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
+82 -1
View File
@@ -8,9 +8,16 @@ import {
function createPool() {
let row = null;
let createTableCalls = 0;
return {
get createTableCalls() {
return createTableCalls;
},
async query(sql, params = []) {
if (sql.includes('CREATE TABLE')) return [[], []];
if (sql.includes('CREATE TABLE')) {
createTableCalls += 1;
return [[], []];
}
if (sql.includes('SELECT config_json')) return [[...(row ? [row] : [])], []];
if (sql.includes('INSERT INTO')) {
row = {
@@ -33,6 +40,12 @@ test('orchestrator config defaults to a disabled native-safe runtime', async ()
assert.equal(state.config.primaryEngine, 'langgraph');
assert.equal(state.runtime.effective, false);
assert.equal(state.runtime.reason, 'mode_off');
assert.deepEqual(state.runtime.shadowObservation, {
requested: false,
enabled: false,
reason: 'shadow_observation_not_requested',
environmentGate: false,
});
assert.equal(state.engines.find((engine) => engine.id === 'native').configured, true);
assert.equal(state.engines.find((engine) => engine.id === 'langgraph').configured, false);
assert.deepEqual(
@@ -59,6 +72,48 @@ test('orchestrator config normalizes unsafe values and keeps a workflow allowlis
assert.equal(normalized.requestTimeoutMs, 500);
assert.equal(normalized.rolloutPercent, 100);
assert.deepEqual(normalized.workflowAllowlist, ['code-run-v1']);
assert.equal(
normalizeOrchestratorConfig({
serviceUrl: 'https://orchestrator.example/internal?token=secret',
}).serviceUrl,
'',
);
});
test('orchestrator config initializes its schema once per shared pool', async () => {
const pool = createPool();
const first = createOrchestratorAdminConfigService(pool, { env: {} });
const second = createOrchestratorAdminConfigService(pool, { env: {} });
await first.ensureSchema();
await first.getRuntimeState();
await second.getAdminConfig();
assert.equal(pool.createTableCalls, 1);
});
test('orchestrator blocks non-loopback service origins unless explicitly allowed', async () => {
const pool = createPool();
const blockedService = createOrchestratorAdminConfigService(pool, { env: {} });
const blocked = await blockedService.updateAdminConfig({
mode: 'shadow',
serviceUrl: 'https://orchestrator.example',
});
assert.equal(blocked.runtime.effective, false);
assert.equal(blocked.runtime.reason, 'service_url_not_allowed');
assert.deepEqual(blocked.runtime.shadowObservation, {
requested: true,
enabled: false,
reason: 'service_url_not_allowed',
environmentGate: false,
});
const allowedService = createOrchestratorAdminConfigService(pool, {
env: {
MEMIND_ORCHESTRATOR_ALLOWED_ORIGINS: 'https://orchestrator.example',
},
});
const allowed = await allowedService.getRuntimeState();
assert.equal(allowed.runtime.effective, true);
assert.equal(allowed.runtime.reason, null);
});
test('orchestrator config persists versioned admin updates and selects canary users', async () => {
@@ -194,6 +249,32 @@ test('Shadow mode evaluates the Canary rollout while Native remains effective',
assert.equal(native.candidateEngine, 'native');
assert.equal(native.candidateReason, 'canary_not_selected');
assert.equal(native.dryRun, true);
const runtime = await service.getRuntimeState();
assert.equal(runtime.runtime.shadowsLangGraph, false);
assert.deepEqual(runtime.runtime.shadowObservation, {
requested: true,
enabled: false,
reason: 'environment_shadow_observation_gate_disabled',
environmentGate: false,
});
});
test('Shadow runtime reports the independent Portal observation wiring gate', async () => {
const service = createOrchestratorAdminConfigService(createPool(), {
env: { MEMIND_ORCHESTRATOR_SHADOW_OBSERVATION_ENABLED: '1' },
});
await service.updateAdminConfig({
mode: 'shadow',
serviceUrl: 'http://127.0.0.1:8093',
});
const runtime = await service.getRuntimeState();
assert.equal(runtime.runtime.shadowsLangGraph, true);
assert.deepEqual(runtime.runtime.shadowObservation, {
requested: true,
enabled: true,
reason: null,
environmentGate: true,
});
});
test('orchestrator emergency kill switch always forces native selection', async () => {