test(ai-chat): покрыть safety-путь детектора деградации + границы (ревью #454)
Ревью: дизайн LGTM, но safety-фича недотестирована. Добавлено (только тесты,
прод-код не тронут):
- e2e-реакция детектора: streamText эмитит degenerate-чанки → union abortSignal
срабатывает с 'Output degeneration detected' (отличимо от Stop) → onAbort
пишет status:error + OUTPUT_DEGENERATION_ERROR + усечённый content, лиза MCP
закрыта. Именно ДЕЙСТВИЕ на детект (детектит-но-не-действует = защиты нет);
- граница monochar-порога: hasPeriodicTail('x'×59)=false, ('x'×60)=true
(мутация >=→> раньше выживала);
- empty-turn маркер (шаги исчерпаны + без текста → STEP_LIMIT_NO_ANSWER_MARKER;
негативы на нормальный текст-ход и на исчерпание-с-текстом — гардят AND);
- различение degeneration-onAbort vs user-Stop (Stop → status:aborted, без
error/усечения).
Мутационно: (a) >=→> роняет 60-границу; (b) нейтрализация onAbort-ветки роняет
reaction-тест; (c) нейтрализация маркера роняет empty-turn-тест. +7 тестов,
137 passed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -26,6 +26,8 @@ import {
|
||||
STEP_BUDGET_WARNING_LEAD,
|
||||
FINAL_STEP_INSTRUCTION,
|
||||
FINAL_STEP_NUDGE,
|
||||
STEP_LIMIT_NO_ANSWER_MARKER,
|
||||
OUTPUT_DEGENERATION_ERROR,
|
||||
} from './ai-chat.service';
|
||||
import type { AiChatMessage, Workspace } from '@docmost/db/types/entity.types';
|
||||
import { buildSystemPrompt } from './ai-chat.prompt';
|
||||
@@ -1535,3 +1537,348 @@ describe('AiChatService.stream — resumable pipe options (#184 phase 1.5)', ()
|
||||
expect(streamRegistry.abortEntry).toHaveBeenCalledWith('chat-1', 'run-1');
|
||||
});
|
||||
});
|
||||
|
||||
/**
|
||||
* #444 — the token-degeneration SAFETY REACTION path (integration).
|
||||
*
|
||||
* output-degeneration.spec.ts proves the detector DETECTS; this proves the wired
|
||||
* REACTION: a degenerate stream must (1) trip the detector in onChunk, (2) abort
|
||||
* the turn via the INTERNAL degeneration controller (distinct from a user Stop),
|
||||
* (3) truncate the runaway tail before persist in onAbort, (4) persist status
|
||||
* 'error' with the OUTPUT_DEGENERATION_ERROR message (not a bare 'aborted' and not
|
||||
* a swept 'streaming'), and (5) still release the leased external MCP clients.
|
||||
*
|
||||
* Harness: streamText is the SAME jest.fn mocked at the top of this file. Unlike
|
||||
* the pipe-options suite above (which only inspects the pipe call), this mock
|
||||
* CAPTURES the streamText options (onChunk/onAbort/onFinish + abortSignal) so the
|
||||
* test can drive the callbacks exactly as the AI SDK would — feeding degenerate
|
||||
* text-delta chunks through onChunk until the service's own AbortController fires,
|
||||
* then invoking onAbort (which the SDK does on an aborted signal). No new mocking
|
||||
* style is invented; it reuses the makeRes / service-construction shape above.
|
||||
*/
|
||||
describe('AiChatService.stream — token-degeneration reaction (#444)', () => {
|
||||
const streamTextMock = streamText as unknown as jest.Mock;
|
||||
|
||||
beforeEach(() => {
|
||||
streamTextMock.mockReset();
|
||||
jest
|
||||
.spyOn(Logger.prototype, 'log')
|
||||
.mockImplementation(() => undefined as never);
|
||||
jest
|
||||
.spyOn(Logger.prototype, 'error')
|
||||
.mockImplementation(() => undefined as never);
|
||||
jest
|
||||
.spyOn(Logger.prototype, 'warn')
|
||||
.mockImplementation(() => undefined as never);
|
||||
});
|
||||
|
||||
afterEach(() => jest.restoreAllMocks());
|
||||
|
||||
function makeRes() {
|
||||
return {
|
||||
raw: {
|
||||
writeHead: jest.fn(),
|
||||
write: jest.fn(),
|
||||
once: jest.fn(),
|
||||
on: jest.fn(),
|
||||
flushHeaders: jest.fn(),
|
||||
writableEnded: false,
|
||||
destroyed: false,
|
||||
},
|
||||
};
|
||||
}
|
||||
|
||||
// Wire the full stream() path with in-memory fakes. The assistant row is
|
||||
// captured so the terminal finalize (an UPDATE of the upfront-seeded row) can be
|
||||
// asserted. One external MCP client with a close() spy lets us assert leases are
|
||||
// released on the terminal path. lockdown OFF (default) so the detector is the
|
||||
// active guard.
|
||||
function makeService() {
|
||||
// The upfront insert seeds the assistant row; findById/insert stamp a stable
|
||||
// id so planFinalizeAssistant picks the UPDATE path.
|
||||
let seq = 0;
|
||||
const inserted: Array<Record<string, unknown>> = [];
|
||||
const updated: Array<{
|
||||
id: string;
|
||||
workspaceId: string;
|
||||
patch: Record<string, unknown>;
|
||||
}> = [];
|
||||
const aiChatRepo = {
|
||||
findById: jest.fn(async () => ({ id: 'chat-1', workspaceId: 'ws-1' })),
|
||||
insert: jest.fn(),
|
||||
};
|
||||
const aiChatMessageRepo = {
|
||||
insert: jest.fn(async (row: Record<string, unknown>) => {
|
||||
inserted.push(row);
|
||||
return { id: row.role === 'assistant' ? 'assistant-1' : `user-${++seq}` };
|
||||
}),
|
||||
findAllByChat: jest.fn(async () => []),
|
||||
update: jest.fn(
|
||||
async (
|
||||
id: string,
|
||||
workspaceId: string,
|
||||
patch: Record<string, unknown>,
|
||||
) => {
|
||||
updated.push({ id, workspaceId, patch });
|
||||
return { id };
|
||||
},
|
||||
),
|
||||
};
|
||||
const aiSettings = { resolve: jest.fn(async () => ({})) };
|
||||
const tools = { forUser: jest.fn(async () => ({})) };
|
||||
const mcpClose = jest.fn(async () => undefined);
|
||||
const mcpClients = {
|
||||
toolsFor: jest.fn(async () => ({
|
||||
tools: {},
|
||||
clients: [{ close: mcpClose }],
|
||||
outcomes: [],
|
||||
instructions: [],
|
||||
})),
|
||||
};
|
||||
const streamRegistry = { open: jest.fn(), bind: jest.fn(), abortEntry: jest.fn() };
|
||||
const svc = new AiChatService(
|
||||
{} as never,
|
||||
aiChatRepo as never,
|
||||
aiChatMessageRepo as never,
|
||||
{} as never, // aiChatPageSnapshotRepo (no open page -> never touched)
|
||||
aiSettings as never,
|
||||
tools as never,
|
||||
mcpClients as never,
|
||||
{} as never, // aiAgentRoleRepo
|
||||
{} as never, // pageRepo (no open page)
|
||||
{} as never, // pageAccess
|
||||
{
|
||||
isAiChatDeferredToolsEnabled: () => false,
|
||||
// lockdown OFF => the degeneration detector is the anti-babble guard.
|
||||
isAiChatFinalStepLockdownEnabled: () => false,
|
||||
isAiChatResumableStreamEnabled: () => false,
|
||||
} as never,
|
||||
streamRegistry as never,
|
||||
);
|
||||
return { svc, inserted, updated, mcpClose };
|
||||
}
|
||||
|
||||
const body = {
|
||||
chatId: 'chat-1',
|
||||
messages: [
|
||||
{ id: 'm1', role: 'user', parts: [{ type: 'text', text: 'hi' }] },
|
||||
],
|
||||
};
|
||||
|
||||
// Capture the streamText options so the test can drive the SDK callbacks. The
|
||||
// returned result stub is enough for the post-streamText wiring (consumeStream +
|
||||
// pipeUIMessageStreamToResponse are no-ops here).
|
||||
function captureStreamText(): { opts: () => Record<string, any> } {
|
||||
let captured: Record<string, any> | undefined;
|
||||
streamTextMock.mockImplementation((options: Record<string, any>) => {
|
||||
captured = options;
|
||||
return {
|
||||
consumeStream: jest.fn(),
|
||||
pipeUIMessageStreamToResponse: jest.fn(),
|
||||
};
|
||||
});
|
||||
return {
|
||||
opts: () => {
|
||||
if (!captured) throw new Error('streamText was not called');
|
||||
return captured;
|
||||
},
|
||||
};
|
||||
}
|
||||
|
||||
async function drive(svc: AiChatService): Promise<void> {
|
||||
await svc.stream({
|
||||
user: { id: 'u1' } as never,
|
||||
workspace: { id: 'ws-1' } as never,
|
||||
sessionId: 's1',
|
||||
body: body as never,
|
||||
res: makeRes() as never,
|
||||
signal: new AbortController().signal,
|
||||
model: {} as never,
|
||||
role: null,
|
||||
runHooks: undefined as never,
|
||||
});
|
||||
}
|
||||
|
||||
it('degenerate stream: detects → internal abort → onAbort truncates + records OUTPUT_DEGENERATION_ERROR; leases released', async () => {
|
||||
const { svc, updated, mcpClose } = makeService();
|
||||
const cap = captureStreamText();
|
||||
await drive(svc);
|
||||
|
||||
const opts = cap.opts();
|
||||
// The turn's abort signal is the UNION of the socket/run signal and the
|
||||
// internal degeneration controller — untripped before any output.
|
||||
expect(opts.abortSignal.aborted).toBe(false);
|
||||
|
||||
// Feed a runaway "loadTools.\n" loop the way the SDK streams it: many small
|
||||
// text-delta chunks. The onChunk throttle only re-checks every ~2000 chars, so
|
||||
// deliver well past that so the detector's identical-line rule (>=25 lines)
|
||||
// and the ~2000-char throttle both fire.
|
||||
const line = 'loadTools.\n';
|
||||
let delivered = 0;
|
||||
for (let i = 0; i < 400 && !opts.abortSignal.aborted; i++) {
|
||||
opts.onChunk({ chunk: { type: 'text-delta', text: line } });
|
||||
delivered += line.length;
|
||||
}
|
||||
|
||||
// The detector must have tripped and aborted via the INTERNAL controller — the
|
||||
// reason carries the degeneration message, distinguishing it from a user Stop
|
||||
// (which aborts with no such reason) or a socket disconnect.
|
||||
expect(opts.abortSignal.aborted).toBe(true);
|
||||
expect(delivered).toBeGreaterThan(2000);
|
||||
expect(String(opts.abortSignal.reason)).toContain(
|
||||
'Output degeneration detected',
|
||||
);
|
||||
|
||||
// The SDK reacts to the aborted signal by invoking onAbort. `steps` is empty
|
||||
// (the runaway never finished a step); the in-progress runaway text is what
|
||||
// gets truncated + persisted.
|
||||
await opts.onAbort({ steps: [] });
|
||||
|
||||
// Terminal finalize = an UPDATE of the upfront-seeded assistant row (assistant
|
||||
// row was inserted upfront, so planFinalizeAssistant -> UPDATE).
|
||||
expect(updated).toHaveLength(1);
|
||||
const patch = updated[0].patch as {
|
||||
status: string;
|
||||
content: string;
|
||||
metadata: Record<string, unknown>;
|
||||
};
|
||||
// (4) status 'error' with the degeneration message — NOT 'aborted' and NOT a
|
||||
// swept 'streaming'. This distinguishes it from a user Stop / server restart.
|
||||
expect(patch.status).toBe('error');
|
||||
expect(patch.metadata.error).toBe(OUTPUT_DEGENERATION_ERROR);
|
||||
expect(patch.metadata.finishReason).toBe('error');
|
||||
// (3) the runaway tail is TRUNCATED, not the full multi-KB babble: the marker
|
||||
// is present and the persisted content is far shorter than what was streamed.
|
||||
expect(patch.content).toContain('output truncated');
|
||||
expect(patch.content.length).toBeLessThan(delivered);
|
||||
// Only a few loop reps survive (truncateDegeneratedTail keeps a handful).
|
||||
expect((patch.content.match(/loadTools\./g) ?? []).length).toBeLessThan(10);
|
||||
|
||||
// (5) the leased external MCP client is still released on this terminal path.
|
||||
expect(mcpClose).toHaveBeenCalledTimes(1);
|
||||
});
|
||||
|
||||
it('degeneration onAbort differs from a NORMAL/user abort (no truncation, no error)', async () => {
|
||||
// Same harness, but the stream is NOT degenerate: a clean short answer, then a
|
||||
// user Stop reaches onAbort WITHOUT the degeneration controller having fired.
|
||||
const { svc, updated, mcpClose } = makeService();
|
||||
const cap = captureStreamText();
|
||||
await drive(svc);
|
||||
const opts = cap.opts();
|
||||
|
||||
opts.onChunk({ chunk: { type: 'text-delta', text: 'A normal partial answer.' } });
|
||||
// The detector never tripped -> the union signal is NOT aborted by us.
|
||||
expect(opts.abortSignal.aborted).toBe(false);
|
||||
|
||||
// A user Stop / disconnect drives onAbort with the partial (clean) text.
|
||||
await opts.onAbort({ steps: [] });
|
||||
|
||||
expect(updated).toHaveLength(1);
|
||||
const patch = updated[0].patch as {
|
||||
status: string;
|
||||
content: string;
|
||||
metadata: Record<string, unknown>;
|
||||
};
|
||||
// A normal abort persists status 'aborted' with NO error and NO truncation
|
||||
// marker — the branch is genuinely distinguished from the degeneration path.
|
||||
expect(patch.status).toBe('aborted');
|
||||
expect('error' in patch.metadata).toBe(false);
|
||||
expect(patch.content).toBe('A normal partial answer.');
|
||||
expect(patch.content).not.toContain('output truncated');
|
||||
// Cleanup still runs on the normal abort path too.
|
||||
expect(mcpClose).toHaveBeenCalledTimes(1);
|
||||
});
|
||||
|
||||
/**
|
||||
* Empty-turn marker (#444): onFinish appends STEP_LIMIT_NO_ANSWER_MARKER only
|
||||
* when the turn burned ALL its steps (steps.length >= MAX_AGENT_STEPS) AND never
|
||||
* produced any text. The negative: a normal turn ending WITH text is left alone.
|
||||
*/
|
||||
it('empty turn (no text + steps exhausted) persists the STEP_LIMIT_NO_ANSWER_MARKER', async () => {
|
||||
const { svc, updated } = makeService();
|
||||
const cap = captureStreamText();
|
||||
await drive(svc);
|
||||
const opts = cap.opts();
|
||||
|
||||
// MAX_AGENT_STEPS text-less steps (only tool calls) => step-exhausted, no text.
|
||||
const steps = Array.from({ length: MAX_AGENT_STEPS }, () => ({
|
||||
text: '',
|
||||
toolCalls: [{ toolCallId: 'c1', toolName: 'searchPages', input: {} }],
|
||||
toolResults: [
|
||||
{ toolCallId: 'c1', toolName: 'searchPages', output: { hits: [] } },
|
||||
],
|
||||
}));
|
||||
await opts.onFinish({
|
||||
text: '',
|
||||
finishReason: 'tool-calls',
|
||||
totalUsage: { inputTokens: 1, outputTokens: 1, totalTokens: 2 },
|
||||
usage: { inputTokens: 1, outputTokens: 1 },
|
||||
steps,
|
||||
});
|
||||
|
||||
expect(updated).toHaveLength(1);
|
||||
const patch = updated[0].patch as { status: string; content: string };
|
||||
expect(patch.status).toBe('completed');
|
||||
// The synthetic marker is the trailing text of the persisted content.
|
||||
expect(patch.content).toContain(STEP_LIMIT_NO_ANSWER_MARKER);
|
||||
});
|
||||
|
||||
it('normal turn ending WITH text does NOT get the empty-turn marker', async () => {
|
||||
const { svc, updated } = makeService();
|
||||
const cap = captureStreamText();
|
||||
await drive(svc);
|
||||
const opts = cap.opts();
|
||||
|
||||
// A single step that produced a real answer, well under the step cap.
|
||||
const steps = [
|
||||
{ text: 'Here is the finished answer.', toolCalls: [], toolResults: [] },
|
||||
];
|
||||
await opts.onFinish({
|
||||
text: 'Here is the finished answer.',
|
||||
finishReason: 'stop',
|
||||
totalUsage: { inputTokens: 1, outputTokens: 1, totalTokens: 2 },
|
||||
usage: { inputTokens: 1, outputTokens: 1 },
|
||||
steps,
|
||||
});
|
||||
|
||||
expect(updated).toHaveLength(1);
|
||||
const patch = updated[0].patch as { status: string; content: string };
|
||||
expect(patch.status).toBe('completed');
|
||||
expect(patch.content).toBe('Here is the finished answer.');
|
||||
expect(patch.content).not.toContain(STEP_LIMIT_NO_ANSWER_MARKER);
|
||||
});
|
||||
|
||||
it('step-exhausted turn that DID produce text keeps the text, no marker (guards the AND)', async () => {
|
||||
// Exhausting the step budget alone must NOT append the marker when SOME step
|
||||
// produced text — the marker keys off "no text" too. Drive the real onFinish
|
||||
// with MAX_AGENT_STEPS steps where the last one carries the answer.
|
||||
const { svc, updated } = makeService();
|
||||
const cap = captureStreamText();
|
||||
await drive(svc);
|
||||
const opts = cap.opts();
|
||||
|
||||
const steps = Array.from({ length: MAX_AGENT_STEPS }, (_, i) => ({
|
||||
text: i === MAX_AGENT_STEPS - 1 ? 'Final synthesized answer.' : '',
|
||||
toolCalls:
|
||||
i === MAX_AGENT_STEPS - 1
|
||||
? []
|
||||
: [{ toolCallId: `c${i}`, toolName: 'searchPages', input: {} }],
|
||||
toolResults:
|
||||
i === MAX_AGENT_STEPS - 1
|
||||
? []
|
||||
: [{ toolCallId: `c${i}`, toolName: 'searchPages', output: {} }],
|
||||
}));
|
||||
await opts.onFinish({
|
||||
text: 'Final synthesized answer.',
|
||||
finishReason: 'stop',
|
||||
totalUsage: { inputTokens: 1, outputTokens: 1, totalTokens: 2 },
|
||||
usage: { inputTokens: 1, outputTokens: 1 },
|
||||
steps,
|
||||
});
|
||||
|
||||
expect(updated).toHaveLength(1);
|
||||
const patch = updated[0].patch as { content: string };
|
||||
expect(patch.content).toContain('Final synthesized answer.');
|
||||
expect(patch.content).not.toContain(STEP_LIMIT_NO_ANSWER_MARKER);
|
||||
});
|
||||
});
|
||||
|
||||
@@ -124,6 +124,23 @@ const config = { retries: 3, timeout: 5000, backoff: 'exp' };
|
||||
expect(isDegenerateOutput(text)).toBe(false);
|
||||
});
|
||||
|
||||
// TRIVIAL_MIN_REPEATS boundary (#444 review). The monochar-tail branch fires at
|
||||
// EXACTLY 60 identical trailing chars (`run >= TRIVIAL_MIN_REPEATS`), so 59 is
|
||||
// clean and 60 trips. These pin the `>=` and MUST fail if the comparison is
|
||||
// flipped to `>` (the surviving mutation). The value 60 is HARD-CODED here on
|
||||
// purpose: TRIVIAL_MIN_REPEATS is a private constant and the assert must lock
|
||||
// the literal boundary the reviewer named, not track a constant edit.
|
||||
it('NEGATIVE: 59 identical trailing chars is one below the monochar threshold', () => {
|
||||
expect(hasPeriodicTail('x'.repeat(59))).toBe(false);
|
||||
expect(isDegenerateOutput('x'.repeat(59))).toBe(false);
|
||||
});
|
||||
|
||||
it('POSITIVE: 60 identical trailing chars hits the monochar threshold exactly', () => {
|
||||
// Fails if `run >= TRIVIAL_MIN_REPEATS` is mutated to `run > …`.
|
||||
expect(hasPeriodicTail('x'.repeat(60))).toBe(true);
|
||||
expect(isDegenerateOutput('x'.repeat(60))).toBe(true);
|
||||
});
|
||||
|
||||
// Positive counterparts: a GENUINE single-char runaway (hundreds+ of repeats)
|
||||
// and the real incident (period>=2, "loadTools." ×N) must still fire.
|
||||
it('POSITIVE: a genuine single-char runaway is still flagged', () => {
|
||||
|
||||
Reference in New Issue
Block a user