feat(routing): add pi canary guardrails and backend audit telemetry

2026-02-23 21:13:01 -08:00
parent ac61c9c3fb
commit a389cd659c
4 changed files with 224 additions and 13 deletions
@@ -1279,6 +1279,151 @@ describe('daemon external backend integration', () => {
    expect(processSpy).toHaveBeenCalled();
    expect(reply).toHaveBeenCalledWith(expect.objectContaining({ text: 'native fallback response' }));
  });
+
+  it('uses pi_embedded backend for plain text canary turns', async () => {
+    const processSpy = vi.spyOn(AgentOrchestrator.prototype, 'process');
+    const history: Array<{ role: 'user' | 'assistant'; content: string }> = [];
+    const session = {
+      id: 'telegram:pi-canary',
+      addMessage: vi.fn((msg: { role: 'user' | 'assistant'; content: string }) => {
+        history.push(msg);
+        return msg;
+      }),
+      getHistory: vi.fn(() => [...history]),
+      clear: vi.fn(),
+      replaceHistory: vi.fn(),
+      getConfig: vi.fn(() => undefined),
+      setConfig: vi.fn(),
+      deleteConfig: vi.fn(),
+    };
+
+    const piBackend = {
+      name: 'pi_embedded',
+      process: vi.fn(async () => 'pi embedded response'),
+    };
+
+    const router = createMessageRouter({
+      sessionManager: { getSession: vi.fn(() => session) } as unknown as MessageRouterDeps['sessionManager'],
+      modelRouter: {
+        getAvailableTiers: () => ['fast', 'default', 'complex', 'local'],
+        getAllLabels: () => ({ fast: 'fast', default: 'default', complex: 'complex', local: 'local' }),
+        getLabel: (tier: string) => tier,
+      } as unknown as MessageRouterDeps['modelRouter'],
+      systemPrompt: 'test prompt',
+      toolRegistry: {
+        clone() { return this; },
+        register: vi.fn(),
+      } as unknown as MessageRouterDeps['toolRegistry'],
+      toolExecutor: {} as unknown as MessageRouterDeps['toolExecutor'],
+      config: {
+        agents: {
+          primary_tier: 'default',
+          delegation: {
+            compaction: 'fast',
+            memory_extraction: 'fast',
+            classification: 'fast',
+            tool_summarisation: 'fast',
+            complex_reasoning: 'complex',
+          },
+          max_delegation_depth: 3,
+          max_iterations: 10,
+        },
+        backends: {
+          pi_embedded: { no_tools_mode: false },
+        },
+        compaction: { enabled: false },
+        models: { default: { provider: 'anthropic', model: 'claude' } },
+      } as unknown as MessageRouterDeps['config'],
+      externalBackends: { pi_embedded: piBackend } as unknown as MessageRouterDeps['externalBackends'],
+      defaultName: 'pi_embedded',
+    });
+
+    const reply = vi.fn(async (_message: OutboundMessage) => {});
+    await router.handler({
+      id: 'm-pi-canary',
+      channel: 'telegram',
+      senderId: 'pi-canary',
+      text: 'just chat with me',
+      timestamp: Date.now(),
+    } as MessageRouterInput, reply);
+
+    expect(piBackend.process).toHaveBeenCalled();
+    expect(processSpy).not.toHaveBeenCalled();
+    expect(reply).toHaveBeenCalledWith(expect.objectContaining({ text: 'pi embedded response' }));
+  });
+
+  it('forces native processing for pi_embedded no-tools mode when prompt appears tool-oriented', async () => {
+    const processSpy = vi.spyOn(AgentOrchestrator.prototype, 'process')
+      .mockResolvedValue('native guarded response');
+    const history: Array<{ role: 'user' | 'assistant'; content: string }> = [];
+    const session = {
+      id: 'telegram:pi-no-tools',
+      addMessage: vi.fn((msg: { role: 'user' | 'assistant'; content: string }) => {
+        history.push(msg);
+        return msg;
+      }),
+      getHistory: vi.fn(() => [...history]),
+      clear: vi.fn(),
+      replaceHistory: vi.fn(),
+      getConfig: vi.fn(() => undefined),
+      setConfig: vi.fn(),
+      deleteConfig: vi.fn(),
+    };
+
+    const piBackend = {
+      name: 'pi_embedded',
+      process: vi.fn(async () => 'pi embedded response'),
+    };
+
+    const router = createMessageRouter({
+      sessionManager: { getSession: vi.fn(() => session) } as unknown as MessageRouterDeps['sessionManager'],
+      modelRouter: {
+        getAvailableTiers: () => ['fast', 'default', 'complex', 'local'],
+        getAllLabels: () => ({ fast: 'fast', default: 'default', complex: 'complex', local: 'local' }),
+        getLabel: (tier: string) => tier,
+      } as unknown as MessageRouterDeps['modelRouter'],
+      systemPrompt: 'test prompt',
+      toolRegistry: {
+        clone() { return this; },
+        register: vi.fn(),
+      } as unknown as MessageRouterDeps['toolRegistry'],
+      toolExecutor: {} as unknown as MessageRouterDeps['toolExecutor'],
+      config: {
+        agents: {
+          primary_tier: 'default',
+          delegation: {
+            compaction: 'fast',
+            memory_extraction: 'fast',
+            classification: 'fast',
+            tool_summarisation: 'fast',
+            complex_reasoning: 'complex',
+          },
+          max_delegation_depth: 3,
+          max_iterations: 10,
+        },
+        backends: {
+          pi_embedded: { no_tools_mode: true },
+        },
+        compaction: { enabled: false },
+        models: { default: { provider: 'anthropic', model: 'claude' } },
+      } as unknown as MessageRouterDeps['config'],
+      externalBackends: { pi_embedded: piBackend } as unknown as MessageRouterDeps['externalBackends'],
+      defaultName: 'pi_embedded',
+    });
+
+    const reply = vi.fn(async (_message: OutboundMessage) => {});
+    await router.handler({
+      id: 'm-pi-no-tools',
+      channel: 'telegram',
+      senderId: 'pi-no-tools',
+      text: 'please read the file and run a shell command',
+      timestamp: Date.now(),
+    } as MessageRouterInput, reply);
+
+    expect(piBackend.process).not.toHaveBeenCalled();
+    expect(processSpy).toHaveBeenCalled();
+    expect(reply).toHaveBeenCalledWith(expect.objectContaining({ text: 'native guarded response' }));
+  });
 });

 describe('daemon audio routing integration', () => {