Q131: remove review agents and eval suites whose pass only counted configuration rows
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
1 parent
074a1fc8e3
commit
fe45cd4d46
20 files changed
+32
-476
No files matched your search
@@ -32,11 +32,6 @@ const LEGACY_TABLES = [
|
||||
"operational_events",
|
||||
"chaos_drills",
|
||||
"chaos_drill_runs",
|
||||
"sdlc_review_agents",
|
||||
"sdlc_review_runs",
|
||||
"eval_suites",
|
||||
"eval_cases",
|
||||
"eval_runs",
|
||||
"pr_bot_gates",
|
||||
"chat_spaces",
|
||||
"chat_messages",
|
||||
@@ -164,14 +159,6 @@ const RELATION_CHECKS: RelationCheck[] = [
|
||||
{ table: "chaos_drills", column: "org_id", referencedTable: "organizations", referencedColumn: "id", reason: "dangling_org_id" },
|
||||
{ table: "chaos_drill_runs", column: "org_id", referencedTable: "organizations", referencedColumn: "id", reason: "dangling_org_id" },
|
||||
{ table: "chaos_drill_runs", column: "drill_id", referencedTable: "chaos_drills", referencedColumn: "id", reason: "dangling_drill_id" },
|
||||
{ table: "sdlc_review_agents", column: "org_id", referencedTable: "organizations", referencedColumn: "id", reason: "dangling_org_id" },
|
||||
{ table: "sdlc_review_runs", column: "org_id", referencedTable: "organizations", referencedColumn: "id", reason: "dangling_org_id" },
|
||||
{ table: "sdlc_review_runs", column: "review_agent_id", referencedTable: "sdlc_review_agents", referencedColumn: "id", reason: "dangling_review_agent_id" },
|
||||
{ table: "eval_suites", column: "org_id", referencedTable: "organizations", referencedColumn: "id", reason: "dangling_org_id" },
|
||||
{ table: "eval_cases", column: "org_id", referencedTable: "organizations", referencedColumn: "id", reason: "dangling_org_id" },
|
||||
{ table: "eval_cases", column: "suite_id", referencedTable: "eval_suites", referencedColumn: "id", reason: "dangling_suite_id" },
|
||||
{ table: "eval_runs", column: "org_id", referencedTable: "organizations", referencedColumn: "id", reason: "dangling_org_id" },
|
||||
{ table: "eval_runs", column: "suite_id", referencedTable: "eval_suites", referencedColumn: "id", reason: "dangling_suite_id" },
|
||||
{ table: "pr_bot_gates", column: "org_id", referencedTable: "organizations", referencedColumn: "id", reason: "dangling_org_id" },
|
||||
{ table: "chat_spaces", column: "org_id", referencedTable: "organizations", referencedColumn: "id", reason: "dangling_org_id" },
|
||||
{ table: "chat_messages", column: "org_id", referencedTable: "organizations", referencedColumn: "id", reason: "dangling_org_id" },
|
||||
|
||||
@@ -1,40 +1,8 @@
|
||||
import type { DatabaseSync } from "node:sqlite";
|
||||
import type { RunEvalSuiteInput, RunSdlcReviewInput, SetPrBotGateInput } from "../../shared/commands";
|
||||
import {
|
||||
EVAL_SUITE_KEY,
|
||||
PR_BOT_GATE_KEY,
|
||||
SDLC_REVIEW_AGENT_KEY,
|
||||
SDLC_REVIEW_TARGET_TYPE,
|
||||
SDLC_RUN_STATE,
|
||||
type EvalSuiteKey,
|
||||
type SdlcReviewAgentKey,
|
||||
type SdlcReviewTargetType,
|
||||
} from "../../shared/domain";
|
||||
import { PROVIDER_CATALOG } from "../../shared/provider-catalog";
|
||||
import type { SetPrBotGateInput } from "../../shared/commands";
|
||||
import { PR_BOT_GATE_KEY } from "../../shared/domain";
|
||||
import { createId } from "../../shared/ids";
|
||||
import { AuditService } from "./audit-service";
|
||||
import { providerAdapters } from "./provider-adapters";
|
||||
|
||||
const REVIEW_AGENTS = [
|
||||
{ key: SDLC_REVIEW_AGENT_KEY.CODE_REVIEWER, label: "Code Review Agent", specialty: "static quality and architecture" },
|
||||
{ key: SDLC_REVIEW_AGENT_KEY.MIGRATION_REVIEWER, label: "Migration Review Agent", specialty: "schema diff and rollback evidence" },
|
||||
{ key: SDLC_REVIEW_AGENT_KEY.POLICY_REVIEWER, label: "Policy Review Agent", specialty: "provider, approval, and access policy" },
|
||||
] as const;
|
||||
|
||||
const EVAL_SUITES = [
|
||||
{ key: EVAL_SUITE_KEY.CONTRACT, label: "Contract checks", scope: "provider adapters, event schemas, slash commands" },
|
||||
{ key: EVAL_SUITE_KEY.INTEGRATION, label: "Integration checks", scope: "chat, approval, file, skill resolution flows" },
|
||||
{ key: EVAL_SUITE_KEY.REPLAY, label: "Replay checks", scope: "audit timeline and historical decision regression" },
|
||||
] as const;
|
||||
|
||||
const EVAL_CASES = [
|
||||
{ suiteKey: EVAL_SUITE_KEY.CONTRACT, title: "Provider catalog has adapters", sourceType: "provider_gateway", sourceRef: "providers", expectedSignal: "adapter_coverage" },
|
||||
{ suiteKey: EVAL_SUITE_KEY.CONTRACT, title: "Schema ledger is current", sourceType: "database", sourceRef: "schema_migrations", expectedSignal: "migration_ledger" },
|
||||
{ suiteKey: EVAL_SUITE_KEY.INTEGRATION, title: "Approval routes are assigned", sourceType: "org_governance", sourceRef: "approval_assignments", expectedSignal: "route_coverage" },
|
||||
{ suiteKey: EVAL_SUITE_KEY.INTEGRATION, title: "Skills can resolve runtime snapshots", sourceType: "skills_platform", sourceRef: "skill_runtime_snapshots", expectedSignal: "skill_resolution" },
|
||||
{ suiteKey: EVAL_SUITE_KEY.REPLAY, title: "Audit timeline has domain events", sourceType: "audit", sourceRef: "audit_events", expectedSignal: "replayable_timeline" },
|
||||
{ suiteKey: EVAL_SUITE_KEY.REPLAY, title: "Operational signals are captured", sourceType: "operations", sourceRef: "operational_events", expectedSignal: "slo_signal" },
|
||||
] as const;
|
||||
|
||||
const PR_BOT_GATES = [
|
||||
{ key: PR_BOT_GATE_KEY.STATIC_CHECKS, label: "Static checks", gateType: "typecheck_lint_test" },
|
||||
@@ -43,23 +11,6 @@ const PR_BOT_GATES = [
|
||||
{ key: PR_BOT_GATE_KEY.HUMAN_APPROVAL, label: "Human approval", gateType: "approval_required" },
|
||||
] as const;
|
||||
|
||||
interface ReviewAgentRow {
|
||||
id: string;
|
||||
key: SdlcReviewAgentKey;
|
||||
enabled: number;
|
||||
}
|
||||
|
||||
interface EvalSuiteRow {
|
||||
id: string;
|
||||
key: EvalSuiteKey;
|
||||
}
|
||||
|
||||
interface GateCheck {
|
||||
key: string;
|
||||
passed: boolean;
|
||||
value: number | string | boolean;
|
||||
}
|
||||
|
||||
export class SdlcService {
|
||||
private readonly audit: AuditService;
|
||||
|
||||
@@ -68,28 +19,6 @@ export class SdlcService {
|
||||
}
|
||||
|
||||
seedOrganizationSdlc(orgId: string, now: string): void {
|
||||
for (const agent of REVIEW_AGENTS) {
|
||||
this.db.prepare(`
|
||||
insert into sdlc_review_agents(id, org_id, key, label, specialty, enabled, created_at, updated_at)
|
||||
values (?, ?, ?, ?, ?, 1, ?, ?)
|
||||
on conflict(org_id, key) do nothing
|
||||
`).run(createId("sra"), orgId, agent.key, agent.label, agent.specialty, now, now);
|
||||
}
|
||||
for (const suite of EVAL_SUITES) {
|
||||
this.db.prepare(`
|
||||
insert into eval_suites(id, org_id, key, label, scope, created_at, updated_at)
|
||||
values (?, ?, ?, ?, ?, ?, ?)
|
||||
on conflict(org_id, key) do nothing
|
||||
`).run(createId("evs"), orgId, suite.key, suite.label, suite.scope, now, now);
|
||||
}
|
||||
for (const item of EVAL_CASES) {
|
||||
const suite = this.requireSuiteByKey(orgId, item.suiteKey);
|
||||
this.db.prepare(`
|
||||
insert into eval_cases(id, org_id, suite_id, title, source_type, source_ref, expected_signal, created_at)
|
||||
values (?, ?, ?, ?, ?, ?, ?, ?)
|
||||
on conflict(suite_id, title) do nothing
|
||||
`).run(createId("evc"), orgId, suite.id, item.title, item.sourceType, item.sourceRef, item.expectedSignal, now);
|
||||
}
|
||||
for (const gate of PR_BOT_GATES) {
|
||||
this.db.prepare(`
|
||||
insert into pr_bot_gates(id, org_id, key, label, gate_type, enabled, created_at, updated_at)
|
||||
@@ -99,58 +28,6 @@ export class SdlcService {
|
||||
}
|
||||
}
|
||||
|
||||
runReview(input: RunSdlcReviewInput): string {
|
||||
const agent = this.db.prepare("select id, key, enabled from sdlc_review_agents where id = ? and org_id = ?")
|
||||
.get(input.reviewAgentId, input.orgId) as ReviewAgentRow | undefined;
|
||||
if (!agent) throw new Error("SDLC review agent was not found.");
|
||||
if (agent.enabled !== 1) throw new Error("SDLC review agent is disabled.");
|
||||
this.assertTarget(input.orgId, input.targetType, input.targetRef);
|
||||
const checks = this.reviewChecks(input.orgId, agent.key, input.targetType);
|
||||
const state = checks.every((check) => check.passed) ? SDLC_RUN_STATE.PASSED : SDLC_RUN_STATE.FAILED;
|
||||
const id = createId("srr");
|
||||
const now = new Date().toISOString();
|
||||
this.db.prepare(`
|
||||
insert into sdlc_review_runs(id, org_id, review_agent_id, target_type, target_ref, state, result_json, created_at, completed_at)
|
||||
values (?, ?, ?, ?, ?, ?, ?, ?, ?)
|
||||
`).run(id, input.orgId, input.reviewAgentId, input.targetType, input.targetRef, state, JSON.stringify({ checks }), now, now);
|
||||
this.audit.record({
|
||||
orgId: input.orgId,
|
||||
entityType: "sdlc_review_run",
|
||||
entityId: id,
|
||||
actorType: "system",
|
||||
actorId: null,
|
||||
action: "created",
|
||||
payload: { reviewAgentKey: agent.key, state, targetType: input.targetType },
|
||||
});
|
||||
return id;
|
||||
}
|
||||
|
||||
runEvalSuite(input: RunEvalSuiteInput): string {
|
||||
const suite = this.db.prepare("select id, key from eval_suites where id = ? and org_id = ?")
|
||||
.get(input.suiteId, input.orgId) as EvalSuiteRow | undefined;
|
||||
if (!suite) throw new Error("Eval suite was not found.");
|
||||
const checks = this.evalChecks(input.orgId, suite.key);
|
||||
const passedCount = checks.filter((check) => check.passed).length;
|
||||
const failedCount = checks.length - passedCount;
|
||||
const state = failedCount === 0 ? SDLC_RUN_STATE.PASSED : SDLC_RUN_STATE.FAILED;
|
||||
const id = createId("evr");
|
||||
const now = new Date().toISOString();
|
||||
this.db.prepare(`
|
||||
insert into eval_runs(id, org_id, suite_id, state, passed_count, failed_count, result_json, created_at, completed_at)
|
||||
values (?, ?, ?, ?, ?, ?, ?, ?, ?)
|
||||
`).run(id, input.orgId, input.suiteId, state, passedCount, failedCount, JSON.stringify({ checks }), now, now);
|
||||
this.audit.record({
|
||||
orgId: input.orgId,
|
||||
entityType: "eval_run",
|
||||
entityId: id,
|
||||
actorType: "system",
|
||||
actorId: null,
|
||||
action: "created",
|
||||
payload: { suiteKey: suite.key, state, passedCount, failedCount },
|
||||
});
|
||||
return id;
|
||||
}
|
||||
|
||||
setPrBotGate(input: SetPrBotGateInput): void {
|
||||
const gate = this.db.prepare("select key from pr_bot_gates where id = ? and org_id = ?")
|
||||
.get(input.gateId, input.orgId) as { key: string } | undefined;
|
||||
@@ -169,141 +46,4 @@ export class SdlcService {
|
||||
payload: { key: gate.key, enabled: input.enabled },
|
||||
});
|
||||
}
|
||||
|
||||
private requireSuiteByKey(orgId: string, key: EvalSuiteKey): { id: string } {
|
||||
const suite = this.db.prepare("select id from eval_suites where org_id = ? and key = ?")
|
||||
.get(orgId, key) as { id: string } | undefined;
|
||||
if (!suite) throw new Error("Eval suite is required.");
|
||||
return suite;
|
||||
}
|
||||
|
||||
private assertTarget(orgId: string, type: SdlcReviewTargetType, ref: string): void {
|
||||
if (!Object.values(SDLC_REVIEW_TARGET_TYPE).includes(type)) throw new Error("SDLC review target type is invalid.");
|
||||
const targetRef = ref.trim();
|
||||
if (!targetRef) throw new Error("SDLC review target is required.");
|
||||
if (type === SDLC_REVIEW_TARGET_TYPE.WORKSPACE) {
|
||||
const org = this.db.prepare("select id from organizations where id = ?").get(targetRef);
|
||||
if (!org || targetRef !== orgId) throw new Error("SDLC workspace target is invalid.");
|
||||
return;
|
||||
}
|
||||
if (type === SDLC_REVIEW_TARGET_TYPE.MIGRATION) {
|
||||
const report = this.db.prepare("select id from migration_reports where id = ? and org_id = ?").get(targetRef, orgId);
|
||||
if (!report) throw new Error("SDLC migration target was not found.");
|
||||
return;
|
||||
}
|
||||
const policy = this.db.prepare("select id from policy_rules where id = ? and org_id = ?").get(targetRef, orgId);
|
||||
if (!policy) throw new Error("SDLC policy target was not found.");
|
||||
}
|
||||
|
||||
private reviewChecks(orgId: string, agentKey: SdlcReviewAgentKey, targetType: SdlcReviewTargetType): GateCheck[] {
|
||||
if (agentKey === SDLC_REVIEW_AGENT_KEY.MIGRATION_REVIEWER) return this.migrationReviewChecks(orgId);
|
||||
if (agentKey === SDLC_REVIEW_AGENT_KEY.POLICY_REVIEWER) return this.policyReviewChecks(orgId);
|
||||
return this.codeReviewChecks(orgId, targetType);
|
||||
}
|
||||
|
||||
private codeReviewChecks(orgId: string, targetType: SdlcReviewTargetType): GateCheck[] {
|
||||
const enabledGates = count(this.db, "select count(*) as count from pr_bot_gates where org_id = ? and enabled = 1", orgId);
|
||||
const auditEvents = count(this.db, "select count(*) as count from audit_events where org_id = ?", orgId);
|
||||
return [
|
||||
{ key: "target_type", passed: targetType === SDLC_REVIEW_TARGET_TYPE.WORKSPACE, value: targetType },
|
||||
{ key: "enabled_pr_gates", passed: enabledGates >= 3, value: enabledGates },
|
||||
{ key: "recent_audit_events", passed: auditEvents > 0, value: auditEvents },
|
||||
];
|
||||
}
|
||||
|
||||
private migrationReviewChecks(orgId: string): GateCheck[] {
|
||||
const reports = this.db.prepare("select report_json from migration_reports where org_id = ?").all(orgId) as Array<{ report_json: string }>;
|
||||
const rollbackPacks = reports.filter((report) => hasObjectField(report.report_json, "rollbackPack")).length;
|
||||
const migrationGateEnabled = this.isGateEnabled(orgId, PR_BOT_GATE_KEY.MIGRATION_DIFF);
|
||||
return [
|
||||
{ key: "migration_gate_enabled", passed: migrationGateEnabled, value: migrationGateEnabled },
|
||||
{ key: "rollback_pack_coverage", passed: reports.length === rollbackPacks, value: `${rollbackPacks}/${reports.length}` },
|
||||
];
|
||||
}
|
||||
|
||||
private policyReviewChecks(orgId: string): GateCheck[] {
|
||||
const providerConflicts = count(this.db, "select count(*) as count from providers where org_id = ? and data_policy = 'zero_data_retention' and background_enabled = 1", orgId);
|
||||
const routeRules = count(this.db, "select count(*) as count from approval_routing_rules where org_id = ?", orgId);
|
||||
const policyGateEnabled = this.isGateEnabled(orgId, PR_BOT_GATE_KEY.POLICY_DIFF);
|
||||
return [
|
||||
{ key: "policy_gate_enabled", passed: policyGateEnabled, value: policyGateEnabled },
|
||||
{ key: "provider_policy_conflicts", passed: providerConflicts === 0, value: providerConflicts },
|
||||
{ key: "approval_routes", passed: routeRules > 0, value: routeRules },
|
||||
];
|
||||
}
|
||||
|
||||
private evalChecks(orgId: string, suiteKey: EvalSuiteKey): GateCheck[] {
|
||||
if (suiteKey === EVAL_SUITE_KEY.CONTRACT) return this.contractEvalChecks();
|
||||
if (suiteKey === EVAL_SUITE_KEY.INTEGRATION) return this.integrationEvalChecks(orgId);
|
||||
return this.replayEvalChecks(orgId);
|
||||
}
|
||||
|
||||
private contractEvalChecks(): GateCheck[] {
|
||||
const catalogKeys = PROVIDER_CATALOG.map((provider) => provider.key);
|
||||
const adapterKeys = Object.keys(providerAdapters);
|
||||
const missing = catalogKeys.filter((key) => !adapterKeys.includes(key));
|
||||
const runtimeContractReady = PROVIDER_CATALOG.every((provider) => {
|
||||
const adapter = providerAdapters[provider.key];
|
||||
return typeof adapter.authorize === "function"
|
||||
&& typeof adapter.refresh === "function"
|
||||
&& typeof adapter.revoke === "function"
|
||||
&& typeof adapter.listModels === "function"
|
||||
&& typeof adapter.getRateLimitState === "function"
|
||||
&& typeof adapter.streamResponse === "function"
|
||||
&& typeof adapter.cancelResponse === "function";
|
||||
});
|
||||
const schemaVersion = count(this.db, "select coalesce(max(version), 0) as count from schema_migrations");
|
||||
const rateLimitRegistry = count(this.db, `
|
||||
select count(*) as count
|
||||
from sqlite_master
|
||||
where type = 'table' and name = 'provider_rate_limit_observations'
|
||||
`);
|
||||
return [
|
||||
{ key: "provider_adapter_coverage", passed: missing.length === 0, value: `${adapterKeys.length}/${catalogKeys.length}` },
|
||||
{ key: "provider_rate_limit_registry", passed: rateLimitRegistry === 1, value: rateLimitRegistry },
|
||||
{ key: "provider_runtime_contract", passed: runtimeContractReady, value: runtimeContractReady },
|
||||
{ key: "schema_version", passed: schemaVersion >= 31, value: schemaVersion },
|
||||
];
|
||||
}
|
||||
|
||||
private integrationEvalChecks(orgId: string): GateCheck[] {
|
||||
const chatSpaces = count(this.db, "select count(*) as count from chat_spaces where org_id = ?", orgId);
|
||||
const approvalRoutes = count(this.db, "select count(*) as count from approval_routing_rules where org_id = ?", orgId);
|
||||
const skillSnapshots = count(this.db, `
|
||||
select count(*) as count
|
||||
from skill_runtime_snapshots
|
||||
join agents on agents.id = skill_runtime_snapshots.agent_id
|
||||
where agents.org_id = ?
|
||||
`, orgId);
|
||||
return [
|
||||
{ key: "chat_spaces", passed: chatSpaces > 0, value: chatSpaces },
|
||||
{ key: "approval_routes", passed: approvalRoutes > 0, value: approvalRoutes },
|
||||
{ key: "skill_runtime_snapshots", passed: skillSnapshots > 0, value: skillSnapshots },
|
||||
];
|
||||
}
|
||||
|
||||
private replayEvalChecks(orgId: string): GateCheck[] {
|
||||
const auditEvents = count(this.db, "select count(*) as count from audit_events where org_id = ?", orgId);
|
||||
const operationalEvents = count(this.db, "select count(*) as count from operational_events where org_id = ?", orgId);
|
||||
return [
|
||||
{ key: "audit_events", passed: auditEvents > 0, value: auditEvents },
|
||||
{ key: "operational_events", passed: operationalEvents > 0, value: operationalEvents },
|
||||
];
|
||||
}
|
||||
|
||||
private isGateEnabled(orgId: string, key: string): boolean {
|
||||
const row = this.db.prepare("select enabled from pr_bot_gates where org_id = ? and key = ?")
|
||||
.get(orgId, key) as { enabled: number } | undefined;
|
||||
if (!row) throw new Error("PR bot gate is required.");
|
||||
return row.enabled === 1;
|
||||
}
|
||||
}
|
||||
|
||||
function count(db: DatabaseSync, sql: string, ...params: string[]): number {
|
||||
return (db.prepare(sql).get(...params) as { count: number }).count;
|
||||
}
|
||||
|
||||
function hasObjectField(json: string, field: string): boolean {
|
||||
const parsed = JSON.parse(json) as Record<string, unknown>;
|
||||
return typeof parsed[field] === "object" && parsed[field] !== null;
|
||||
}
|
||||
@@ -81,8 +81,6 @@ export class SnapshotService {
|
||||
.map((event) => ({ ...event, success: toBool(event.success) }));
|
||||
const chaosDrills = list<AppSnapshot["chaosDrills"][number]>(this.db, "chaos_drills", "key asc")
|
||||
.map((drill) => ({ ...drill, enabled: toBool(drill.enabled) }));
|
||||
const sdlcReviewAgents = list<AppSnapshot["sdlcReviewAgents"][number]>(this.db, "sdlc_review_agents", "key asc")
|
||||
.map((agent) => ({ ...agent, enabled: toBool(agent.enabled) }));
|
||||
const prBotGates = list<AppSnapshot["prBotGates"][number]>(this.db, "pr_bot_gates", "key asc")
|
||||
.map((gate) => ({ ...gate, enabled: toBool(gate.enabled) }));
|
||||
const rollbackSnapshots = list<AppSnapshot["rollbackSnapshots"][number]>(this.db, "rollback_snapshots", "created_at desc")
|
||||
@@ -164,11 +162,6 @@ export class SnapshotService {
|
||||
operationalEvents,
|
||||
chaosDrills,
|
||||
chaosDrillRuns: list(this.db, "chaos_drill_runs", "started_at desc"),
|
||||
sdlcReviewAgents,
|
||||
sdlcReviewRuns: list(this.db, "sdlc_review_runs", "created_at desc"),
|
||||
evalSuites: list(this.db, "eval_suites", "key asc"),
|
||||
evalCases: list(this.db, "eval_cases", "title asc"),
|
||||
evalRuns: list(this.db, "eval_runs", "created_at desc"),
|
||||
prBotGates,
|
||||
auditReplayRuns: list(this.db, "audit_replay_runs", "created_at desc"),
|
||||
auditEvents: recentList<AppSnapshot["auditEvents"][number]>(this.db, "audit_events", "created_at desc", 200),
|
||||
|
||||
Reference in new issue
Block a user