@principles/core 1.287.1 → 1.287.3

package/dist/runtime-v2/adapter/__tests__/pi-ai-runtime-adapter.test.js

@@ -1,2128 +0,0 @@
-/**
- * PiAiRuntimeAdapter unit tests.
- *
- * Verifies all PDRuntimeAdapter interface methods, pi-ai API usage,
- * error mapping, DiagnosticianOutputV1 validation, retry logic,
- * and telemetry emission.
- *
- * All pi-ai calls are mocked — no real API keys needed.
- */
-import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
-import { PDRuntimeError } from '../../error-categories.js';
-// Mock @earendil-works/pi-ai/compat (value facade used by the adapter) at module level
-vi.mock('@earendil-works/pi-ai/compat', () => ({
-    getModel: vi.fn(),
-    getProviders: vi.fn(() => ['openrouter', 'anthropic', 'openai', 'google']),
-    completeSimple: vi.fn(),
-}));
-// Mock store/event-emitter to capture telemetry calls
-vi.mock('../../store/event-emitter.js', () => ({
-    storeEmitter: { emitTelemetry: vi.fn() },
-}));
-import { getModel, completeSimple } from '@earendil-works/pi-ai/compat';
-import { storeEmitter } from '../../store/event-emitter.js';
-import { PiAiRuntimeAdapter } from '../pi-ai-runtime-adapter.js';
-const mockGetModel = getModel;
-const mockComplete = completeSimple;
-const mockEmitTelemetry = storeEmitter.emitTelemetry;
-// ── Fixtures ──
-const VALID_DIAGNOSIS = {
-    valid: true,
-    diagnosisId: 'diag-test-1',
-    summary: 'Test summary',
-    rootCause: 'Test root cause',
-    violatedPrinciples: [],
-    evidence: [],
-    recommendations: [{ kind: 'defer', description: 'no actionable recommendation' }],
-    confidence: 0.9,
-};
-function makeAssistantMessage(text, overrides = {}) {
-    return {
-        content: [{ type: 'text', text }],
-        role: 'assistant',
-        stopReason: 'stop',
-        api: 'openai-completions',
-        provider: 'openrouter',
-        model: 'anthropic/claude-sonnet-4',
-        usage: { input: 10, output: 5, cacheRead: 0, cacheWrite: 0, totalTokens: 15, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } },
-        timestamp: Date.now(),
-        ...overrides,
-    };
-}
-/** Create an AssistantMessage that resolved with stopReason:'error'. */
-function makeErrorResponse(errorMessage, overrides = {}) {
-    return makeAssistantMessage('', {
-        stopReason: 'error',
-        errorMessage,
-        content: [],
-        ...overrides,
-    });
-}
-/** Create an AssistantMessage that resolved with stopReason:'aborted'. */
-function makeAbortedResponse(errorMessage = 'The operation was aborted') {
-    return makeAssistantMessage('', {
-        stopReason: 'aborted',
-        errorMessage,
-        content: [],
-    });
-}
-function makeStartRunInput(overrides = {}) {
-    return {
-        agentSpec: { agentId: 'diagnostician', schemaVersion: 'v1' },
-        inputPayload: 'Diagnose this pain signal',
-        contextItems: [],
-        timeoutMs: 60_000,
-        ...overrides,
-    };
-}
-function makeAdapter(overrides = {}) {
-    return new PiAiRuntimeAdapter({
-        provider: 'openrouter',
-        model: 'anthropic/claude-sonnet-4',
-        apiKeyEnv: 'TEST_API_KEY',
-        _testBackoffDelayMs: 0,
-        outputPathStrategy: 'free_form_only', // Default to free_form for existing tests
-        ...overrides,
-    });
-}
-/**
- * Helper: assert that startRun rejects with a PDRuntimeError of the given category.
- * Uses a fresh adapter and mock setup to avoid mock-consumption issues.
- */
-async function expectStartRunError(category, mockSetup, adapterOverrides = {}) {
-    mockSetup();
-    const adapter = makeAdapter(adapterOverrides);
-    let caught = undefined;
-    try {
-        await adapter.startRun(makeStartRunInput());
-    }
-    catch (err) {
-        caught = err instanceof PDRuntimeError ? err : undefined;
-    }
-    expect(caught).toBeInstanceOf(PDRuntimeError);
-    expect(caught?.category).toBe(category);
-}
-/** Extract the first emitted telemetry event matching the given eventType. */
-function findTelemetryEvent(eventType) {
-    // Runtime Contract Rule 2 (no `as` bypass): validate the mock-call shape
-    // before treating it as a telemetry payload.
-    const call = mockEmitTelemetry.mock.calls.find((c) => {
-        const [payload] = c;
-        return (typeof payload === 'object' &&
-            payload !== null &&
-            Object.hasOwn(payload, 'eventType') &&
-            payload.eventType === eventType);
-    });
-    if (!call)
-        return undefined;
-    const [payload] = call;
-    return typeof payload === 'object' && payload !== null
-        ? payload
-        : undefined;
-}
-/**
- * Type guard for the tool-context shape passed as the 2nd arg to completeSimple.
- * Avoids `as` casts on mock-call data (Runtime Contract Rule 2 / ERR-001).
- */
-function isToolContext(value) {
-    if (typeof value !== 'object' || value === null || !Object.hasOwn(value, 'tools')) {
-        return false;
-    }
-    const { tools } = value;
-    if (!Array.isArray(tools))
-        return false;
-    return tools.every((tool) => {
-        if (typeof tool !== 'object' || tool === null)
-            return false;
-        return (Object.hasOwn(tool, 'name') &&
-            typeof tool.name === 'string' &&
-            Object.hasOwn(tool, 'parameters'));
-    });
-}
-// ── Tests ──
-describe('PiAiRuntimeAdapter', () => {
-    beforeEach(() => {
-        vi.clearAllMocks();
-        process.env.TEST_API_KEY = 'test-key-123';
-        // Default mock: getModel returns a model object, complete returns valid diagnosis
-        mockGetModel.mockReturnValue({ id: 'anthropic/claude-sonnet-4' });
-        mockComplete.mockResolvedValue(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-    });
-    afterEach(() => {
-        delete process.env.TEST_API_KEY;
-    });
-    // ── kind() ──
-    describe('kind()', () => {
-        it('returns "pi-ai" (RS-02)', () => {
-            const adapter = makeAdapter();
-            expect(adapter.kind()).toBe('pi-ai');
-        });
-    });
-    // ── getCapabilities() ──
-    describe('getCapabilities()', () => {
-        it('returns correct RuntimeCapabilities shape', async () => {
-            const adapter = makeAdapter();
-            const caps = await adapter.getCapabilities();
-            expect(caps.supportsStructuredJsonOutput).toBe(true);
-            expect(caps.supportsToolUse).toBe(false);
-            expect(caps.supportsWorkingDirectory).toBe(false);
-            expect(caps.supportsModelSelection).toBe(true);
-            expect(caps.supportsLongRunningSessions).toBe(false);
-            expect(caps.supportsCancellation).toBe(true);
-            expect(caps.supportsArtifactWriteBack).toBe(false);
-            expect(caps.supportsConcurrentRuns).toBe(false);
-            expect(caps.supportsStreaming).toBe(false);
-        });
-    });
-    // ── healthCheck() ──
-    describe('healthCheck()', () => {
-        it('returns healthy=true when apiKey exists, getModel succeeds, and complete probe returns {"ok":true}', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage('{"ok":true}'));
-            const adapter = makeAdapter();
-            const health = await adapter.healthCheck();
-            expect(health.healthy).toBe(true);
-            expect(health.degraded).toBe(false);
-            expect(health.warnings).toEqual([]);
-            expect(mockGetModel).toHaveBeenCalledWith('openrouter', 'anthropic/claude-sonnet-4');
-        });
-        it('returns healthy=false when apiKeyEnv is missing', async () => {
-            delete process.env.TEST_API_KEY;
-            const adapter = makeAdapter();
-            const health = await adapter.healthCheck();
-            expect(health.healthy).toBe(false);
-            expect(health.degraded).toBe(false);
-            expect(health.warnings.some(w => w.includes('TEST_API_KEY'))).toBe(true);
-        });
-        it('returns healthy=false when getModel throws', async () => {
-            mockGetModel.mockImplementationOnce(() => { throw new Error('invalid provider'); });
-            const adapter = makeAdapter();
-            const health = await adapter.healthCheck();
-            expect(health.healthy).toBe(false);
-            expect(health.degraded).toBe(false);
-            expect(health.warnings.some(w => w.includes('invalid provider'))).toBe(true);
-        });
-        it('returns healthy=false with degraded=true when complete probe returns unexpected result', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage('{"unexpected": true}'));
-            const adapter = makeAdapter();
-            const health = await adapter.healthCheck();
-            expect(health.healthy).toBe(false);
-            expect(health.degraded).toBe(true);
-            expect(health.warnings.some(w => w.includes('unexpected result'))).toBe(true);
-        });
-        it('returns healthy=false with degraded=true when probe times out', async () => {
-            mockComplete.mockImplementationOnce(() => {
-                const err = new DOMException('The operation was aborted', 'AbortError');
-                throw err;
-            });
-            const adapter = makeAdapter();
-            const health = await adapter.healthCheck();
-            expect(health.healthy).toBe(false);
-            expect(health.degraded).toBe(true);
-            expect(health.warnings.some(w => w.includes('timed out'))).toBe(true);
-        });
-        it('returns healthy=false with degraded=true when complete resolves with stopReason:error + timeout message', async () => {
-            mockComplete.mockResolvedValueOnce(makeErrorResponse('Request timed out.'));
-            const adapter = makeAdapter();
-            const health = await adapter.healthCheck();
-            expect(health.healthy).toBe(false);
-            expect(health.degraded).toBe(true);
-            expect(health.warnings.some(w => w.includes('timed out'))).toBe(true);
-            // Must NOT say "no text content"
-            expect(health.warnings.every(w => !w.includes('no text content'))).toBe(true);
-        });
-        it('returns healthy=false with degraded=false when complete resolves with stopReason:error + auth error', async () => {
-            mockComplete.mockResolvedValueOnce(makeErrorResponse('401 Unauthorized'));
-            const adapter = makeAdapter();
-            const health = await adapter.healthCheck();
-            expect(health.healthy).toBe(false);
-            expect(health.degraded).toBe(false);
-            expect(health.warnings.some(w => w.includes('execution failed') || w.includes('401'))).toBe(true);
-            // Must NOT misclassify as "no text content"
-            expect(health.warnings.every(w => !w.includes('no text content'))).toBe(true);
-        });
-        it('respects config.timeoutMs for probe timeout', async () => {
-            // Capture the timeoutMs passed to complete
-            let capturedTimeoutMs = undefined;
-            mockComplete.mockImplementationOnce(async (_model, _ctx, opts) => {
-                capturedTimeoutMs = opts.timeoutMs;
-                return makeAssistantMessage('{"ok":true}');
-            });
-            const adapter = makeAdapter({ timeoutMs: 15_000 });
-            await adapter.healthCheck();
-            expect(capturedTimeoutMs).toBe(15_000);
-        });
-    });
-    // ── startRun() — success path ──
-    describe('startRun()', () => {
-        it('calls getModel with provider and model from config', async () => {
-            const adapter = makeAdapter();
-            await adapter.startRun(makeStartRunInput());
-            expect(mockGetModel).toHaveBeenCalledWith('openrouter', 'anthropic/claude-sonnet-4');
-        });
-        it('calls complete with correct context including UserMessage with timestamp', async () => {
-            const adapter = makeAdapter();
-            await adapter.startRun(makeStartRunInput());
-            expect(mockComplete).toHaveBeenCalledTimes(1);
-            const [, context] = mockComplete.mock.calls[0];
-            expect(context.messages).toHaveLength(1);
-            expect(context.messages[0]).toBeDefined();
-            if (context.messages[0]) {
-                expect(context.messages[0].role).toBe('user');
-                expect(context.messages[0].content).toBe('Diagnose this pain signal');
-                expect(typeof context.messages[0].timestamp).toBe('number');
-            }
-        });
-        it('passes apiKey, effectiveTimeoutMs (input > config > default), and maxRetries: 0 to complete options', async () => {
-            const adapter = makeAdapter({ maxRetries: 3, timeoutMs: 120_000 });
-            await adapter.startRun(makeStartRunInput({ timeoutMs: 90_000 }));
-            const [, , options] = mockComplete.mock.calls[0];
-            expect(options.apiKey).toBe('test-key-123');
-            expect(options.maxRetries).toBe(0);
-            expect(options.timeoutMs).toBe(90_000);
-        });
-        it('returns RunHandle with runId, runtimeKind="pi-ai", and valid ISO startedAt', async () => {
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeStartRunInput());
-            expect(handle.runId).toBeTruthy();
-            expect(handle.runtimeKind).toBe('pi-ai');
-            expect(new Date(handle.startedAt).toISOString()).toBe(handle.startedAt);
-        });
-        it('stores output in memory (fetchOutput works after startRun completes)', async () => {
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeStartRunInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output).not.toBeNull();
-            expect(output?.runId).toBe(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-        });
-        it('serializes non-string inputPayload to JSON for the UserMessage', async () => {
-            const adapter = makeAdapter();
-            await adapter.startRun(makeStartRunInput({ inputPayload: { pain: 'signal', severity: 0.8 } }));
-            const [, context] = mockComplete.mock.calls[0];
-            expect(context.messages[0]).toBeDefined();
-            if (context.messages[0]) {
-                expect(context.messages[0].content).toBe('{"pain":"signal","severity":0.8}');
-            }
-        });
-        // ── systemPrompt support (PRI-501 follow-up) ──
-        it('TC1: passes config.systemPrompt to Context.systemPrompt when set', async () => {
-            const adapter = makeAdapter({ systemPrompt: 'You are a diagnostician.' });
-            await adapter.startRun(makeStartRunInput());
-            const [, context] = mockComplete.mock.calls[0];
-            expect(context.systemPrompt).toBe('You are a diagnostician.');
-        });
-        it('TC2: omits Context.systemPrompt when config.systemPrompt is unset (backward compat)', async () => {
-            const adapter = makeAdapter();
-            await adapter.startRun(makeStartRunInput());
-            const [, context] = mockComplete.mock.calls[0];
-            expect(context.systemPrompt).toBeUndefined();
-        });
-        it('TC3: omits Context.systemPrompt when config.systemPrompt is empty string', async () => {
-            const adapter = makeAdapter({ systemPrompt: '' });
-            await adapter.startRun(makeStartRunInput());
-            const [, context] = mockComplete.mock.calls[0];
-            expect(context.systemPrompt).toBeUndefined();
-        });
-        // ── layered systemPrompt (PRI-633) ──
-        it('PRI-633: StartRunInput.systemPrompt rides as Context.systemPrompt (base layer)', async () => {
-            const adapter = makeAdapter();
-            await adapter.startRun(makeStartRunInput({ systemPrompt: 'You are a root cause analysis expert.' }));
-            const [, context] = mockComplete.mock.calls[0];
-            expect(context.systemPrompt).toBe('You are a root cause analysis expert.');
-        });
-        it('PRI-633: profile config.systemPrompt is appended AFTER the run base layer', async () => {
-            const adapter = makeAdapter({ systemPrompt: 'PROFILE APPEND LAYER' });
-            await adapter.startRun(makeStartRunInput({ systemPrompt: 'BASE ROLE LAYER' }));
-            const [, context] = mockComplete.mock.calls[0];
-            expect(context.systemPrompt).toBe('BASE ROLE LAYER\n\nPROFILE APPEND LAYER');
-        });
-        it('PRI-633: omit Context.systemPrompt when both layers are absent/blank (unchanged shape)', async () => {
-            const adapter = makeAdapter({ systemPrompt: '   ' });
-            await adapter.startRun(makeStartRunInput({ systemPrompt: undefined }));
-            const [, context] = mockComplete.mock.calls[0];
-            expect(context.systemPrompt).toBeUndefined();
-        });
-    });
-    // ── JSON extraction (balanced parsing) ──
-    describe('JSON extraction (balanced parsing)', () => {
-        it('parses plain JSON from LLM response', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeStartRunInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-        });
-        it('parses JSON wrapped in prose text', async () => {
-            const proseWrapped = `Here is the analysis:\n${JSON.stringify(VALID_DIAGNOSIS)}\nDone.`;
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(proseWrapped));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeStartRunInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-        });
-        it('parses JSON inside code fences', async () => {
-            const codeFenced = '```json\n' + JSON.stringify(VALID_DIAGNOSIS) + '\n```';
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(codeFenced));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeStartRunInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-        });
-        it('throws output_invalid when no JSON object found in response', async () => {
-            await expectStartRunError('output_invalid', () => {
-                mockComplete.mockReset();
-                mockComplete.mockResolvedValue(makeAssistantMessage('No JSON here at all'));
-            });
-        });
-    });
-    // ── startRun() error mapping ──
-    describe('startRun() error mapping', () => {
-        it('throws PDRuntimeError("runtime_unavailable") when apiKeyEnv missing from process.env', async () => {
-            await expectStartRunError('runtime_unavailable', () => {
-                delete process.env.TEST_API_KEY;
-            });
-        });
-        it('throws PDRuntimeError("timeout") when LLM request is aborted/timed out', async () => {
-            await expectStartRunError('timeout', () => {
-                mockComplete.mockReset();
-                mockComplete.mockImplementation(() => {
-                    throw new DOMException('The operation was aborted', 'AbortError');
-                });
-            });
-        });
-        it('throws PDRuntimeError("output_invalid") when LLM response contains no parseable JSON', async () => {
-            await expectStartRunError('output_invalid', () => {
-                mockComplete.mockReset();
-                mockComplete.mockResolvedValue(makeAssistantMessage('Just plain text, no JSON at all'));
-            });
-        });
-        it('skips schema validation when no outputSchemaRef is provided (backward compatible)', async () => {
-            mockComplete.mockReset();
-            mockComplete.mockResolvedValue(makeAssistantMessage(JSON.stringify({ valid: true, missing: 'fields' })));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeStartRunInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ valid: true, missing: 'fields' });
-        });
-        it('throws PDRuntimeError("output_invalid") when parsed JSON does not match schema for given outputSchemaRef', async () => {
-            mockComplete.mockReset();
-            mockComplete.mockResolvedValue(makeAssistantMessage(JSON.stringify({ valid: true, missing: 'fields' })));
-            const adapter = makeAdapter();
-            await expect(adapter.startRun(makeStartRunInput({
-                outputSchemaRef: 'diagnostician-output-v1',
-            }))).rejects.toMatchObject({ category: 'output_invalid' });
-        });
-        it('throws PDRuntimeError("execution_failed") after retry exhaustion on network error', async () => {
-            await expectStartRunError('execution_failed', () => {
-                mockComplete.mockReset();
-                mockComplete.mockRejectedValue(new Error('ECONNREFUSED'));
-            }, { maxRetries: 1 });
-        });
-        it('throws PDRuntimeError("timeout") when complete resolves with stopReason:error + timeout message', async () => {
-            await expectStartRunError('timeout', () => {
-                mockComplete.mockReset();
-                mockComplete.mockResolvedValue(makeErrorResponse('Request timed out.'));
-            });
-        });
-        it('throws PDRuntimeError("execution_failed") when complete resolves with stopReason:error + 401 Unauthorized', async () => {
-            await expectStartRunError('execution_failed', () => {
-                mockComplete.mockReset();
-                mockComplete.mockResolvedValue(makeErrorResponse('401 Unauthorized'));
-            });
-        });
-        it('throws PDRuntimeError("timeout") when complete resolves with stopReason:aborted', async () => {
-            await expectStartRunError('timeout', () => {
-                mockComplete.mockReset();
-                mockComplete.mockResolvedValue(makeAbortedResponse());
-            });
-        });
-        it('retries resolved stopReason:error (execution_failed) before giving up', async () => {
-            // First call: resolves with error, second call: succeeds
-            mockComplete
-                .mockResolvedValueOnce(makeErrorResponse('503 Service Unavailable'))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ maxRetries: 2 });
-            const handle = await adapter.startRun(makeStartRunInput());
-            expect(mockComplete).toHaveBeenCalledTimes(2);
-            expect(handle.runId).toBeTruthy();
-        });
-        it('RunStatus.reason includes bounded errorMessage from resolved-error response', async () => {
-            mockComplete.mockReset();
-            mockComplete.mockResolvedValue(makeErrorResponse('401 Unauthorized: invalid API key'));
-            const adapter = makeAdapter({ maxRetries: 0 });
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch {
-                // startRun throws, but internally run state should have the reason
-            }
-            // Telemetry should include the error message
-            const failedEvent = findTelemetryEvent('runtime_invocation_failed');
-            expect(failedEvent).toBeDefined();
-            const payload = failedEvent?.payload;
-            expect(payload.errorCategory).toBe('execution_failed');
-            expect(typeof payload.errorMessage).toBe('string');
-            expect(payload.errorMessage.includes('401')).toBe(true);
-        });
-    });
-    // ── startRun() retry logic ──
-    describe('startRun() retry logic', () => {
-        it('retries transient failures up to maxRetries times', async () => {
-            // Fail twice, succeed on third
-            mockComplete
-                .mockRejectedValueOnce(new Error('ECONNREFUSED'))
-                .mockRejectedValueOnce(new Error('ETIMEDOUT'))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ maxRetries: 2 });
-            const handle = await adapter.startRun(makeStartRunInput());
-            expect(mockComplete).toHaveBeenCalledTimes(3);
-            expect(handle.runtimeKind).toBe('pi-ai');
-        });
-        it('succeeds on second attempt after first transient failure', async () => {
-            mockComplete
-                .mockRejectedValueOnce(new Error('socket hang up'))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ maxRetries: 2 });
-            const handle = await adapter.startRun(makeStartRunInput());
-            expect(mockComplete).toHaveBeenCalledTimes(2);
-            expect(handle.runId).toBeTruthy();
-        });
-        it('does not retry on PDRuntimeError (output_invalid)', async () => {
-            mockComplete.mockReset();
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify({ valid: true })))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify({ valid: true })))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify({ valid: true })))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify({ valid: true })));
-            const adapter = makeAdapter({ maxRetries: 3 });
-            await expect(adapter.startRun(makeStartRunInput({
-                outputSchemaRef: 'diagnostician-output-v1',
-            }))).rejects.toThrow(PDRuntimeError);
-            // 1 original + 3 repair attempts = 4 calls (not infinite retries)
-            expect(mockComplete).toHaveBeenCalledTimes(4);
-        });
-        it('does not retry on PDRuntimeError (runtime_unavailable)', async () => {
-            delete process.env.TEST_API_KEY;
-            const adapter = makeAdapter({ maxRetries: 3 });
-            await expect(adapter.startRun(makeStartRunInput())).rejects.toMatchObject({
-                category: 'runtime_unavailable',
-            });
-            // complete should not be called at all — error happens before LLM call
-            expect(mockComplete).not.toHaveBeenCalled();
-        });
-    });
-    // ── pollRun() ──
-    describe('pollRun()', () => {
-        it('returns terminal status (succeeded) for completed run', async () => {
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeStartRunInput());
-            const status = await adapter.pollRun(handle.runId);
-            expect(status.runId).toBe(handle.runId);
-            expect(status.status).toBe('succeeded');
-            expect(status.startedAt).toBeTruthy();
-            expect(status.endedAt).toBeTruthy();
-        });
-        it('returns failed status with reason when startRun fails', async () => {
-            mockComplete.mockReset();
-            mockComplete.mockImplementation(async () => {
-                throw new DOMException('aborted', 'AbortError');
-            });
-            const adapter = makeAdapter();
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch (err) {
-                // Verify the error path doesn't crash and emits telemetry
-                expect(err).toBeInstanceOf(PDRuntimeError);
-            }
-        });
-        it('throws PDRuntimeError("input_invalid") for unknown runId', async () => {
-            const adapter = makeAdapter();
-            await expect(adapter.pollRun('nonexistent-run-id')).rejects.toThrow(PDRuntimeError);
-            await expect(adapter.pollRun('nonexistent-run-id')).rejects.toMatchObject({
-                category: 'input_invalid',
-            });
-        });
-    });
-    // ── fetchOutput() ──
-    describe('fetchOutput()', () => {
-        it('returns StructuredRunOutput with runId and payload after successful run', async () => {
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeStartRunInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output).not.toBeNull();
-            expect(output?.runId).toBe(handle.runId);
-            expect(output?.payload).toMatchObject({
-                valid: true,
-                diagnosisId: 'diag-test-1',
-            });
-        });
-        it('returns null for unknown runId', async () => {
-            const adapter = makeAdapter();
-            const output = await adapter.fetchOutput('nonexistent-run-id');
-            expect(output).toBeNull();
-        });
-    });
-    // ── cancelRun() ──
-    describe('cancelRun()', () => {
-        it('resolves without error', async () => {
-            const adapter = makeAdapter();
-            await expect(adapter.cancelRun('any-run-id')).resolves.toBeUndefined();
-        });
-    });
-    // ── fetchArtifacts() ──
-    describe('fetchArtifacts()', () => {
-        it('returns empty array for a run that was started', async () => {
-            const adapter = makeAdapter();
-            const input = makeStartRunInput({ taskRef: { taskId: 'test-run-id' } });
-            const handle = await adapter.startRun(input);
-            const artifacts = await adapter.fetchArtifacts(handle.runId);
-            expect(artifacts).toEqual([]);
-        });
-        it('throws for an unknown runId', async () => {
-            const adapter = makeAdapter();
-            await expect(adapter.fetchArtifacts('unknown-run-id')).rejects.toThrow("Run 'unknown-run-id' not found");
-        });
-    });
-    // ── telemetry (AD-15) ──
-    describe('telemetry (AD-15)', () => {
-        it('emits runtime_invocation_started event on startRun', async () => {
-            const adapter = makeAdapter();
-            await adapter.startRun(makeStartRunInput());
-            const startedEvent = findTelemetryEvent('runtime_invocation_started');
-            expect(startedEvent).toBeDefined();
-            const payload = startedEvent?.payload;
-            expect(payload.runtimeKind).toBe('pi-ai');
-            expect(payload.provider).toBe('openrouter');
-            expect(payload.model).toBe('anthropic/claude-sonnet-4');
-        });
-        it('emits runtime_invocation_succeeded on successful completion', async () => {
-            const adapter = makeAdapter();
-            await adapter.startRun(makeStartRunInput());
-            const succeededEvent = findTelemetryEvent('runtime_invocation_succeeded');
-            expect(succeededEvent).toBeDefined();
-            const payload = succeededEvent?.payload;
-            expect(payload.runtimeKind).toBe('pi-ai');
-        });
-        it('emits runtime_invocation_failed on failure', async () => {
-            mockComplete.mockReset();
-            mockComplete.mockRejectedValue(new Error('network down'));
-            const adapter = makeAdapter({ maxRetries: 0 });
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch { /* expected */ }
-            const failedEvent = findTelemetryEvent('runtime_invocation_failed');
-            expect(failedEvent).toBeDefined();
-            const payload = failedEvent?.payload;
-            expect(payload.runtimeKind).toBe('pi-ai');
-            expect(payload.errorCategory).toBe('execution_failed');
-        });
-        it('emits runtime_invocation_started before LLM call (even if call fails)', async () => {
-            mockComplete.mockReset();
-            mockComplete.mockRejectedValue(new Error('fail'));
-            const adapter = makeAdapter({ maxRetries: 0 });
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch { /* expected */ }
-            const startedEvent = findTelemetryEvent('runtime_invocation_started');
-            expect(startedEvent).toBeDefined();
-        });
-    });
-    // ── Error category coverage ──
-    describe('error category coverage', () => {
-        it('all 5 PDRuntimeError categories are reachable', () => {
-            // This test documents which categories are tested where:
-            // - runtime_unavailable: apiKeyEnv missing test
-            // - timeout: AbortError test
-            // - output_invalid: JSON parse failure + schema mismatch tests
-            // - execution_failed: retry exhaustion test
-            // - input_invalid: pollRun with unknown runId test
-            const categories = ['runtime_unavailable', 'timeout', 'output_invalid', 'execution_failed', 'input_invalid'];
-            expect(categories).toHaveLength(5);
-        });
-    });
-    // ── Run state management ──
-    describe('run state management', () => {
-        it('stores failed run state on error', async () => {
-            mockComplete.mockReset();
-            mockComplete.mockImplementation(() => {
-                throw new DOMException('aborted', 'AbortError');
-            });
-            const adapter = makeAdapter();
-            // startRun will throw, but internally the run state should be stored
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch {
-                // We don't have the runId since startRun threw.
-                // This test verifies the error path doesn't crash.
-            }
-            // Verify the adapter didn't crash and telemetry was emitted
-            expect(mockEmitTelemetry).toHaveBeenCalled();
-        });
-        it('creates run state before LLM call', async () => {
-            let resolveComplete;
-            const blockedPromise = new Promise(resolve => { resolveComplete = resolve; });
-            mockComplete.mockReturnValueOnce(blockedPromise);
-            const adapter = makeAdapter();
-            // Start the run — it will block on complete()
-            const handlePromise = adapter.startRun(makeStartRunInput());
-            // Unblock the LLM call
-            if (resolveComplete) {
-                resolveComplete(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            }
-            const handle = await handlePromise;
-            // After completion, pollRun should return succeeded
-            const status = await adapter.pollRun(handle.runId);
-            expect(status.status).toBe('succeeded');
-        });
-    });
-    // ── startRun() output repair (PRI-71) ──
-    describe('startRun() output repair (PRI-71)', () => {
-        /** Invalid output with string confidence and wrong-case kind. */
-        const INVALID_DIAGNOSIS = {
-            valid: true,
-            diagnosisId: 'diag-repair-1',
-            taskId: 'task-repair-1',
-            summary: 'Test repair',
-            rootCause: 'Test root cause',
-            violatedPrinciples: [],
-            evidence: [],
-            recommendations: [{ kind: 'Rule', description: 'Fix casing' }],
-            confidence: '85%',
-        };
-        /** Make a startRunInput with diagnostician output schema ref. */
-        function makeDiagnosticianInput(overrides = {}) {
-            return makeStartRunInput({
-                outputSchemaRef: 'diagnostician-output-v1',
-                taskRef: { taskId: 'task-repair-1' },
-                ...overrides,
-            });
-        }
-        it('repairs confidence "85%" to 0.85 via second LLM call', async () => {
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeDiagnosticianInput());
-            expect(mockComplete).toHaveBeenCalledTimes(2);
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ confidence: 0.9 });
-        });
-        it('PRI-633: repair calls carry the merged layered system prompt', async () => {
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ systemPrompt: 'PROFILE APPEND LAYER' });
-            await adapter.startRun(makeDiagnosticianInput({ systemPrompt: 'BASE ROLE LAYER' }));
-            expect(mockComplete).toHaveBeenCalledTimes(2);
-            // Call 2 is the repair call — its Context must keep the layered prompt.
-            const [, repairContext] = mockComplete.mock.calls[1];
-            expect(repairContext.systemPrompt).toBe('BASE ROLE LAYER\n\nPROFILE APPEND LAYER');
-        });
-        it('repair fails — still throws output_invalid after repair attempt', async () => {
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)));
-            const adapter = makeAdapter();
-            await expect(adapter.startRun(makeDiagnosticianInput())).rejects.toMatchObject({
-                category: 'output_invalid',
-            });
-            // 1 original + 3 repair attempts = 4 calls
-            expect(mockComplete).toHaveBeenCalledTimes(4);
-        });
-        it('unknown outputSchemaRef fails loud before repair is attempted (rc-3)', async () => {
-            // CodeRabbit PR #1259: unknown non-empty ref now fails loud at schema
-            // resolution (in startRun), so repair is never attempted.
-            // NOTE: no mockComplete setup — resolution fails BEFORE any LLM call,
-            // so no mock value is queued to leak into subsequent tests.
-            const adapter = makeAdapter();
-            await expect(adapter.startRun(makeStartRunInput({
-                outputSchemaRef: 'custom-unknown-v1',
-            }))).rejects.toMatchObject({
-                category: 'output_invalid',
-                message: expect.stringMatching(/Unknown outputSchemaRef: custom-unknown-v1/i),
-            });
-            // LLM was never called because resolution failed first.
-            expect(mockComplete).toHaveBeenCalledTimes(0);
-        });
-        it('skips repair for non-JSON output (no schema errors to report)', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage('Not JSON at all'));
-            const adapter = makeAdapter();
-            await expect(adapter.startRun(makeDiagnosticianInput())).rejects.toMatchObject({
-                category: 'output_invalid',
-            });
-            expect(mockComplete).toHaveBeenCalledTimes(1);
-        });
-        it('emits output_extraction_failed telemetry when extractJsonObject returns null', async () => {
-            const proseOutput = 'I evaluated the plan and it looks good. The implementation is solid.';
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(proseOutput));
-            const adapter = makeAdapter();
-            try {
-                await adapter.startRun(makeDiagnosticianInput());
-            }
-            catch { /* expected */ }
-            const extractionEvent = findTelemetryEvent('output_extraction_failed');
-            expect(extractionEvent).toBeDefined();
-            const payload = extractionEvent?.payload;
-            expect(payload.outputSchemaRef).toBe('diagnostician-output-v1');
-            expect(payload.provider).toBe('openrouter');
-            expect(payload.model).toBe('anthropic/claude-sonnet-4');
-            expect(payload.rawOutputPreview).toBe(proseOutput);
-        });
-        it('emits output_repair_attempted telemetry on repair attempt', async () => {
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter();
-            await adapter.startRun(makeDiagnosticianInput());
-            const repairEvent = findTelemetryEvent('output_repair_attempted');
-            expect(repairEvent).toBeDefined();
-            const payload = repairEvent?.payload;
-            expect(payload.repaired).toBe(true);
-            expect(payload.attemptsUsed).toBe(1);
-        });
-        it('emits output_repair_attempted with repaired=false on failure', async () => {
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)));
-            const adapter = makeAdapter();
-            try {
-                await adapter.startRun(makeDiagnosticianInput());
-            }
-            catch { /* expected */ }
-            const repairEvent = findTelemetryEvent('output_repair_attempted');
-            expect(repairEvent).toBeDefined();
-            const payload = repairEvent?.payload;
-            expect(payload.repaired).toBe(false);
-        });
-        it('reuses same provider/model for repair call', async () => {
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter();
-            await adapter.startRun(makeDiagnosticianInput());
-            // Both calls should use same model
-            expect(mockComplete).toHaveBeenCalledTimes(2);
-            const [model1] = mockComplete.mock.calls[0];
-            const [model2] = mockComplete.mock.calls[1];
-            expect(model1).toEqual(model2);
-        });
-        it('repair call includes the repair prompt with schema errors', async () => {
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter();
-            await adapter.startRun(makeDiagnosticianInput());
-            // Second call's context should contain the repair prompt
-            const [, context] = mockComplete.mock.calls[1];
-            const repairMessage = context.messages[0]?.content ?? '';
-            expect(repairMessage).toContain('schema');
-            expect(repairMessage).toContain('confidence');
-        });
-        it('repair budget exhausted — default maxRepairAttempts=3, no infinite loop', async () => {
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS)));
-            const adapter = makeAdapter();
-            await expect(adapter.startRun(makeDiagnosticianInput())).rejects.toMatchObject({
-                category: 'output_invalid',
-            });
-            // 1 original + 3 repair attempts = 4 calls total (not infinite)
-            expect(mockComplete).toHaveBeenCalledTimes(4);
-        });
-        it('repairs evaluator-output-v1 with prose-wrapped JSON', async () => {
-            const EVALUATOR_OUTPUT_INVALID = {
-                taskId: 'task-eval-repair-1',
-                sourceArtificerArtifactId: 'pi-art-artificer-repair-1',
-                evaluation: {
-                    decision: 'approved',
-                    summary: 'Good plan',
-                    score: '0.85',
-                    strengths: ['Clear'],
-                    concerns: [],
-                    requiredChanges: [],
-                },
-                sourceTrace: {
-                    artificerArtifactId: 'pi-art-artificer-repair-1',
-                },
-                risks: [],
-                generatedAt: '2026-05-11T12:00:00.000Z',
-            };
-            const EVALUATOR_OUTPUT_VALID = {
-                taskId: 'task-eval-repair-1',
-                sourceArtificerArtifactId: 'pi-art-artificer-repair-1',
-                evaluation: {
-                    decision: 'approved',
-                    summary: 'Good plan',
-                    score: 0.85,
-                    strengths: ['Clear'],
-                    concerns: [],
-                    requiredChanges: [],
-                },
-                sourceTrace: {
-                    artificerArtifactId: 'pi-art-artificer-repair-1',
-                },
-                risks: [],
-                generatedAt: '2026-05-11T12:00:00.000Z',
-            };
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(EVALUATOR_OUTPUT_INVALID)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(EVALUATOR_OUTPUT_VALID)));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeStartRunInput({
-                outputSchemaRef: 'evaluator-output-v1',
-                taskRef: { taskId: 'task-eval-repair-1' },
-            }));
-            expect(mockComplete).toHaveBeenCalledTimes(2);
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ evaluation: { score: 0.85 } });
-        });
-    });
-    // ── startRun() schema dispatch by outputSchemaRef ──
-    describe('startRun() schema dispatch by outputSchemaRef', () => {
-        const VALID_DREAMER_OUTPUT = {
-            valid: true,
-            taskId: 'task-dreamer-1',
-            candidates: [
-                {
-                    candidateIndex: 0,
-                    badDecision: 'Ignored error handling',
-                    betterDecision: 'Add try/catch around API calls',
-                    rationale: 'Prevents unhandled rejections',
-                    confidence: 0.85,
-                    riskLevel: 'low',
-                    strategicPerspective: 'Defensive programming',
-                },
-            ],
-            contextRefs: [],
-            generatedAt: '2025-01-01T00:00:00.000Z',
-        };
-        const VALID_PHILOSOPHER_OUTPUT = {
-            taskId: 'task-philosopher-1',
-            sourceDreamerArtifactId: 'pi-art-dreamer-1',
-            thesis: 'Error handling is essential for reliability',
-            principleCandidate: {
-                title: 'Always handle API errors',
-                rationale: 'Unhandled errors cause cascading failures',
-                scope: 'All external API calls',
-                confidence: 0.9,
-            },
-            risks: ['May add boilerplate'],
-            generatedAt: '2025-01-01T00:00:00.000Z',
-        };
-        it('validates dreamer-output-v1 with DreamerOutputV1Schema', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DREAMER_OUTPUT)));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeStartRunInput({
-                outputSchemaRef: 'dreamer-output-v1',
-                taskRef: { taskId: 'task-dreamer-1' },
-            }));
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ valid: true });
-        });
-        it('rejects invalid dreamer output with output_invalid', async () => {
-            const invalidDreamerOutput = { valid: true, taskId: 'task-dreamer-1', candidates: 'not-array' };
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(invalidDreamerOutput)));
-            const adapter = makeAdapter();
-            await expect(adapter.startRun(makeStartRunInput({
-                outputSchemaRef: 'dreamer-output-v1',
-                taskRef: { taskId: 'task-dreamer-1' },
-            }))).rejects.toMatchObject({ category: 'output_invalid' });
-        });
-        it('validates philosopher-output-v1 with PhilosopherOutputV1Schema', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_PHILOSOPHER_OUTPUT)));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeStartRunInput({
-                outputSchemaRef: 'philosopher-output-v1',
-                taskRef: { taskId: 'task-philosopher-1' },
-            }));
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ thesis: 'Error handling is essential for reliability' });
-        });
-        it('rejects invalid philosopher output with output_invalid', async () => {
-            const invalidPhilosopherOutput = { taskId: 'task-philosopher-1', principleCandidate: null };
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(invalidPhilosopherOutput)));
-            const adapter = makeAdapter();
-            await expect(adapter.startRun(makeStartRunInput({
-                outputSchemaRef: 'philosopher-output-v1',
-                taskRef: { taskId: 'task-philosopher-1' },
-            }))).rejects.toMatchObject({ category: 'output_invalid' });
-        });
-        it('falls back to DiagnosticianOutputV1Schema for diagnostician-output-v1', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeStartRunInput({
-                outputSchemaRef: 'diagnostician-output-v1',
-                taskRef: { taskId: 'task-test-1' },
-            }));
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ valid: true, diagnosisId: 'diag-test-1' });
-        });
-        it('unknown outputSchemaRef fails loud with output_invalid (rc-3, was silent skip)', async () => {
-            // CodeRabbit PR #1259: previously an unknown non-empty outputSchemaRef
-            // resolved to undefined and silently skipped schema-gated validation.
-            // That is an rc-9/rc-3 silent-fallback. Now it fails loud, consistent
-            // with OpenClawCliRuntimeAdapter.fetchOutput.
-            // NOTE: no mockComplete setup — resolution fails BEFORE any LLM call.
-            const adapter = makeAdapter();
-            await expect(adapter.startRun(makeStartRunInput({
-                outputSchemaRef: 'custom-output-v1',
-            }))).rejects.toMatchObject({
-                category: 'output_invalid',
-                message: expect.stringMatching(/Unknown outputSchemaRef: custom-output-v1/i),
-            });
-            expect(mockComplete).toHaveBeenCalledTimes(0);
-        });
-        it('no outputSchemaRef skips schema validation (backward compatible)', async () => {
-            const arbitraryOutput = { anything: 'goes' };
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(arbitraryOutput)));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeStartRunInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject(arbitraryOutput);
-        });
-        it('attempts repair for dreamer-output-v1 schema errors', async () => {
-            const invalidDreamerOutput = {
-                valid: true,
-                taskId: 'task-dreamer-1',
-                candidates: [{ candidateIndex: 0, badDecision: 'test', betterDecision: 'test', rationale: 'test', confidence: '85%', riskLevel: 'low', strategicPerspective: 'test' }],
-                contextRefs: [],
-                generatedAt: '2025-01-01T00:00:00.000Z',
-            };
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(invalidDreamerOutput)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DREAMER_OUTPUT)));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeStartRunInput({
-                outputSchemaRef: 'dreamer-output-v1',
-                taskRef: { taskId: 'task-dreamer-1' },
-            }));
-            expect(mockComplete).toHaveBeenCalledTimes(2);
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ valid: true });
-        });
-    });
-    // ── PRI-200: Structured output repair loop with evidence pack ──
-    describe('PRI-200: structured output repair loop with evidence pack', () => {
-        const INVALID_DIAGNOSIS_SCHEMA = {
-            valid: true,
-            diagnosisId: 'diag-pri200-1',
-            taskId: 'task-pri200-1',
-            summary: 'Test PRI-200',
-            rootCause: 'Test root cause',
-            violatedPrinciples: [],
-            evidence: [],
-            recommendations: [{ kind: 'Rule', description: 'Fix casing' }],
-            confidence: '85%',
-        };
-        function makeDiagnosticianInput(overrides = {}) {
-            return makeStartRunInput({
-                outputSchemaRef: 'diagnostician-output-v1',
-                taskRef: { taskId: 'task-pri200-1' },
-                ...overrides,
-            });
-        }
-        function resetMock() {
-            mockComplete.mockReset();
-            mockComplete.mockResolvedValue(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-        }
-        it('1. valid JSON + valid schema → no repair triggered', async () => {
-            resetMock();
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeDiagnosticianInput());
-            expect(mockComplete).toHaveBeenCalledTimes(1);
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ confidence: 0.9 });
-            const repairEvent = findTelemetryEvent('output_repair_attempted');
-            expect(repairEvent).toBeUndefined();
-            const schemaInvalidEvent = findTelemetryEvent('output_schema_invalid');
-            expect(schemaInvalidEvent).toBeUndefined();
-        });
-        it('2. prose-wrapped JSON extraction succeeds → no repair triggered', async () => {
-            resetMock();
-            const proseWrapped = `Here is the analysis:\n${JSON.stringify(VALID_DIAGNOSIS)}\nDone.`;
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(proseWrapped));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeDiagnosticianInput());
-            expect(mockComplete).toHaveBeenCalledTimes(1);
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-        });
-        it('3. extraction failed → output_extraction_failed telemetry + output_invalid', async () => {
-            resetMock();
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage('No JSON here at all'));
-            const adapter = makeAdapter();
-            try {
-                await adapter.startRun(makeDiagnosticianInput());
-            }
-            catch { /* expected */ }
-            const extractionEvent = findTelemetryEvent('output_extraction_failed');
-            expect(extractionEvent).toBeDefined();
-            const payload = extractionEvent?.payload;
-            expect(payload.outputSchemaRef).toBe('diagnostician-output-v1');
-            expect(payload.provider).toBe('openrouter');
-            expect(payload.model).toBe('anthropic/claude-sonnet-4');
-            expect(typeof payload.rawOutputPreview).toBe('string');
-        });
-        it('4. schema invalid once, repair returns valid → success + repairAttempts[0].repaired=true', async () => {
-            resetMock();
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS_SCHEMA)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeDiagnosticianInput());
-            expect(mockComplete).toHaveBeenCalledTimes(2);
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ confidence: 0.9 });
-            const repairEvent = findTelemetryEvent('output_repair_attempted');
-            expect(repairEvent).toBeDefined();
-            const repairPayload = repairEvent?.payload;
-            expect(repairPayload.repaired).toBe(true);
-            expect(repairPayload.attemptsUsed).toBe(1);
-        });
-        it('5. schema invalid, repair also invalid → output_repair_exhausted + output_invalid', async () => {
-            resetMock();
-            mockComplete.mockResolvedValue(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS_SCHEMA)));
-            const adapter = makeAdapter();
-            await expect(adapter.startRun(makeDiagnosticianInput())).rejects.toMatchObject({
-                category: 'output_invalid',
-            });
-            const exhaustedEvent = findTelemetryEvent('output_repair_exhausted');
-            expect(exhaustedEvent).toBeDefined();
-            const exhaustedPayload = exhaustedEvent?.payload;
-            expect(exhaustedPayload.outputSchemaRef).toBe('diagnostician-output-v1');
-            expect(exhaustedPayload.provider).toBe('openrouter');
-            expect(exhaustedPayload.model).toBe('anthropic/claude-sonnet-4');
-            expect(typeof exhaustedPayload.rawOutputPreview).toBe('string');
-            expect(Array.isArray(exhaustedPayload.validationErrors)).toBe(true);
-            expect(Array.isArray(exhaustedPayload.repairAttempts)).toBe(true);
-            expect(typeof exhaustedPayload.finalFailureReason).toBe('string');
-        });
-        it('6. repair loop never exceeds configured max attempts', async () => {
-            resetMock();
-            mockComplete.mockResolvedValue(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS_SCHEMA)));
-            const adapter = makeAdapter();
-            await expect(adapter.startRun(makeDiagnosticianInput())).rejects.toMatchObject({
-                category: 'output_invalid',
-            });
-            // 1 original + 3 repair (default maxRepairAttempts=3) = 4 calls total
-            expect(mockComplete).toHaveBeenCalledTimes(4);
-        });
-        it('7. repair prompt includes schemaRef and error paths', async () => {
-            resetMock();
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS_SCHEMA)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter();
-            await adapter.startRun(makeDiagnosticianInput());
-            const [, context] = mockComplete.mock.calls[1];
-            const repairMessage = context.messages[0]?.content ?? '';
-            expect(repairMessage).toContain('diagnostician-output-v1');
-            expect(repairMessage).toContain('/confidence');
-        });
-        it('8. repair attempt must not override lineage fields silently', async () => {
-            resetMock();
-            const originalWithLineage = {
-                ...INVALID_DIAGNOSIS_SCHEMA,
-                taskId: 'task-original-1',
-            };
-            const repairedWithChangedLineage = {
-                ...VALID_DIAGNOSIS,
-                taskId: 'task-CHANGED-by-llm',
-            };
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(originalWithLineage)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(repairedWithChangedLineage)));
-            const adapter = makeAdapter();
-            const handle = await adapter.startRun(makeDiagnosticianInput({
-                taskRef: { taskId: 'task-original-1' },
-            }));
-            const output = await adapter.fetchOutput(handle.runId);
-            const payload = output?.payload;
-            // taskId is a lineage field — it must be stripped from output (PRI-272).
-            // Downstream consumers get taskId from RunnerContext/TaskRecord.
-            expect(payload.taskId).toBeUndefined();
-        });
-        it('9. provider/model/rawOutputPreview are present in evidence pack', async () => {
-            resetMock();
-            mockComplete.mockResolvedValue(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS_SCHEMA)));
-            const adapter = makeAdapter();
-            try {
-                await adapter.startRun(makeDiagnosticianInput());
-            }
-            catch { /* expected */ }
-            const exhaustedEvent = findTelemetryEvent('output_repair_exhausted');
-            expect(exhaustedEvent).toBeDefined();
-            const payload = exhaustedEvent?.payload;
-            expect(payload.provider).toBe('openrouter');
-            expect(payload.model).toBe('anthropic/claude-sonnet-4');
-            expect(typeof payload.rawOutputPreview).toBe('string');
-            expect(payload.rawOutputPreview.length).toBeGreaterThan(0);
-        });
-        it('10. unknown schemaRef fails loud (rc-3; was backward-compatible silent skip)', async () => {
-            // CodeRabbit PR #1259: previously an unknown non-empty outputSchemaRef
-            // silently skipped schema validation (the "backward-compatible behavior").
-            // That silent skip is an rc-9/rc-3 defect — now it fails loud, matching
-            // OpenClawCliRuntimeAdapter. NOTE: no mockComplete setup because schema
-            // resolution fails before any LLM call.
-            resetMock();
-            const adapter = makeAdapter();
-            await expect(adapter.startRun(makeStartRunInput({
-                outputSchemaRef: 'custom-unknown-v1',
-            }))).rejects.toMatchObject({
-                category: 'output_invalid',
-                message: expect.stringMatching(/Unknown outputSchemaRef: custom-unknown-v1/i),
-            });
-            expect(mockComplete).toHaveBeenCalledTimes(0);
-        });
-        it('emits output_schema_invalid telemetry before repair attempt', async () => {
-            resetMock();
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS_SCHEMA)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter();
-            await adapter.startRun(makeDiagnosticianInput());
-            const schemaInvalidEvent = findTelemetryEvent('output_schema_invalid');
-            expect(schemaInvalidEvent).toBeDefined();
-            const payload = schemaInvalidEvent?.payload;
-            expect(payload.outputSchemaRef).toBe('diagnostician-output-v1');
-            expect(payload.provider).toBe('openrouter');
-            expect(payload.model).toBe('anthropic/claude-sonnet-4');
-            expect(typeof payload.rawOutputPreview).toBe('string');
-            expect(Array.isArray(payload.validationErrors)).toBe(true);
-            expect(payload.validationErrors.length).toBeGreaterThan(0);
-        });
-        it('evidence pack includes repairAttempts with schemaRef and attempt number', async () => {
-            resetMock();
-            mockComplete.mockResolvedValue(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS_SCHEMA)));
-            const adapter = makeAdapter();
-            try {
-                await adapter.startRun(makeDiagnosticianInput());
-            }
-            catch { /* expected */ }
-            const exhaustedEvent = findTelemetryEvent('output_repair_exhausted');
-            expect(exhaustedEvent).toBeDefined();
-            const payload = exhaustedEvent?.payload;
-            const attempts = payload.repairAttempts;
-            expect(attempts.length).toBeGreaterThan(0);
-            expect(attempts[0]?.schemaRef).toBe('diagnostician-output-v1');
-            expect(typeof attempts[0]?.attempt).toBe('number');
-            expect(typeof attempts[0]?.repaired).toBe('boolean');
-        });
-        it('output_schema_invalid telemetry includes validation error paths', async () => {
-            resetMock();
-            mockComplete
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(INVALID_DIAGNOSIS_SCHEMA)))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter();
-            await adapter.startRun(makeDiagnosticianInput());
-            const schemaInvalidEvent = findTelemetryEvent('output_schema_invalid');
-            expect(schemaInvalidEvent).toBeDefined();
-            const payload = schemaInvalidEvent?.payload;
-            const errors = payload.validationErrors;
-            expect(errors.length).toBeGreaterThan(0);
-            const errorPaths = errors.map(e => e.path);
-            expect(errorPaths.some(p => typeof p === 'string' && p.includes('confidence'))).toBe(true);
-        });
-    });
-    // ── PRI-220: Provider Timeout and Transient Retry Policy ──
-    describe('PRI-220: Provider Timeout and Transient Retry Policy', () => {
-        it('1. provider timeout before client budget is provider_transient_timeout', async () => {
-            // LLM call fails with timeout-like error, client signal is NOT aborted
-            mockComplete.mockReset();
-            mockComplete.mockRejectedValue(new Error('connection timeout'));
-            const adapter = makeAdapter({ maxRetries: 0 });
-            let caughtErr = undefined;
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch (err) {
-                if (err instanceof PDRuntimeError)
-                    caughtErr = err;
-            }
-            expect(caughtErr).toBeDefined();
-            expect(caughtErr?.category).toBe('timeout');
-            expect(caughtErr?.details?.timeoutClassification).toBe('provider_transient_timeout');
-            expect(caughtErr?.details?.effectiveTimeoutMs).toBe(60_000);
-            expect(typeof caughtErr?.details?.elapsedMs).toBe('number');
-        });
-        it('2. client timeout budget exhausted when signal is aborted', async () => {
-            // LLM call aborts, client signal IS aborted
-            mockComplete.mockReset();
-            mockComplete.mockImplementation(async (_model, _ctx, options) => {
-                const signal = options.signal;
-                const err = new DOMException('The operation was aborted', 'AbortError');
-                Object.defineProperty(signal, 'aborted', { value: true, writable: true });
-                throw err;
-            });
-            const adapter = makeAdapter({ maxRetries: 0 });
-            let caughtErr = undefined;
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch (err) {
-                if (err instanceof PDRuntimeError)
-                    caughtErr = err;
-            }
-            expect(caughtErr).toBeDefined();
-            expect(caughtErr?.category).toBe('timeout');
-            expect(caughtErr?.details?.timeoutClassification).toBe('client_timeout_budget_exhausted');
-        });
-        it('3. unrecognized generic non-timeout error is classified as not_applicable', async () => {
-            mockComplete.mockReset();
-            mockComplete.mockRejectedValue(new Error('Generic database connection failed'));
-            const adapter = makeAdapter({ maxRetries: 0 });
-            let caughtErr = undefined;
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch (err) {
-                if (err instanceof PDRuntimeError)
-                    caughtErr = err;
-            }
-            expect(caughtErr).toBeDefined();
-            expect(caughtErr?.category).toBe('execution_failed');
-            expect(caughtErr?.details?.timeoutClassification).toBeUndefined();
-        });
-        it('5. provider_transient_timeout retries and then succeeds within max limits', async () => {
-            mockComplete.mockReset();
-            // First attempt: transient timeout, Second attempt: success
-            mockComplete
-                .mockRejectedValueOnce(new Error('socket timeout'))
-                .mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ maxRetries: 2 });
-            const handle = await adapter.startRun(makeStartRunInput());
-            expect(mockComplete).toHaveBeenCalledTimes(2);
-            expect(handle.runId).toBeDefined();
-            // Check telemetry has retry attempt
-            const retryEvent = findTelemetryEvent('runtime_invocation_failed');
-            expect(retryEvent).toBeDefined();
-            const payload = retryEvent?.payload;
-            expect(payload?.isRetryAttempt).toBe(true);
-            expect(payload?.retryAttempt).toBe(0);
-        });
-        it('6. provider_transient_timeout exhausted maxRetries and fails permanently', async () => {
-            mockComplete.mockReset();
-            mockComplete.mockRejectedValue(new Error('API read timeout'));
-            const adapter = makeAdapter({ maxRetries: 1 });
-            let caughtErr = undefined;
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch (err) {
-                if (err instanceof PDRuntimeError)
-                    caughtErr = err;
-            }
-            expect(mockComplete).toHaveBeenCalledTimes(2); // attempt 0 and attempt 1
-            expect(caughtErr).toBeDefined();
-            expect(caughtErr?.category).toBe('timeout');
-            expect(caughtErr?.details?.retryAttempt).toBe(1);
-        });
-        it('7. client_timeout_budget_exhausted is never retried', async () => {
-            mockComplete.mockReset();
-            mockComplete.mockImplementation(async (_model, _ctx, options) => {
-                const signal = options.signal;
-                Object.defineProperty(signal, 'aborted', { value: true, writable: true });
-                throw new DOMException('The operation was aborted', 'AbortError');
-            });
-            const adapter = makeAdapter({ maxRetries: 2 });
-            let caughtErr = undefined;
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch (err) {
-                if (err instanceof PDRuntimeError)
-                    caughtErr = err;
-            }
-            expect(mockComplete).toHaveBeenCalledTimes(1); // Fails immediately, no retry
-            expect(caughtErr).toBeDefined();
-            expect(caughtErr?.category).toBe('timeout');
-            expect(caughtErr?.details?.timeoutClassification).toBe('client_timeout_budget_exhausted');
-        });
-        it('10. circular objects, BigInt, and null-prototype objects in errors do not crash telemetry', async () => {
-            mockComplete.mockReset();
-            mockComplete.mockRejectedValueOnce(new Error('fail'));
-            const nullProto = Object.create(null);
-            nullProto.message = 'null prototype error';
-            const circular = { message: 'circular error' };
-            circular.self = circular;
-            const bigIntErr = { message: 'bigint error', value: 9007199254740991n };
-            const adapter = makeAdapter({ maxRetries: 0 });
-            try {
-                await adapter.startRun(makeStartRunInput({ inputPayload: nullProto }));
-            }
-            catch { /* expected */ }
-            try {
-                await adapter.startRun(makeStartRunInput({ inputPayload: circular }));
-            }
-            catch { /* expected */ }
-            try {
-                await adapter.startRun(makeStartRunInput({ inputPayload: bigIntErr }));
-            }
-            catch { /* expected */ }
-            expect(true).toBe(true);
-        });
-        it('12. telemetry includes provider/model/elapsedMs/timeoutMs/timeoutSource/classification/attempt', async () => {
-            mockComplete.mockReset();
-            mockComplete.mockRejectedValue(new Error('request abort'));
-            const adapter = makeAdapter({ maxRetries: 0 });
-            try {
-                await adapter.startRun(makeStartRunInput({ timeoutMs: 33_000 }));
-            }
-            catch { /* expected */ }
-            const failedEvent = findTelemetryEvent('runtime_invocation_failed');
-            expect(failedEvent).toBeDefined();
-            const payload = failedEvent?.payload;
-            expect(payload.provider).toBe('openrouter');
-            expect(payload.model).toBe('anthropic/claude-sonnet-4');
-            expect(typeof payload.elapsedMs).toBe('number');
-            expect(payload.effectiveTimeoutMs).toBe(33_000);
-            expect(payload.timeoutSource).toBe('runner_input');
-            expect(payload.timeoutClassification).toBe('provider_transient_timeout');
-            expect(payload.retryAttempt).toBe(0);
-        });
-        it('13. caps subsequent retry attempt timeoutMs to remaining overall budget', async () => {
-            mockComplete.mockReset();
-            mockComplete.mockImplementationOnce(async (_model, _ctx, _options) => {
-                await new Promise(resolve => setTimeout(resolve, 40));
-                throw new Error('transient timeout');
-            });
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ maxRetries: 2 });
-            await adapter.startRun(makeStartRunInput({ timeoutMs: 100 }));
-            expect(mockComplete).toHaveBeenCalledTimes(2);
-            const secondCallOpts = mockComplete.mock.calls[1]?.[2];
-            expect(secondCallOpts).toBeDefined();
-            const tc = secondCallOpts.timeoutMs;
-            expect(typeof tc).toBe('number');
-            expect(tc).toBeLessThanOrEqual(70);
-            expect(tc).toBeGreaterThan(0);
-        });
-        it('14. fails fast with client_timeout_budget_exhausted when budget is exhausted', async () => {
-            mockComplete.mockReset();
-            mockComplete.mockImplementation(async () => {
-                await new Promise(resolve => setTimeout(resolve, 50));
-                throw new Error('transient timeout');
-            });
-            const adapter = makeAdapter({ maxRetries: 2 });
-            let caughtErr = undefined;
-            try {
-                await adapter.startRun(makeStartRunInput({ timeoutMs: 30 }));
-            }
-            catch (err) {
-                if (err instanceof PDRuntimeError)
-                    caughtErr = err;
-            }
-            expect(mockComplete).toHaveBeenCalledTimes(1);
-            expect(caughtErr).toBeDefined();
-            expect(caughtErr?.category).toBe('timeout');
-            expect(caughtErr?.details?.timeoutClassification).toBe('client_timeout_budget_exhausted');
-        });
-        it('15. untrusted error classification ignores inherited properties and handles malformed fields safely', async () => {
-            mockComplete.mockReset();
-            const proto = {
-                message: 'timeout - inherited but should be ignored if not own property',
-                name: 'AbortError',
-            };
-            const malformedErr = Object.create(proto);
-            malformedErr.message = 'ordinary own message';
-            malformedErr.name = 'OrdinaryError';
-            mockComplete.mockRejectedValueOnce(malformedErr);
-            const adapter = makeAdapter({ maxRetries: 0 });
-            let caughtErr = undefined;
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch (err) {
-                if (err instanceof PDRuntimeError)
-                    caughtErr = err;
-            }
-            expect(caughtErr).toBeDefined();
-            expect(caughtErr?.category).toBe('execution_failed');
-            expect(caughtErr?.details?.timeoutClassification).toBeUndefined();
-        });
-    });
-    // ── PRI-400: Reasoning-model output handling (BUG-007a/b) ──
-    describe('PRI-400: reasoning-model output handling (BUG-007a/b)', () => {
-        /** Create an AssistantMessage with empty text content but ThinkingContent with valid JSON. */
-        function makeThinkingOnlyResponse(thinkingText, overrides = {}) {
-            return {
-                content: [
-                    { type: 'text', text: '' },
-                    { type: 'thinking', thinking: thinkingText },
-                ],
-                role: 'assistant',
-                stopReason: 'length',
-                api: 'openai-completions',
-                provider: 'lmstudio',
-                model: 'qwen3.6-27b-mtp',
-                usage: { input: 10, output: 5, cacheRead: 0, cacheWrite: 0, totalTokens: 15, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } },
-                timestamp: Date.now(),
-                ...overrides,
-            };
-        }
-        /** Create an AssistantMessage with empty content and no thinking. */
-        function makeEmptyContentResponse(overrides = {}) {
-            return {
-                content: [{ type: 'text', text: '' }],
-                role: 'assistant',
-                stopReason: 'length',
-                api: 'openai-completions',
-                provider: 'lmstudio',
-                model: 'qwen3.6-27b-mtp',
-                usage: { input: 10, output: 0, cacheRead: 0, cacheWrite: 0, totalTokens: 10, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } },
-                timestamp: Date.now(),
-                ...overrides,
-            };
-        }
-        it('(a) content="" + ThinkingContent with valid JSON + stopReason="length" → extracts JSON from thinking', async () => {
-            // Simulates the exact dogfood scenario: reasoning model spends all tokens on thinking
-            const thinkingWithJson = `Let me analyze this step by step...\n\n${JSON.stringify(VALID_DIAGNOSIS)}`;
-            mockComplete.mockResolvedValueOnce(makeThinkingOnlyResponse(thinkingWithJson));
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only' });
-            const handle = await adapter.startRun(makeStartRunInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-        });
-        it('(b) content="" + no ThinkingContent + stopReason="length" → fail-loud with truncation error', async () => {
-            mockComplete.mockResolvedValueOnce(makeEmptyContentResponse());
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only' });
-            let caughtErr;
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch (err) {
-                if (err instanceof PDRuntimeError)
-                    caughtErr = err;
-            }
-            expect(caughtErr).toBeInstanceOf(PDRuntimeError);
-            expect(caughtErr?.category).toBe('output_invalid');
-            expect(caughtErr?.message).toContain('truncated');
-            expect(caughtErr?.details?.nextAction).toBeDefined();
-        });
-        it('content="" + no ThinkingContent + stopReason="stop" → fail-loud with "No text or reasoning"', async () => {
-            mockComplete.mockResolvedValueOnce(makeEmptyContentResponse({ stopReason: 'stop' }));
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only' });
-            let caughtErr;
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch (err) {
-                if (err instanceof PDRuntimeError)
-                    caughtErr = err;
-            }
-            expect(caughtErr).toBeInstanceOf(PDRuntimeError);
-            expect(caughtErr?.category).toBe('output_invalid');
-            expect(caughtErr?.message).toContain('No text or reasoning content');
-            expect(caughtErr?.details?.nextAction).toBeDefined();
-        });
-        it('(c) content="normal JSON" + no ThinkingContent → existing path unchanged (regression)', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only' });
-            const handle = await adapter.startRun(makeStartRunInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-        });
-        it('content="" + ThinkingContent with no extractable JSON → output_invalid', async () => {
-            mockComplete.mockResolvedValueOnce(makeThinkingOnlyResponse('Just thinking prose, no JSON here'));
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only' });
-            let caughtErr;
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch (err) {
-                if (err instanceof PDRuntimeError)
-                    caughtErr = err;
-            }
-            expect(caughtErr).toBeInstanceOf(PDRuntimeError);
-            expect(caughtErr?.category).toBe('output_invalid');
-        });
-        it('stopReason="length" with empty content and empty thinking → fail-loud with nextAction', async () => {
-            mockComplete.mockResolvedValueOnce(makeEmptyContentResponse());
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only' });
-            let caughtErr;
-            try {
-                await adapter.startRun(makeStartRunInput());
-            }
-            catch (err) {
-                if (err instanceof PDRuntimeError)
-                    caughtErr = err;
-            }
-            expect(caughtErr).toBeInstanceOf(PDRuntimeError);
-            expect(caughtErr?.category).toBe('output_invalid');
-            expect(caughtErr?.message).toContain('finish_reason=length');
-            expect(caughtErr?.details?.nextAction).toContain('maxTokens');
-        });
-        it('stopReason="length" with valid thinking content → succeeds despite truncation', async () => {
-            const thinkingWithJson = JSON.stringify(VALID_DIAGNOSIS);
-            mockComplete.mockResolvedValueOnce(makeThinkingOnlyResponse(thinkingWithJson));
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only' });
-            const handle = await adapter.startRun(makeStartRunInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-        });
-        it('completeWithRetry omits maxTokens by default (PRI-621: pi-ai native model ceiling)', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only' });
-            await adapter.startRun(makeStartRunInput());
-            const [, , options] = mockComplete.mock.calls[0];
-            expect(options.maxTokens).toBeUndefined();
-        });
-        it('completeWithRetry uses config.maxTokens override when provided', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only', maxTokens: 8192 });
-            await adapter.startRun(makeStartRunInput());
-            const [, , options] = mockComplete.mock.calls[0];
-            expect(options.maxTokens).toBe(8192);
-        });
-        it('completeWithRetry omits maxTokens for deepseek too — provider-name heuristic retired (PRI-621)', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only', provider: 'deepseek', model: 'deepseek-v4-flash', baseUrl: 'https://api.deepseek.com' });
-            await adapter.startRun(makeStartRunInput());
-            const [, , options] = mockComplete.mock.calls[0];
-            expect(options.maxTokens).toBeUndefined();
-        });
-        it('PRI-621: custom provider relaying a catalog-known model keeps catalog metadata (catalog-first resolveModel)', async () => {
-            // Bai relay serving deepseek-v4-flash: the catalog entry (reasoning
-            // model, 384K output ceiling) must survive with only transport fields
-            // overridden — the old hardcoded fallback misclassified it as
-            // non-reasoning with a 32K ceiling and fed the RC0 budget bug.
-            mockGetModel.mockImplementation((_provider, id) => id === 'deepseek-v4-flash'
-                ? {
-                    id: 'deepseek-v4-flash',
-                    name: 'DeepSeek V4 Flash',
-                    api: 'openai-completions',
-                    provider: 'deepseek',
-                    baseUrl: 'https://api.deepseek.com',
-                    reasoning: true,
-                    input: ['text'],
-                    cost: { input: 0.14, output: 0.28, cacheRead: 0, cacheWrite: 0 },
-                    contextWindow: 1_000_000,
-                    maxTokens: 384_000,
-                }
-                : undefined);
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ provider: 'Bai', model: 'deepseek-v4-flash', baseUrl: 'https://api.b.ai/v1', outputPathStrategy: 'free_form_only' });
-            await adapter.startRun(makeStartRunInput());
-            const [model] = mockComplete.mock.calls[0];
-            expect(model.provider).toBe('Bai');
-            expect(model.baseUrl).toBe('https://api.b.ai/v1');
-            expect(model.reasoning).toBe(true);
-            expect(model.maxTokens).toBe(384_000);
-        });
-        it('PRI-621: custom provider with unknown model id falls back to conservative hardcoded metadata', async () => {
-            mockGetModel.mockReturnValue(undefined);
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ provider: 'Bai', model: 'totally-unknown-model', baseUrl: 'https://api.b.ai/v1', outputPathStrategy: 'free_form_only' });
-            await adapter.startRun(makeStartRunInput());
-            const [model] = mockComplete.mock.calls[0];
-            expect(model.reasoning).toBe(false);
-            expect(model.maxTokens).toBe(32_000);
-        });
-        it('PRI-621: built-in provider with unknown model id fails loud (runtime_unavailable)', async () => {
-            mockGetModel.mockReturnValue(undefined);
-            const adapter = makeAdapter({ provider: 'openai', model: 'gpt-not-a-model' });
-            await expect(adapter.startRun(makeStartRunInput())).rejects.toMatchObject({ category: 'runtime_unavailable' });
-        });
-        it('PRI-621 RC3: schema-aware extraction picks the full object over an earlier inner fragment', async () => {
-            // Recurrence of 2026-08-28: the answer contained a small lineage
-            // fragment BEFORE the complete output; first-object extraction kept
-            // validating the fragment ("all required properties undefined") and
-            // the repair loop fixed the wrong object.
-            const response = `Here is the trace reference {"diagnosisId":"diag-fragment-1"} and the complete output below:\n${JSON.stringify(VALID_DIAGNOSIS)}`;
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(response));
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only' });
-            const handle = await adapter.startRun(makeStartRunInput({ outputSchemaRef: 'diagnostician-output-v1' }));
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-        });
-        it('mixed content: text + thinking → prefers text content (regression)', async () => {
-            // When both text and thinking are present, text should be preferred
-            const mixedResponse = {
-                content: [
-                    { type: 'text', text: JSON.stringify(VALID_DIAGNOSIS) },
-                    { type: 'thinking', thinking: 'Some reasoning here' },
-                ],
-                role: 'assistant',
-                stopReason: 'stop',
-                api: 'openai-completions',
-                provider: 'openrouter',
-                model: 'anthropic/claude-sonnet-4',
-                usage: { input: 10, output: 5, cacheRead: 0, cacheWrite: 0, totalTokens: 15, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } },
-                timestamp: Date.now(),
-            };
-            mockComplete.mockResolvedValueOnce(mixedResponse);
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only' });
-            const handle = await adapter.startRun(makeStartRunInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-        });
-    });
-    // ── PRI-405: Reasoning-model E2E stub tests ───────────────────────────────
-    describe('PRI-405: reasoning-model E2E stub tests', () => {
-        /** Valid DiagRootCauseOutputV1 for split pipeline Stage A. */
-        const VALID_ROOT_CAUSE_OUTPUT = {
-            valid: true,
-            diagnosisId: 'diag-pri405-1',
-            taskId: 'diag_rootcause-diagnosis_pain-pri405',
-            summary: 'AI助手在修改代码前未阅读错误手册，导致重复了类型断言绕过校验问题',
-            causalChain: [
-                {
-                    why: 1,
-                    statement: 'AI助手直接使用as类型断言绕过运行时校验',
-                    evidenceRefs: ['owner_reported:cli'],
-                },
-                {
-                    why: 2,
-                    statement: '修改代码前未阅读AGENTS.md中的错误手册和ERR-001相关条目',
-                    evidenceRefs: ['owner_reported:cli'],
-                },
-                {
-                    why: 3,
-                    statement: '缺乏强制性的代码修改前必读规范检查机制',
-                    evidenceRefs: ['owner_reported:cli'],
-                },
-            ],
-            rootCause: 'Design: 缺乏强制性的代码修改前必读规范检查机制，导致AI助手绕过关键校验步骤',
-            rootCauseCategory: 'Design',
-            evidence: [
-                {
-                    sourceRef: 'owner_reported:cli',
-                    note: 'AI助手在修改代码时使用了as类型断言绕过运行时校验，违反了AGENTS.md中ERR-001的规定',
-                },
-            ],
-            confidence: 0.85,
-            ambiguityNotes: [
-                '关联核心公理: T-01 (先理解结构再修改)',
-            ],
-        };
-        /** Create a thinking-only response (no text block) simulating reasoning model output. */
-        function makeThinkingOnlyResponse(thinkingText, overrides = {}) {
-            return {
-                content: [
-                    { type: 'thinking', thinking: thinkingText },
-                ],
-                role: 'assistant',
-                stopReason: 'length',
-                api: 'openai-completions',
-                provider: 'lmstudio',
-                model: 'qwen3.6-27b-mtp',
-                usage: { input: 10, output: 5, cacheRead: 0, cacheWrite: 0, totalTokens: 15, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } },
-                timestamp: Date.now(),
-                ...overrides,
-            };
-        }
-        // ── Scenario A: thinking fallback + schema validation passes ──
-        it('Scenario A: thinking-only response with valid JSON passes diag-rootcause-output-v1 schema validation', async () => {
-            // Simulates reasoning model that puts all structured output in thinking block
-            const thinkingWithJson = `让我逐步分析这个pain信号...\n\n${JSON.stringify(VALID_ROOT_CAUSE_OUTPUT)}`;
-            mockComplete.mockResolvedValueOnce(makeThinkingOnlyResponse(thinkingWithJson));
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only' });
-            const handle = await adapter.startRun(makeStartRunInput({
-                outputSchemaRef: 'diag-rootcause-output-v1',
-                taskRef: { taskId: 'diag_rootcause-diagnosis_pain-pri405' },
-            }));
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output).not.toBeNull();
-            // Schema validation passed — output was extracted from thinking and validated
-            expect(output?.payload).toMatchObject({
-                valid: true,
-                diagnosisId: 'diag-pri405-1',
-                rootCauseCategory: 'Design',
-            });
-            // taskId is a lineage field — stripped by adapter (PRI-272)
-            const payload = output?.payload;
-            expect(payload.taskId).toBeUndefined();
-        });
-        // ── Scenario B: thinking with prose (no JSON) + truncated → fail-loud ──
-        it('Scenario B: thinking with prose but no JSON + stopReason=length → fail-loud output_invalid', async () => {
-            // Simulates reasoning model that spent all tokens on thinking prose, no JSON output.
-            // extractAssistantTextOrThrow extracts the thinking text (it has content),
-            // but extractJsonObject finds no JSON → output_invalid.
-            // This is the correct fail-loud path: the adapter does NOT silently succeed.
-            const thinkingProseOnly = 'Here is a thinking process: 1. analyze the pain signal 2. identify root cause 3. classify the category...';
-            mockComplete.mockResolvedValueOnce(makeThinkingOnlyResponse(thinkingProseOnly));
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only' });
-            let caughtErr;
-            try {
-                await adapter.startRun(makeStartRunInput({
-                    outputSchemaRef: 'diag-rootcause-output-v1',
-                    taskRef: { taskId: 'diag_rootcause-diagnosis_pain-pri405' },
-                }));
-            }
-            catch (err) {
-                if (err instanceof PDRuntimeError)
-                    caughtErr = err;
-            }
-            // EP-03: fail-loud with structured reason — adapter does NOT silently succeed
-            expect(caughtErr).toBeInstanceOf(PDRuntimeError);
-            expect(caughtErr?.category).toBe('output_invalid');
-            // The error must indicate why: either no JSON found, truncated, or finish_reason=length
-            const reason = caughtErr?.message ?? '';
-            expect(reason.includes('No valid JSON') || reason.includes('truncated') || reason.includes('finish_reason=length'), `Expected reason to indicate extraction failure, got: ${reason}`).toBe(true);
-        });
-        it('Scenario B (variant): empty thinking + stopReason=length → fail-loud with truncated + nextAction containing maxTokens', async () => {
-            // When thinking is empty and stopReason=length, extractAssistantTextOrThrow
-            // throws the specific truncated error with nextAction guidance.
-            mockComplete.mockResolvedValueOnce(makeThinkingOnlyResponse('', { stopReason: 'length' }));
-            const adapter = makeAdapter({ outputPathStrategy: 'free_form_only' });
-            let caughtErr;
-            try {
-                await adapter.startRun(makeStartRunInput({
-                    outputSchemaRef: 'diag-rootcause-output-v1',
-                    taskRef: { taskId: 'diag_rootcause-diagnosis_pain-pri405' },
-                }));
-            }
-            catch (err) {
-                if (err instanceof PDRuntimeError)
-                    caughtErr = err;
-            }
-            // EP-03: fail-loud with structured reason and nextAction
-            expect(caughtErr).toBeInstanceOf(PDRuntimeError);
-            expect(caughtErr?.category).toBe('output_invalid');
-            expect(caughtErr?.message).toContain('truncated');
-            expect(caughtErr?.details?.nextAction).toBeDefined();
-            const nextAction = String(caughtErr?.details?.nextAction ?? '');
-            expect(nextAction.includes('maxTokens')).toBe(true);
-        });
-    });
-    // ── PRI-284: tool_call_first / json_mode_first path tests ──
-    describe('output path strategy: tool_call_first (PRI-284)', () => {
-        function makeToolCallResponse(args, toolName = 'record_diagnostician_output_v1') {
-            return {
-                content: [
-                    { type: 'toolCall', id: 'call-1', name: toolName, arguments: args },
-                ],
-                role: 'assistant',
-                stopReason: 'toolUse',
-                api: 'openai-completions',
-                provider: 'openrouter',
-                model: 'anthropic/claude-sonnet-4',
-                usage: { input: 10, output: 5, cacheRead: 0, cacheWrite: 0, totalTokens: 15, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } },
-                timestamp: Date.now(),
-            };
-        }
-        function makeToolCallInput(overrides = {}) {
-            return {
-                agentSpec: { agentId: 'diagnostician', schemaVersion: 'v1' },
-                inputPayload: 'Diagnose this pain signal',
-                contextItems: [],
-                timeoutMs: 60_000,
-                outputSchemaRef: 'diagnostician-output-v1',
-                ...overrides,
-            };
-        }
-        it('tool_call success: provider returns toolUse with valid schema → output extracted from tool args', async () => {
-            mockComplete.mockResolvedValueOnce(makeToolCallResponse(VALID_DIAGNOSIS));
-            const adapter = makeAdapter({ outputPathStrategy: 'tool_call_first' });
-            const handle = await adapter.startRun(makeToolCallInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1', valid: true });
-            // Verify telemetry includes correct outputPath
-            const pathEvent = findTelemetryEvent('output_path_chosen');
-            expect(pathEvent?.payload).toMatchObject({
-                outputPath: 'tool_call',
-                outputSchemaRef: 'diagnostician-output-v1',
-            });
-        });
-        it('tool_call strips lineage fields from tool args (ERR-008)', async () => {
-            const outputWithLineage = { ...VALID_DIAGNOSIS, sourcePainId: 'pain-injected', sourceTaskId: 'task-injected' };
-            mockComplete.mockResolvedValueOnce(makeToolCallResponse(outputWithLineage));
-            const adapter = makeAdapter({ outputPathStrategy: 'tool_call_first' });
-            const handle = await adapter.startRun(makeToolCallInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).not.toHaveProperty('sourcePainId');
-            expect(output?.payload).not.toHaveProperty('sourceTaskId');
-            // Non-lineage fields preserved
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-        });
-        it('tool_call schema invalid → falls through to json_mode → free_form', async () => {
-            // Path 1: tool call with invalid schema args
-            const invalidToolArgs = { invalid: 'data' };
-            mockComplete.mockResolvedValueOnce(makeToolCallResponse(invalidToolArgs));
-            // Path 2: json_mode also fails (non-JSON text)
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage('not json at all'));
-            // Path 3: free_form succeeds
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ outputPathStrategy: 'tool_call_first' });
-            const handle = await adapter.startRun(makeToolCallInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-            // Verify fallback telemetry was emitted for tool_call path.
-            // Runtime Contract Rule 2: validate shape before reading eventType.
-            const fallbackEvents = mockEmitTelemetry.mock.calls.filter((c) => {
-                const [payload] = c;
-                return (typeof payload === 'object' &&
-                    payload !== null &&
-                    Object.hasOwn(payload, 'eventType') &&
-                    payload.eventType === 'output_path_fallback');
-            });
-            expect(fallbackEvents.length).toBeGreaterThanOrEqual(1);
-        });
-        it('tool_call not supported (stopReason != toolUse) → falls to json_mode', async () => {
-            // Path 1 (tool_call): provider returns non-toolUse response
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage('not json'));
-            // Path 2 (json_mode): provider returns valid JSON
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ outputPathStrategy: 'tool_call_first' });
-            const handle = await adapter.startRun(makeToolCallInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-        });
-        it('tool_call uses per-runner schema from registry, not hardcoded (PRI-284)', async () => {
-            mockComplete.mockResolvedValueOnce(makeToolCallResponse(VALID_DIAGNOSIS));
-            const adapter = makeAdapter({ outputPathStrategy: 'tool_call_first' });
-            await adapter.startRun(makeToolCallInput());
-            // Verify the context passed to completeSimple includes a tool
-            // whose name is derived from the schemaRef, not hardcoded.
-            // Runtime Contract Rule 2 (no `as` bypass): use a type guard on the
-            // unknown mock-call shape instead of asserting the type.
-            const [firstCallArgs] = mockComplete.mock.calls;
-            expect(firstCallArgs).toBeDefined();
-            const rawContext = firstCallArgs?.[1];
-            expect(isToolContext(rawContext)).toBe(true);
-            if (!isToolContext(rawContext))
-                throw new Error('invalid tool context');
-            expect(rawContext.tools.length).toBe(1);
-            const [tool] = rawContext.tools;
-            if (!tool)
-                throw new Error('expected exactly one tool');
-            // Name derived from schemaRef 'diagnostician-output-v1'
-            expect(tool.name).toBe('record_diagnostician_output_v1');
-            // Not the old hardcoded name
-            expect(tool.name).not.toBe('record_diagnosis_v1');
-            // Parameters should be the schema from registry
-            expect(tool.parameters).toBeDefined();
-        });
-    });
-    describe('output path strategy: json_mode_first (PRI-284)', () => {
-        function makeJsonModeInput(overrides = {}) {
-            return {
-                agentSpec: { agentId: 'diagnostician', schemaVersion: 'v1' },
-                inputPayload: 'Diagnose this pain signal',
-                contextItems: [],
-                timeoutMs: 60_000,
-                outputSchemaRef: 'diagnostician-output-v1',
-                ...overrides,
-            };
-        }
-        it('json_mode success: provider returns valid JSON text → output extracted', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ outputPathStrategy: 'json_mode_first' });
-            const handle = await adapter.startRun(makeJsonModeInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-            const pathEvent = findTelemetryEvent('output_path_chosen');
-            expect(pathEvent?.payload).toMatchObject({
-                outputPath: 'json_object_mode',
-            });
-        });
-        it('json_mode strips lineage fields (ERR-008)', async () => {
-            const outputWithLineage = { ...VALID_DIAGNOSIS, sourcePainId: 'pain-llm-injected' };
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(outputWithLineage)));
-            const adapter = makeAdapter({ outputPathStrategy: 'json_mode_first' });
-            const handle = await adapter.startRun(makeJsonModeInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).not.toHaveProperty('sourcePainId');
-        });
-        it('json_mode parse fails → falls to free_form', async () => {
-            // Path 2: json_mode returns non-parseable text
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage('this is not json at all'));
-            // Path 3: free_form succeeds
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ outputPathStrategy: 'json_mode_first' });
-            const handle = await adapter.startRun(makeJsonModeInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-            // Verify fallback telemetry.
-            // Runtime Contract Rule 2: validate shape before reading eventType.
-            const fallbackEvents = mockEmitTelemetry.mock.calls.filter((c) => {
-                const [payload] = c;
-                return (typeof payload === 'object' &&
-                    payload !== null &&
-                    Object.hasOwn(payload, 'eventType') &&
-                    payload.eventType === 'output_path_fallback');
-            });
-            expect(fallbackEvents.length).toBeGreaterThanOrEqual(1);
-        });
-        it('json_mode schema invalid → falls to free_form', async () => {
-            // Path 2: json_mode returns valid JSON but wrong schema
-            const wrongSchema = { wrongField: 'wrong value' };
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(wrongSchema)));
-            // Path 3: free_form succeeds with correct schema
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage(JSON.stringify(VALID_DIAGNOSIS)));
-            const adapter = makeAdapter({ outputPathStrategy: 'json_mode_first' });
-            const handle = await adapter.startRun(makeJsonModeInput());
-            const output = await adapter.fetchOutput(handle.runId);
-            expect(output?.payload).toMatchObject({ diagnosisId: 'diag-test-1' });
-        });
-    });
-    // ── PRI-707: finish metadata preservation + truncation classification ──
-    //
-    // Before PRI-707 the provider's stopReason/usage were dropped whenever any
-    // text was extractable, so a token-limit cut was indistinguishable from a
-    // true format error ("No valid JSON found in LLM response"). These tests
-    // pin the preserved evidence surfaces: error details (→ durable
-    // diagnosticJson via the runner's classifyError path), telemetry payloads,
-    // and the repair-prompt truncation notice.
-    describe('PRI-707: finish metadata preservation and truncation classification', () => {
-        function makeSchemaInput(overrides = {}) {
-            return {
-                agentSpec: { agentId: 'diagnostician', schemaVersion: 'v1' },
-                inputPayload: 'Diagnose this pain signal',
-                contextItems: [],
-                timeoutMs: 60_000,
-                outputSchemaRef: 'diagnostician-output-v1',
-                ...overrides,
-            };
-        }
-        /** Type-guard the repair prompt out of a completeSimple mock call (rc-2). */
-        function repairPromptFromCall(call) {
-            if (!Array.isArray(call) || call.length < 2)
-                return undefined;
-            const [context] = call.slice(1, 2);
-            if (typeof context !== 'object' || context === null || !Object.hasOwn(context, 'messages'))
-                return undefined;
-            const { messages } = context;
-            if (!Array.isArray(messages) || messages.length === 0)
-                return undefined;
-            const [first] = messages;
-            if (typeof first !== 'object' || first === null || !Object.hasOwn(first, 'content'))
-                return undefined;
-            const { content } = first;
-            return typeof content === 'string' ? content : undefined;
-        }
-        function usageWith(output) {
-            return { input: 120, output, cacheRead: 0, cacheWrite: 0, totalTokens: 120 + output, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } };
-        }
-        it('T3+T4: finish_reason=length + usage → extraction failure carries truncated evidence (details + telemetry)', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage('{"diagnosisId":"diag-trunc', {
-                stopReason: 'length',
-                usage: usageWith(64),
-            }));
-            const adapter = makeAdapter(); // free_form_only → Path 3 directly
-            let caught;
-            try {
-                await adapter.startRun(makeSchemaInput());
-            }
-            catch (err) {
-                caught = err instanceof PDRuntimeError ? err : undefined;
-            }
-            expect(caught?.category).toBe('output_invalid');
-            expect(caught?.message).toContain('finish_reason=length');
-            // The runner persists err.details into diagnosticJson.output_failure_details
-            // (PRI-559 P0-2) — so these fields become durable failure evidence.
-            expect(caught?.details).toMatchObject({
-                truncated: true,
-                stopReason: 'length',
-                outputTokens: 64,
-            });
-            expect(typeof caught?.details?.nextAction).toBe('string');
-            const extractionEvent = findTelemetryEvent('output_extraction_failed');
-            expect(extractionEvent?.payload).toMatchObject({
-                stopReason: 'length',
-                outputTokens: 64,
-                truncated: true,
-            });
-        });
-        it('T5: finish_reason=stop + no extractable JSON → conservative legacy classification, no truncation claim', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage('I cannot answer that.'));
-            const adapter = makeAdapter();
-            let caught;
-            try {
-                await adapter.startRun(makeSchemaInput());
-            }
-            catch (err) {
-                caught = err instanceof PDRuntimeError ? err : undefined;
-            }
-            expect(caught?.category).toBe('output_invalid');
-            expect(caught?.message).toBe('[output_invalid] No valid JSON found in LLM response');
-            expect(caught?.details).toBeUndefined();
-            const extractionEvent = findTelemetryEvent('output_extraction_failed');
-            const payload = extractionEvent?.payload;
-            expect(payload?.truncated).toBeUndefined();
-        });
-        it('T2: valid JSON + schema violation with finish_reason=stop → schema-invalid, never claimed as truncation', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage('{"unexpected":true}'));
-            // Repair attempts fail (unparseable) → exhausted.
-            mockComplete.mockResolvedValue(makeAssistantMessage('still not json'));
-            const adapter = makeAdapter();
-            let caught;
-            try {
-                await adapter.startRun(makeSchemaInput());
-            }
-            catch (err) {
-                caught = err instanceof PDRuntimeError ? err : undefined;
-            }
-            expect(caught?.category).toBe('output_invalid');
-            expect(caught?.message).toContain('does not match');
-            const evidencePack = caught?.details?.evidencePack;
-            expect(evidencePack).toBeDefined();
-            // The PRI-621 RC3 heuristic fires (object matches no required key)…
-            expect(evidencePack?.truncationSuspected).toBe(true);
-            // …but the definitive finish-metadata signal must NOT claim truncation.
-            expect(evidencePack?.truncated).toBeUndefined();
-            expect(evidencePack?.stopReason).toBe('stop');
-        });
-        it('T8: finish_reason=length + schema violation → repair prompt carries truncation notice and evidencePack marks truncated', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage('{"unexpected":true}', { stopReason: 'length' }));
-            mockComplete.mockResolvedValue(makeAssistantMessage('still not json'));
-            const adapter = makeAdapter();
-            let caught;
-            try {
-                await adapter.startRun(makeSchemaInput());
-            }
-            catch (err) {
-                caught = err instanceof PDRuntimeError ? err : undefined;
-            }
-            const evidencePack = caught?.details?.evidencePack;
-            expect(evidencePack?.truncated).toBe(true);
-            expect(evidencePack?.stopReason).toBe('length');
-            // Terminal telemetry must carry the same truncation evidence so
-            // monitoring can distinguish a token-limit cut from an ordinary
-            // schema failure (PR #1574 review finding). Default helper usage
-            // reports output: 5.
-            const exhaustedEvent = findTelemetryEvent('output_repair_exhausted');
-            expect(exhaustedEvent?.payload).toMatchObject({
-                stopReason: 'length',
-                truncated: true,
-                outputTokens: 5,
-            });
-            // The FIRST completeSimple call is the original generation; repair
-            // calls follow. Their user message must carry the truncation notice so
-            // the repair LLM stops treating a cut fragment as a misunderstanding.
-            const repairCalls = mockComplete.mock.calls.slice(1);
-            expect(repairCalls.length).toBeGreaterThan(0);
-            const firstRepairPrompt = repairPromptFromCall(repairCalls[0]);
-            expect(firstRepairPrompt).toContain('TRUNCATION NOTICE');
-            expect(firstRepairPrompt).toContain('finish_reason=length');
-        });
-        it('truncation notice absent when finish metadata shows a clean stop (repair prompt unchanged)', async () => {
-            mockComplete.mockResolvedValueOnce(makeAssistantMessage('{"unexpected":true}'));
-            mockComplete.mockResolvedValue(makeAssistantMessage('still not json'));
-            const adapter = makeAdapter();
-            try {
-                await adapter.startRun(makeSchemaInput());
-            }
-            catch {
-                // output_invalid expected — assertion is on the prompt below.
-            }
-            const repairCalls = mockComplete.mock.calls.slice(1);
-            expect(repairCalls.length).toBeGreaterThan(0);
-            const firstRepairPrompt = repairPromptFromCall(repairCalls[0]);
-            expect(firstRepairPrompt).toBeDefined();
-            expect(firstRepairPrompt).not.toContain('TRUNCATION NOTICE');
-        });
-    });
-});
-//# sourceMappingURL=pi-ai-runtime-adapter.test.js.map