@rulemetric/cli 0.6.8 → 0.7.0

package/dist/commands/crons/list.js.map

@@ -2,6 +2,6 @@
   "version": 3,
   "sources": ["../../../src/commands/crons/list.ts"],
   "sourcesContent": ["import { printTable } from '@oclif/table';\nimport { BaseCommand } from '../../base-command.js';\n// Metadata-only import: pulling the full manual-tasks module here would drag the\n// heavy worker handlers (postgres/OTel/Sentry) into this command's bundle chunk\n// and crash a standalone npm install at load. See manual-tasks-meta.ts.\nimport { MANUAL_TASK_META } from '../../lib/manual-tasks-meta.js';\n\nexport default class CronsList extends BaseCommand {\n    static override description = 'List manually-runnable maintenance tasks';\n\n    static override examples = [\n        '<%= config.bin %> crons list',\n    ];\n\n    async run(): Promise<void> {\n        await this.parse(CronsList);\n        printTable({\n            data: MANUAL_TASK_META.map((t) => ({\n                Name: t.name,\n                'Scheduled by': t.scheduledBy,\n                Description: t.description,\n            })),\n            columns: ['Name', 'Scheduled by', 'Description'],\n        });\n        this.log('');\n        this.log('Scheduled crons run in Supabase pg_cron \u2014 audit them with:');\n        this.log(\"  psql \\\"$DATABASE_URL\\\" -c 'select jobname, schedule, active from cron.job;'\");\n        this.log('Run any task above on demand with: rulemetric crons run <name>');\n    }\n}\n"],
-  "mappings": ";;;;;;;;;;;;;;AAAA,SAAS,kBAAkB;AAO3B,IAAqB,YAArB,MAAqB,mBAAkB,YAAY;AAAA,EAC/C,OAAgB,cAAc;AAAA,EAE9B,OAAgB,WAAW;AAAA,IACvB;AAAA,EACJ;AAAA,EAEA,MAAM,MAAqB;AACvB,UAAM,KAAK,MAAM,UAAS;AAC1B,eAAW;AAAA,MACP,MAAM,iBAAiB,IAAI,CAAC,OAAO;AAAA,QAC/B,MAAM,EAAE;AAAA,QACR,gBAAgB,EAAE;AAAA,QAClB,aAAa,EAAE;AAAA,MACnB,EAAE;AAAA,MACF,SAAS,CAAC,QAAQ,gBAAgB,aAAa;AAAA,IACnD,CAAC;AACD,SAAK,IAAI,EAAE;AACX,SAAK,IAAI,iEAA4D;AACrE,SAAK,IAAI,6EAA+E;AACxF,SAAK,IAAI,gEAAgE;AAAA,EAC7E;AACJ;",
+  "mappings": ";;;;;;;;;;;;;;;AAAA,SAAS,kBAAkB;AAO3B,IAAqB,YAArB,MAAqB,mBAAkB,YAAY;AAAA,EAC/C,OAAgB,cAAc;AAAA,EAE9B,OAAgB,WAAW;AAAA,IACvB;AAAA,EACJ;AAAA,EAEA,MAAM,MAAqB;AACvB,UAAM,KAAK,MAAM,UAAS;AAC1B,eAAW;AAAA,MACP,MAAM,iBAAiB,IAAI,CAAC,OAAO;AAAA,QAC/B,MAAM,EAAE;AAAA,QACR,gBAAgB,EAAE;AAAA,QAClB,aAAa,EAAE;AAAA,MACnB,EAAE;AAAA,MACF,SAAS,CAAC,QAAQ,gBAAgB,aAAa;AAAA,IACnD,CAAC;AACD,SAAK,IAAI,EAAE;AACX,SAAK,IAAI,iEAA4D;AACrE,SAAK,IAAI,6EAA+E;AACxF,SAAK,IAAI,gEAAgE;AAAA,EAC7E;AACJ;",
   "names": []
 }

package/dist/commands/crons/run.js

@@ -3,11 +3,12 @@ import {
 } from "../../chunk-JFDWKLQC.js";
 import {
   BaseCommand
-} from "../../chunk-SZ7VDCD6.js";
+} from "../../chunk-2HK7RGZB.js";
 import "../../chunk-R56PH7YE.js";
-import "../../chunk-SLL3OBKN.js";
+import "../../chunk-GQNJO2P2.js";
+import "../../chunk-37LN5O7M.js";
 import "../../chunk-EKJ2CABV.js";
-import "../../chunk-ZRLYHBPN.js";
+import "../../chunk-ZUJBM6HE.js";
 import "../../chunk-W2YERO7E.js";
 import "../../chunk-NSBPE2FW.js";
 

package/dist/commands/crons/run.js.map

@@ -2,6 +2,6 @@
   "version": 3,
   "sources": ["../../../src/commands/crons/run.ts"],
   "sourcesContent": ["import { Args, Flags } from '@oclif/core';\nimport { BaseCommand } from '../../base-command.js';\n// Metadata-only static import (dependency-free). The executor (runManualTask) and\n// its heavy db/OTel/Sentry handlers are lazy-imported inside run(), so loading\n// this command never pulls `postgres` into its bundle chunk and a standalone\n// npm install no longer crashes at module-load. See manual-tasks-meta.ts.\nimport { MANUAL_TASK_META } from '../../lib/manual-tasks-meta.js';\n\nexport default class CronsRun extends BaseCommand {\n    static override description = 'Run a maintenance task synchronously, right now';\n\n    static override examples = [\n        '<%= config.bin %> crons run cron_check_changelog',\n        '<%= config.bin %> crons run cron_suggest_instructions --quiet',\n    ];\n\n    static override args = {\n        name: Args.string({\n            description: 'Task name (run `rulemetric crons list` to see all)',\n            required: true,\n        }),\n    };\n\n    static override flags = {\n        quiet: Flags.boolean({\n            char: 'q',\n            description: 'Suppress per-line log output; print summary only',\n        }),\n    };\n\n    async run(): Promise<void> {\n        const { args, flags } = await this.parse(CronsRun);\n        this.requireAuth();\n\n        // Validate against the dependency-free metadata first, so an unknown task\n        // never triggers the lazy load of the heavy executor below.\n        const def = MANUAL_TASK_META.find((t) => t.name === args.name);\n        if (!def) {\n            this.error(`Unknown task: ${args.name}\\nRun \\`rulemetric crons list\\` to see available tasks.`);\n        }\n\n        // Lazy-load the executor only when a task is actually run. This keeps the\n        // heavy handler chunk (postgres/OTel/Sentry) out of this command's static\n        // import graph. Executing a task still needs the worker runtime\n        // (DATABASE_URL + those deps); a standalone npm install lacks them, so\n        // turn the raw MODULE_NOT_FOUND into an actionable message.\n        let runManualTask: typeof import('../../lib/manual-tasks.js').runManualTask;\n        try {\n            ({ runManualTask } = await import('../../lib/manual-tasks.js'));\n        } catch (err) {\n            const code = (err as { code?: string } | null)?.code;\n            if (code === 'ERR_MODULE_NOT_FOUND' || code === 'MODULE_NOT_FOUND') {\n                this.error(\n                    `\\`crons run\\` needs the worker runtime, which isn't available in this install.\\n` +\n                        `These maintenance tasks connect to the database directly \u2014 run them from a full ` +\n                        `RuleMetric checkout with DATABASE_URL and the worker dependencies installed.`,\n                );\n            }\n            throw err;\n        }\n\n        this.log(`\u25B6 Running ${def.name} \u2026`);\n        const startedAt = Date.now();\n        const result = await runManualTask(args.name);\n\n        if (!flags.quiet) {\n            for (const line of result.logs) {\n                const prefix = line.level === 'error' || line.level === 'warn' ? `[${line.level}]` : '';\n                this.log(`  ${prefix} ${line.message}`);\n            }\n        }\n\n        const seconds = ((result.durationMs ?? Date.now() - startedAt) / 1000).toFixed(1);\n        if (result.ok) {\n            this.log(`\u2713 ${def.name} completed in ${seconds}s (${result.logs.length} log line${result.logs.length === 1 ? '' : 's'})`);\n        } else {\n            this.error(`\u2717 ${def.name} failed after ${seconds}s: ${result.error}`);\n        }\n    }\n}\n"],
-  "mappings": ";;;;;;;;;;;;;;AAAA,SAAS,MAAM,aAAa;AAQ5B,IAAqB,WAArB,MAAqB,kBAAiB,YAAY;AAAA,EAC9C,OAAgB,cAAc;AAAA,EAE9B,OAAgB,WAAW;AAAA,IACvB;AAAA,IACA;AAAA,EACJ;AAAA,EAEA,OAAgB,OAAO;AAAA,IACnB,MAAM,KAAK,OAAO;AAAA,MACd,aAAa;AAAA,MACb,UAAU;AAAA,IACd,CAAC;AAAA,EACL;AAAA,EAEA,OAAgB,QAAQ;AAAA,IACpB,OAAO,MAAM,QAAQ;AAAA,MACjB,MAAM;AAAA,MACN,aAAa;AAAA,IACjB,CAAC;AAAA,EACL;AAAA,EAEA,MAAM,MAAqB;AACvB,UAAM,EAAE,MAAM,MAAM,IAAI,MAAM,KAAK,MAAM,SAAQ;AACjD,SAAK,YAAY;AAIjB,UAAM,MAAM,iBAAiB,KAAK,CAAC,MAAM,EAAE,SAAS,KAAK,IAAI;AAC7D,QAAI,CAAC,KAAK;AACN,WAAK,MAAM,iBAAiB,KAAK,IAAI;AAAA,sDAAyD;AAAA,IAClG;AAOA,QAAI;AACJ,QAAI;AACA,OAAC,EAAE,cAAc,IAAI,MAAM,OAAO,2BAA2B;AAAA,IACjE,SAAS,KAAK;AACV,YAAM,OAAQ,KAAkC;AAChD,UAAI,SAAS,0BAA0B,SAAS,oBAAoB;AAChE,aAAK;AAAA,UACD;AAAA;AAAA,QAGJ;AAAA,MACJ;AACA,YAAM;AAAA,IACV;AAEA,SAAK,IAAI,kBAAa,IAAI,IAAI,SAAI;AAClC,UAAM,YAAY,KAAK,IAAI;AAC3B,UAAM,SAAS,MAAM,cAAc,KAAK,IAAI;AAE5C,QAAI,CAAC,MAAM,OAAO;AACd,iBAAW,QAAQ,OAAO,MAAM;AAC5B,cAAM,SAAS,KAAK,UAAU,WAAW,KAAK,UAAU,SAAS,IAAI,KAAK,KAAK,MAAM;AACrF,aAAK,IAAI,KAAK,MAAM,IAAI,KAAK,OAAO,EAAE;AAAA,MAC1C;AAAA,IACJ;AAEA,UAAM,YAAY,OAAO,cAAc,KAAK,IAAI,IAAI,aAAa,KAAM,QAAQ,CAAC;AAChF,QAAI,OAAO,IAAI;AACX,WAAK,IAAI,UAAK,IAAI,IAAI,iBAAiB,OAAO,MAAM,OAAO,KAAK,MAAM,YAAY,OAAO,KAAK,WAAW,IAAI,KAAK,GAAG,GAAG;AAAA,IAC5H,OAAO;AACH,WAAK,MAAM,UAAK,IAAI,IAAI,iBAAiB,OAAO,MAAM,OAAO,KAAK,EAAE;AAAA,IACxE;AAAA,EACJ;AACJ;",
+  "mappings": ";;;;;;;;;;;;;;;AAAA,SAAS,MAAM,aAAa;AAQ5B,IAAqB,WAArB,MAAqB,kBAAiB,YAAY;AAAA,EAC9C,OAAgB,cAAc;AAAA,EAE9B,OAAgB,WAAW;AAAA,IACvB;AAAA,IACA;AAAA,EACJ;AAAA,EAEA,OAAgB,OAAO;AAAA,IACnB,MAAM,KAAK,OAAO;AAAA,MACd,aAAa;AAAA,MACb,UAAU;AAAA,IACd,CAAC;AAAA,EACL;AAAA,EAEA,OAAgB,QAAQ;AAAA,IACpB,OAAO,MAAM,QAAQ;AAAA,MACjB,MAAM;AAAA,MACN,aAAa;AAAA,IACjB,CAAC;AAAA,EACL;AAAA,EAEA,MAAM,MAAqB;AACvB,UAAM,EAAE,MAAM,MAAM,IAAI,MAAM,KAAK,MAAM,SAAQ;AACjD,SAAK,YAAY;AAIjB,UAAM,MAAM,iBAAiB,KAAK,CAAC,MAAM,EAAE,SAAS,KAAK,IAAI;AAC7D,QAAI,CAAC,KAAK;AACN,WAAK,MAAM,iBAAiB,KAAK,IAAI;AAAA,sDAAyD;AAAA,IAClG;AAOA,QAAI;AACJ,QAAI;AACA,OAAC,EAAE,cAAc,IAAI,MAAM,OAAO,2BAA2B;AAAA,IACjE,SAAS,KAAK;AACV,YAAM,OAAQ,KAAkC;AAChD,UAAI,SAAS,0BAA0B,SAAS,oBAAoB;AAChE,aAAK;AAAA,UACD;AAAA;AAAA,QAGJ;AAAA,MACJ;AACA,YAAM;AAAA,IACV;AAEA,SAAK,IAAI,kBAAa,IAAI,IAAI,SAAI;AAClC,UAAM,YAAY,KAAK,IAAI;AAC3B,UAAM,SAAS,MAAM,cAAc,KAAK,IAAI;AAE5C,QAAI,CAAC,MAAM,OAAO;AACd,iBAAW,QAAQ,OAAO,MAAM;AAC5B,cAAM,SAAS,KAAK,UAAU,WAAW,KAAK,UAAU,SAAS,IAAI,KAAK,KAAK,MAAM;AACrF,aAAK,IAAI,KAAK,MAAM,IAAI,KAAK,OAAO,EAAE;AAAA,MAC1C;AAAA,IACJ;AAEA,UAAM,YAAY,OAAO,cAAc,KAAK,IAAI,IAAI,aAAa,KAAM,QAAQ,CAAC;AAChF,QAAI,OAAO,IAAI;AACX,WAAK,IAAI,UAAK,IAAI,IAAI,iBAAiB,OAAO,MAAM,OAAO,KAAK,MAAM,YAAY,OAAO,KAAK,WAAW,IAAI,KAAK,GAAG,GAAG;AAAA,IAC5H,OAAO;AACH,WAAK,MAAM,UAAK,IAAI,IAAI,iBAAiB,OAAO,MAAM,OAAO,KAAK,EAAE;AAAA,IACxE;AAAA,EACJ;AACJ;",
   "names": []
 }

package/dist/commands/dashboard.js

@@ -1,10 +1,11 @@
 import {
   BaseCommand
-} from "../chunk-SZ7VDCD6.js";
+} from "../chunk-2HK7RGZB.js";
 import "../chunk-R56PH7YE.js";
-import "../chunk-SLL3OBKN.js";
+import "../chunk-GQNJO2P2.js";
+import "../chunk-37LN5O7M.js";
 import "../chunk-EKJ2CABV.js";
-import "../chunk-ZRLYHBPN.js";
+import "../chunk-ZUJBM6HE.js";
 import "../chunk-W2YERO7E.js";
 import "../chunk-NSBPE2FW.js";
 

package/dist/commands/dashboard.js.map

@@ -2,6 +2,6 @@
   "version": 3,
   "sources": ["../../src/commands/dashboard.ts"],
   "sourcesContent": ["import { Flags } from '@oclif/core';\nimport { BaseCommand } from '../base-command.js';\n\nexport default class Dashboard extends BaseCommand {\n  static override description = 'Interactive TUI dashboard showing session analytics, outcomes, friction, and insights';\n\n  static override examples = [\n    '<%= config.bin %> dashboard',\n    '<%= config.bin %> dashboard --period 30days',\n    '<%= config.bin %> dashboard --launcher',\n  ];\n\n  static override flags = {\n    period: Flags.string({\n      char: 'p',\n      default: 'week',\n      options: ['today', 'week', '30days', 'month', 'all'],\n      description: 'Initial time period',\n    }),\n    launcher: Flags.boolean({\n      char: 'l',\n      default: false,\n      description: 'Open the launcher overlay immediately',\n    }),\n  };\n\n  async run(): Promise<void> {\n    const { flags } = await this.parse(Dashboard);\n    const { renderDashboard } = await import('../dashboard/Dashboard.js');\n    await renderDashboard(\n      flags.period as 'today' | 'week' | '30days' | 'month' | 'all',\n      { openLauncher: flags.launcher }\n    );\n  }\n}\n"],
-  "mappings": ";;;;;;;;;;;AAAA,SAAS,aAAa;AAGtB,IAAqB,YAArB,MAAqB,mBAAkB,YAAY;AAAA,EACjD,OAAgB,cAAc;AAAA,EAE9B,OAAgB,WAAW;AAAA,IACzB;AAAA,IACA;AAAA,IACA;AAAA,EACF;AAAA,EAEA,OAAgB,QAAQ;AAAA,IACtB,QAAQ,MAAM,OAAO;AAAA,MACnB,MAAM;AAAA,MACN,SAAS;AAAA,MACT,SAAS,CAAC,SAAS,QAAQ,UAAU,SAAS,KAAK;AAAA,MACnD,aAAa;AAAA,IACf,CAAC;AAAA,IACD,UAAU,MAAM,QAAQ;AAAA,MACtB,MAAM;AAAA,MACN,SAAS;AAAA,MACT,aAAa;AAAA,IACf,CAAC;AAAA,EACH;AAAA,EAEA,MAAM,MAAqB;AACzB,UAAM,EAAE,MAAM,IAAI,MAAM,KAAK,MAAM,UAAS;AAC5C,UAAM,EAAE,gBAAgB,IAAI,MAAM,OAAO,2BAA2B;AACpE,UAAM;AAAA,MACJ,MAAM;AAAA,MACN,EAAE,cAAc,MAAM,SAAS;AAAA,IACjC;AAAA,EACF;AACF;",
+  "mappings": ";;;;;;;;;;;;AAAA,SAAS,aAAa;AAGtB,IAAqB,YAArB,MAAqB,mBAAkB,YAAY;AAAA,EACjD,OAAgB,cAAc;AAAA,EAE9B,OAAgB,WAAW;AAAA,IACzB;AAAA,IACA;AAAA,IACA;AAAA,EACF;AAAA,EAEA,OAAgB,QAAQ;AAAA,IACtB,QAAQ,MAAM,OAAO;AAAA,MACnB,MAAM;AAAA,MACN,SAAS;AAAA,MACT,SAAS,CAAC,SAAS,QAAQ,UAAU,SAAS,KAAK;AAAA,MACnD,aAAa;AAAA,IACf,CAAC;AAAA,IACD,UAAU,MAAM,QAAQ;AAAA,MACtB,MAAM;AAAA,MACN,SAAS;AAAA,MACT,aAAa;AAAA,IACf,CAAC;AAAA,EACH;AAAA,EAEA,MAAM,MAAqB;AACzB,UAAM,EAAE,MAAM,IAAI,MAAM,KAAK,MAAM,UAAS;AAC5C,UAAM,EAAE,gBAAgB,IAAI,MAAM,OAAO,2BAA2B;AACpE,UAAM;AAAA,MACJ,MAAM;AAAA,MACN,EAAE,cAAc,MAAM,SAAS;AAAA,IACjC;AAAA,EACF;AACF;",
   "names": []
 }

package/dist/commands/diagnostics.js

@@ -0,0 +1,117 @@
+import {
+  diagnose,
+  isProxyBreakingNode
+} from "../chunk-J6QFW4LD.js";
+import {
+  GATEWAY_PORT,
+  MITMPROXY_PORT,
+  isGatewayListening
+} from "../chunk-4LYCDS2N.js";
+import "../chunk-7LVVKQMZ.js";
+import "../chunk-5EDMGK3W.js";
+import {
+  BaseCommand
+} from "../chunk-2HK7RGZB.js";
+import "../chunk-R56PH7YE.js";
+import "../chunk-GQNJO2P2.js";
+import "../chunk-37LN5O7M.js";
+import "../chunk-EKJ2CABV.js";
+import {
+  DEFAULT_API_URL
+} from "../chunk-ZUJBM6HE.js";
+import {
+  isAuthenticated,
+  loadEnvFileAuth
+} from "../chunk-W2YERO7E.js";
+import "../chunk-NSBPE2FW.js";
+
+// src/commands/diagnostics.ts
+import { spawnSync } from "node:child_process";
+import { existsSync, readFileSync } from "node:fs";
+import { join } from "node:path";
+var Diagnostics = class _Diagnostics extends BaseCommand {
+  static description = 'Diagnose capture / proxy / connectivity problems (e.g. a client-side "502 Bad Gateway")';
+  static examples = ["<%= config.bin %> diagnostics"];
+  baseUrl() {
+    return process.env.RULEMETRIC_API_URL ?? loadEnvFileAuth().apiUrl ?? DEFAULT_API_URL;
+  }
+  /** GET /health with the proxy bypassed — "is production reachable at all?" */
+  async checkDirect() {
+    const savedHttps = process.env.HTTPS_PROXY;
+    const savedHttp = process.env.HTTP_PROXY;
+    delete process.env.HTTPS_PROXY;
+    delete process.env.HTTP_PROXY;
+    const controller = new AbortController();
+    const timer = setTimeout(() => controller.abort(), 8e3);
+    try {
+      const res = await globalThis.fetch(`${this.baseUrl()}/health`, { signal: controller.signal });
+      return res.ok;
+    } catch {
+      return false;
+    } finally {
+      clearTimeout(timer);
+      if (savedHttps !== void 0) process.env.HTTPS_PROXY = savedHttps;
+      if (savedHttp !== void 0) process.env.HTTP_PROXY = savedHttp;
+    }
+  }
+  /** Resolve the configured proxy: env HTTPS_PROXY, else the project settings files. */
+  configuredProxy() {
+    if (process.env.HTTPS_PROXY) return process.env.HTTPS_PROXY;
+    for (const fileName of ["settings.local.json", "settings.json"]) {
+      const settingsPath = join(process.cwd(), ".claude", fileName);
+      if (!existsSync(settingsPath)) continue;
+      try {
+        const parsed = JSON.parse(readFileSync(settingsPath, "utf8"));
+        if (parsed.env?.HTTPS_PROXY) return parsed.env.HTTPS_PROXY;
+      } catch {
+      }
+    }
+    return null;
+  }
+  /** GET /health THROUGH the proxy via curl (Node fetch won't honor HTTPS_PROXY). */
+  checkThroughProxy(proxyUrl) {
+    const res = spawnSync(
+      "curl",
+      ["-sS", "-x", proxyUrl, "--max-time", "8", "-o", "/dev/null", "-w", "%{http_code}", `${this.baseUrl()}/health`],
+      { encoding: "utf8", stdio: ["ignore", "pipe", "ignore"] }
+    );
+    if (res.error || typeof res.stdout !== "string") return null;
+    return res.stdout.trim() === "200";
+  }
+  async run() {
+    await this.parse(_Diagnostics);
+    const nodeVersion = process.version;
+    const proxyUrl = this.configuredProxy();
+    const httpsProxySet = proxyUrl != null;
+    const [directApiOk, gatewayAlive, mitmproxyAlive] = await Promise.all([
+      this.checkDirect(),
+      isGatewayListening(GATEWAY_PORT),
+      isGatewayListening(MITMPROXY_PORT)
+    ]);
+    const proxyPathOk = httpsProxySet && proxyUrl ? this.checkThroughProxy(proxyUrl) : null;
+    const line = (label, ok, detail) => {
+      const mark = ok === null ? "\u2013" : ok ? "[OK]" : "[!!]";
+      this.log(`  ${mark} ${label.padEnd(22)} ${detail}`);
+    };
+    this.log("RuleMetric diagnostics");
+    this.log("");
+    line("Node", !isProxyBreakingNode(nodeVersion), `${nodeVersion}${isProxyBreakingNode(nodeVersion) ? " (< 20.11 \u2014 breaks the bundled proxy)" : ""}`);
+    line("API (direct)", directApiOk, `${this.baseUrl()}/health ${directApiOk ? "reachable" : "UNREACHABLE"}`);
+    line("Authenticated", isAuthenticated(), isAuthenticated() ? "yes" : "no \u2014 run `rulemetric auth login`");
+    line("Proxy configured", httpsProxySet ? true : null, httpsProxySet ? proxyUrl : "no HTTPS_PROXY set");
+    if (httpsProxySet) {
+      line("Gateway :" + GATEWAY_PORT, gatewayAlive, gatewayAlive ? "listening" : "NOT listening");
+      line("mitmproxy :" + MITMPROXY_PORT, mitmproxyAlive, mitmproxyAlive ? "listening" : "not listening (gateway should route direct)");
+      line("API (via proxy)", proxyPathOk, proxyPathOk === null ? "inconclusive (curl unavailable)" : proxyPathOk ? "reachable" : "FAILED \u2014 502/blocked through your proxy");
+    }
+    const d = diagnose({ nodeVersion, directApiOk, httpsProxySet, gatewayAlive, mitmproxyAlive, proxyPathOk });
+    this.log("");
+    const badge = d.severity === "ok" ? "[OK]" : d.severity === "warn" ? "[!!]" : "[XX]";
+    this.log(`${badge} ${d.headline}`);
+    for (const h of d.hints) this.log(`   \u2022 ${h}`);
+  }
+};
+export {
+  Diagnostics as default
+};
+//# sourceMappingURL=diagnostics.js.map

package/dist/commands/diagnostics.js.map

@@ -0,0 +1,7 @@
+{
+  "version": 3,
+  "sources": ["../../src/commands/diagnostics.ts"],
+  "sourcesContent": ["import { spawnSync } from 'node:child_process';\nimport { existsSync, readFileSync } from 'node:fs';\nimport { join } from 'node:path';\nimport { BaseCommand } from '../base-command.js';\nimport { DEFAULT_API_URL } from '../lib/api-client.js';\nimport { loadEnvFileAuth, isAuthenticated } from '../lib/auth.js';\nimport { isGatewayListening, GATEWAY_PORT, MITMPROXY_PORT } from '../lib/gateway-lifecycle.js';\nimport { diagnose, isProxyBreakingNode } from '../lib/diagnostics.js';\n\n/**\n * `rulemetric diagnostics` \u2014 a doctor for the \"capture isn't working / API\n * returns 502\" class. Its core move is comparing DIRECT reachability against\n * reachability THROUGH the local proxy: when direct works but the proxy path\n * fails, the user's own gateway (:8787) is producing the 502, not production.\n */\nexport default class Diagnostics extends BaseCommand {\n  static override description =\n    'Diagnose capture / proxy / connectivity problems (e.g. a client-side \"502 Bad Gateway\")';\n\n  static override examples = ['<%= config.bin %> diagnostics'];\n\n  private baseUrl(): string {\n    return process.env.RULEMETRIC_API_URL ?? loadEnvFileAuth().apiUrl ?? DEFAULT_API_URL;\n  }\n\n  /** GET /health with the proxy bypassed \u2014 \"is production reachable at all?\" */\n  private async checkDirect(): Promise<boolean> {\n    const savedHttps = process.env.HTTPS_PROXY;\n    const savedHttp = process.env.HTTP_PROXY;\n    delete process.env.HTTPS_PROXY;\n    delete process.env.HTTP_PROXY;\n    const controller = new AbortController();\n    const timer = setTimeout(() => controller.abort(), 8000);\n    try {\n      const res = await globalThis.fetch(`${this.baseUrl()}/health`, { signal: controller.signal });\n      return res.ok;\n    } catch {\n      return false;\n    } finally {\n      clearTimeout(timer);\n      if (savedHttps !== undefined) process.env.HTTPS_PROXY = savedHttps;\n      if (savedHttp !== undefined) process.env.HTTP_PROXY = savedHttp;\n    }\n  }\n\n  /** Resolve the configured proxy: env HTTPS_PROXY, else the project settings files. */\n  private configuredProxy(): string | null {\n    if (process.env.HTTPS_PROXY) return process.env.HTTPS_PROXY;\n    // settings.local.json (machine-local, where current installs write the\n    // proxy env) wins over the shared settings.json (legacy installs) \u2014\n    // mirrors Claude Code's merge order.\n    for (const fileName of ['settings.local.json', 'settings.json']) {\n      const settingsPath = join(process.cwd(), '.claude', fileName);\n      if (!existsSync(settingsPath)) continue;\n      try {\n        const parsed = JSON.parse(readFileSync(settingsPath, 'utf8')) as {\n          env?: { HTTPS_PROXY?: string };\n        };\n        if (parsed.env?.HTTPS_PROXY) return parsed.env.HTTPS_PROXY;\n      } catch {\n        /* ignore malformed settings */\n      }\n    }\n    return null;\n  }\n\n  /** GET /health THROUGH the proxy via curl (Node fetch won't honor HTTPS_PROXY). */\n  private checkThroughProxy(proxyUrl: string): boolean | null {\n    const res = spawnSync(\n      'curl',\n      ['-sS', '-x', proxyUrl, '--max-time', '8', '-o', '/dev/null', '-w', '%{http_code}', `${this.baseUrl()}/health`],\n      { encoding: 'utf8', stdio: ['ignore', 'pipe', 'ignore'] },\n    );\n    if (res.error || typeof res.stdout !== 'string') return null; // curl missing \u2192 inconclusive\n    return res.stdout.trim() === '200';\n  }\n\n  async run(): Promise<void> {\n    await this.parse(Diagnostics);\n\n    const nodeVersion = process.version;\n    const proxyUrl = this.configuredProxy();\n    const httpsProxySet = proxyUrl != null;\n\n    const [directApiOk, gatewayAlive, mitmproxyAlive] = await Promise.all([\n      this.checkDirect(),\n      isGatewayListening(GATEWAY_PORT),\n      isGatewayListening(MITMPROXY_PORT),\n    ]);\n    const proxyPathOk = httpsProxySet && proxyUrl ? this.checkThroughProxy(proxyUrl) : null;\n\n    const line = (label: string, ok: boolean | null, detail: string) => {\n      const mark = ok === null ? '\u2013' : ok ? '[OK]' : '[!!]';\n      this.log(`  ${mark} ${label.padEnd(22)} ${detail}`);\n    };\n\n    this.log('RuleMetric diagnostics');\n    this.log('');\n    line('Node', !isProxyBreakingNode(nodeVersion), `${nodeVersion}${isProxyBreakingNode(nodeVersion) ? ' (< 20.11 \u2014 breaks the bundled proxy)' : ''}`);\n    line('API (direct)', directApiOk, `${this.baseUrl()}/health ${directApiOk ? 'reachable' : 'UNREACHABLE'}`);\n    line('Authenticated', isAuthenticated(), isAuthenticated() ? 'yes' : 'no \u2014 run `rulemetric auth login`');\n    line('Proxy configured', httpsProxySet ? true : null, httpsProxySet ? (proxyUrl as string) : 'no HTTPS_PROXY set');\n    if (httpsProxySet) {\n      line('Gateway :' + GATEWAY_PORT, gatewayAlive, gatewayAlive ? 'listening' : 'NOT listening');\n      line('mitmproxy :' + MITMPROXY_PORT, mitmproxyAlive, mitmproxyAlive ? 'listening' : 'not listening (gateway should route direct)');\n      line('API (via proxy)', proxyPathOk, proxyPathOk === null ? 'inconclusive (curl unavailable)' : proxyPathOk ? 'reachable' : 'FAILED \u2014 502/blocked through your proxy');\n    }\n\n    const d = diagnose({ nodeVersion, directApiOk, httpsProxySet, gatewayAlive, mitmproxyAlive, proxyPathOk });\n    this.log('');\n    const badge = d.severity === 'ok' ? '[OK]' : d.severity === 'warn' ? '[!!]' : '[XX]';\n    this.log(`${badge} ${d.headline}`);\n    for (const h of d.hints) this.log(`   \u2022 ${h}`);\n  }\n}\n"],
+  "mappings": ";;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA,SAAS,iBAAiB;AAC1B,SAAS,YAAY,oBAAoB;AACzC,SAAS,YAAY;AAarB,IAAqB,cAArB,MAAqB,qBAAoB,YAAY;AAAA,EACnD,OAAgB,cACd;AAAA,EAEF,OAAgB,WAAW,CAAC,+BAA+B;AAAA,EAEnD,UAAkB;AACxB,WAAO,QAAQ,IAAI,sBAAsB,gBAAgB,EAAE,UAAU;AAAA,EACvE;AAAA;AAAA,EAGA,MAAc,cAAgC;AAC5C,UAAM,aAAa,QAAQ,IAAI;AAC/B,UAAM,YAAY,QAAQ,IAAI;AAC9B,WAAO,QAAQ,IAAI;AACnB,WAAO,QAAQ,IAAI;AACnB,UAAM,aAAa,IAAI,gBAAgB;AACvC,UAAM,QAAQ,WAAW,MAAM,WAAW,MAAM,GAAG,GAAI;AACvD,QAAI;AACF,YAAM,MAAM,MAAM,WAAW,MAAM,GAAG,KAAK,QAAQ,CAAC,WAAW,EAAE,QAAQ,WAAW,OAAO,CAAC;AAC5F,aAAO,IAAI;AAAA,IACb,QAAQ;AACN,aAAO;AAAA,IACT,UAAE;AACA,mBAAa,KAAK;AAClB,UAAI,eAAe,OAAW,SAAQ,IAAI,cAAc;AACxD,UAAI,cAAc,OAAW,SAAQ,IAAI,aAAa;AAAA,IACxD;AAAA,EACF;AAAA;AAAA,EAGQ,kBAAiC;AACvC,QAAI,QAAQ,IAAI,YAAa,QAAO,QAAQ,IAAI;AAIhD,eAAW,YAAY,CAAC,uBAAuB,eAAe,GAAG;AAC/D,YAAM,eAAe,KAAK,QAAQ,IAAI,GAAG,WAAW,QAAQ;AAC5D,UAAI,CAAC,WAAW,YAAY,EAAG;AAC/B,UAAI;AACF,cAAM,SAAS,KAAK,MAAM,aAAa,cAAc,MAAM,CAAC;AAG5D,YAAI,OAAO,KAAK,YAAa,QAAO,OAAO,IAAI;AAAA,MACjD,QAAQ;AAAA,MAER;AAAA,IACF;AACA,WAAO;AAAA,EACT;AAAA;AAAA,EAGQ,kBAAkB,UAAkC;AAC1D,UAAM,MAAM;AAAA,MACV;AAAA,MACA,CAAC,OAAO,MAAM,UAAU,cAAc,KAAK,MAAM,aAAa,MAAM,gBAAgB,GAAG,KAAK,QAAQ,CAAC,SAAS;AAAA,MAC9G,EAAE,UAAU,QAAQ,OAAO,CAAC,UAAU,QAAQ,QAAQ,EAAE;AAAA,IAC1D;AACA,QAAI,IAAI,SAAS,OAAO,IAAI,WAAW,SAAU,QAAO;AACxD,WAAO,IAAI,OAAO,KAAK,MAAM;AAAA,EAC/B;AAAA,EAEA,MAAM,MAAqB;AACzB,UAAM,KAAK,MAAM,YAAW;AAE5B,UAAM,cAAc,QAAQ;AAC5B,UAAM,WAAW,KAAK,gBAAgB;AACtC,UAAM,gBAAgB,YAAY;AAElC,UAAM,CAAC,aAAa,cAAc,cAAc,IAAI,MAAM,QAAQ,IAAI;AAAA,MACpE,KAAK,YAAY;AAAA,MACjB,mBAAmB,YAAY;AAAA,MAC/B,mBAAmB,cAAc;AAAA,IACnC,CAAC;AACD,UAAM,cAAc,iBAAiB,WAAW,KAAK,kBAAkB,QAAQ,IAAI;AAEnF,UAAM,OAAO,CAAC,OAAe,IAAoB,WAAmB;AAClE,YAAM,OAAO,OAAO,OAAO,WAAM,KAAK,SAAS;AAC/C,WAAK,IAAI,KAAK,IAAI,IAAI,MAAM,OAAO,EAAE,CAAC,IAAI,MAAM,EAAE;AAAA,IACpD;AAEA,SAAK,IAAI,wBAAwB;AACjC,SAAK,IAAI,EAAE;AACX,SAAK,QAAQ,CAAC,oBAAoB,WAAW,GAAG,GAAG,WAAW,GAAG,oBAAoB,WAAW,IAAI,+CAA0C,EAAE,EAAE;AAClJ,SAAK,gBAAgB,aAAa,GAAG,KAAK,QAAQ,CAAC,WAAW,cAAc,cAAc,aAAa,EAAE;AACzG,SAAK,iBAAiB,gBAAgB,GAAG,gBAAgB,IAAI,QAAQ,uCAAkC;AACvG,SAAK,oBAAoB,gBAAgB,OAAO,MAAM,gBAAiB,WAAsB,oBAAoB;AACjH,QAAI,eAAe;AACjB,WAAK,cAAc,cAAc,cAAc,eAAe,cAAc,eAAe;AAC3F,WAAK,gBAAgB,gBAAgB,gBAAgB,iBAAiB,cAAc,6CAA6C;AACjI,WAAK,mBAAmB,aAAa,gBAAgB,OAAO,oCAAoC,cAAc,cAAc,8CAAyC;AAAA,IACvK;AAEA,UAAM,IAAI,SAAS,EAAE,aAAa,aAAa,eAAe,cAAc,gBAAgB,YAAY,CAAC;AACzG,SAAK,IAAI,EAAE;AACX,UAAM,QAAQ,EAAE,aAAa,OAAO,SAAS,EAAE,aAAa,SAAS,SAAS;AAC9E,SAAK,IAAI,GAAG,KAAK,IAAI,EAAE,QAAQ,EAAE;AACjC,eAAW,KAAK,EAAE,MAAO,MAAK,IAAI,aAAQ,CAAC,EAAE;AAAA,EAC/C;AACF;",
+  "names": []
+}

package/dist/commands/eval-targets/create.js

@@ -1,12 +1,13 @@
 import {
   BaseCommand
-} from "../../chunk-SZ7VDCD6.js";
+} from "../../chunk-2HK7RGZB.js";
 import "../../chunk-R56PH7YE.js";
-import "../../chunk-SLL3OBKN.js";
+import "../../chunk-GQNJO2P2.js";
+import "../../chunk-37LN5O7M.js";
 import "../../chunk-EKJ2CABV.js";
 import {
   apiPost
-} from "../../chunk-ZRLYHBPN.js";
+} from "../../chunk-ZUJBM6HE.js";
 import "../../chunk-W2YERO7E.js";
 import "../../chunk-NSBPE2FW.js";
 

package/dist/commands/eval-targets/create.js.map

@@ -2,6 +2,6 @@
   "version": 3,
   "sources": ["../../../src/commands/eval-targets/create.ts"],
   "sourcesContent": ["import { Flags } from '@oclif/core';\nimport { readFileSync } from 'node:fs';\nimport { BaseCommand } from '../../base-command.js';\nimport { apiPost } from '../../lib/api-client.js';\n\nexport default class EvalTargetCreate extends BaseCommand {\n  static override description = 'Create an eval target (instruction, skill, prompt, or plan to test)';\n\n  static override examples = [\n    '<%= config.bin %> eval-targets create --name my-rules --type instruction --file ./CLAUDE.md',\n    '<%= config.bin %> eval-targets create --name my-prompt --type prompt --content \"Always use TypeScript\"',\n  ];\n\n  static override flags = {\n    name: Flags.string({ char: 'n', description: 'Name for the eval target', required: true }),\n    type: Flags.string({ char: 't', description: 'Target type', options: ['instruction', 'skill', 'prompt', 'plan'], required: true }),\n    file: Flags.string({ char: 'f', description: 'Path to content file' }),\n    content: Flags.string({ char: 'c', description: 'Content string (alternative to --file)' }),\n    description: Flags.string({ char: 'd', description: 'Description of the target', default: '' }),\n  };\n\n  async run(): Promise<void> {\n    const { flags } = await this.parse(EvalTargetCreate);\n    this.requireAuth();\n\n    let content: string;\n    if (flags.file) {\n      content = readFileSync(flags.file, 'utf-8');\n    } else if (flags.content) {\n      content = flags.content;\n    } else {\n      this.error('Provide either --file or --content');\n    }\n\n    const result = await apiPost<{ id: string; name: string; version: number }>('/api/eval-targets', {\n      name: flags.name,\n      type: flags.type,\n      content,\n      description: flags.description,\n    });\n\n    this.log(`Created eval target: ${result.name} (${result.id}) v${result.version}`);\n  }\n}\n"],
-  "mappings": ";;;;;;;;;;;;;AAAA,SAAS,aAAa;AACtB,SAAS,oBAAoB;AAI7B,IAAqB,mBAArB,MAAqB,0BAAyB,YAAY;AAAA,EACxD,OAAgB,cAAc;AAAA,EAE9B,OAAgB,WAAW;AAAA,IACzB;AAAA,IACA;AAAA,EACF;AAAA,EAEA,OAAgB,QAAQ;AAAA,IACtB,MAAM,MAAM,OAAO,EAAE,MAAM,KAAK,aAAa,4BAA4B,UAAU,KAAK,CAAC;AAAA,IACzF,MAAM,MAAM,OAAO,EAAE,MAAM,KAAK,aAAa,eAAe,SAAS,CAAC,eAAe,SAAS,UAAU,MAAM,GAAG,UAAU,KAAK,CAAC;AAAA,IACjI,MAAM,MAAM,OAAO,EAAE,MAAM,KAAK,aAAa,uBAAuB,CAAC;AAAA,IACrE,SAAS,MAAM,OAAO,EAAE,MAAM,KAAK,aAAa,yCAAyC,CAAC;AAAA,IAC1F,aAAa,MAAM,OAAO,EAAE,MAAM,KAAK,aAAa,6BAA6B,SAAS,GAAG,CAAC;AAAA,EAChG;AAAA,EAEA,MAAM,MAAqB;AACzB,UAAM,EAAE,MAAM,IAAI,MAAM,KAAK,MAAM,iBAAgB;AACnD,SAAK,YAAY;AAEjB,QAAI;AACJ,QAAI,MAAM,MAAM;AACd,gBAAU,aAAa,MAAM,MAAM,OAAO;AAAA,IAC5C,WAAW,MAAM,SAAS;AACxB,gBAAU,MAAM;AAAA,IAClB,OAAO;AACL,WAAK,MAAM,oCAAoC;AAAA,IACjD;AAEA,UAAM,SAAS,MAAM,QAAuD,qBAAqB;AAAA,MAC/F,MAAM,MAAM;AAAA,MACZ,MAAM,MAAM;AAAA,MACZ;AAAA,MACA,aAAa,MAAM;AAAA,IACrB,CAAC;AAED,SAAK,IAAI,wBAAwB,OAAO,IAAI,KAAK,OAAO,EAAE,MAAM,OAAO,OAAO,EAAE;AAAA,EAClF;AACF;",
+  "mappings": ";;;;;;;;;;;;;;AAAA,SAAS,aAAa;AACtB,SAAS,oBAAoB;AAI7B,IAAqB,mBAArB,MAAqB,0BAAyB,YAAY;AAAA,EACxD,OAAgB,cAAc;AAAA,EAE9B,OAAgB,WAAW;AAAA,IACzB;AAAA,IACA;AAAA,EACF;AAAA,EAEA,OAAgB,QAAQ;AAAA,IACtB,MAAM,MAAM,OAAO,EAAE,MAAM,KAAK,aAAa,4BAA4B,UAAU,KAAK,CAAC;AAAA,IACzF,MAAM,MAAM,OAAO,EAAE,MAAM,KAAK,aAAa,eAAe,SAAS,CAAC,eAAe,SAAS,UAAU,MAAM,GAAG,UAAU,KAAK,CAAC;AAAA,IACjI,MAAM,MAAM,OAAO,EAAE,MAAM,KAAK,aAAa,uBAAuB,CAAC;AAAA,IACrE,SAAS,MAAM,OAAO,EAAE,MAAM,KAAK,aAAa,yCAAyC,CAAC;AAAA,IAC1F,aAAa,MAAM,OAAO,EAAE,MAAM,KAAK,aAAa,6BAA6B,SAAS,GAAG,CAAC;AAAA,EAChG;AAAA,EAEA,MAAM,MAAqB;AACzB,UAAM,EAAE,MAAM,IAAI,MAAM,KAAK,MAAM,iBAAgB;AACnD,SAAK,YAAY;AAEjB,QAAI;AACJ,QAAI,MAAM,MAAM;AACd,gBAAU,aAAa,MAAM,MAAM,OAAO;AAAA,IAC5C,WAAW,MAAM,SAAS;AACxB,gBAAU,MAAM;AAAA,IAClB,OAAO;AACL,WAAK,MAAM,oCAAoC;AAAA,IACjD;AAEA,UAAM,SAAS,MAAM,QAAuD,qBAAqB;AAAA,MAC/F,MAAM,MAAM;AAAA,MACZ,MAAM,MAAM;AAAA,MACZ;AAAA,MACA,aAAa,MAAM;AAAA,IACrB,CAAC;AAED,SAAK,IAAI,wBAAwB,OAAO,IAAI,KAAK,OAAO,EAAE,MAAM,OAAO,OAAO,EAAE;AAAA,EAClF;AACF;",
   "names": []
 }

package/dist/commands/eval-targets/list.js

@@ -4,13 +4,14 @@ import {
 } from "../../chunk-BZDKEQKU.js";
 import {
   BaseCommand
-} from "../../chunk-SZ7VDCD6.js";
+} from "../../chunk-2HK7RGZB.js";
 import "../../chunk-R56PH7YE.js";
-import "../../chunk-SLL3OBKN.js";
+import "../../chunk-GQNJO2P2.js";
+import "../../chunk-37LN5O7M.js";
 import "../../chunk-EKJ2CABV.js";
 import {
   apiGet
-} from "../../chunk-ZRLYHBPN.js";
+} from "../../chunk-ZUJBM6HE.js";
 import "../../chunk-W2YERO7E.js";
 import "../../chunk-NSBPE2FW.js";
 

package/dist/commands/eval-targets/list.js.map

@@ -2,6 +2,6 @@
   "version": 3,
   "sources": ["../../../src/commands/eval-targets/list.ts"],
   "sourcesContent": ["import { Flags } from '@oclif/core';\nimport { printTable } from '@oclif/table';\nimport { BaseCommand } from '../../base-command.js';\nimport { apiGet } from '../../lib/api-client.js';\nimport { formatDate, truncate } from '../../lib/output.js';\n\ninterface EvalTarget {\n  id: string;\n  name: string;\n  type: string;\n  version: number;\n  description: string;\n  updatedAt: string;\n}\n\nexport default class EvalTargetList extends BaseCommand {\n  static override description = 'List eval targets';\n\n  static override examples = [\n    '<%= config.bin %> eval-targets list',\n    '<%= config.bin %> eval-targets list --type instruction',\n  ];\n\n  static override flags = {\n    type: Flags.string({ char: 't', description: 'Filter by type', options: ['instruction', 'skill', 'prompt', 'plan'] }),\n  };\n\n  async run(): Promise<void> {\n    const { flags } = await this.parse(EvalTargetList);\n    this.requireAuth();\n\n    const query = flags.type ? `?type=${flags.type}` : '';\n    const results = await apiGet<EvalTarget[]>(`/api/eval-targets${query}`);\n\n    if (results.length === 0) {\n      this.log('No eval targets found. Create one with `rulemetric eval-targets create`.');\n      return;\n    }\n\n    printTable({\n      data: results.map((t) => ({\n        Name: truncate(t.name, 30),\n        Type: t.type,\n        Version: `v${t.version}`,\n        Description: truncate(t.description, 40),\n        Updated: formatDate(t.updatedAt),\n      })),\n      columns: ['Name', 'Type', 'Version', 'Description', 'Updated'],\n    });\n  }\n}\n"],
-  "mappings": ";;;;;;;;;;;;;;;;;AAAA,SAAS,aAAa;AACtB,SAAS,kBAAkB;AAc3B,IAAqB,iBAArB,MAAqB,wBAAuB,YAAY;AAAA,EACtD,OAAgB,cAAc;AAAA,EAE9B,OAAgB,WAAW;AAAA,IACzB;AAAA,IACA;AAAA,EACF;AAAA,EAEA,OAAgB,QAAQ;AAAA,IACtB,MAAM,MAAM,OAAO,EAAE,MAAM,KAAK,aAAa,kBAAkB,SAAS,CAAC,eAAe,SAAS,UAAU,MAAM,EAAE,CAAC;AAAA,EACtH;AAAA,EAEA,MAAM,MAAqB;AACzB,UAAM,EAAE,MAAM,IAAI,MAAM,KAAK,MAAM,eAAc;AACjD,SAAK,YAAY;AAEjB,UAAM,QAAQ,MAAM,OAAO,SAAS,MAAM,IAAI,KAAK;AACnD,UAAM,UAAU,MAAM,OAAqB,oBAAoB,KAAK,EAAE;AAEtE,QAAI,QAAQ,WAAW,GAAG;AACxB,WAAK,IAAI,0EAA0E;AACnF;AAAA,IACF;AAEA,eAAW;AAAA,MACT,MAAM,QAAQ,IAAI,CAAC,OAAO;AAAA,QACxB,MAAM,SAAS,EAAE,MAAM,EAAE;AAAA,QACzB,MAAM,EAAE;AAAA,QACR,SAAS,IAAI,EAAE,OAAO;AAAA,QACtB,aAAa,SAAS,EAAE,aAAa,EAAE;AAAA,QACvC,SAAS,WAAW,EAAE,SAAS;AAAA,MACjC,EAAE;AAAA,MACF,SAAS,CAAC,QAAQ,QAAQ,WAAW,eAAe,SAAS;AAAA,IAC/D,CAAC;AAAA,EACH;AACF;",
+  "mappings": ";;;;;;;;;;;;;;;;;;AAAA,SAAS,aAAa;AACtB,SAAS,kBAAkB;AAc3B,IAAqB,iBAArB,MAAqB,wBAAuB,YAAY;AAAA,EACtD,OAAgB,cAAc;AAAA,EAE9B,OAAgB,WAAW;AAAA,IACzB;AAAA,IACA;AAAA,EACF;AAAA,EAEA,OAAgB,QAAQ;AAAA,IACtB,MAAM,MAAM,OAAO,EAAE,MAAM,KAAK,aAAa,kBAAkB,SAAS,CAAC,eAAe,SAAS,UAAU,MAAM,EAAE,CAAC;AAAA,EACtH;AAAA,EAEA,MAAM,MAAqB;AACzB,UAAM,EAAE,MAAM,IAAI,MAAM,KAAK,MAAM,eAAc;AACjD,SAAK,YAAY;AAEjB,UAAM,QAAQ,MAAM,OAAO,SAAS,MAAM,IAAI,KAAK;AACnD,UAAM,UAAU,MAAM,OAAqB,oBAAoB,KAAK,EAAE;AAEtE,QAAI,QAAQ,WAAW,GAAG;AACxB,WAAK,IAAI,0EAA0E;AACnF;AAAA,IACF;AAEA,eAAW;AAAA,MACT,MAAM,QAAQ,IAAI,CAAC,OAAO;AAAA,QACxB,MAAM,SAAS,EAAE,MAAM,EAAE;AAAA,QACzB,MAAM,EAAE;AAAA,QACR,SAAS,IAAI,EAAE,OAAO;AAAA,QACtB,aAAa,SAAS,EAAE,aAAa,EAAE;AAAA,QACvC,SAAS,WAAW,EAAE,SAAS;AAAA,MACjC,EAAE;AAAA,MACF,SAAS,CAAC,QAAQ,QAAQ,WAAW,eAAe,SAAS;AAAA,IAC/D,CAAC;AAAA,EACH;AACF;",
   "names": []
 }

package/dist/commands/evals/agent.js

@@ -1,29 +1,29 @@
 import {
   runAgentLoop
-} from "../../chunk-BBLTN35F.js";
-import "../../chunk-JEJ3J5J6.js";
-import "../../chunk-RCMXTLQF.js";
-import "../../chunk-N5A2IQAK.js";
-import "../../chunk-OTOWLUOJ.js";
-import "../../chunk-5ADHO5QF.js";
+} from "../../chunk-7UDETV3P.js";
+import "../../chunk-GIUBARWS.js";
+import "../../chunk-YOBA3NAM.js";
+import "../../chunk-OO7JDFS4.js";
+import "../../chunk-QSN77T7C.js";
+import "../../chunk-XK42ZBDE.js";
+import "../../chunk-3LKVS6W5.js";
+import "../../chunk-4YAP7RT7.js";
+import "../../chunk-XZXS2W24.js";
+import "../../chunk-JULOLTZS.js";
+import "../../chunk-6NBS5XFS.js";
+import {
+  closeDb
+} from "../../chunk-EI5BHWXA.js";
 import "../../chunk-Y4BJXXYV.js";
 import "../../chunk-W53GKIZQ.js";
-import "../../chunk-3QVMIJUL.js";
-import "../../chunk-XZXS2W24.js";
-import "../../chunk-RQ2TMLKG.js";
-import "../../chunk-VBGUNFKY.js";
-import "../../chunk-FZKLLNDS.js";
-import "../../chunk-AHDRN6NE.js";
 import "../../chunk-DGHWRQXL.js";
+import "../../chunk-52WIYRZH.js";
+import "../../chunk-RQ2TMLKG.js";
+import "../../chunk-PMI56ED5.js";
 import "../../chunk-E3BIT53W.js";
-import "../../chunk-QSN77T7C.js";
-import {
-  closeDb
-} from "../../chunk-6EMZI4Y3.js";
-import "../../chunk-3POYC5NA.js";
-import "../../chunk-DR5WLIOU.js";
+import "../../chunk-FZKLLNDS.js";
+import "../../chunk-KJKBDQ4D.js";
 import "../../chunk-J7N3DLH6.js";
-import "../../chunk-52WIYRZH.js";
 import "../../chunk-OQSQC7VB.js";
 import "../../chunk-EKP32DLN.js";
 import "../../chunk-F5N6CLJJ.js";
@@ -33,11 +33,12 @@ import "../../chunk-YNIH3KMU.js";
 import "../../chunk-QN34TGD3.js";
 import {
   BaseCommand
-} from "../../chunk-SZ7VDCD6.js";
+} from "../../chunk-2HK7RGZB.js";
 import "../../chunk-R56PH7YE.js";
-import "../../chunk-SLL3OBKN.js";
+import "../../chunk-GQNJO2P2.js";
+import "../../chunk-37LN5O7M.js";
 import "../../chunk-EKJ2CABV.js";
-import "../../chunk-ZRLYHBPN.js";
+import "../../chunk-ZUJBM6HE.js";
 import {
   getDefaultConfigDir,
   loadToken
@@ -270,12 +271,32 @@ var EvalAgentCommand = class _EvalAgentCommand extends BaseCommand {
     process.on("SIGTERM", () => {
       void shutdown("SIGTERM");
     });
+    let restartForUpgrade = false;
     try {
       await runAgentLoop({
         signal: controller.signal,
         log: (msg) => this.log(msg),
-        version: this.config.version
+        version: this.config.version,
+        readInstalledVersion: () => {
+          try {
+            const pkg = JSON.parse(
+              readFileSync(join(this.config.root, "package.json"), "utf-8")
+            );
+            return pkg.version ?? null;
+          } catch {
+            return null;
+          }
+        },
+        onVersionDrift: () => {
+          restartForUpgrade = true;
+        }
       });
+      if (restartForUpgrade) {
+        this.log("[upgrade] Installed CLI version changed \u2014 restarting worker onto the new code.");
+        await Promise.allSettled([flushSentry(2e3), shutdownTelemetry(), closeDb()]);
+        releaseAgentLock();
+        process.exit(1);
+      }
     } finally {
       releaseAgentLock();
       await closeDb();

package/dist/commands/evals/agent.js.map

@@ -1,7 +1,7 @@
 {
   "version": 3,
   "sources": ["../../../../../packages/telemetry/src/init.ts", "../../../../../packages/telemetry/src/otlp-headers.ts", "../../../../../packages/telemetry/src/hono-middleware.ts", "../../../../../packages/telemetry/src/span-attributes.ts", "../../../../../packages/telemetry/src/sentry.ts", "../../../src/commands/evals/agent.ts"],
-  "sourcesContent": ["import { trace, context, type Tracer } from '@opentelemetry/api';\nimport { NodeSDK } from '@opentelemetry/sdk-node';\nimport { OTLPTraceExporter } from '@opentelemetry/exporter-trace-otlp-http';\nimport {\n  BatchSpanProcessor,\n  ConsoleSpanExporter,\n  SimpleSpanProcessor,\n} from '@opentelemetry/sdk-trace-node';\nimport { Resource } from '@opentelemetry/resources';\nimport { ATTR_SERVICE_NAME, ATTR_SERVICE_VERSION } from '@opentelemetry/semantic-conventions';\nimport { parseOtlpHeaders } from './otlp-headers.js';\n\nlet sdk: NodeSDK | null = null;\n\nexport interface TelemetryConfig {\n  serviceName: string;\n  serviceVersion?: string;\n}\n\n/**\n * Boot states for `initTelemetry`. Returned discriminator lets callers\n * emit a precise boot log:\n *   - 'otlp'    \u2014 spans ship over OTLP to a real backend (Honeycomb etc.)\n *   - 'console' \u2014 spans print to stdout as JSON (Fly logs ingest them);\n *                 the \"open + zero third-party\" default\n *   - 'off'     \u2014 telemetry globally disabled (TELEMETRY_ENABLED=false)\n *                 OR a prior `initTelemetry()` already booted the SDK\n *                 in this process (idempotent skip)\n */\nexport type TelemetryBootState = 'otlp' | 'console' | 'silent' | 'off';\n\nexport function isTelemetryEnabled(): boolean {\n  return process.env.TELEMETRY_ENABLED !== 'false';\n}\n\n/**\n * Initialize OpenTelemetry tracing. Two transport modes, both built on\n * the same OTel SDK \u2014 only the exporter changes:\n *\n *   - OTLP backend  \u2192 set `OTEL_EXPORTER_OTLP_ENDPOINT` (+ optional\n *                     `OTEL_EXPORTER_OTLP_HEADERS` for vendor auth)\n *   - Console (stdout) \u2192 the default when no endpoint is set\n *\n * The console path is intentionally on-by-default. Without a real backend\n * configured, spans are written to stdout as JSON and end up in Fly's log\n * stream alongside everything else \u2014 searchable, structured, and you\n * stay inside the open OpenTelemetry spec without paying a third party.\n * Swap to OTLP later by setting one env var; no code change needed.\n *\n * Returns `'off'` when:\n *   - TELEMETRY_ENABLED=false (operator kill switch \u2014 beats any config)\n *   - The SDK was already started in this process (idempotent skip)\n *\n * See packages/telemetry/src/init.test.ts for the full state table.\n */\nexport function initTelemetry(config: TelemetryConfig): TelemetryBootState {\n  if (!isTelemetryEnabled()) {\n    return 'off';\n  }\n  if (sdk !== null) {\n    // OTel's global tracer provider gets set on first start(); calling\n    // start() twice triggers an SDK warning. Treat second-call as a no-op.\n    return 'off';\n  }\n\n  const resource = new Resource({\n    [ATTR_SERVICE_NAME]: config.serviceName,\n    [ATTR_SERVICE_VERSION]: config.serviceVersion ?? '0.1.0',\n  });\n\n  const endpoint = process.env.OTEL_EXPORTER_OTLP_ENDPOINT;\n  let mode: TelemetryBootState;\n  let spanProcessor: BatchSpanProcessor | SimpleSpanProcessor | undefined;\n\n  if (endpoint) {\n    // OTLP path. Headers via OTEL_EXPORTER_OTLP_HEADERS (W3C convention,\n    // comma-separated key=value pairs). Parsed explicitly rather than\n    // relying on SDK-version-specific auto-discovery \u2014 predictable boot\n    // matters more than tracking whatever the current default is.\n    //   Honeycomb:     x-honeycomb-team=hcaik_xxx (plus optional dataset)\n    //   Grafana Cloud: authorization=Basic <base64creds>\n    const headers = parseOtlpHeaders(process.env.OTEL_EXPORTER_OTLP_HEADERS);\n    const exporter = new OTLPTraceExporter({\n      url: `${endpoint}/v1/traces`,\n      headers,\n    });\n    spanProcessor = new BatchSpanProcessor(exporter);\n    mode = 'otlp';\n  } else if (process.env.OTEL_SPAN_EXPORT === 'none') {\n    // Silent path: trace context still propagates (so trace IDs are available\n    // for correlation and an OTLP endpoint can be added later with zero code\n    // change), but NO span processor is registered \u2014 nothing is written to\n    // stdout. This exists because the ConsoleSpanExporter below writes one\n    // JSON span per ended span, and on the local launchd API (no OTLP\n    // endpoint, fixed log file, no Fly-managed rotation) that grew api-stdout\n    // to 370MB / ~400k spans. Set OTEL_SPAN_EXPORT=none on that process; Fly\n    // leaves it unset so console spans stay grep-able in its managed log stream.\n    spanProcessor = undefined;\n    mode = 'silent';\n  } else {\n    // Console (stdout) path. SimpleSpanProcessor \u2014 not batched \u2014 so spans\n    // appear in chronological order in the log stream. Acceptable cost for\n    // tracing: a span is written per ended span, vs. a batched POST for OTLP.\n    // Logs that would otherwise need a UI to make sense (trace tree\n    // visualization) are still queryable via `flyctl logs | grep traceId=...`\n    // or any log-aggregator that ingests Fly's stream.\n    const exporter = new ConsoleSpanExporter();\n    spanProcessor = new SimpleSpanProcessor(exporter);\n    mode = 'console';\n  }\n\n  sdk = new NodeSDK({\n    resource,\n    spanProcessors: spanProcessor ? [spanProcessor] : [],\n  });\n\n  sdk.start();\n  return mode;\n}\n\nexport async function shutdownTelemetry(): Promise<void> {\n  if (sdk) {\n    await sdk.shutdown();\n    sdk = null;\n  }\n}\n\nexport function getTracer(name?: string): Tracer {\n  return trace.getTracer(name ?? 'rulemetric');\n}\n\nexport function getTraceId(): string | undefined {\n  const span = trace.getSpan(context.active());\n  return span?.spanContext().traceId;\n}\n", "// Parse the `OTEL_EXPORTER_OTLP_HEADERS` env var into a plain headers object\n// that the OTLP exporter constructor accepts.\n//\n// OpenTelemetry SDK env-vars spec (\u00A710) defines this as a comma-separated list\n// of `key=value` pairs. We parse it ourselves rather than relying on the\n// `@opentelemetry/exporter-trace-otlp-http` package's auto-discovery, because\n// that auto-discovery has shifted between SDK versions and we want predictable\n// behaviour at boot.\n//\n// Real-world usage:\n//   - Honeycomb:   `x-honeycomb-team=hcaik_xxx` (and optionally `,x-honeycomb-dataset=...`)\n//   - Grafana Cloud: `authorization=Basic <base64>` (note: base64 padding can be `==`)\n//\n// Edge cases (covered by `otlp-headers.test.ts`):\n//   - undefined or empty \u2192 {} (caller will get a no-headers exporter)\n//   - whitespace around `=` and `,` is trimmed (operator-friendly)\n//   - only the first `=` is the separator, so base64 padding (==) survives in\n//     values\n//   - malformed entries (no `=`) are skipped silently\n//   - empty keys are dropped (never inject a \"\" header)\n//   - duplicate keys: last value wins, matching URL search-params semantics\n\nexport function parseOtlpHeaders(raw: string | undefined): Record<string, string> {\n  if (!raw) return {};\n  const out: Record<string, string> = {};\n  for (const entry of raw.split(',')) {\n    const eq = entry.indexOf('=');\n    if (eq === -1) continue;\n    const key = entry.slice(0, eq).trim();\n    const value = entry.slice(eq + 1).trim();\n    if (!key) continue;\n    out[key] = value;\n  }\n  return out;\n}\n", "import { SpanKind, SpanStatusCode } from '@opentelemetry/api';\nimport type { MiddlewareHandler } from 'hono';\nimport { getTracer, isTelemetryEnabled } from './init.js';\n\nexport function telemetryMiddleware(): MiddlewareHandler {\n  return async (c, next) => {\n    if (!isTelemetryEnabled()) {\n      await next();\n      return;\n    }\n\n    const tracer = getTracer();\n    const method = c.req.method;\n    const path = c.req.path;\n\n    await tracer.startActiveSpan(\n      `${method} ${path}`,\n      {\n        kind: SpanKind.SERVER,\n        attributes: {\n          'http.method': method,\n          'http.url': c.req.url,\n          'http.route': path,\n        },\n      },\n      async (span) => {\n        try {\n          await next();\n\n          span.setAttribute('http.status_code', c.res.status);\n\n          const userId = c.get('userId' as never);\n          if (userId) {\n            span.setAttribute('rulemetric.user_id', userId as string);\n          }\n\n          if (c.res.status >= 400) {\n            span.setStatus({ code: SpanStatusCode.ERROR });\n          }\n        } catch (err) {\n          span.setStatus({ code: SpanStatusCode.ERROR });\n          span.recordException(err as Error);\n          throw err;\n        } finally {\n          span.end();\n        }\n      },\n    );\n  };\n}\n", "import { trace, context as otelContext } from '@opentelemetry/api';\n\n/**\n * Set attributes on the currently-active span, if one exists. Safe no-op\n * when no span is active (e.g. running outside of a Hono request handler,\n * or with telemetry disabled).\n *\n * Use this from business-logic code that wants to expose a discriminator\n * for telemetry queries \u2014 e.g. \"which branch did `resolveCallerOrgId` take\"\n * in `apps/api/src/routes/sessions.ts`. Span attributes are searchable in\n * any OpenTelemetry backend (Honeycomb, Grafana Cloud, ...) and in\n * `flyctl logs` when the console exporter is active.\n *\n * Keeping this thin wrapper in the telemetry package means application\n * code never imports `@opentelemetry/api` directly \u2014 the package boundary\n * stays the only place that knows about the OTel SDK shape.\n */\nexport function setSpanAttributes(\n  attributes: Record<string, string | number | boolean>,\n): void {\n  const span = trace.getSpan(otelContext.active());\n  if (!span) return;\n  for (const [key, value] of Object.entries(attributes)) {\n    span.setAttribute(key, value);\n  }\n}\n", "// Thin wrapper around @sentry/node so callers don't have to guard every\n// captureException with `if (process.env.SENTRY_DSN)`. The wrapper handles\n// three concerns:\n//\n//   1. Init only when SENTRY_DSN is set. Otherwise every entry point becomes\n//      a silent no-op \u2014 never throws, never logs noise, never opens a\n//      network transport.\n//   2. Idempotent init. Double-init is a Sentry SDK foot-gun (the second\n//      call wins, but the warning lands at runtime). The module-level\n//      `enabled` flag short-circuits the second call.\n//   3. tracesSampleRate = 0 by default. We already have OpenTelemetry\n//      shipping spans to a tracing backend; paying Sentry's quota for the\n//      same data isn't useful. Sentry's job here is errors-only.\n//\n// Wiring:\n//   - `apps/api/src/index.ts`           \u2192 initSentry({ serviceName: 'rulemetric-api' })\n//   - `apps/api/src/middleware/auth.ts` \u2192 setSentryUser(userId) after JWT verify\n//   - `apps/cli/src/commands/evals/agent.ts` \u2192 initSentry({ serviceName: 'rulemetric-agent' })\n//     (the thin-agent worker \u2014 Phase 4c moved the init here from the old\n//     graphile `listen.ts`; that path no longer inits Sentry)\n//\n// All four hooks are no-op-safe when DSN is unset, so the wiring doesn't\n// require feature-flag plumbing.\n\nimport * as Sentry from '@sentry/node';\n\nlet enabled = false;\n\nexport interface InitSentryConfig {\n  serviceName: string;\n}\n\n/**\n * Initialize Sentry if SENTRY_DSN is set. Presence of the DSN is now the\n * single switch: set it (incl. as a Fly secret in prod) to enable error\n * capture, leave it unset for a silent no-op.\n *\n * Previously gated to non-production by policy; that gate was dropped so\n * production errors actually get reported. Errors-only (`tracesSampleRate:\n * 0`) keeps Sentry quota bounded \u2014 OpenTelemetry still owns tracing.\n *\n * Returns true when init actually ran, false when it was skipped (DSN unset\n * or a prior call already initialized). Idempotent.\n */\nexport function initSentry(config: InitSentryConfig): boolean {\n  if (enabled) return false;\n  const dsn = process.env.SENTRY_DSN;\n  if (!dsn) return false;\n\n  Sentry.init({\n    dsn,\n    serverName: config.serviceName,\n    environment: process.env.NODE_ENV ?? 'development',\n    // We rely on OpenTelemetry for tracing \u2014 Sentry stays errors-only here.\n    // Override at the env-var level if you ever want Sentry performance too:\n    // `SENTRY_TRACES_SAMPLE_RATE=0.05` (Sentry SDK reads it natively).\n    tracesSampleRate: 0,\n  });\n\n  enabled = true;\n  return true;\n}\n\n/**\n * Report an exception. Coerces non-Error throws to Error so Sentry's\n * grouping/stack-trace logic has something to work with. Silent no-op when\n * SENTRY_DSN is unset.\n */\nexport function captureException(\n  err: unknown,\n  context?: { extra?: Record<string, unknown>; tags?: Record<string, string> },\n): void {\n  if (!enabled) return;\n  const e = err instanceof Error ? err : new Error(typeof err === 'string' ? err : JSON.stringify(err));\n  Sentry.captureException(e, context);\n}\n\n/**\n * Attach (or clear, with null) the active user on the Sentry scope. Called\n * from the API auth middleware so any captured exception during the request\n * carries `user.id`. Silent no-op when SENTRY_DSN is unset.\n */\nexport function setSentryUser(userId: string | null): void {\n  if (!enabled) return;\n  if (userId === null) {\n    Sentry.setUser(null);\n    return;\n  }\n  Sentry.setUser({ id: userId });\n}\n\n/**\n * Flush any queued events with a timeout. Critical to call before\n * `process.exit()` in worker shutdown \u2014 Sentry batches events and the SIGTERM\n * path was previously dropping the final exception (the one we'd most want\n * to see). Resolves immediately when SENTRY_DSN is unset.\n */\nexport async function flushSentry(timeoutMs: number): Promise<void> {\n  if (!enabled) return;\n  await Sentry.flush(timeoutMs);\n}\n\n/**\n * Reset the module-level state. Test-only \u2014 exported so test suites can\n * unwind the singleton between cases without `vi.resetModules` plumbing in\n * every block.\n */\nexport function _resetSentryForTests(): void {\n  enabled = false;\n}\n", "import { initSentry, flushSentry, initTelemetry, shutdownTelemetry } from '@rulemetric/telemetry';\nimport { BaseCommand } from '../../base-command.js';\nimport { loadToken, getDefaultConfigDir } from '../../lib/auth.js';\nimport { runAgentLoop } from '../../lib/agent-loop.js';\nimport { closeDb } from '@rulemetric/db';\nimport { existsSync, mkdirSync, readFileSync, unlinkSync, writeFileSync } from 'node:fs';\nimport { dirname, join } from 'node:path';\nimport { execFileSync } from 'node:child_process';\n\n// Anchored under the config dir (honors RULEMETRIC_CONFIG_DIR; default\n// ~/.config/rulemetric, like gateway.pid), NOT os.tmpdir(), to avoid\n// launch-context TMPDIR divergence (launchd vs SSH/sudo/CI). Config-dir-scoped\n// so an isolated config (e.g. a cold-start test harness) can run its own agent\n// without colliding with the machine's real launchd-worker lock; two agents\n// sharing one config dir \u2014 the orphaned-fleet case this guard exists for \u2014 are\n// still refused.\nconst AGENT_LOCK_PATH = join(getDefaultConfigDir(), 'evals-agent.lock');\n\nfunction pidAlive(pid: number): boolean {\n  try {\n    process.kill(pid, 0);\n    return true;\n  } catch (e) {\n    // ESRCH = no such process (stale); EPERM = exists but owned by another user\n    return (e as { code?: string }).code === 'EPERM';\n  }\n}\n\n// Defends against PID reuse: a SIGKILL/OOM leaves the lock behind and the OS may\n// later assign that PID to an unrelated process \u2014 `pidAlive` alone would then\n// wedge `evals agent` (refuse to start forever). Confirm the holder is actually\n// an evals-agent before treating the lock as held; otherwise it's reclaimable.\nfunction isLiveAgent(pid: number): boolean {\n  if (!pidAlive(pid)) return false;\n  try {\n    const cmd = execFileSync('ps', ['-p', String(pid), '-o', 'command='], {\n      encoding: 'utf-8',\n      stdio: ['ignore', 'pipe', 'ignore'],\n    });\n    return /run\\.js\\s+evals\\s+agent|evals\\s+agent/.test(cmd);\n  } catch {\n    // ps unavailable/failed \u2014 fall back to liveness only (conservative: held)\n    return true;\n  }\n}\n\n// Single-instance guard. Each `evals agent` daemon holds its own postgres-js\n// pool; an orphaned, unsupervised fleet (4 stale agents found alive 3+ days on\n// 2026-06-21) was a confirmed driver of the recurring Supavisor 200-client\n// saturation. ATOMIC acquire via O_EXCL (flag 'wx') closes the TOCTOU window\n// where two simultaneously-starting agents could both acquire. On EEXIST the\n// holder is liveness-checked; a stale lock (holder gone) is reclaimed once.\n// See docs/incidents/2026-06-21-pool-exhaustion.md.\nfunction acquireAgentLock(): { ok: true } | { ok: false; holder: number } {\n  try {\n    mkdirSync(dirname(AGENT_LOCK_PATH), { recursive: true });\n  } catch {\n    /* ignore \u2014 writeFileSync will surface a real problem */\n  }\n  for (let attempt = 0; attempt < 2; attempt++) {\n    try {\n      // 'wx' = O_CREAT | O_EXCL: fails atomically with EEXIST if it exists.\n      writeFileSync(AGENT_LOCK_PATH, String(process.pid), { flag: 'wx' });\n      return { ok: true };\n    } catch (e) {\n      if ((e as { code?: string }).code !== 'EEXIST') throw e;\n      let holder = 0;\n      try {\n        holder = Number(readFileSync(AGENT_LOCK_PATH, 'utf-8').trim());\n      } catch {\n        /* unreadable \u2014 treat as stale */\n      }\n      if (holder && holder !== process.pid && isLiveAgent(holder)) {\n        return { ok: false, holder };\n      }\n      // stale (holder gone) / ours / unreadable \u2014 drop it and retry the atomic create once\n      try { unlinkSync(AGENT_LOCK_PATH); } catch { /* raced with another reclaimer */ }\n    }\n  }\n  // Last resort after a reclaim race \u2014 best-effort non-exclusive write.\n  try { writeFileSync(AGENT_LOCK_PATH, String(process.pid)); } catch { /* ignore */ }\n  return { ok: true };\n}\n\nfunction releaseAgentLock(): void {\n  try {\n    if (\n      existsSync(AGENT_LOCK_PATH) &&\n      Number(readFileSync(AGENT_LOCK_PATH, 'utf-8').trim()) === process.pid\n    ) {\n      unlinkSync(AGENT_LOCK_PATH);\n    }\n  } catch {\n    /* best-effort */\n  }\n}\n\n/**\n * `rulemetric evals agent` \u2014 Phase 4b parallel-run thin agent.\n *\n * Polls the /api/work endpoints (Phase 4a) instead of binding a Postgres\n * LISTEN/NOTIFY connection. This is the scaffold for the cutover described\n * in docs/plans/2026-05-25-thin-agent-migration.md; the default worker is\n * the sole worker mode since Phase 4c deleted the graphile path.\n *\n * Both commands can run side-by-side on the same user during the cutover \u2014\n * the API's atomic SKIP-LOCKED claim guarantees a job is handed to exactly\n * one agent. Picking which one runs is a launchd / shell choice.\n */\nexport default class EvalAgentCommand extends BaseCommand {\n  static override description =\n    'The worker daemon: long-polls /api/work/next, claims jobs, and runs them locally. Installed as a launchd service by `rulemetric service install`.';\n\n  static override examples = [\n    '<%= config.bin %> evals agent',\n  ];\n\n  async run(): Promise<void> {\n    await this.parse(EvalAgentCommand);\n    this.requireAuth();\n\n    // Single-instance guard \u2014 refuse to start a second agent. Each extra agent\n    // holds its own DB pool; an orphaned fleet was a confirmed cause of pooler\n    // saturation (pool-exhaustion RCA 2026-06-21).\n    const lock = acquireAgentLock();\n    if (!lock.ok) {\n      this.log(\n        `Another rulemetric evals agent is already running (pid ${lock.holder}). ` +\n          'Refusing to start a second instance.',\n      );\n      return;\n    }\n\n    process.env.PG_APPLICATION_NAME ||= 'rulemetric-agent';\n\n    const sentryEnabled = initSentry({ serviceName: 'rulemetric-agent' });\n    if (sentryEnabled) this.log('[sentry] error capture enabled');\n\n    const otelMode = initTelemetry({ serviceName: 'rulemetric-agent' });\n    if (otelMode === 'otlp') {\n      this.log(`[otel] traces exporting to ${process.env.OTEL_EXPORTER_OTLP_ENDPOINT}`);\n    } else if (otelMode === 'console') {\n      this.log('[otel] cron spans printing to stdout (set OTEL_EXPORTER_OTLP_ENDPOINT to ship to a backend)');\n    } else {\n      this.log('[otel] disabled (TELEMETRY_ENABLED=false)');\n    }\n\n    // Early auth sanity check \u2014 the agent loop also reads the token on every\n    // request via api-client, but failing here gives a friendlier error than\n    // a 401 thirty seconds into the first long-poll.\n    const token = loadToken();\n    if (!token && !process.env.RULEMETRIC_API_KEY) {\n      this.error('No auth token found. Run `rulemetric auth login` first.');\n    }\n\n    this.log('Starting thin agent (Phase 4b parallel-run)...');\n    this.log('Long-polling /api/work/next. Press Ctrl+C to stop.\\n');\n\n    const controller = new AbortController();\n\n    const shutdown = async (signal: string) => {\n      this.log(`\\n[${signal}] shutting down agent...`);\n      controller.abort();\n      // Flush telemetry + drain the pool on the way out. All members run\n      // concurrently under Promise.allSettled (flushSentry 2s, closeDb 5s), so\n      // the total stays well under launchd's ~20s shutdown grace.\n      await Promise.allSettled([\n        flushSentry(2_000),\n        shutdownTelemetry(),\n        // Drain the postgres-js pool so this agent's Supavisor client sessions\n        // are released (Terminate) instead of orphaned as ghosts on every\n        // launchd bounce. Pool-exhaustion RCA 2026-06-21.\n        closeDb(),\n      ]);\n      releaseAgentLock();\n      process.exit(0);\n    };\n    process.on('SIGINT', () => { void shutdown('SIGINT'); });\n    process.on('SIGTERM', () => { void shutdown('SIGTERM'); });\n\n    try {\n      await runAgentLoop({\n        signal: controller.signal,\n        log: (msg) => this.log(msg),\n        version: this.config.version,\n      });\n    } finally {\n      releaseAgentLock();\n      // Drain on the throw path too (defense-in-depth). The signal-handler drain\n      // covers normal shutdown; a thrown runAgentLoop must not skip it. closeDb\n      // is idempotent, so the happy-path SIGTERM drain stays a no-op.\n      await closeDb();\n    }\n  }\n}\n"],
-  "mappings": ";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA,SAAS,OAAO,eAA4B;AAC5C,SAAS,eAAe;AACxB,SAAS,yBAAyB;AAClC,SACE,oBACA,qBACA,2BACK;AACP,SAAS,gBAAgB;AACzB,SAAS,mBAAmB,4BAA4B;;;ACalD,SAAU,iBAAiB,KAAuB;AACtD,MAAI,CAAC;AAAK,WAAO,CAAA;AACjB,QAAM,MAA8B,CAAA;AACpC,aAAW,SAAS,IAAI,MAAM,GAAG,GAAG;AAClC,UAAM,KAAK,MAAM,QAAQ,GAAG;AAC5B,QAAI,OAAO;AAAI;AACf,UAAM,MAAM,MAAM,MAAM,GAAG,EAAE,EAAE,KAAI;AACnC,UAAM,QAAQ,MAAM,MAAM,KAAK,CAAC,EAAE,KAAI;AACtC,QAAI,CAAC;AAAK;AACV,QAAI,GAAG,IAAI;EACb;AACA,SAAO;AACT;;;ADtBA,IAAI,MAAsB;AAmBpB,SAAU,qBAAkB;AAChC,SAAO,QAAQ,IAAI,sBAAsB;AAC3C;AAsBM,SAAU,cAAc,QAAuB;AACnD,MAAI,CAAC,mBAAkB,GAAI;AACzB,WAAO;EACT;AACA,MAAI,QAAQ,MAAM;AAGhB,WAAO;EACT;AAEA,QAAM,WAAW,IAAI,SAAS;IAC5B,CAAC,iBAAiB,GAAG,OAAO;IAC5B,CAAC,oBAAoB,GAAG,OAAO,kBAAkB;GAClD;AAED,QAAM,WAAW,QAAQ,IAAI;AAC7B,MAAI;AACJ,MAAI;AAEJ,MAAI,UAAU;AAOZ,UAAM,UAAU,iBAAiB,QAAQ,IAAI,0BAA0B;AACvE,UAAM,WAAW,IAAI,kBAAkB;MACrC,KAAK,GAAG,QAAQ;MAChB;KACD;AACD,oBAAgB,IAAI,mBAAmB,QAAQ;AAC/C,WAAO;EACT,WAAW,QAAQ,IAAI,qBAAqB,QAAQ;AASlD,oBAAgB;AAChB,WAAO;EACT,OAAO;AAOL,UAAM,WAAW,IAAI,oBAAmB;AACxC,oBAAgB,IAAI,oBAAoB,QAAQ;AAChD,WAAO;EACT;AAEA,QAAM,IAAI,QAAQ;IAChB;IACA,gBAAgB,gBAAgB,CAAC,aAAa,IAAI,CAAA;GACnD;AAED,MAAI,MAAK;AACT,SAAO;AACT;AAEA,eAAsB,oBAAiB;AACrC,MAAI,KAAK;AACP,UAAM,IAAI,SAAQ;AAClB,UAAM;EACR;AACF;;;AE7HA,SAAS,UAAU,sBAAsB;;;ACAzC,SAAS,SAAAA,QAAO,WAAW,mBAAmB;;;ACwB9C,YAAY,YAAY;AAExB,IAAI,UAAU;AAkBR,SAAU,WAAW,QAAwB;AACjD,MAAI;AAAS,WAAO;AACpB,QAAM,MAAM,QAAQ,IAAI;AACxB,MAAI,CAAC;AAAK,WAAO;AAEjB,EAAO,YAAK;IACV;IACA,YAAY,OAAO;IACnB,aAAa,QAAQ,IAAI,YAAY;;;;IAIrC,kBAAkB;GACnB;AAED,YAAU;AACV,SAAO;AACT;AAoCA,eAAsB,YAAY,WAAiB;AACjD,MAAI,CAAC;AAAS;AACd,QAAa,aAAM,SAAS;AAC9B;;;AC/FA,SAAS,YAAY,WAAW,cAAc,YAAY,qBAAqB;AAC/E,SAAS,SAAS,YAAY;AAC9B,SAAS,oBAAoB;AAS7B,IAAM,kBAAkB,KAAK,oBAAoB,GAAG,kBAAkB;AAEtE,SAAS,SAAS,KAAsB;AACtC,MAAI;AACF,YAAQ,KAAK,KAAK,CAAC;AACnB,WAAO;AAAA,EACT,SAAS,GAAG;AAEV,WAAQ,EAAwB,SAAS;AAAA,EAC3C;AACF;AAMA,SAAS,YAAY,KAAsB;AACzC,MAAI,CAAC,SAAS,GAAG,EAAG,QAAO;AAC3B,MAAI;AACF,UAAM,MAAM,aAAa,MAAM,CAAC,MAAM,OAAO,GAAG,GAAG,MAAM,UAAU,GAAG;AAAA,MACpE,UAAU;AAAA,MACV,OAAO,CAAC,UAAU,QAAQ,QAAQ;AAAA,IACpC,CAAC;AACD,WAAO,wCAAwC,KAAK,GAAG;AAAA,EACzD,QAAQ;AAEN,WAAO;AAAA,EACT;AACF;AASA,SAAS,mBAAiE;AACxE,MAAI;AACF,cAAU,QAAQ,eAAe,GAAG,EAAE,WAAW,KAAK,CAAC;AAAA,EACzD,QAAQ;AAAA,EAER;AACA,WAAS,UAAU,GAAG,UAAU,GAAG,WAAW;AAC5C,QAAI;AAEF,oBAAc,iBAAiB,OAAO,QAAQ,GAAG,GAAG,EAAE,MAAM,KAAK,CAAC;AAClE,aAAO,EAAE,IAAI,KAAK;AAAA,IACpB,SAAS,GAAG;AACV,UAAK,EAAwB,SAAS,SAAU,OAAM;AACtD,UAAI,SAAS;AACb,UAAI;AACF,iBAAS,OAAO,aAAa,iBAAiB,OAAO,EAAE,KAAK,CAAC;AAAA,MAC/D,QAAQ;AAAA,MAER;AACA,UAAI,UAAU,WAAW,QAAQ,OAAO,YAAY,MAAM,GAAG;AAC3D,eAAO,EAAE,IAAI,OAAO,OAAO;AAAA,MAC7B;AAEA,UAAI;AAAE,mBAAW,eAAe;AAAA,MAAG,QAAQ;AAAA,MAAqC;AAAA,IAClF;AAAA,EACF;AAEA,MAAI;AAAE,kBAAc,iBAAiB,OAAO,QAAQ,GAAG,CAAC;AAAA,EAAG,QAAQ;AAAA,EAAe;AAClF,SAAO,EAAE,IAAI,KAAK;AACpB;AAEA,SAAS,mBAAyB;AAChC,MAAI;AACF,QACE,WAAW,eAAe,KAC1B,OAAO,aAAa,iBAAiB,OAAO,EAAE,KAAK,CAAC,MAAM,QAAQ,KAClE;AACA,iBAAW,eAAe;AAAA,IAC5B;AAAA,EACF,QAAQ;AAAA,EAER;AACF;AAcA,IAAqB,mBAArB,MAAqB,0BAAyB,YAAY;AAAA,EACxD,OAAgB,cACd;AAAA,EAEF,OAAgB,WAAW;AAAA,IACzB;AAAA,EACF;AAAA,EAEA,MAAM,MAAqB;AACzB,UAAM,KAAK,MAAM,iBAAgB;AACjC,SAAK,YAAY;AAKjB,UAAM,OAAO,iBAAiB;AAC9B,QAAI,CAAC,KAAK,IAAI;AACZ,WAAK;AAAA,QACH,0DAA0D,KAAK,MAAM;AAAA,MAEvE;AACA;AAAA,IACF;AAEA,YAAQ,IAAI,wBAAwB;AAEpC,UAAM,gBAAgB,WAAW,EAAE,aAAa,mBAAmB,CAAC;AACpE,QAAI,cAAe,MAAK,IAAI,gCAAgC;AAE5D,UAAM,WAAW,cAAc,EAAE,aAAa,mBAAmB,CAAC;AAClE,QAAI,aAAa,QAAQ;AACvB,WAAK,IAAI,8BAA8B,QAAQ,IAAI,2BAA2B,EAAE;AAAA,IAClF,WAAW,aAAa,WAAW;AACjC,WAAK,IAAI,6FAA6F;AAAA,IACxG,OAAO;AACL,WAAK,IAAI,2CAA2C;AAAA,IACtD;AAKA,UAAM,QAAQ,UAAU;AACxB,QAAI,CAAC,SAAS,CAAC,QAAQ,IAAI,oBAAoB;AAC7C,WAAK,MAAM,yDAAyD;AAAA,IACtE;AAEA,SAAK,IAAI,gDAAgD;AACzD,SAAK,IAAI,sDAAsD;AAE/D,UAAM,aAAa,IAAI,gBAAgB;AAEvC,UAAM,WAAW,OAAO,WAAmB;AACzC,WAAK,IAAI;AAAA,GAAM,MAAM,0BAA0B;AAC/C,iBAAW,MAAM;AAIjB,YAAM,QAAQ,WAAW;AAAA,QACvB,YAAY,GAAK;AAAA,QACjB,kBAAkB;AAAA;AAAA;AAAA;AAAA,QAIlB,QAAQ;AAAA,MACV,CAAC;AACD,uBAAiB;AACjB,cAAQ,KAAK,CAAC;AAAA,IAChB;AACA,YAAQ,GAAG,UAAU,MAAM;AAAE,WAAK,SAAS,QAAQ;AAAA,IAAG,CAAC;AACvD,YAAQ,GAAG,WAAW,MAAM;AAAE,WAAK,SAAS,SAAS;AAAA,IAAG,CAAC;AAEzD,QAAI;AACF,YAAM,aAAa;AAAA,QACjB,QAAQ,WAAW;AAAA,QACnB,KAAK,CAAC,QAAQ,KAAK,IAAI,GAAG;AAAA,QAC1B,SAAS,KAAK,OAAO;AAAA,MACvB,CAAC;AAAA,IACH,UAAE;AACA,uBAAiB;AAIjB,YAAM,QAAQ;AAAA,IAChB;AAAA,EACF;AACF;",
+  "sourcesContent": ["import { trace, context, type Tracer } from '@opentelemetry/api';\nimport { NodeSDK } from '@opentelemetry/sdk-node';\nimport { OTLPTraceExporter } from '@opentelemetry/exporter-trace-otlp-http';\nimport {\n  BatchSpanProcessor,\n  ConsoleSpanExporter,\n  SimpleSpanProcessor,\n} from '@opentelemetry/sdk-trace-node';\nimport { Resource } from '@opentelemetry/resources';\nimport { ATTR_SERVICE_NAME, ATTR_SERVICE_VERSION } from '@opentelemetry/semantic-conventions';\nimport { parseOtlpHeaders } from './otlp-headers.js';\n\nlet sdk: NodeSDK | null = null;\n\nexport interface TelemetryConfig {\n  serviceName: string;\n  serviceVersion?: string;\n}\n\n/**\n * Boot states for `initTelemetry`. Returned discriminator lets callers\n * emit a precise boot log:\n *   - 'otlp'    \u2014 spans ship over OTLP to a real backend (Honeycomb etc.)\n *   - 'console' \u2014 spans print to stdout as JSON (Fly logs ingest them);\n *                 the \"open + zero third-party\" default\n *   - 'off'     \u2014 telemetry globally disabled (TELEMETRY_ENABLED=false)\n *                 OR a prior `initTelemetry()` already booted the SDK\n *                 in this process (idempotent skip)\n */\nexport type TelemetryBootState = 'otlp' | 'console' | 'silent' | 'off';\n\nexport function isTelemetryEnabled(): boolean {\n  return process.env.TELEMETRY_ENABLED !== 'false';\n}\n\n/**\n * Initialize OpenTelemetry tracing. Two transport modes, both built on\n * the same OTel SDK \u2014 only the exporter changes:\n *\n *   - OTLP backend  \u2192 set `OTEL_EXPORTER_OTLP_ENDPOINT` (+ optional\n *                     `OTEL_EXPORTER_OTLP_HEADERS` for vendor auth)\n *   - Console (stdout) \u2192 the default when no endpoint is set\n *\n * The console path is intentionally on-by-default. Without a real backend\n * configured, spans are written to stdout as JSON and end up in Fly's log\n * stream alongside everything else \u2014 searchable, structured, and you\n * stay inside the open OpenTelemetry spec without paying a third party.\n * Swap to OTLP later by setting one env var; no code change needed.\n *\n * Returns `'off'` when:\n *   - TELEMETRY_ENABLED=false (operator kill switch \u2014 beats any config)\n *   - The SDK was already started in this process (idempotent skip)\n *\n * See packages/telemetry/src/init.test.ts for the full state table.\n */\nexport function initTelemetry(config: TelemetryConfig): TelemetryBootState {\n  if (!isTelemetryEnabled()) {\n    return 'off';\n  }\n  if (sdk !== null) {\n    // OTel's global tracer provider gets set on first start(); calling\n    // start() twice triggers an SDK warning. Treat second-call as a no-op.\n    return 'off';\n  }\n\n  const resource = new Resource({\n    [ATTR_SERVICE_NAME]: config.serviceName,\n    [ATTR_SERVICE_VERSION]: config.serviceVersion ?? '0.1.0',\n  });\n\n  const endpoint = process.env.OTEL_EXPORTER_OTLP_ENDPOINT;\n  let mode: TelemetryBootState;\n  let spanProcessor: BatchSpanProcessor | SimpleSpanProcessor | undefined;\n\n  if (endpoint) {\n    // OTLP path. Headers via OTEL_EXPORTER_OTLP_HEADERS (W3C convention,\n    // comma-separated key=value pairs). Parsed explicitly rather than\n    // relying on SDK-version-specific auto-discovery \u2014 predictable boot\n    // matters more than tracking whatever the current default is.\n    //   Honeycomb:     x-honeycomb-team=hcaik_xxx (plus optional dataset)\n    //   Grafana Cloud: authorization=Basic <base64creds>\n    const headers = parseOtlpHeaders(process.env.OTEL_EXPORTER_OTLP_HEADERS);\n    const exporter = new OTLPTraceExporter({\n      url: `${endpoint}/v1/traces`,\n      headers,\n    });\n    spanProcessor = new BatchSpanProcessor(exporter);\n    mode = 'otlp';\n  } else if (process.env.OTEL_SPAN_EXPORT === 'none') {\n    // Silent path: trace context still propagates (so trace IDs are available\n    // for correlation and an OTLP endpoint can be added later with zero code\n    // change), but NO span processor is registered \u2014 nothing is written to\n    // stdout. This exists because the ConsoleSpanExporter below writes one\n    // JSON span per ended span, and on the local launchd API (no OTLP\n    // endpoint, fixed log file, no Fly-managed rotation) that grew api-stdout\n    // to 370MB / ~400k spans. Set OTEL_SPAN_EXPORT=none on that process; Fly\n    // leaves it unset so console spans stay grep-able in its managed log stream.\n    spanProcessor = undefined;\n    mode = 'silent';\n  } else {\n    // Console (stdout) path. SimpleSpanProcessor \u2014 not batched \u2014 so spans\n    // appear in chronological order in the log stream. Acceptable cost for\n    // tracing: a span is written per ended span, vs. a batched POST for OTLP.\n    // Logs that would otherwise need a UI to make sense (trace tree\n    // visualization) are still queryable via `flyctl logs | grep traceId=...`\n    // or any log-aggregator that ingests Fly's stream.\n    const exporter = new ConsoleSpanExporter();\n    spanProcessor = new SimpleSpanProcessor(exporter);\n    mode = 'console';\n  }\n\n  sdk = new NodeSDK({\n    resource,\n    spanProcessors: spanProcessor ? [spanProcessor] : [],\n  });\n\n  sdk.start();\n  return mode;\n}\n\nexport async function shutdownTelemetry(): Promise<void> {\n  if (sdk) {\n    await sdk.shutdown();\n    sdk = null;\n  }\n}\n\nexport function getTracer(name?: string): Tracer {\n  return trace.getTracer(name ?? 'rulemetric');\n}\n\nexport function getTraceId(): string | undefined {\n  const span = trace.getSpan(context.active());\n  return span?.spanContext().traceId;\n}\n", "// Parse the `OTEL_EXPORTER_OTLP_HEADERS` env var into a plain headers object\n// that the OTLP exporter constructor accepts.\n//\n// OpenTelemetry SDK env-vars spec (\u00A710) defines this as a comma-separated list\n// of `key=value` pairs. We parse it ourselves rather than relying on the\n// `@opentelemetry/exporter-trace-otlp-http` package's auto-discovery, because\n// that auto-discovery has shifted between SDK versions and we want predictable\n// behaviour at boot.\n//\n// Real-world usage:\n//   - Honeycomb:   `x-honeycomb-team=hcaik_xxx` (and optionally `,x-honeycomb-dataset=...`)\n//   - Grafana Cloud: `authorization=Basic <base64>` (note: base64 padding can be `==`)\n//\n// Edge cases (covered by `otlp-headers.test.ts`):\n//   - undefined or empty \u2192 {} (caller will get a no-headers exporter)\n//   - whitespace around `=` and `,` is trimmed (operator-friendly)\n//   - only the first `=` is the separator, so base64 padding (==) survives in\n//     values\n//   - malformed entries (no `=`) are skipped silently\n//   - empty keys are dropped (never inject a \"\" header)\n//   - duplicate keys: last value wins, matching URL search-params semantics\n\nexport function parseOtlpHeaders(raw: string | undefined): Record<string, string> {\n  if (!raw) return {};\n  const out: Record<string, string> = {};\n  for (const entry of raw.split(',')) {\n    const eq = entry.indexOf('=');\n    if (eq === -1) continue;\n    const key = entry.slice(0, eq).trim();\n    const value = entry.slice(eq + 1).trim();\n    if (!key) continue;\n    out[key] = value;\n  }\n  return out;\n}\n", "import { SpanKind, SpanStatusCode } from '@opentelemetry/api';\nimport type { MiddlewareHandler } from 'hono';\nimport { getTracer, isTelemetryEnabled } from './init.js';\n\nexport function telemetryMiddleware(): MiddlewareHandler {\n  return async (c, next) => {\n    if (!isTelemetryEnabled()) {\n      await next();\n      return;\n    }\n\n    const tracer = getTracer();\n    const method = c.req.method;\n    const path = c.req.path;\n\n    await tracer.startActiveSpan(\n      `${method} ${path}`,\n      {\n        kind: SpanKind.SERVER,\n        attributes: {\n          'http.method': method,\n          'http.url': c.req.url,\n          'http.route': path,\n        },\n      },\n      async (span) => {\n        try {\n          await next();\n\n          span.setAttribute('http.status_code', c.res.status);\n\n          const userId = c.get('userId' as never);\n          if (userId) {\n            span.setAttribute('rulemetric.user_id', userId as string);\n          }\n\n          if (c.res.status >= 400) {\n            span.setStatus({ code: SpanStatusCode.ERROR });\n          }\n        } catch (err) {\n          span.setStatus({ code: SpanStatusCode.ERROR });\n          span.recordException(err as Error);\n          throw err;\n        } finally {\n          span.end();\n        }\n      },\n    );\n  };\n}\n", "import { trace, context as otelContext } from '@opentelemetry/api';\n\n/**\n * Set attributes on the currently-active span, if one exists. Safe no-op\n * when no span is active (e.g. running outside of a Hono request handler,\n * or with telemetry disabled).\n *\n * Use this from business-logic code that wants to expose a discriminator\n * for telemetry queries \u2014 e.g. \"which branch did `resolveCallerOrgId` take\"\n * in `apps/api/src/routes/sessions.ts`. Span attributes are searchable in\n * any OpenTelemetry backend (Honeycomb, Grafana Cloud, ...) and in\n * `flyctl logs` when the console exporter is active.\n *\n * Keeping this thin wrapper in the telemetry package means application\n * code never imports `@opentelemetry/api` directly \u2014 the package boundary\n * stays the only place that knows about the OTel SDK shape.\n */\nexport function setSpanAttributes(\n  attributes: Record<string, string | number | boolean>,\n): void {\n  const span = trace.getSpan(otelContext.active());\n  if (!span) return;\n  for (const [key, value] of Object.entries(attributes)) {\n    span.setAttribute(key, value);\n  }\n}\n", "// Thin wrapper around @sentry/node so callers don't have to guard every\n// captureException with `if (process.env.SENTRY_DSN)`. The wrapper handles\n// three concerns:\n//\n//   1. Init only when SENTRY_DSN is set. Otherwise every entry point becomes\n//      a silent no-op \u2014 never throws, never logs noise, never opens a\n//      network transport.\n//   2. Idempotent init. Double-init is a Sentry SDK foot-gun (the second\n//      call wins, but the warning lands at runtime). The module-level\n//      `enabled` flag short-circuits the second call.\n//   3. tracesSampleRate = 0 by default. We already have OpenTelemetry\n//      shipping spans to a tracing backend; paying Sentry's quota for the\n//      same data isn't useful. Sentry's job here is errors-only.\n//\n// Wiring:\n//   - `apps/api/src/index.ts`           \u2192 initSentry({ serviceName: 'rulemetric-api' })\n//   - `apps/api/src/middleware/auth.ts` \u2192 setSentryUser(userId) after JWT verify\n//   - `apps/cli/src/commands/evals/agent.ts` \u2192 initSentry({ serviceName: 'rulemetric-agent' })\n//     (the thin-agent worker \u2014 Phase 4c moved the init here from the old\n//     graphile `listen.ts`; that path no longer inits Sentry)\n//\n// All four hooks are no-op-safe when DSN is unset, so the wiring doesn't\n// require feature-flag plumbing.\n\nimport * as Sentry from '@sentry/node';\n\nlet enabled = false;\n\nexport interface InitSentryConfig {\n  serviceName: string;\n}\n\n/**\n * Initialize Sentry if SENTRY_DSN is set. Presence of the DSN is now the\n * single switch: set it (incl. as a Fly secret in prod) to enable error\n * capture, leave it unset for a silent no-op.\n *\n * Previously gated to non-production by policy; that gate was dropped so\n * production errors actually get reported. Errors-only (`tracesSampleRate:\n * 0`) keeps Sentry quota bounded \u2014 OpenTelemetry still owns tracing.\n *\n * Returns true when init actually ran, false when it was skipped (DSN unset\n * or a prior call already initialized). Idempotent.\n */\nexport function initSentry(config: InitSentryConfig): boolean {\n  if (enabled) return false;\n  const dsn = process.env.SENTRY_DSN;\n  if (!dsn) return false;\n\n  Sentry.init({\n    dsn,\n    serverName: config.serviceName,\n    environment: process.env.NODE_ENV ?? 'development',\n    // We rely on OpenTelemetry for tracing \u2014 Sentry stays errors-only here.\n    // Override at the env-var level if you ever want Sentry performance too:\n    // `SENTRY_TRACES_SAMPLE_RATE=0.05` (Sentry SDK reads it natively).\n    tracesSampleRate: 0,\n  });\n\n  enabled = true;\n  return true;\n}\n\n/**\n * Report an exception. Coerces non-Error throws to Error so Sentry's\n * grouping/stack-trace logic has something to work with. Silent no-op when\n * SENTRY_DSN is unset.\n */\nexport function captureException(\n  err: unknown,\n  context?: { extra?: Record<string, unknown>; tags?: Record<string, string> },\n): void {\n  if (!enabled) return;\n  const e = err instanceof Error ? err : new Error(typeof err === 'string' ? err : JSON.stringify(err));\n  Sentry.captureException(e, context);\n}\n\n/**\n * Attach (or clear, with null) the active user on the Sentry scope. Called\n * from the API auth middleware so any captured exception during the request\n * carries `user.id`. Silent no-op when SENTRY_DSN is unset.\n */\nexport function setSentryUser(userId: string | null): void {\n  if (!enabled) return;\n  if (userId === null) {\n    Sentry.setUser(null);\n    return;\n  }\n  Sentry.setUser({ id: userId });\n}\n\n/**\n * Flush any queued events with a timeout. Critical to call before\n * `process.exit()` in worker shutdown \u2014 Sentry batches events and the SIGTERM\n * path was previously dropping the final exception (the one we'd most want\n * to see). Resolves immediately when SENTRY_DSN is unset.\n */\nexport async function flushSentry(timeoutMs: number): Promise<void> {\n  if (!enabled) return;\n  await Sentry.flush(timeoutMs);\n}\n\n/**\n * Reset the module-level state. Test-only \u2014 exported so test suites can\n * unwind the singleton between cases without `vi.resetModules` plumbing in\n * every block.\n */\nexport function _resetSentryForTests(): void {\n  enabled = false;\n}\n", "import { initSentry, flushSentry, initTelemetry, shutdownTelemetry } from '@rulemetric/telemetry';\nimport { BaseCommand } from '../../base-command.js';\nimport { loadToken, getDefaultConfigDir } from '../../lib/auth.js';\nimport { runAgentLoop } from '../../lib/agent-loop.js';\nimport { closeDb } from '@rulemetric/db';\nimport { existsSync, mkdirSync, readFileSync, unlinkSync, writeFileSync } from 'node:fs';\nimport { dirname, join } from 'node:path';\nimport { execFileSync } from 'node:child_process';\n\n// Anchored under the config dir (honors RULEMETRIC_CONFIG_DIR; default\n// ~/.config/rulemetric, like gateway.pid), NOT os.tmpdir(), to avoid\n// launch-context TMPDIR divergence (launchd vs SSH/sudo/CI). Config-dir-scoped\n// so an isolated config (e.g. a cold-start test harness) can run its own agent\n// without colliding with the machine's real launchd-worker lock; two agents\n// sharing one config dir \u2014 the orphaned-fleet case this guard exists for \u2014 are\n// still refused.\nconst AGENT_LOCK_PATH = join(getDefaultConfigDir(), 'evals-agent.lock');\n\nfunction pidAlive(pid: number): boolean {\n  try {\n    process.kill(pid, 0);\n    return true;\n  } catch (e) {\n    // ESRCH = no such process (stale); EPERM = exists but owned by another user\n    return (e as { code?: string }).code === 'EPERM';\n  }\n}\n\n// Defends against PID reuse: a SIGKILL/OOM leaves the lock behind and the OS may\n// later assign that PID to an unrelated process \u2014 `pidAlive` alone would then\n// wedge `evals agent` (refuse to start forever). Confirm the holder is actually\n// an evals-agent before treating the lock as held; otherwise it's reclaimable.\nfunction isLiveAgent(pid: number): boolean {\n  if (!pidAlive(pid)) return false;\n  try {\n    const cmd = execFileSync('ps', ['-p', String(pid), '-o', 'command='], {\n      encoding: 'utf-8',\n      stdio: ['ignore', 'pipe', 'ignore'],\n    });\n    return /run\\.js\\s+evals\\s+agent|evals\\s+agent/.test(cmd);\n  } catch {\n    // ps unavailable/failed \u2014 fall back to liveness only (conservative: held)\n    return true;\n  }\n}\n\n// Single-instance guard. Each `evals agent` daemon holds its own postgres-js\n// pool; an orphaned, unsupervised fleet (4 stale agents found alive 3+ days on\n// 2026-06-21) was a confirmed driver of the recurring Supavisor 200-client\n// saturation. ATOMIC acquire via O_EXCL (flag 'wx') closes the TOCTOU window\n// where two simultaneously-starting agents could both acquire. On EEXIST the\n// holder is liveness-checked; a stale lock (holder gone) is reclaimed once.\n// See docs/incidents/2026-06-21-pool-exhaustion.md.\nfunction acquireAgentLock(): { ok: true } | { ok: false; holder: number } {\n  try {\n    mkdirSync(dirname(AGENT_LOCK_PATH), { recursive: true });\n  } catch {\n    /* ignore \u2014 writeFileSync will surface a real problem */\n  }\n  for (let attempt = 0; attempt < 2; attempt++) {\n    try {\n      // 'wx' = O_CREAT | O_EXCL: fails atomically with EEXIST if it exists.\n      writeFileSync(AGENT_LOCK_PATH, String(process.pid), { flag: 'wx' });\n      return { ok: true };\n    } catch (e) {\n      if ((e as { code?: string }).code !== 'EEXIST') throw e;\n      let holder = 0;\n      try {\n        holder = Number(readFileSync(AGENT_LOCK_PATH, 'utf-8').trim());\n      } catch {\n        /* unreadable \u2014 treat as stale */\n      }\n      if (holder && holder !== process.pid && isLiveAgent(holder)) {\n        return { ok: false, holder };\n      }\n      // stale (holder gone) / ours / unreadable \u2014 drop it and retry the atomic create once\n      try { unlinkSync(AGENT_LOCK_PATH); } catch { /* raced with another reclaimer */ }\n    }\n  }\n  // Last resort after a reclaim race \u2014 best-effort non-exclusive write.\n  try { writeFileSync(AGENT_LOCK_PATH, String(process.pid)); } catch { /* ignore */ }\n  return { ok: true };\n}\n\nfunction releaseAgentLock(): void {\n  try {\n    if (\n      existsSync(AGENT_LOCK_PATH) &&\n      Number(readFileSync(AGENT_LOCK_PATH, 'utf-8').trim()) === process.pid\n    ) {\n      unlinkSync(AGENT_LOCK_PATH);\n    }\n  } catch {\n    /* best-effort */\n  }\n}\n\n/**\n * `rulemetric evals agent` \u2014 Phase 4b parallel-run thin agent.\n *\n * Polls the /api/work endpoints (Phase 4a) instead of binding a Postgres\n * LISTEN/NOTIFY connection. This is the scaffold for the cutover described\n * in docs/plans/2026-05-25-thin-agent-migration.md; the default worker is\n * the sole worker mode since Phase 4c deleted the graphile path.\n *\n * Both commands can run side-by-side on the same user during the cutover \u2014\n * the API's atomic SKIP-LOCKED claim guarantees a job is handed to exactly\n * one agent. Picking which one runs is a launchd / shell choice.\n */\nexport default class EvalAgentCommand extends BaseCommand {\n  static override description =\n    'The worker daemon: long-polls /api/work/next, claims jobs, and runs them locally. Installed as a launchd service by `rulemetric service install`.';\n\n  static override examples = [\n    '<%= config.bin %> evals agent',\n  ];\n\n  async run(): Promise<void> {\n    await this.parse(EvalAgentCommand);\n    this.requireAuth();\n\n    // Single-instance guard \u2014 refuse to start a second agent. Each extra agent\n    // holds its own DB pool; an orphaned fleet was a confirmed cause of pooler\n    // saturation (pool-exhaustion RCA 2026-06-21).\n    const lock = acquireAgentLock();\n    if (!lock.ok) {\n      this.log(\n        `Another rulemetric evals agent is already running (pid ${lock.holder}). ` +\n          'Refusing to start a second instance.',\n      );\n      return;\n    }\n\n    process.env.PG_APPLICATION_NAME ||= 'rulemetric-agent';\n\n    const sentryEnabled = initSentry({ serviceName: 'rulemetric-agent' });\n    if (sentryEnabled) this.log('[sentry] error capture enabled');\n\n    const otelMode = initTelemetry({ serviceName: 'rulemetric-agent' });\n    if (otelMode === 'otlp') {\n      this.log(`[otel] traces exporting to ${process.env.OTEL_EXPORTER_OTLP_ENDPOINT}`);\n    } else if (otelMode === 'console') {\n      this.log('[otel] cron spans printing to stdout (set OTEL_EXPORTER_OTLP_ENDPOINT to ship to a backend)');\n    } else {\n      this.log('[otel] disabled (TELEMETRY_ENABLED=false)');\n    }\n\n    // Early auth sanity check \u2014 the agent loop also reads the token on every\n    // request via api-client, but failing here gives a friendlier error than\n    // a 401 thirty seconds into the first long-poll.\n    const token = loadToken();\n    if (!token && !process.env.RULEMETRIC_API_KEY) {\n      this.error('No auth token found. Run `rulemetric auth login` first.');\n    }\n\n    this.log('Starting thin agent (Phase 4b parallel-run)...');\n    this.log('Long-polling /api/work/next. Press Ctrl+C to stop.\\n');\n\n    const controller = new AbortController();\n\n    const shutdown = async (signal: string) => {\n      this.log(`\\n[${signal}] shutting down agent...`);\n      controller.abort();\n      // Flush telemetry + drain the pool on the way out. All members run\n      // concurrently under Promise.allSettled (flushSentry 2s, closeDb 5s), so\n      // the total stays well under launchd's ~20s shutdown grace.\n      await Promise.allSettled([\n        flushSentry(2_000),\n        shutdownTelemetry(),\n        // Drain the postgres-js pool so this agent's Supavisor client sessions\n        // are released (Terminate) instead of orphaned as ghosts on every\n        // launchd bounce. Pool-exhaustion RCA 2026-06-21.\n        closeDb(),\n      ]);\n      releaseAgentLock();\n      process.exit(0);\n    };\n    process.on('SIGINT', () => { void shutdown('SIGINT'); });\n    process.on('SIGTERM', () => { void shutdown('SIGTERM'); });\n\n    // Set when the loop stops because `npm i -g` replaced the CLI under this\n    // running process \u2014 exit NON-ZERO after cleanup so the supervisor respawns\n    // onto the new code (systemd is Restart=on-failure; launchd KeepAlive\n    // restarts on any exit). Without this, a remote worker keeps executing the\n    // old code until someone manually restarts the unit.\n    let restartForUpgrade = false;\n    try {\n      await runAgentLoop({\n        signal: controller.signal,\n        log: (msg) => this.log(msg),\n        version: this.config.version,\n        readInstalledVersion: () => {\n          try {\n            const pkg = JSON.parse(\n              readFileSync(join(this.config.root, 'package.json'), 'utf-8'),\n            ) as { version?: string };\n            return pkg.version ?? null;\n          } catch {\n            return null;\n          }\n        },\n        onVersionDrift: () => {\n          restartForUpgrade = true;\n        },\n      });\n      if (restartForUpgrade) {\n        this.log('[upgrade] Installed CLI version changed \u2014 restarting worker onto the new code.');\n        await Promise.allSettled([flushSentry(2_000), shutdownTelemetry(), closeDb()]);\n        releaseAgentLock();\n        process.exit(1);\n      }\n    } finally {\n      releaseAgentLock();\n      // Drain on the throw path too (defense-in-depth). The signal-handler drain\n      // covers normal shutdown; a thrown runAgentLoop must not skip it. closeDb\n      // is idempotent, so the happy-path SIGTERM drain stays a no-op.\n      await closeDb();\n    }\n  }\n}\n"],
+  "mappings": ";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA,SAAS,OAAO,eAA4B;AAC5C,SAAS,eAAe;AACxB,SAAS,yBAAyB;AAClC,SACE,oBACA,qBACA,2BACK;AACP,SAAS,gBAAgB;AACzB,SAAS,mBAAmB,4BAA4B;;;ACalD,SAAU,iBAAiB,KAAuB;AACtD,MAAI,CAAC;AAAK,WAAO,CAAA;AACjB,QAAM,MAA8B,CAAA;AACpC,aAAW,SAAS,IAAI,MAAM,GAAG,GAAG;AAClC,UAAM,KAAK,MAAM,QAAQ,GAAG;AAC5B,QAAI,OAAO;AAAI;AACf,UAAM,MAAM,MAAM,MAAM,GAAG,EAAE,EAAE,KAAI;AACnC,UAAM,QAAQ,MAAM,MAAM,KAAK,CAAC,EAAE,KAAI;AACtC,QAAI,CAAC;AAAK;AACV,QAAI,GAAG,IAAI;EACb;AACA,SAAO;AACT;;;ADtBA,IAAI,MAAsB;AAmBpB,SAAU,qBAAkB;AAChC,SAAO,QAAQ,IAAI,sBAAsB;AAC3C;AAsBM,SAAU,cAAc,QAAuB;AACnD,MAAI,CAAC,mBAAkB,GAAI;AACzB,WAAO;EACT;AACA,MAAI,QAAQ,MAAM;AAGhB,WAAO;EACT;AAEA,QAAM,WAAW,IAAI,SAAS;IAC5B,CAAC,iBAAiB,GAAG,OAAO;IAC5B,CAAC,oBAAoB,GAAG,OAAO,kBAAkB;GAClD;AAED,QAAM,WAAW,QAAQ,IAAI;AAC7B,MAAI;AACJ,MAAI;AAEJ,MAAI,UAAU;AAOZ,UAAM,UAAU,iBAAiB,QAAQ,IAAI,0BAA0B;AACvE,UAAM,WAAW,IAAI,kBAAkB;MACrC,KAAK,GAAG,QAAQ;MAChB;KACD;AACD,oBAAgB,IAAI,mBAAmB,QAAQ;AAC/C,WAAO;EACT,WAAW,QAAQ,IAAI,qBAAqB,QAAQ;AASlD,oBAAgB;AAChB,WAAO;EACT,OAAO;AAOL,UAAM,WAAW,IAAI,oBAAmB;AACxC,oBAAgB,IAAI,oBAAoB,QAAQ;AAChD,WAAO;EACT;AAEA,QAAM,IAAI,QAAQ;IAChB;IACA,gBAAgB,gBAAgB,CAAC,aAAa,IAAI,CAAA;GACnD;AAED,MAAI,MAAK;AACT,SAAO;AACT;AAEA,eAAsB,oBAAiB;AACrC,MAAI,KAAK;AACP,UAAM,IAAI,SAAQ;AAClB,UAAM;EACR;AACF;;;AE7HA,SAAS,UAAU,sBAAsB;;;ACAzC,SAAS,SAAAA,QAAO,WAAW,mBAAmB;;;ACwB9C,YAAY,YAAY;AAExB,IAAI,UAAU;AAkBR,SAAU,WAAW,QAAwB;AACjD,MAAI;AAAS,WAAO;AACpB,QAAM,MAAM,QAAQ,IAAI;AACxB,MAAI,CAAC;AAAK,WAAO;AAEjB,EAAO,YAAK;IACV;IACA,YAAY,OAAO;IACnB,aAAa,QAAQ,IAAI,YAAY;;;;IAIrC,kBAAkB;GACnB;AAED,YAAU;AACV,SAAO;AACT;AAoCA,eAAsB,YAAY,WAAiB;AACjD,MAAI,CAAC;AAAS;AACd,QAAa,aAAM,SAAS;AAC9B;;;AC/FA,SAAS,YAAY,WAAW,cAAc,YAAY,qBAAqB;AAC/E,SAAS,SAAS,YAAY;AAC9B,SAAS,oBAAoB;AAS7B,IAAM,kBAAkB,KAAK,oBAAoB,GAAG,kBAAkB;AAEtE,SAAS,SAAS,KAAsB;AACtC,MAAI;AACF,YAAQ,KAAK,KAAK,CAAC;AACnB,WAAO;AAAA,EACT,SAAS,GAAG;AAEV,WAAQ,EAAwB,SAAS;AAAA,EAC3C;AACF;AAMA,SAAS,YAAY,KAAsB;AACzC,MAAI,CAAC,SAAS,GAAG,EAAG,QAAO;AAC3B,MAAI;AACF,UAAM,MAAM,aAAa,MAAM,CAAC,MAAM,OAAO,GAAG,GAAG,MAAM,UAAU,GAAG;AAAA,MACpE,UAAU;AAAA,MACV,OAAO,CAAC,UAAU,QAAQ,QAAQ;AAAA,IACpC,CAAC;AACD,WAAO,wCAAwC,KAAK,GAAG;AAAA,EACzD,QAAQ;AAEN,WAAO;AAAA,EACT;AACF;AASA,SAAS,mBAAiE;AACxE,MAAI;AACF,cAAU,QAAQ,eAAe,GAAG,EAAE,WAAW,KAAK,CAAC;AAAA,EACzD,QAAQ;AAAA,EAER;AACA,WAAS,UAAU,GAAG,UAAU,GAAG,WAAW;AAC5C,QAAI;AAEF,oBAAc,iBAAiB,OAAO,QAAQ,GAAG,GAAG,EAAE,MAAM,KAAK,CAAC;AAClE,aAAO,EAAE,IAAI,KAAK;AAAA,IACpB,SAAS,GAAG;AACV,UAAK,EAAwB,SAAS,SAAU,OAAM;AACtD,UAAI,SAAS;AACb,UAAI;AACF,iBAAS,OAAO,aAAa,iBAAiB,OAAO,EAAE,KAAK,CAAC;AAAA,MAC/D,QAAQ;AAAA,MAER;AACA,UAAI,UAAU,WAAW,QAAQ,OAAO,YAAY,MAAM,GAAG;AAC3D,eAAO,EAAE,IAAI,OAAO,OAAO;AAAA,MAC7B;AAEA,UAAI;AAAE,mBAAW,eAAe;AAAA,MAAG,QAAQ;AAAA,MAAqC;AAAA,IAClF;AAAA,EACF;AAEA,MAAI;AAAE,kBAAc,iBAAiB,OAAO,QAAQ,GAAG,CAAC;AAAA,EAAG,QAAQ;AAAA,EAAe;AAClF,SAAO,EAAE,IAAI,KAAK;AACpB;AAEA,SAAS,mBAAyB;AAChC,MAAI;AACF,QACE,WAAW,eAAe,KAC1B,OAAO,aAAa,iBAAiB,OAAO,EAAE,KAAK,CAAC,MAAM,QAAQ,KAClE;AACA,iBAAW,eAAe;AAAA,IAC5B;AAAA,EACF,QAAQ;AAAA,EAER;AACF;AAcA,IAAqB,mBAArB,MAAqB,0BAAyB,YAAY;AAAA,EACxD,OAAgB,cACd;AAAA,EAEF,OAAgB,WAAW;AAAA,IACzB;AAAA,EACF;AAAA,EAEA,MAAM,MAAqB;AACzB,UAAM,KAAK,MAAM,iBAAgB;AACjC,SAAK,YAAY;AAKjB,UAAM,OAAO,iBAAiB;AAC9B,QAAI,CAAC,KAAK,IAAI;AACZ,WAAK;AAAA,QACH,0DAA0D,KAAK,MAAM;AAAA,MAEvE;AACA;AAAA,IACF;AAEA,YAAQ,IAAI,wBAAwB;AAEpC,UAAM,gBAAgB,WAAW,EAAE,aAAa,mBAAmB,CAAC;AACpE,QAAI,cAAe,MAAK,IAAI,gCAAgC;AAE5D,UAAM,WAAW,cAAc,EAAE,aAAa,mBAAmB,CAAC;AAClE,QAAI,aAAa,QAAQ;AACvB,WAAK,IAAI,8BAA8B,QAAQ,IAAI,2BAA2B,EAAE;AAAA,IAClF,WAAW,aAAa,WAAW;AACjC,WAAK,IAAI,6FAA6F;AAAA,IACxG,OAAO;AACL,WAAK,IAAI,2CAA2C;AAAA,IACtD;AAKA,UAAM,QAAQ,UAAU;AACxB,QAAI,CAAC,SAAS,CAAC,QAAQ,IAAI,oBAAoB;AAC7C,WAAK,MAAM,yDAAyD;AAAA,IACtE;AAEA,SAAK,IAAI,gDAAgD;AACzD,SAAK,IAAI,sDAAsD;AAE/D,UAAM,aAAa,IAAI,gBAAgB;AAEvC,UAAM,WAAW,OAAO,WAAmB;AACzC,WAAK,IAAI;AAAA,GAAM,MAAM,0BAA0B;AAC/C,iBAAW,MAAM;AAIjB,YAAM,QAAQ,WAAW;AAAA,QACvB,YAAY,GAAK;AAAA,QACjB,kBAAkB;AAAA;AAAA;AAAA;AAAA,QAIlB,QAAQ;AAAA,MACV,CAAC;AACD,uBAAiB;AACjB,cAAQ,KAAK,CAAC;AAAA,IAChB;AACA,YAAQ,GAAG,UAAU,MAAM;AAAE,WAAK,SAAS,QAAQ;AAAA,IAAG,CAAC;AACvD,YAAQ,GAAG,WAAW,MAAM;AAAE,WAAK,SAAS,SAAS;AAAA,IAAG,CAAC;AAOzD,QAAI,oBAAoB;AACxB,QAAI;AACF,YAAM,aAAa;AAAA,QACjB,QAAQ,WAAW;AAAA,QACnB,KAAK,CAAC,QAAQ,KAAK,IAAI,GAAG;AAAA,QAC1B,SAAS,KAAK,OAAO;AAAA,QACrB,sBAAsB,MAAM;AAC1B,cAAI;AACF,kBAAM,MAAM,KAAK;AAAA,cACf,aAAa,KAAK,KAAK,OAAO,MAAM,cAAc,GAAG,OAAO;AAAA,YAC9D;AACA,mBAAO,IAAI,WAAW;AAAA,UACxB,QAAQ;AACN,mBAAO;AAAA,UACT;AAAA,QACF;AAAA,QACA,gBAAgB,MAAM;AACpB,8BAAoB;AAAA,QACtB;AAAA,MACF,CAAC;AACD,UAAI,mBAAmB;AACrB,aAAK,IAAI,qFAAgF;AACzF,cAAM,QAAQ,WAAW,CAAC,YAAY,GAAK,GAAG,kBAAkB,GAAG,QAAQ,CAAC,CAAC;AAC7E,yBAAiB;AACjB,gBAAQ,KAAK,CAAC;AAAA,MAChB;AAAA,IACF,UAAE;AACA,uBAAiB;AAIjB,YAAM,QAAQ;AAAA,IAChB;AAAA,EACF;AACF;",
   "names": ["trace"]
 }

package/dist/commands/evals/benchmark.js

@@ -1,16 +1,17 @@
 import {
   aggregateBenchmark,
   postBenchmark
-} from "../../chunk-3QH73QOQ.js";
+} from "../../chunk-PED7L5GX.js";
 import {
   BaseCommand
-} from "../../chunk-SZ7VDCD6.js";
+} from "../../chunk-2HK7RGZB.js";
 import "../../chunk-R56PH7YE.js";
-import "../../chunk-SLL3OBKN.js";
+import "../../chunk-GQNJO2P2.js";
+import "../../chunk-37LN5O7M.js";
 import "../../chunk-EKJ2CABV.js";
 import {
   apiGet
-} from "../../chunk-ZRLYHBPN.js";
+} from "../../chunk-ZUJBM6HE.js";
 import "../../chunk-W2YERO7E.js";
 import "../../chunk-NSBPE2FW.js";
 

package/dist/commands/evals/benchmark.js.map

@@ -2,6 +2,6 @@
   "version": 3,
   "sources": ["../../../src/commands/evals/benchmark.ts"],
   "sourcesContent": ["import { Args, Flags } from '@oclif/core';\nimport { BaseCommand } from '../../base-command.js';\nimport { apiGet } from '../../lib/api-client.js';\nimport { aggregateBenchmark, postBenchmark } from '../../lib/eval-benchmark.js';\n\nexport default class EvalBenchmarkCommand extends BaseCommand {\n  static override description = 'Aggregate eval runs into a benchmark for a target';\n\n  static override examples = [\n    '<%= config.bin %> evals benchmark <target-id>',\n    '<%= config.bin %> evals benchmark <target-id> --iteration 2',\n  ];\n\n  static override args = {\n    targetId: Args.string({ description: 'Eval target ID', required: true }),\n  };\n\n  static override flags = {\n    iteration: Flags.integer({ description: 'Filter runs by iteration', default: 1 }),\n  };\n\n  async run(): Promise<void> {\n    const { args, flags } = await this.parse(EvalBenchmarkCommand);\n    this.requireAuth();\n\n    const target = await apiGet<{ name: string; version: number }>(`/api/eval-targets/${args.targetId}`);\n    this.log(`Aggregating benchmark for \"${target.name}\" v${target.version}, iteration ${flags.iteration}`);\n\n    const result = await aggregateBenchmark(args.targetId, flags.iteration);\n\n    if (result.totalRuns === 0) {\n      this.error('No completed runs found. Run evals first with `rulemetric evals run`.');\n    }\n\n    this.log(`\\nAggregated ${result.totalRuns} runs\\n`);\n\n    // Display summary\n    for (const [config, stats] of Object.entries(result.runSummary)) {\n      if (config === 'delta') {\n        this.log(`\\nDelta (with_target - without_target):`);\n        for (const [metric, value] of Object.entries(stats as unknown as Record<string, string>)) {\n          this.log(`  ${metric}: ${value}`);\n        }\n        continue;\n      }\n      this.log(`${config}:`);\n      for (const [metric, stat] of Object.entries(stats)) {\n        const s = stat as { mean: number; stddev: number; min: number; max: number };\n        this.log(`  ${metric}: mean=${s.mean} stddev=${s.stddev} [${s.min}, ${s.max}]`);\n      }\n    }\n\n    if (result.notes.length > 0) {\n      this.log('\\nNotes:');\n      for (const note of result.notes) {\n        this.log(`  - ${note}`);\n      }\n    }\n\n    // Post to API\n    const benchmark = await postBenchmark(args.targetId, flags.iteration, result);\n    this.log(`\\nBenchmark saved: ${benchmark.id}`);\n  }\n}\n"],
-  "mappings": ";;;;;;;;;;;;;;;;;AAAA,SAAS,MAAM,aAAa;AAK5B,IAAqB,uBAArB,MAAqB,8BAA6B,YAAY;AAAA,EAC5D,OAAgB,cAAc;AAAA,EAE9B,OAAgB,WAAW;AAAA,IACzB;AAAA,IACA;AAAA,EACF;AAAA,EAEA,OAAgB,OAAO;AAAA,IACrB,UAAU,KAAK,OAAO,EAAE,aAAa,kBAAkB,UAAU,KAAK,CAAC;AAAA,EACzE;AAAA,EAEA,OAAgB,QAAQ;AAAA,IACtB,WAAW,MAAM,QAAQ,EAAE,aAAa,4BAA4B,SAAS,EAAE,CAAC;AAAA,EAClF;AAAA,EAEA,MAAM,MAAqB;AACzB,UAAM,EAAE,MAAM,MAAM,IAAI,MAAM,KAAK,MAAM,qBAAoB;AAC7D,SAAK,YAAY;AAEjB,UAAM,SAAS,MAAM,OAA0C,qBAAqB,KAAK,QAAQ,EAAE;AACnG,SAAK,IAAI,8BAA8B,OAAO,IAAI,MAAM,OAAO,OAAO,eAAe,MAAM,SAAS,EAAE;AAEtG,UAAM,SAAS,MAAM,mBAAmB,KAAK,UAAU,MAAM,SAAS;AAEtE,QAAI,OAAO,cAAc,GAAG;AAC1B,WAAK,MAAM,uEAAuE;AAAA,IACpF;AAEA,SAAK,IAAI;AAAA,aAAgB,OAAO,SAAS;AAAA,CAAS;AAGlD,eAAW,CAAC,QAAQ,KAAK,KAAK,OAAO,QAAQ,OAAO,UAAU,GAAG;AAC/D,UAAI,WAAW,SAAS;AACtB,aAAK,IAAI;AAAA,sCAAyC;AAClD,mBAAW,CAAC,QAAQ,KAAK,KAAK,OAAO,QAAQ,KAA0C,GAAG;AACxF,eAAK,IAAI,KAAK,MAAM,KAAK,KAAK,EAAE;AAAA,QAClC;AACA;AAAA,MACF;AACA,WAAK,IAAI,GAAG,MAAM,GAAG;AACrB,iBAAW,CAAC,QAAQ,IAAI,KAAK,OAAO,QAAQ,KAAK,GAAG;AAClD,cAAM,IAAI;AACV,aAAK,IAAI,KAAK,MAAM,UAAU,EAAE,IAAI,WAAW,EAAE,MAAM,KAAK,EAAE,GAAG,KAAK,EAAE,GAAG,GAAG;AAAA,MAChF;AAAA,IACF;AAEA,QAAI,OAAO,MAAM,SAAS,GAAG;AAC3B,WAAK,IAAI,UAAU;AACnB,iBAAW,QAAQ,OAAO,OAAO;AAC/B,aAAK,IAAI,OAAO,IAAI,EAAE;AAAA,MACxB;AAAA,IACF;AAGA,UAAM,YAAY,MAAM,cAAc,KAAK,UAAU,MAAM,WAAW,MAAM;AAC5E,SAAK,IAAI;AAAA,mBAAsB,UAAU,EAAE,EAAE;AAAA,EAC/C;AACF;",
+  "mappings": ";;;;;;;;;;;;;;;;;;AAAA,SAAS,MAAM,aAAa;AAK5B,IAAqB,uBAArB,MAAqB,8BAA6B,YAAY;AAAA,EAC5D,OAAgB,cAAc;AAAA,EAE9B,OAAgB,WAAW;AAAA,IACzB;AAAA,IACA;AAAA,EACF;AAAA,EAEA,OAAgB,OAAO;AAAA,IACrB,UAAU,KAAK,OAAO,EAAE,aAAa,kBAAkB,UAAU,KAAK,CAAC;AAAA,EACzE;AAAA,EAEA,OAAgB,QAAQ;AAAA,IACtB,WAAW,MAAM,QAAQ,EAAE,aAAa,4BAA4B,SAAS,EAAE,CAAC;AAAA,EAClF;AAAA,EAEA,MAAM,MAAqB;AACzB,UAAM,EAAE,MAAM,MAAM,IAAI,MAAM,KAAK,MAAM,qBAAoB;AAC7D,SAAK,YAAY;AAEjB,UAAM,SAAS,MAAM,OAA0C,qBAAqB,KAAK,QAAQ,EAAE;AACnG,SAAK,IAAI,8BAA8B,OAAO,IAAI,MAAM,OAAO,OAAO,eAAe,MAAM,SAAS,EAAE;AAEtG,UAAM,SAAS,MAAM,mBAAmB,KAAK,UAAU,MAAM,SAAS;AAEtE,QAAI,OAAO,cAAc,GAAG;AAC1B,WAAK,MAAM,uEAAuE;AAAA,IACpF;AAEA,SAAK,IAAI;AAAA,aAAgB,OAAO,SAAS;AAAA,CAAS;AAGlD,eAAW,CAAC,QAAQ,KAAK,KAAK,OAAO,QAAQ,OAAO,UAAU,GAAG;AAC/D,UAAI,WAAW,SAAS;AACtB,aAAK,IAAI;AAAA,sCAAyC;AAClD,mBAAW,CAAC,QAAQ,KAAK,KAAK,OAAO,QAAQ,KAA0C,GAAG;AACxF,eAAK,IAAI,KAAK,MAAM,KAAK,KAAK,EAAE;AAAA,QAClC;AACA;AAAA,MACF;AACA,WAAK,IAAI,GAAG,MAAM,GAAG;AACrB,iBAAW,CAAC,QAAQ,IAAI,KAAK,OAAO,QAAQ,KAAK,GAAG;AAClD,cAAM,IAAI;AACV,aAAK,IAAI,KAAK,MAAM,UAAU,EAAE,IAAI,WAAW,EAAE,MAAM,KAAK,EAAE,GAAG,KAAK,EAAE,GAAG,GAAG;AAAA,MAChF;AAAA,IACF;AAEA,QAAI,OAAO,MAAM,SAAS,GAAG;AAC3B,WAAK,IAAI,UAAU;AACnB,iBAAW,QAAQ,OAAO,OAAO;AAC/B,aAAK,IAAI,OAAO,IAAI,EAAE;AAAA,MACxB;AAAA,IACF;AAGA,UAAM,YAAY,MAAM,cAAc,KAAK,UAAU,MAAM,WAAW,MAAM;AAC5E,SAAK,IAAI;AAAA,mBAAsB,UAAU,EAAE,EAAE;AAAA,EAC/C;AACF;",
   "names": []
 }

package/dist/commands/evals/compare.js

@@ -3,13 +3,14 @@ import {
 } from "../../chunk-QN34TGD3.js";
 import {
   BaseCommand
-} from "../../chunk-SZ7VDCD6.js";
+} from "../../chunk-2HK7RGZB.js";
 import "../../chunk-R56PH7YE.js";
-import "../../chunk-SLL3OBKN.js";
+import "../../chunk-GQNJO2P2.js";
+import "../../chunk-37LN5O7M.js";
 import "../../chunk-EKJ2CABV.js";
 import {
   apiGet
-} from "../../chunk-ZRLYHBPN.js";
+} from "../../chunk-ZUJBM6HE.js";
 import "../../chunk-W2YERO7E.js";
 import "../../chunk-NSBPE2FW.js";