25. Within a Gateway process, unchanged backlog counts do not repeat the warning every sweep. A count change at or above 25, or a return to that threshold after recovery, produces a new warning. The backlog size does not discard results or change their retention.

A suspended delivery expires after seven days. The warning records its run and session identifiers, original delivery error, and suspension time. Expiry ends automatic delivery; it does not turn a successful child execution into a failure. For a retained run, /subagents info <runId> shows the captured result, discarded delivery status, and original error, even after the run leaves the recent list. Inspect that result or available child session history and ask the requester to continue. Cleanup and retention still apply: cleanup: "delete" can remove the run, and hidden session cleanup can remove its transcript. Expiry does not promise that deleted results remain recoverable.

Liveness and recovery

OpenClaw does not treat endedAt absence as permanent proof that a sub-agent is still alive. When the reading process can verify a current execution owner or an exact queued collector reservation, an unended run keeps counting regardless of age. Persisted metadata alone does not establish that ownership in another process. Other unended runs stop counting as active/pending after the stale-run window (2 hours, or the configured run timeout plus a short grace period, whichever is longer). These retained counts govern /subagents list, status summaries, descendant completion gating, and per-session concurrency checks; they are not proof that an executor is live.

During a graceful restart or Gateway suspension, lifecycle events from already-admitted runs remain tracked through restart preservation and completion processing. An already-admitted replacement run can finish refreshing a deferred child result before shutdown. The refresh remains tracked until capture and persistence finish; it does not admit a new run.

Completion waits retire with their original registry database during shutdown. Late cleanup recovery leaves retained state for the next Gateway instead of retrying against a closed store. Failed writes with an unknown outcome still report an error and remain fenced until recovery reads the canonical state.

After a Gateway restart, the parent owns continuation of the user's task. Interrupted sub-agents are finalized through their normal completion path instead of automatically relaunched. Their results tell the parent that execution was interrupted and that partially completed actions need checking. A parent waiting for its child batch receives the settled results, including children that finished before the restart, and decides what work remains.

An already-admitted completion turn for a yielded nested requester is different from an orphaned child launch. Its frozen child-result batch retains the exact saved continuation across restart. Registry recovery waits for that owner instead of reporting interruption while the same continuation is being replayed. This does not authorize automatic relaunch of unrelated interrupted child work.

Recovery retires superseded requester-transfer generations before restoring the remaining claim. The current generation keeps its completion custody; obsolete transfers do not retry indefinitely. Transient persistence failures still retry.

Recovery handles both sessions marked abortedLastRun: true and hard kills that prevented the shutdown marker from being written. For a hard kill, the child session must still identify the exact run from the retired Gateway, with no newer run or admitted work owning that session. Live executions and queued collectors keep their existing owners. Orphaned runs settle their native execution state before cleanup, so retained child sessions do not leave phantom running activity. If settlement fails, completion remains available for retry.

Recovery remembers live session ownership without retaining saved prompt snapshots. Unchanged owned rows skip session reads; ownership release or a published session change requests another sweep within one second. Unavailable recovery reads retry per row with exponential backoff from one second to one minute. Registry lifecycle changes and session publications clear that backoff. Persisted rows without a live owner still require a fresh read before recovery can change their state.

Startup session maintenance reports retained run/task owners in one informational summary. Those rows remain with registry recovery; the session-only orphan repair does not compete for their ownership. Ownership changes during a repair and failed ownership checks still produce warnings.

The parent can inspect a retained child transcript and use sessions_send to continue that session, or spawn a replacement after confirming the old execution has stopped. Reusing a child restores its conversation context; it does not replay an interrupted command. Existing cleanup and retention settings still apply.

When a detached cleanup attempt logs subagent cleanup finalize failed, its retries use bounded backoff in the current Gateway process. If those retries are exhausted, the run remains recorded with incomplete cleanup; inspect the warning to identify the failing operation. Unrelated sub-agent completions do not restart failed cleanup or reset its retry budget. Descendant completion still wakes the current requester ancestors waiting on that work. These cleanup retries are separate from completion delivery.

When required registration has an unknown outcome or retained registry state forbids deleting the child session, a Gateway-hosted ordinary spawn's error keeps the child's session and run identifiers. If the first cancellation attempt also fails, the error reports unconfirmed termination and whether Gateway cleanup was scheduled. The Gateway retains the child's admission slot while retrying and rechecks the exact run owner before each attempt. If database admission retires while that child still runs, cancellation stops but its slot stays reserved until the child controller retires or Gateway shutdown takes over. The child session stays intact. Provisional session rollback, collector FIFO cleanup, and local embedded cleanup remain joined. Inspect the retained child before retrying the spawn.

Stopping

An explicit Stop targeting a parent run cancels the children associated with that run and their descendants, including ordinary sub-agents and Swarm collectors. Successful cancellation keeps selected queued collectors from starting while running children stop. Exact-run cancellation does not cancel unrelated turns or clear unrelated session-wide queues.

For raw child session keys such as global, registration records the known owning agent as childAgentId. Child cancellation derives that agent's session store from current configuration, just as it does for agent-qualified keys, which need no recorded binding. Clearing the child's queued follow-ups and commands stays within that agent, even when another agent uses the same raw session key. Legacy runs without a recorded owner use current configuration to resolve one agent and clear only that agent's queues. The optional binding stays in payload_json when an older build rewrites the run.

Owner-aware child lookups keep watched follow-ups, steering, run generations, completion transcripts, and timeout reconciliation separate when agents share a raw key. Terminal events and session timing updates use that same child owner. Rows without a recorded owner and callers without an explicit owner retain their existing key-only lookup behavior.

Stop also retires pending completion continuations for the selected work, even when a child has already finished. Cancelling a completion turn retires its matching child batch, so automatic delivery retries cannot start it again under a new run ID. Captured child results and their execution outcomes remain intact; you can inspect them or send a new instruction afterward.

A parent Stop also drains ordinary commands retained by its selected native children, even after their model execution has completed. Completed child results remain completed; command cleanup does not add canceled model runs to the result. Commands deliberately started with background: true remain independent. A live command continuation retains the original request's child ownership when it runs in another session. Stop selects only that request's bound native children, leaving newer human work and replacement child generations alone. A fresh session Stop can still select those children after the continuation has completed and its notification has been consumed. It uses the current original session's authority and generation; a historical exact-run Stop does not regain authority after that request's active, process, and event ownership has ended.

For Gateway callers, chat.abort with a runId uses this exact-parent scope. sessions.abort with a runId also targets that run. When it resolves a recovered native run without a chat controller, it cancels children only if the captured active parent accepts Stop; a declined or no-active-run result, including an already-finalizing parent, leaves those children alone.

Sending /stop in the requester chat has broader scope: it aborts requester session work, clears its queues, and cancels its active child tree. Session-wide sessions.abort also requests descendant cancellation; clearing queued follow-ups requires clearQueued: true. Ordinary chat.abort without a runId does not cascade to children. These operations retain their normal authorization checks.

A typed /stop sent through chat.send honors expectedLeafEntryId and, when that branch check is present, sessionId. If the check fails during descendant cancellation, the Gateway refuses further cancellation and reports active-leaf-changed. Cancellation already accepted by a child still settles. The cancellation result waits for pending child-session metadata writes before checking that the original session still owns the outcome.

Incomplete cancellation is reported as an error, not a clean success. /stop reports actual stopped and failed child counts. A committed child cancellation remains in the stopped count if later cleanup fails or the parent is replaced. The cleanup error remains visible; a replacement run requires its own Stop request. Inspect the remaining native subagent runs with subagents and retry their cancellation; request acknowledgment does not mean all runtime cleanup is instantaneous.

Accepted children remain independent after ordinary parent completion, yield, or timeout. Those events do not automatically cancel them.

Limitations

• Direct announce attempts are best-effort, but admitted session-queued completion handoffs and their native subagent records survive gateway restarts in the shared SQLite state database.
• Sub-agents still share the same Gateway process resources; concurrency caps apply per spawning session or Swarm group, not to total Gateway concurrency.
• sessions_spawn returns { status: "accepted", runId, childSessionKey } when startup is accepted, without waiting for the child task to finish. Cloud-worker spawns can wait for provisioning before returning this receipt.
• Sub-agent context only injects AGENTS.md (no SOUL.md, IDENTITY.md, USER.md, MEMORY.md, or BOOTSTRAP.md). Its ## Tools section carries environment-specific notes. Codex-native subagents follow the same boundary through native AGENTS.md discovery, while parent-only persona, identity, and user files are injected as turn-scoped collaboration instructions so children do not clone them.
• Recursive spawning is enabled through depth 5 by default. Set maxSpawnDepth from 1 through 5 to lower the boundary.
• maxChildrenPerAgent caps active children per session (default 5, range 1-20).