fix(agents): clamp subagent spawn thinking overrides

Fixes #92412.

Subagent spawns that request an unsupported explicit thinking level now clamp through the existing provider/model thinking fallback instead of hard-failing after the orchestrator has already received an accepted ack. The exception is limited to trusted subagent spawn runs by requiring both the subagent lane and a subagent-shaped session key, so interactive and non-subagent explicit `--thinking` validation still fails loudly.

Verification:
- `node scripts/run-vitest.mjs src/agents/agent-command.live-model-switch.test.ts --maxWorkers=1`
- `.agents/skills/autoreview/scripts/autoreview --mode branch --base origin/main`
- Testbox-through-Crabbox `tbx_01kv2wt0nqavsmnvzzzy2antrc`: `OPENCLAW_CHECK_CHANGED_REMOTE_CHILD=1 OPENCLAW_CHANGED_LANES_RAW_SYNC=1 corepack pnpm check:changed`
- GitHub PR checks clean on `c71186863337d9dfb9a18e5349ebef634a7d5ccd`

src/agents/agent-command.live-model-switch.test.ts

@@ -328,7 +328,6 @@ vi.mock("../routing/session-key.js", async () => {
   );
   return {
     ...actual,
-    isSubagentSessionKey: () => false,
     normalizeAgentId: (id: string) => id,
     normalizeMainKey: (key?: string | null) => key?.trim() || "main",
   };
@@ -1340,6 +1339,75 @@ describe("agentCommand – LiveSessionModelSwitchError retry", () => {
     });
   });
 
+  it("clamps unsupported explicit thinking for subagent spawns instead of throwing", async () => {
+    setupSingleAttemptFallback();
+    state.runAgentAttemptMock.mockResolvedValue(makeSuccessResult("anthropic", "claude-fable-5"));
+    state.resolvedSessionKeyMock = "agent:planner:subagent:00000000-0000-4000-8000-000000000000";
+    state.isThinkingLevelSupportedMock.mockReturnValue(false);
+    state.resolveSupportedThinkingLevelMock.mockReturnValue("high");
+
+    await agentCommand({
+      message: "hello",
+      sessionKey: state.resolvedSessionKeyMock,
+      thinking: "xhigh",
+      lane: "subagent",
+    });
+
+    expect(state.resolveSupportedThinkingLevelMock).toHaveBeenCalled();
+    expectRecordFields(mockCallArg(state.runAgentAttemptMock), {
+      resolvedThinkLevel: "high",
+    });
+  });
+
+  it("rejects unsupported explicit thinking for interactive subagent-key runs", async () => {
+    setupSingleAttemptFallback();
+    state.runAgentAttemptMock.mockResolvedValue(makeSuccessResult("anthropic", "claude-fable-5"));
+    state.resolvedSessionKeyMock = "agent:planner:subagent:00000000-0000-4000-8000-000000000000";
+    state.isThinkingLevelSupportedMock.mockReturnValue(false);
+
+    await expect(
+      agentCommand({
+        message: "hello",
+        sessionKey: state.resolvedSessionKeyMock,
+        thinking: "xhigh",
+      }),
+    ).rejects.toThrow(/is not supported/u);
+    expect(state.runAgentAttemptMock).not.toHaveBeenCalled();
+  });
+
+  it("rejects unsupported explicit thinking for non-subagent sessions on the subagent lane", async () => {
+    setupSingleAttemptFallback();
+    state.runAgentAttemptMock.mockResolvedValue(makeSuccessResult("anthropic", "claude-fable-5"));
+    state.resolvedSessionKeyMock = "agent:main:main";
+    state.isThinkingLevelSupportedMock.mockReturnValue(false);
+
+    await expect(
+      agentCommand({
+        message: "hello",
+        sessionKey: state.resolvedSessionKeyMock,
+        thinking: "xhigh",
+        lane: "subagent",
+      }),
+    ).rejects.toThrow(/is not supported/u);
+    expect(state.runAgentAttemptMock).not.toHaveBeenCalled();
+  });
+
+  it("rejects unsupported explicit thinking for direct interactive runs", async () => {
+    setupSingleAttemptFallback();
+    state.runAgentAttemptMock.mockResolvedValue(makeSuccessResult("anthropic", "claude-fable-5"));
+    state.resolvedSessionKeyMock = "agent:main:main";
+    state.isThinkingLevelSupportedMock.mockReturnValue(false);
+
+    await expect(
+      agentCommand({
+        message: "hello",
+        to: "+1234567890",
+        thinking: "xhigh",
+      }),
+    ).rejects.toThrow(/is not supported/u);
+    expect(state.runAgentAttemptMock).not.toHaveBeenCalled();
+  });
+
   it("skips the initial session touch after gateway ingress already persisted activity", async () => {
     setupSingleAttemptFallback();
     state.runAgentAttemptMock.mockResolvedValue(makeSuccessResult("openai", "gpt-5.4"));

src/agents/agent-command.ts

@@ -819,6 +819,7 @@ async function prepareAgentCommandExecution(opts: AgentCommandOpts, runtime: Run
     manifestMetadataSnapshot,
     modelManifestContext,
     runId,
+    isSubagentLane,
     acpManager,
     acpResolution,
   };
@@ -859,6 +860,7 @@ async function agentCommandInternal(
     cwd,
     agentDir,
     runId,
+    isSubagentLane,
     acpManager,
     acpResolution,
     pluginsEnabled,
@@ -1597,7 +1599,11 @@ async function agentCommandInternal(
       })
     ) {
       const explicitThink = Boolean(thinkOnce || thinkOverride);
-      if (explicitThink) {
+      const isSubagentSpawnRun = isSubagentLane && isSubagentSessionKey(sessionKey);
+      // Spawn-lane subagents are fire-and-forget; the orchestrator already got
+      // an "accepted" ack, so throwing here strands the run and half-fails fan-outs.
+      // Clamp like the embedded runner; interactive --thinking keeps the throw.
+      if (explicitThink && !isSubagentSpawnRun) {
         throw new Error(
           `Thinking level "${resolvedThinkLevel}" is not supported for ${provider}/${model}. Use one of: ${formatThinkingLevels(provider, model, ", ", thinkingCatalog)}.`,
         );