crewly 1.20.436 → 1.20.438

package/packages/crewly-agent/src/runtime/subagent.test.ts

@@ -0,0 +1,497 @@
+// Short-lived sub-agents (#1317): tool restriction, config derivation, spawn tool.
+import { describe, it, expect, vi, type Mocked } from 'vitest';
+import path from 'node:path';
+import { homedir } from 'node:os';
+import {
+  SUBAGENT_READ_TOOLS,
+  SUBAGENT_WRITE_TOOLS,
+  REPORT_TOOL_NAME,
+  SPAWN_TOOL_NAME,
+  buildSubagentConfig,
+  buildSubagentPrompt,
+  createSpawnSubagentTool,
+  createSubagentRegistry,
+  forwardChildToolCalls,
+  isInsideDir,
+  resolveSubagentModel,
+  resolveSubagentPath,
+  restrictToolsForSubagent,
+  summarizeSubagentRun,
+  type SubagentReport,
+  type SubagentRunner,
+} from './subagent.js';
+import { createTools, getToolNames, TOOL_SENSITIVITY } from './tool-registry.js';
+import { CrewlyApiClient } from './api-client.js';
+import { CREWLY_AGENT_DEFAULTS, WRITE_TOOLS, type AgentRunResult, type CrewlyAgentConfig, type ToolDefinition } from './types.js';
+
+const CWD = '/work/project';
+
+const parentConfig: CrewlyAgentConfig = {
+  model: { provider: 'deepseek', modelId: 'deepseek-chat', temperature: 0.2, maxTokens: 4096 },
+  maxSteps: 500,
+  sessionName: 'think-tank-kai',
+  apiBaseUrl: 'http://localhost:8787',
+  systemPrompt: 'You are Kai, the team lead.',
+  maxHistoryMessages: 40,
+  compactionThreshold: 0.8,
+  projectPath: CWD,
+  memberId: 'member-1',
+  mcpServers: { playwright: { command: 'npx', args: ['playwright-mcp'] } },
+};
+
+/** A run result with the given tool calls and no incompleteness. */
+function runResult(partial: Partial<AgentRunResult> = {}): AgentRunResult {
+  return {
+    text: 'final text',
+    steps: 3,
+    usage: { input: 100, output: 20, cachedInput: 5 },
+    toolCalls: [],
+    finishReason: 'stop',
+    ...partial,
+  };
+}
+
+/** A scripted child runner. */
+function fakeRunner(opts: {
+  result?: AgentRunResult | (() => Promise<AgentRunResult>);
+  report?: SubagentReport | null;
+  throwOn?: 'initialize' | 'run';
+}): SubagentRunner & { shutdown: ReturnType<typeof vi.fn>; run: ReturnType<typeof vi.fn> } {
+  const run = vi.fn(async (_m: string, _c?: string, _md?: Record<string, string>, _o?: { abortSignal?: AbortSignal }) => {
+    if (opts.throwOn === 'run') throw new Error('provider exploded');
+    return typeof opts.result === 'function' ? opts.result() : (opts.result ?? runResult());
+  });
+  return {
+    initialize: vi.fn(async () => {
+      if (opts.throwOn === 'initialize') throw new Error('no api key');
+    }),
+    run,
+    shutdown: vi.fn(async () => undefined),
+    getSubagentReport: () => opts.report ?? null,
+  };
+}
+
+describe('subagent', () => {
+  describe('isInsideDir / resolveSubagentPath', () => {
+    it('accepts the directory itself and anything beneath it', () => {
+      expect(isInsideDir(CWD, CWD)).toBe(true);
+      expect(isInsideDir(`${CWD}/src/a.ts`, CWD)).toBe(true);
+      expect(isInsideDir(`${CWD}/src/../src/b.ts`, CWD)).toBe(true);
+    });
+
+    it('rejects siblings, parents and look-alike prefixes', () => {
+      expect(isInsideDir('/work/project2/a.ts', CWD)).toBe(false);
+      expect(isInsideDir('/work', CWD)).toBe(false);
+      expect(isInsideDir(`${CWD}/../other/x`, CWD)).toBe(false);
+      expect(isInsideDir('/etc/passwd', CWD)).toBe(false);
+    });
+
+    it('resolves relative paths against cwd and expands ~ and $HOME like the file tools', () => {
+      expect(resolveSubagentPath('src/a.ts', CWD)).toBe(`${CWD}/src/a.ts`);
+      expect(resolveSubagentPath('/abs/x', CWD)).toBe('/abs/x');
+      expect(resolveSubagentPath('~/notes.md', CWD)).toBe(path.join(homedir(), 'notes.md'));
+      expect(resolveSubagentPath('$HOME/notes.md', CWD)).toBe(path.join(homedir(), 'notes.md'));
+    });
+  });
+
+  describe('restrictToolsForSubagent', () => {
+    let mockClient: Mocked<CrewlyApiClient>;
+    let full: Record<string, ToolDefinition>;
+
+    const policy = { cwd: CWD, readOnly: false, depth: 1, parentSession: 'think-tank-kai' };
+
+    function freshTools(): Record<string, ToolDefinition> {
+      mockClient = {
+        get: vi.fn<any>().mockResolvedValue({ success: false, data: null, status: 404 }),
+        post: vi.fn<any>(),
+        delete: vi.fn<any>(),
+      } as any;
+      return createTools(mockClient, 'think-tank-kai/sub-1', CWD);
+    }
+
+    it('keeps only file, search and shell tools plus report_to_parent', () => {
+      full = freshTools();
+      const restricted = restrictToolsForSubagent(full, policy, () => undefined);
+      const names = Object.keys(restricted).sort();
+      const expected = [...SUBAGENT_READ_TOOLS, ...SUBAGENT_WRITE_TOOLS, REPORT_TOOL_NAME].filter((n) => n === REPORT_TOOL_NAME || full[n]).sort();
+      expect(names).toEqual(expected);
+      expect(names).toContain('read_file');
+      expect(names).toContain('grep');
+      expect(names).toContain('bash_exec');
+      expect(names).toContain('edit_file');
+    });
+
+    it('has none of the communication, delegation, lifecycle, memory or ticket tools', () => {
+      full = freshTools();
+      const restricted = restrictToolsForSubagent(full, policy, () => undefined);
+      for (const forbidden of [
+        'send_message', 'reply_slack', 'broadcast', 'delegate_task', 'handoff_task', 'report_status', 'complete_task',
+        'get_tasks', 'start_agent', 'stop_agent', 'handle_agent_failure', 'remember', 'recall_memory', 'schedule_check',
+        'register_self', 'subscribe_event', 'get_audit_log', 'compact_memory', 'computer', 'desktop_task', 'git_commit',
+        SPAWN_TOOL_NAME,
+      ]) {
+        expect(restricted[forbidden], forbidden).toBeUndefined();
+      }
+    });
+
+    it('never passes through a spawn_subagent or MCP tool present in the input map', () => {
+      full = freshTools();
+      const withExtras: Record<string, ToolDefinition> = {
+        ...full,
+        [SPAWN_TOOL_NAME]: { description: 'x', inputSchema: {} as any, execute: async () => ({}) },
+        'playwright:click': { description: 'x', inputSchema: {} as any, execute: async () => ({}) },
+      };
+      const restricted = restrictToolsForSubagent(withExtras, policy, () => undefined);
+      expect(restricted[SPAWN_TOOL_NAME]).toBeUndefined();
+      expect(restricted['playwright:click']).toBeUndefined();
+    });
+
+    it('readOnly drops the shell and the write tools', () => {
+      full = freshTools();
+      const restricted = restrictToolsForSubagent(full, { ...policy, readOnly: true }, () => undefined);
+      const names = Object.keys(restricted);
+      for (const w of SUBAGENT_WRITE_TOOLS) expect(names).not.toContain(w);
+      for (const w of WRITE_TOOLS) expect(names).not.toContain(w);
+      expect(names).toContain('read_file');
+      expect(names).toContain(REPORT_TOOL_NAME);
+    });
+
+    it('refuses edit_file and write_file outside cwd and resolves relative paths inside it', async () => {
+      const inner = vi.fn(async () => ({ success: true }));
+      const tools: Record<string, ToolDefinition> = {
+        edit_file: { description: 'e', inputSchema: {} as any, execute: inner },
+        write_file: { description: 'w', inputSchema: {} as any, execute: inner },
+      };
+      const restricted = restrictToolsForSubagent(tools, policy, () => undefined);
+
+      const outside = await restricted.edit_file.execute({ file_path: '/etc/hosts', old_string: 'a', new_string: 'b' }) as Record<string, unknown>;
+      expect(outside.success).toBe(false);
+      expect(String(outside.error)).toMatch(/outside the sub-agent working directory/);
+      const climb = await restricted.write_file.execute({ file_path: '../sibling/x.ts', content: '' }) as Record<string, unknown>;
+      expect(climb.success).toBe(false);
+      const home = await restricted.write_file.execute({ file_path: '~/x.ts', content: '' }) as Record<string, unknown>;
+      expect(home.success).toBe(false);
+      expect(inner).not.toHaveBeenCalled();
+
+      await restricted.edit_file.execute({ file_path: 'src/a.ts', old_string: 'a', new_string: 'b' });
+      expect(inner).toHaveBeenCalledWith(expect.objectContaining({ file_path: `${CWD}/src/a.ts` }));
+    });
+
+    it('forces bash_exec into cwd when the model asks for a directory outside it', async () => {
+      const inner = vi.fn(async () => ({ success: true, stdout: '' }));
+      const tools: Record<string, ToolDefinition> = {
+        bash_exec: { description: 'b', inputSchema: {} as any, execute: inner },
+      };
+      const restricted = restrictToolsForSubagent(tools, policy, () => undefined);
+      await restricted.bash_exec.execute({ command: 'ls' });
+      expect(inner).toHaveBeenLastCalledWith({ command: 'ls', cwd: CWD });
+      await restricted.bash_exec.execute({ command: 'ls', cwd: '/' });
+      expect(inner).toHaveBeenLastCalledWith({ command: 'ls', cwd: CWD });
+      await restricted.bash_exec.execute({ command: 'ls', cwd: 'packages' });
+      expect(inner).toHaveBeenLastCalledWith({ command: 'ls', cwd: `${CWD}/packages` });
+    });
+
+    it('report_to_parent records the report and tells the child to stop', async () => {
+      const reports: SubagentReport[] = [];
+      const restricted = restrictToolsForSubagent({}, policy, (r) => reports.push(r));
+      const out = await restricted[REPORT_TOOL_NAME].execute({ summary: 'found 3 call sites', filesChanged: ['a.ts'] }) as Record<string, unknown>;
+      expect(out.success).toBe(true);
+      expect(String(out.note)).toMatch(/finish your turn/i);
+      expect(reports).toEqual([{ summary: 'found 3 call sites', filesChanged: ['a.ts'] }]);
+      expect(TOOL_SENSITIVITY[REPORT_TOOL_NAME]).toBe('safe');
+      expect(TOOL_SENSITIVITY[SPAWN_TOOL_NAME]).toBe('sensitive');
+      // Neither is a registry tool: the runner adds them per run.
+      expect(getToolNames()).not.toContain(SPAWN_TOOL_NAME);
+    });
+  });
+
+  describe('resolveSubagentModel', () => {
+    const parent = parentConfig.model;
+
+    it('keeps the parent model when none is asked for', () => {
+      expect(resolveSubagentModel(parent)).toEqual({ model: parent });
+      expect(resolveSubagentModel(parent, '  ')).toEqual({ model: parent });
+    });
+
+    it('swaps the model id on the same provider, or takes provider/modelId', () => {
+      expect(resolveSubagentModel(parent, 'deepseek-reasoner')).toEqual({ model: { ...parent, modelId: 'deepseek-reasoner' } });
+      expect(resolveSubagentModel(parent, 'google/gemini-3-flash-preview')).toEqual({
+        model: { ...parent, provider: 'google', modelId: 'gemini-3-flash-preview' },
+      });
+    });
+
+    it('refuses an unknown provider or an empty model id', () => {
+      expect(resolveSubagentModel(parent, 'acme/foo')).toEqual({ error: expect.stringMatching(/unknown model provider "acme"/) });
+      expect(resolveSubagentModel(parent, 'openai/')).toEqual({ error: expect.stringMatching(/names no model/) });
+    });
+  });
+
+  describe('buildSubagentConfig', () => {
+    it('derives a child config: own session name, short prompt, cwd as project path, policy, no MCP or member id', () => {
+      const built = buildSubagentConfig(parentConfig, { task: 'Find every caller of foo()' }, 2);
+      expect('config' in built).toBe(true);
+      const cfg = (built as { config: CrewlyAgentConfig }).config;
+      expect(cfg.sessionName).toBe('think-tank-kai/sub-2');
+      expect(cfg.role).toBe('subagent');
+      expect(cfg.model).toEqual(parentConfig.model);
+      expect(cfg.maxSteps).toBe(CREWLY_AGENT_DEFAULTS.SUBAGENT.MAX_STEPS);
+      expect(cfg.projectPath).toBe(CWD);
+      expect(cfg.apiBaseUrl).toBe(parentConfig.apiBaseUrl);
+      expect(cfg.systemPrompt).toBe(buildSubagentPrompt('think-tank-kai', CWD, false));
+      expect(cfg.systemPrompt).not.toContain('You are Kai');
+      expect(cfg.subagent).toEqual({ cwd: CWD, readOnly: false, depth: 1, parentSession: 'think-tank-kai' });
+      expect(cfg.mcpServers).toBeUndefined();
+      expect(cfg.memberId).toBeUndefined();
+      expect(cfg.evalMode).toBeUndefined();
+    });
+
+    it('caps maxSteps, resolves cwd under the project, and honours readOnly', () => {
+      const built = buildSubagentConfig(parentConfig, { task: 'x', maxSteps: 999, cwd: 'packages/ui', readOnly: true }, 1);
+      const cfg = (built as { config: CrewlyAgentConfig }).config;
+      expect(cfg.maxSteps).toBe(CREWLY_AGENT_DEFAULTS.SUBAGENT.MAX_STEPS_CAP);
+      expect(cfg.projectPath).toBe(`${CWD}/packages/ui`);
+      expect(cfg.subagent?.readOnly).toBe(true);
+      expect(cfg.systemPrompt).toMatch(/read-only tools/);
+      const small = buildSubagentConfig(parentConfig, { task: 'x', maxSteps: 5 }, 1);
+      expect((small as { config: CrewlyAgentConfig }).config.maxSteps).toBe(5);
+    });
+
+    it('refuses an empty or oversized task and a bad model', () => {
+      expect(buildSubagentConfig(parentConfig, { task: '   ' }, 1)).toEqual({ error: 'task is required' });
+      const huge = 'x'.repeat(CREWLY_AGENT_DEFAULTS.SUBAGENT.TASK_MAX_CHARS + 1);
+      expect(buildSubagentConfig(parentConfig, { task: huge }, 1)).toEqual({ error: expect.stringMatching(/too long/) });
+      expect(buildSubagentConfig(parentConfig, { task: 'x', model: 'nope/m' }, 1)).toEqual({ error: expect.stringMatching(/unknown model provider/) });
+    });
+
+    it('a child of a child would be depth 2 — but a child never gets the spawn tool, so this stays theoretical', () => {
+      const child = (buildSubagentConfig(parentConfig, { task: 'x' }, 1) as { config: CrewlyAgentConfig }).config;
+      const grandchild = (buildSubagentConfig(child, { task: 'y' }, 1) as { config: CrewlyAgentConfig }).config;
+      expect(grandchild.subagent?.depth).toBe(2);
+      expect(grandchild.sessionName).toBe('think-tank-kai/sub-1/sub-1');
+    });
+  });
+
+  describe('summarizeSubagentRun', () => {
+    it('zero working tool calls is did_nothing, even when the child reported', () => {
+      const out = summarizeSubagentRun(
+        runResult({ toolCalls: [{ toolName: REPORT_TOOL_NAME, args: { summary: 'all good' }, result: { success: true } }] }),
+        { summary: 'all good', filesChanged: [] },
+        's/sub-1',
+        10,
+        false,
+      );
+      expect(out.stoppedReason).toBe('did_nothing');
+      expect(out.success).toBe(false);
+      expect(out.toolCalls).toBe(0);
+      expect(out.summary).toBe('all good');
+    });
+
+    it('done: counts working calls, lists tool names, collects files from successful writes and the report', () => {
+      const out = summarizeSubagentRun(
+        runResult({
+          toolCalls: [
+            { toolName: 'grep', args: { pattern: 'foo' }, result: { success: true } },
+            { toolName: 'edit_file', args: { file_path: `${CWD}/a.ts` }, result: { success: true } },
+            { toolName: 'edit_file', args: { file_path: `${CWD}/b.ts` }, result: { success: false, error: 'not unique' } },
+            { toolName: 'write_file', args: { file_path: `${CWD}/c.ts` }, result: { success: true } },
+            { toolName: REPORT_TOOL_NAME, args: { summary: 'r' }, result: { success: true } },
+          ],
+        }),
+        { summary: 'changed a and c', filesChanged: [`${CWD}/d.md`] },
+        's/sub-1',
+        1234,
+        false,
+      );
+      expect(out.stoppedReason).toBe('done');
+      expect(out.success).toBe(true);
+      expect(out.toolCalls).toBe(4);
+      expect(out.toolNames).toEqual(['grep', 'edit_file', 'write_file']);
+      expect(out.filesChanged).toEqual([`${CWD}/a.ts`, `${CWD}/c.ts`, `${CWD}/d.md`]);
+      expect(out.usage).toEqual({ input: 100, output: 20, cachedInput: 5 });
+      expect(out.durationMs).toBe(1234);
+      expect(out.sessionName).toBe('s/sub-1');
+    });
+
+    it('maps incompleteness and timeouts, and falls back to the final text without a report', () => {
+      const calls = [{ toolName: 'read_file', args: { file_path: 'x' }, result: { success: true } }];
+      expect(summarizeSubagentRun(runResult({ toolCalls: calls, incomplete: { reason: 'steps-exhausted', detail: '', finishReason: 'tool-calls', recoveryAttempts: 0 } }), null, 's', 1, false).stoppedReason).toBe('steps_exhausted');
+      expect(summarizeSubagentRun(runResult({ toolCalls: calls, incomplete: { reason: 'truncated', detail: '', finishReason: 'length', recoveryAttempts: 3 } }), null, 's', 1, false).stoppedReason).toBe('incomplete');
+      const timed = summarizeSubagentRun(runResult({ toolCalls: calls }), null, 's', 1, true);
+      expect(timed.stoppedReason).toBe('timeout');
+      expect(timed.error).toMatch(/exceeded/);
+      expect(summarizeSubagentRun(runResult({ toolCalls: calls, text: '  the answer  ' }), null, 's', 1, false).summary).toBe('the answer');
+    });
+
+    it('caps a long summary', () => {
+      const long = 'y'.repeat(CREWLY_AGENT_DEFAULTS.SUBAGENT.SUMMARY_MAX_CHARS + 50);
+      const out = summarizeSubagentRun(runResult({ toolCalls: [{ toolName: 'grep', args: {}, result: {} }] }), { summary: long, filesChanged: [] }, 's', 1, false);
+      expect(out.summary.length).toBeLessThan(long.length);
+      expect(out.summary).toMatch(/\[truncated\]$/);
+    });
+  });
+
+  describe('forwardChildToolCalls', () => {
+    it('relays tool start/finish to the parent with a sub-n prefix and forwards nothing else', () => {
+      const parent = { onToolCallStart: vi.fn(), onToolCallFinish: vi.fn(), onTextChunk: vi.fn(), onStepFinish: vi.fn() };
+      const cb = forwardChildToolCalls(() => parent, 'think-tank-kai/sub-2');
+      cb.onToolCallStart!('grep', { pattern: 'x' });
+      cb.onToolCallFinish!('grep', { pattern: 'x' }, { success: true }, 12);
+      expect(parent.onToolCallStart).toHaveBeenCalledWith('sub-2:grep', { pattern: 'x' });
+      expect(parent.onToolCallFinish).toHaveBeenCalledWith('sub-2:grep', { pattern: 'x' }, { success: true }, 12);
+      expect(cb.onTextChunk).toBeUndefined();
+      expect(cb.onStepFinish).toBeUndefined();
+    });
+
+    it('reads the parent callbacks on every call, so a parent with none set is fine', () => {
+      let parent: Record<string, unknown> = {};
+      const cb = forwardChildToolCalls(() => parent, 'p/sub-1');
+      expect(() => cb.onToolCallStart!('read_file', {})).not.toThrow();
+      const start = vi.fn();
+      parent = { onToolCallStart: start };
+      cb.onToolCallStart!('read_file', { file_path: 'a' });
+      expect(start).toHaveBeenCalledWith('sub-1:read_file', { file_path: 'a' });
+    });
+  });
+
+  describe('createSpawnSubagentTool', () => {
+    it('runs a child with the task as its only message, returns its outcome, counts usage and disposes it', async () => {
+      const registry = createSubagentRegistry();
+      const child = fakeRunner({
+        result: runResult({ toolCalls: [{ toolName: 'grep', args: { pattern: 'x' }, result: { success: true } }] }),
+        report: { summary: 'two call sites: a.ts:10, b.ts:42', filesChanged: [] },
+      });
+      const createRunner = vi.fn(() => child);
+      const onUsage = vi.fn();
+      const tool = createSpawnSubagentTool({
+        parentConfig,
+        registry,
+        createRunner,
+        onUsage,
+        childStreaming: (sessionName) => ({ relay: sessionName }) as unknown as import('./types.js').StreamingEventCallbacks,
+      });
+
+      const out = await tool.execute({ task: '  Find callers of foo()  ', readOnly: true }) as Record<string, unknown>;
+
+      expect(createRunner).toHaveBeenCalledTimes(1);
+      const cfg = createRunner.mock.calls[0]![0] as unknown as CrewlyAgentConfig;
+      expect(cfg.sessionName).toBe('think-tank-kai/sub-1');
+      expect(cfg.subagent).toEqual({ cwd: CWD, readOnly: true, depth: 1, parentSession: 'think-tank-kai' });
+      expect(child.initialize).toHaveBeenCalledTimes(1);
+      expect(child.run).toHaveBeenCalledTimes(1);
+      expect(child.run.mock.calls[0]![0]).toBe('Find callers of foo()');
+      expect(child.run.mock.calls[0]![3]?.abortSignal).toBeInstanceOf(AbortSignal);
+      expect((child.run.mock.calls[0]![3] as { streaming?: unknown })?.streaming).toEqual({ relay: 'think-tank-kai/sub-1' });
+      expect(out.success).toBe(true);
+      expect(out.stoppedReason).toBe('done');
+      expect(out.summary).toBe('two call sites: a.ts:10, b.ts:42');
+      expect(out.sessionName).toBe('think-tank-kai/sub-1');
+      expect(onUsage).toHaveBeenCalledWith({ input: 100, output: 20, cachedInput: 5 });
+      expect(child.shutdown).toHaveBeenCalledTimes(1);
+      expect(registry).toEqual({ active: 0, spawned: 1 });
+    });
+
+    it('a child that made no tool call comes back as did_nothing, not success', async () => {
+      const tool = createSpawnSubagentTool({
+        parentConfig,
+        registry: createSubagentRegistry(),
+        createRunner: () => fakeRunner({ result: runResult({ toolCalls: [], text: 'I would look at the files.' }) }),
+      });
+      const out = await tool.execute({ task: 'look' }) as Record<string, unknown>;
+      expect(out.success).toBe(false);
+      expect(out.stoppedReason).toBe('did_nothing');
+      expect(out.summary).toBe('I would look at the files.');
+    });
+
+    it('numbers children in spawn order and refuses a fourth concurrent child', async () => {
+      const registry = createSubagentRegistry();
+      let release!: () => void;
+      const gate = new Promise<void>((r) => { release = r; });
+      const slow = () => fakeRunner({
+        result: async () => { await gate; return runResult({ toolCalls: [{ toolName: 'grep', args: {}, result: {} }] }); },
+      });
+      const createRunner = vi.fn(slow);
+      const tool = createSpawnSubagentTool({ parentConfig, registry, createRunner });
+
+      const running = [tool.execute({ task: 'a' }), tool.execute({ task: 'b' }), tool.execute({ task: 'c' })];
+      await Promise.resolve();
+      expect(registry.active).toBe(CREWLY_AGENT_DEFAULTS.SUBAGENT.MAX_CONCURRENT);
+      const refused = await tool.execute({ task: 'd' }) as Record<string, unknown>;
+      expect(refused.success).toBe(false);
+      expect(refused.stoppedReason).toBe('error');
+      expect(String(refused.error)).toMatch(/at most 3 sub-agents/);
+      expect(createRunner).toHaveBeenCalledTimes(3);
+
+      release();
+      const done = await Promise.all(running) as Array<Record<string, unknown>>;
+      expect(done.map((d) => d.sessionName)).toEqual(['think-tank-kai/sub-1', 'think-tank-kai/sub-2', 'think-tank-kai/sub-3']);
+      expect(registry).toEqual({ active: 0, spawned: 3 });
+
+      // A slot is free again.
+      const fifth = await createSpawnSubagentTool({ parentConfig, registry, createRunner: () => fakeRunner({ result: runResult({ toolCalls: [{ toolName: 'grep', args: {}, result: {} }] }) }) })
+        .execute({ task: 'e' }) as Record<string, unknown>;
+      expect(fifth.sessionName).toBe('think-tank-kai/sub-4');
+    });
+
+    it('a runner that throws yields error, still disposes the child and frees the slot', async () => {
+      const registry = createSubagentRegistry();
+      const child = fakeRunner({ throwOn: 'run' });
+      const tool = createSpawnSubagentTool({ parentConfig, registry, createRunner: () => child });
+      const out = await tool.execute({ task: 'boom' }) as Record<string, unknown>;
+      expect(out.success).toBe(false);
+      expect(out.stoppedReason).toBe('error');
+      expect(out.error).toBe('provider exploded');
+      expect(out.sessionName).toBe('think-tank-kai/sub-1');
+      expect(child.shutdown).toHaveBeenCalledTimes(1);
+      expect(registry.active).toBe(0);
+
+      const bad = fakeRunner({ throwOn: 'initialize' });
+      const out2 = await createSpawnSubagentTool({ parentConfig, registry, createRunner: () => bad }).execute({ task: 'x' }) as Record<string, unknown>;
+      expect(out2.error).toBe('no api key');
+      expect(bad.shutdown).toHaveBeenCalledTimes(1);
+    });
+
+    it('aborts the child when the wall-clock budget runs out and reports timeout', async () => {
+      vi.useFakeTimers();
+      try {
+        const registry = createSubagentRegistry();
+        const child = fakeRunner({
+          result: () => new Promise<AgentRunResult>((_, reject) => {
+            // The real runner rejects on abort; mirror that.
+            setTimeout(() => reject(new Error('aborted')), 60_000);
+          }),
+        });
+        const tool = createSpawnSubagentTool({ parentConfig, registry, createRunner: () => child, timeoutMs: 1_000 });
+        const pending = tool.execute({ task: 'slow' });
+        await vi.advanceTimersByTimeAsync(1_000);
+        const signal = child.run.mock.calls[0]![3]?.abortSignal as AbortSignal;
+        expect(signal.aborted).toBe(true);
+        await vi.advanceTimersByTimeAsync(60_000);
+        const out = await pending as Record<string, unknown>;
+        expect(out.stoppedReason).toBe('timeout');
+        expect(out.success).toBe(false);
+        expect(child.shutdown).toHaveBeenCalledTimes(1);
+        expect(registry.active).toBe(0);
+      } finally {
+        vi.useRealTimers();
+      }
+    });
+
+    it('refuses a bad spawn before creating any runner', async () => {
+      const createRunner = vi.fn();
+      const tool = createSpawnSubagentTool({ parentConfig, registry: createSubagentRegistry(), createRunner });
+      const out = await tool.execute({ task: 'x', model: 'nope/m' }) as Record<string, unknown>;
+      expect(out.success).toBe(false);
+      expect(String(out.error)).toMatch(/unknown model provider/);
+      expect(createRunner).not.toHaveBeenCalled();
+    });
+
+    it('describes when to use it and what did_nothing means, and declares its sensitivity', () => {
+      const tool = createSpawnSubagentTool({ parentConfig, registry: createSubagentRegistry(), createRunner: vi.fn() });
+      expect(tool.sensitivity).toBe('sensitive');
+      expect(tool.description).toMatch(/teammate instead/);
+      expect(tool.description).toMatch(/did_nothing/);
+      const parsed = (tool.inputSchema as any).safeParse({ task: 'x', maxSteps: 0 });
+      expect(parsed.success).toBe(false);
+    });
+  });
+});