Summary

sessions_spawn (subagent spawn) persists the raw canonical provider (anthropic) into the child session entry instead of resolving it to the configured CLI runtime (claude-cli). For depth-2 spawns with no cached CLI session binding, the child run registers, never launches a CLI process, and dies 30 minutes later when the sweeper reaps it:

"outcome": { "status": "error", "error": "subagent run lost active execution context" }
"endedReason": "subagent-error"

This is another unaddressed call-site of the regression family tracked in #57326 — the same missing resolveCliRuntimeExecutionProvider call that has now been fixed in the queued-followup path (#82857), cron executor, attempt-execution, and agent-runner.runtime. The two spawn dispatchers were never patched.

Both files verified still unpatched on current main (checked 2026-06-12) and in the released 2026.6.5 dist.

Environment

• OpenClaw 2026.6.5 (5181e4f), Linux arm64
• CLI-backed runtime: agents.defaults.models["anthropic/..."].agentRuntime.id = "claude-cli", auth profile anthropic:claude-cli (OAuth)
• Agent models use canonical refs (anthropic/claude-...), per the documented config pattern

Live evidence on 2026.6.5 — even successful spawns report the unresolved provider

Every sessions_spawn ack on 2026.6.5 returns the raw provider:

{
  "status": "accepted",
  "resolvedModel": "anthropic/claude-...",
  "resolvedProvider": "anthropic",   // ← never "claude-cli", even when claude-cli runs it
  "modelApplied": true
}

Depth-1 spawns still work because agent-runner.runtime repairs the provider downstream. Depth-2 spawns hit the unrepaired window (below) and intermittently die.

Smoking gun — single-field diff between working and failing spawns

Same agent (planner), same model, same config, same depth=2, same day:

The only persisted difference is whether the provider got resolved. When a previous spawn left a cached CLI session binding, downstream reuse masks the bug; a cold spawn (no binding) launches nothing and is swept at +30 min (sweepStaleRunContexts, maxAgeMs=1800s). That cache-masking is why the failure looks intermittent and why it has been hard to pin down.

Root cause (current main, verified at source)

Call site 1 — src/agents/subagent-spawn.ts

The file (1,773 lines) contains zero imports of resolveCliRuntimeExecutionProvider. The child session entry is persisted with the raw split of the model ref:

// src/agents/subagent-spawn.ts:364-388 (persistInitialChildSessionRuntimeModel)
const { provider, model } = splitModelRef(params.resolvedModel);  // "anthropic"
...
store[target.canonicalKey] = mergeSessionEntry(store[target.canonicalKey], {
  model,
  ...(provider ? { modelProvider: provider } : {}),   // ← persists RAW "anthropic"
});

buildResolvedSubagentModelMetadata (line 288) reports the same raw value back to the caller — which is exactly the "resolvedProvider": "anthropic" we see in every spawn ack.

Call site 2 — src/gateway/server-methods/agent.ts

The handler (2,848 lines) also has zero resolver imports. Provider flows raw end-to-end:

// line 1120
const providerOverride = allowModelOverride ? request.provider : undefined;
// line 1422
const effectiveProvider = providerOverride || baseProvider;
// line 2350 — reads back the session entry that call site 1 persisted raw
const activeModelProvider =
  providerOverride ??
  resolveSessionModelRef(cfgForAgent ?? cfg, sessionEntry, activeSessionAgentId).provider;
// line ~2540 — dispatches with the raw provider
dispatchAgentRunFromGateway({ ingressOpts: { provider: providerOverride, ... } });

Why depth 2 is the victim

The run is registered (registerSubagentRun) before agent-runner.runtime's own resolver fires. Depth-1 spawns from a live main session go through reply-path admission that resolves the runtime independently. Depth-2 spawns (subagent → subagent) depend entirely on the session entry persisted by call site 1 — which holds modelProvider: "anthropic". With no cached CLI binding to fall back on, the launch path tries to honor a provider that is not an executable runtime, silently no-ops, and the sweeper declares "lost execution context" 30 minutes later. Empty transcript, zero tokens, parent pipeline stalls. Related symptom reports: #77720, #47975.

Proposed fix — same pattern as the merged #82857

src/auto-reply/reply/followup-runner.ts:808-823 is the maintainer-accepted template:

const cliExecutionProvider =
  (sessionRuntimeOverride && isCliProvider(sessionRuntimeOverride, runtimeConfig)
    ? sessionRuntimeOverride
    : undefined) ??
  resolveCliRuntimeExecutionProvider({
    provider, cfg: runtimeConfig, agentId: run.agentId,
    modelId: model, authProfileId: selectedAuthProfile.authProfileId,
  }) ??
  provider;

Apply it at both unpatched sites:

1. subagent-spawn.ts — resolve before persisting (in persistInitialChildSessionRuntimeModel, or at its call site where cfg and the target agent id are in scope):

import { resolveCliRuntimeExecutionProvider } from "./model-runtime-aliases.js";

const { provider: rawProvider, model } = splitModelRef(params.resolvedModel);
const provider = rawProvider
  ? (resolveCliRuntimeExecutionProvider({
      provider: rawProvider,
      cfg: params.cfg,
      agentId: params.agentId,   // needs the target agent id threaded in
      modelId: model,
    }) ?? rawProvider)
  : rawProvider;

This makes the child session entry carry the executable runtime (claude-cli) from the moment it exists — closing the registration→runner window for every depth, and fixing the misleading resolvedProvider in the spawn ack as a side effect.

2. server-methods/agent.ts — resolve activeModelProvider before run registration (line ~2350):

const rawActiveModelProvider =
  providerOverride ??
  resolveSessionModelRef(cfgForAgent ?? cfg, sessionEntry, activeSessionAgentId).provider;
const activeModelProvider =
  resolveCliRuntimeExecutionProvider({
    provider: rawActiveModelProvider,
    cfg: cfgForAgent ?? cfg,
    agentId: activeSessionAgentId,
  }) ?? rawActiveModelProvider;

resolveCliRuntimeExecutionProvider returns undefined for non-CLI providers (src/agents/model-runtime-aliases.ts:232), so both changes are no-ops for embedded/API-backed setups — only configs with a CLI runtime mapping are affected.

Acceptance criteria

1. Depth-2 cold spawn (no prior CLI session binding for the target agent) with a canonical anthropic/... model on a claude-cli runtime: child launches a CLI process and produces a transcript.
2. Child session entry persists modelProvider: "claude-cli" (or the configured CLI runtime) — never the raw canonical provider — when a CLI runtime is configured.
3. Spawn ack resolvedProvider reflects the execution runtime.
4. No behavior change when no CLI backend is configured (resolver returns undefined → falls back to raw provider).

Repro

1. Configure a model with a CLI runtime (e.g. agents.defaults.models["anthropic/claude-sonnet-4-6"].agentRuntime.id = "claude-cli"), maxSpawnDepth >= 2.
2. Spawn agent A from the main session; have A sessions_spawn agent B (mode: "run") — B must be an agent with no prior CLI session binding (fresh install or wipe cliSessionBindings).
3. Observe: B registers, no CLI process starts, transcript stays empty, and after ~30 min the run fails with subagent run lost active execution context.
4. Inspect B's session entry: modelProvider: "anthropic", no claudeCliSessionId, no cliSessionBindings — the signature.

Relationship to existing reports

• CLI-backed helper paths still bypass CLI dispatch on latest main #57326 — same root pattern; this issue isolates the two spawn dispatcher call sites that none of the merged fixes covered
• fix: route remaining cli-backed helper runs through CLI paths #72499 — stalled/conflicting PR covering other helper paths (probe, slug); does not touch these two files
• fix(followup): route CLI runtime drains through CLI runner (#82847) #82857 — merged followup fix; its pattern is reused verbatim here
• Bug: Subagent child sessions receive no termination signal when parent dies mid-run #77720 / Subagent sessions persist after completion, main session becomes unresponsive #47975 — plausible downstream symptom reports of swept depth-2 children