wip: desparate models fucking around

This commit is contained in:
2026-02-26 00:13:52 +01:00
parent 021cddefa3
commit 2a923c7e48
16 changed files with 684 additions and 113 deletions

View File

@@ -97,6 +97,12 @@ describe('OpenCodeManager protocol integration', () => {
const telemetryAfter = getProtocolTelemetryService().getSnapshot();
expect(telemetryAfter.totalTurns).toBe(telemetryBefore.totalTurns + 1);
expect(telemetryAfter.validEnvelopeTurns).toBe(telemetryBefore.validEnvelopeTurns + 1);
expect(chatEngineMock.addMessage).toHaveBeenCalledWith(expect.objectContaining({
conversationId: 'conversation-1',
role: 'assistant',
content: 'Please provide a date range.',
}));
});
it('blocks unsupported actions and records blocked-action telemetry', async () => {
@@ -149,4 +155,71 @@ describe('OpenCodeManager protocol integration', () => {
const telemetryAfter = getProtocolTelemetryService().getSnapshot();
expect(telemetryAfter.blockedActionCount).toBe(telemetryBefore.blockedActionCount + 1);
});
it('retries once with protocol repair prompt when first output is non-canonical', async () => {
const conversation: MockConversation = {
id: 'conversation-3',
model: 'gpt-5',
messages: [{ role: 'user', content: 'show chart' }],
};
const chatEngineMock = createChatEngineMock(conversation);
const manager = new OpenCodeManager(
chatEngineMock as never,
{} as never,
{} as never,
() => null,
);
manager.setApiKey('test-api-key');
const sendSpy = vi.spyOn(manager as never, 'sendOpenAIMessage')
.mockResolvedValueOnce({
content: JSON.stringify({
title: 'Legacy JSON',
widgets: [{ type: 'chart', chartType: 'bar' }],
}),
toolCalls: [],
})
.mockResolvedValueOnce({
content: JSON.stringify({
protocolVersion: '2.0',
assistantText: 'Here is your chart.',
ui: {
specVersion: '1',
elements: [
{
type: 'chart',
chartType: 'bar',
series: [{ label: '2015', value: 86 }],
},
],
},
intent: 'summarize',
needsInput: { required: false, fields: [] },
actions: [],
confidence: 0.8,
traceId: 'trace-retry-success',
}),
toolCalls: [],
});
const result = await manager.sendMessage('conversation-3', 'Build chart', {
metadata: { surface: 'tab' },
});
expect(result.success).toBe(true);
expect(result.envelope?.traceId).toBe('trace-retry-success');
expect(sendSpy).toHaveBeenCalledTimes(2);
const retryMessages = sendSpy.mock.calls[1]?.[2] as Array<{ role: string; content?: string }>;
const lastMessage = retryMessages[retryMessages.length - 1]?.content ?? '';
expect(lastMessage).toContain('failed protocol validation');
expect(lastMessage).toContain('Return ONLY one valid protocol envelope JSON object');
expect(chatEngineMock.addMessage).toHaveBeenCalledWith(expect.objectContaining({
conversationId: 'conversation-3',
role: 'assistant',
content: 'Here is your chart.',
}));
});
});

View File

@@ -171,4 +171,92 @@ describe('ProtocolResponseBuilder', () => {
requiresConfirmation: false,
}));
});
it('drops invalid ui payloads from canonical envelopes before renderer consumption', () => {
const builder = new ProtocolResponseBuilder();
const raw = JSON.stringify({
protocolVersion: '2.0',
assistantText: 'Here is the result',
intent: 'summarize',
needsInput: { required: false, fields: [] },
actions: [],
ui: {
specVersion: '1',
elements: [
{
type: 'chart',
chartType: 'bar',
},
],
},
confidence: 0.7,
traceId: 'trace-invalid-ui',
});
const result = builder.build({
rawAssistantOutput: raw,
surface: 'tab',
capabilities: {
widgets: ['chart'],
actions: ['openPost'],
tools: ['search_posts'],
},
});
expect(result.envelope.ui).toBeUndefined();
expect(result.warnings.some((warning) => warning.includes('Invalid ui payload'))).toBe(true);
});
it('normalizes non-canonical ui element fields inside canonical envelopes', () => {
const builder = new ProtocolResponseBuilder();
const raw = JSON.stringify({
protocolVersion: '2.0',
assistantText: 'Distribution chart ready.',
ui: {
specVersion: '1',
elements: [
{
type: 'chart',
chartType: 'bar',
data: {
labels: ['aside', 'article'],
datasets: [{ data: [181, 53] }],
},
},
{
type: 'text',
content: 'Category breakdown',
},
],
},
intent: 'summarize',
needsInput: { required: false, fields: [] },
actions: [],
confidence: 0.95,
traceId: 'trace-normalize-ui',
});
const result = builder.build({
rawAssistantOutput: raw,
surface: 'tab',
capabilities: {
widgets: ['chart', 'text'],
actions: ['openPost'],
tools: ['search_posts'],
},
});
const elements = result.envelope.ui?.elements as Array<{ type: string; series?: Array<{ label: string; value: number }>; text?: string }>;
expect(elements).toHaveLength(2);
expect(elements[0]?.type).toBe('chart');
expect(elements[0]?.series).toEqual([
{ label: 'aside', value: 181 },
{ label: 'article', value: 53 },
]);
expect(elements[1]).toEqual({ type: 'text', text: 'Category breakdown' });
expect(result.warnings.some((warning) => warning.includes('Normalized non-canonical ui payload'))).toBe(true);
});
});

View File

@@ -192,4 +192,49 @@ describe('assistantPanelSpec', () => {
expect(result).not.toBeNull();
expect(result?.elements).toHaveLength(7);
});
it('parses canonical protocol envelope JSON and extracts assistant text plus ui spec', () => {
const raw = JSON.stringify({
protocolVersion: '2.0',
assistantText: 'Here is your chart.',
ui: {
specVersion: '1',
elements: [
{
type: 'chart',
chartType: 'bar',
data: {
labels: ['aside', 'article'],
datasets: [{ data: [181, 53] }],
},
},
{
type: 'text',
content: 'Breakdown details',
},
],
},
intent: 'summarize',
needsInput: { required: false, fields: [] },
actions: [],
confidence: 0.9,
traceId: 'trace-1',
});
const result = extractAssistantResponseContent(raw);
expect(result.displayText).toBe('Here is your chart.');
expect(result.panelSpec).not.toBeNull();
expect(result.panelSpec?.elements[0]).toMatchObject({
type: 'chart',
series: [
{ label: 'aside', value: 181 },
{ label: 'article', value: 53 },
],
});
expect(result.panelSpec?.elements[1]).toEqual({
type: 'text',
text: 'Breakdown details',
});
});
});

View File

@@ -0,0 +1,56 @@
import { describe, expect, it } from 'vitest';
import { buildActionPoliciesFromEnvelope } from '../../../src/renderer/navigation/protocolActionPolicies';
describe('buildActionPoliciesFromEnvelope', () => {
it('preserves server-provided action policies', () => {
const result = buildActionPoliciesFromEnvelope({
actions: [
{
id: 'a1',
action: 'openSettings',
policy: 'confirm',
requiresConfirmation: true,
},
],
needsInput: {
required: false,
fields: [],
},
});
expect(result).toEqual({
openSettings: 'confirm',
});
});
it('adds confirm policy for submitNeedsInput when clarification is required', () => {
const result = buildActionPoliciesFromEnvelope({
actions: [],
needsInput: {
required: true,
fields: [{ key: 'date', label: 'Date', inputType: 'date' }],
},
});
expect(result.submitNeedsInput).toBe('confirm');
});
it('does not override explicit server policy for submitNeedsInput', () => {
const result = buildActionPoliciesFromEnvelope({
actions: [
{
id: 'a1',
action: 'submitNeedsInput',
policy: 'danger',
requiresConfirmation: true,
},
],
needsInput: {
required: true,
fields: [{ key: 'title', label: 'Title', inputType: 'text' }],
},
});
expect(result.submitNeedsInput).toBe('danger');
});
});

View File

@@ -44,9 +44,34 @@ describe('pythonApiContractV1', () => {
});
});
it('documents chat.sendMessage protocol envelope return contract and metadata input', () => {
expect(getPythonApiMethodContract('chat.sendMessage')).toEqual({
method: 'chat.sendMessage',
description: 'Send message to chat conversation.',
params: [
{
name: 'conversationId',
type: 'string',
required: true,
},
{
name: 'message',
type: 'string',
required: true,
},
{
name: 'metadata',
type: 'object',
required: false,
},
],
returns: "{ success: boolean; message?: string; envelope?: ProtocolResponseEnvelope; protocolVersion?: '2.0'; traceId?: string; warnings?: string[]; error?: string }",
});
});
it('contains semantic version metadata for compatibility checks', () => {
expect(BDS_PYTHON_API_CONTRACT_V1).toMatchObject({
version: '1.4.0',
version: '1.5.0',
generatedAt: expect.any(String),
});
});
@@ -56,6 +81,7 @@ describe('pythonApiContractV1', () => {
expect.objectContaining({ name: 'PostData' }),
expect.objectContaining({ name: 'MediaData' }),
expect.objectContaining({ name: 'ProjectData' }),
expect.objectContaining({ name: 'ProtocolResponseEnvelope' }),
expect.objectContaining({ name: 'ProtocolTelemetrySnapshot' }),
]));
});
@@ -76,7 +102,7 @@ describe('generatePythonApiModuleV1', () => {
expect(moduleCode).toContain('async def search(self, query):');
expect(moduleCode).toContain('async def get_project_metadata(self):');
expect(moduleCode).toContain('async def get_conversations(self):');
expect(moduleCode).toContain('async def send_message(self, conversation_id, message):');
expect(moduleCode).toContain('async def send_message(self, conversation_id, message, metadata=None):');
expect(moduleCode).toContain('class BdsApi:');
expect(moduleCode).toContain('bds = BdsApi(_transport)');
});