feat(orchestrator): harden zero-impact shadow rollout

Gate and bound Portal shadow observations while preserving Native execution. Add fail-closed service boundaries, terminal retention controls, Canary readiness telemetry, ops visibility, and isolated regression coverage.
This commit is contained in:
john
2026-07-25 07:28:37 +08:00
parent 08a48e4849
commit 6df82818c5
33 changed files with 1569 additions and 108 deletions
+110 -2
View File
@@ -1,5 +1,8 @@
import assert from 'node:assert/strict';
import test from 'node:test';
import { MemorySaver } from '@langchain/langgraph';
import { createOrchestratorApp } from './app.mjs';
import { createLangGraphOrchestratorRuntime } from './runtime.mjs';
import { createWorkflowShadowObserver } from './shadow-observer.mjs';
function jsonResponse(body, status = 200) {
@@ -9,6 +12,19 @@ function jsonResponse(body, status = 200) {
});
}
async function listen(app) {
const server = await new Promise((resolve) => {
const candidate = app.listen(0, '127.0.0.1', () => resolve(candidate));
});
const address = server.address();
return {
baseUrl: `http://127.0.0.1:${address.port}`,
close: () => new Promise((resolve, reject) => {
server.close((error) => (error ? reject(error) : resolve()));
}),
};
}
test('shadow observer skips without creating a remote client when mode does not select shadow', async () => {
let fetchCalls = 0;
const observer = createWorkflowShadowObserver({
@@ -127,7 +143,7 @@ test('shadow boundary records non-selected canary decisions for a complete rate
assert.equal(result.executionPlan.taskType, 'code_analysis');
});
test('shadow observer sends a bounded observe-only RunSpec to LangGraph service', async () => {
test('shadow observer sends a control-plane-only RunSpec without user content or identity', async () => {
let capturedUrl = null;
let capturedInit = null;
const observer = createWorkflowShadowObserver({
@@ -175,6 +191,98 @@ test('shadow observer sends a bounded observe-only RunSpec to LangGraph service'
assert.equal(body.version, 'orchestrator-run-v1');
assert.equal(body.policy.executionMode, 'observe-only');
assert.equal(body.policy.sideEffectsAllowed, false);
assert.deepEqual(body.input.sessionRef, { kind: 'goose-session', id: 'session-2' });
assert.equal(body.input.instruction, '[shadow-control-plane-only]');
assert.equal('sessionRef' in body.input, false);
assert.deepEqual(body.subject, { tenantId: null, userId: null });
assert.equal(body.metadata.configVersion, 7);
assert.equal(body.metadata.dataPolicy, 'control-plane-only-v1');
assert.equal(
capturedInit.body.includes('Implement the service boundary'),
false,
);
assert.equal(capturedInit.body.includes('user-2'), false);
assert.equal(capturedInit.body.includes('session-2'), false);
});
test('isolated Portal shadow flow reaches LangGraph over HTTP and removes all terminal state', async (t) => {
const checkpointer = new MemorySaver();
const runtime = createLangGraphOrchestratorRuntime({
checkpointer,
checkpointKind: 'memory',
durable: false,
});
const server = await listen(createOrchestratorApp({
runtime,
serviceToken: 'isolated-shadow-token',
}));
t.after(server.close);
const observer = createWorkflowShadowObserver({
configService: {
async selectEngine() {
return {
shadowEngine: 'langgraph',
reason: 'shadow',
mode: 'shadow',
configVersion: 11,
};
},
async getRuntimeState() {
return {
config: {
serviceUrl: server.baseUrl,
requestTimeoutMs: 2000,
},
};
},
},
serviceToken: 'isolated-shadow-token',
});
const secretUserContent = 'private user content must never cross the shadow boundary';
const result = await observer({
runId: 'isolated-shadow-run-1',
requestId: 'isolated-shadow-request-1',
userId: 'private-user-id',
sessionId: 'private-session-id',
taskType: 'repo_refactor',
userMessage: {
role: 'user',
content: [{ type: 'text', text: secretUserContent }],
},
});
assert.equal(result.observed, true);
assert.equal(result.shadowRun.status, 'succeeded');
assert.equal(result.shadowRun.result.executed, false);
const tuple = await checkpointer.getTuple({
configurable: {
thread_id: 'isolated-shadow-run-1',
checkpoint_ns: '',
},
});
assert.equal(
tuple.checkpoint.channel_values.spec.input.instruction,
'[shadow-control-plane-only]',
);
const checkpointJson = JSON.stringify(tuple.checkpoint.channel_values.spec);
assert.equal(checkpointJson.includes(secretUserContent), false);
assert.equal(checkpointJson.includes('private-user-id'), false);
assert.equal(checkpointJson.includes('private-session-id'), false);
const deleted = await fetch(`${server.baseUrl}/v1/runs/isolated-shadow-run-1`, {
method: 'DELETE',
headers: { authorization: 'Bearer isolated-shadow-token' },
});
assert.equal(deleted.status, 200);
assert.deepEqual(await deleted.json(), {
runId: 'isolated-shadow-run-1',
deleted: true,
executorJobDeleted: true,
});
assert.equal(await runtime.getState('isolated-shadow-run-1'), null);
assert.equal(
await runtime.getExecutorJob('isolated-shadow-run-1:executor-preview'),
null,
);
});