@integrity-labs/agt-cli 0.28.104 → 0.28.105

package/dist/mcp/augmented-admin.js

@@ -21131,6 +21131,10 @@ var AdminDebugClient = class _AdminDebugClient {
   tailLogs(args) {
     return this.get("/admin/debug/tail-logs", _AdminDebugClient.cleanQuery(args));
   }
+  /** Conversation eval-FAILURE breakdown: why conversations couldn't be scored (ENG-6661). */
+  evalFailures(args) {
+    return this.get("/admin/debug/eval-failures", _AdminDebugClient.cleanQuery(args));
+  }
   /** Time-windowed read of ONE allowlisted host log across rotated files (ENG-6515). */
   queryLogs(args) {
     return this.get("/admin/debug/query-logs", _AdminDebugClient.cleanQuery(args));
@@ -21298,6 +21302,10 @@ var inspectFlagsSchema = external_exports.object({
   host: external_exports.string().min(1).max(128).optional().describe("The host id (uuid) or its exact, case-insensitive name. Pass this OR agent_id."),
   agent_id: external_exports.string().min(1).max(64).optional().describe("The agent UUID \u2014 flags are inspected on the agent's current host. Pass this OR host.")
 });
+var evalFailuresSchema = external_exports.object({
+  period: external_exports.enum(["24h", "7d", "30d"]).optional().describe("Window to report over (24h | 7d | 30d). Default 7d."),
+  agent_id: external_exports.string().min(1).max(64).optional().describe("Filter to one agent (UUID); omit for a fleet-wide breakdown across your authorized orgs.")
+});
 var alertIdField = external_exports.string().min(1).max(64).describe("The alert id (uuid), from debug_list_alerts.");
 var triageReasonField = external_exports.string().max(2e3).optional().describe("Why \u2014 recorded on the cross-org audit trail. Be specific.");
 var ackAlertSchema = external_exports.object({
@@ -21679,6 +21687,19 @@ server.tool(
     }
   }
 );
+server.tool(
+  "debug_eval_failures",
+  'Report WHY conversations could not be scored by the host-side success evaluator: the failure counterpart to the "conversation success" trend, which only ever shows EVALUATED conversations (so an agent whose conversations all fail to reconstruct silently vanishes from it). Use when conversation scores look thin/absent for an agent, or to audit eval coverage fleet-wide. Returns { scope, period, scanned, truncated, report, recent }. `report` is the aggregate: { total, gaveUp (subset that hit the retry cap, will never be scored), byReason[] (org-wide count + gaveUp per reason), perAgent[] (worst-first standing) }. The three reasons are data-shaped: `no_transcript` (the host had no local transcript at all), `not_reconstructable` (transcripts exist but none carry the conversation\'s channel ref), `empty_transcript` (turns reconstructed but rendered to nothing). `recent` is the newest failing conversations (boundary + reason only, NEVER transcript content). NOTE: transient/misconfigured EVAL BACKEND failures are NOT here; those show on the host (hosts.eval_backend_*), check debug_get_host. Pass { period? (24h|7d|30d, default 7d), agent_id? (omit for fleet-wide) }. Org-walled in SQL; every call is audited as a cross-org access.',
+  evalFailuresSchema.shape,
+  async (args) => {
+    try {
+      const result = await client.evalFailures(args);
+      return { content: [{ type: "text", text: JSON.stringify(result, null, 2) }] };
+    } catch (err) {
+      return { content: [{ type: "text", text: formatError2(err) }], isError: true };
+    }
+  }
+);
 server.tool(
   "debug_probe_integration",
   "Force a FRESH host-side connectivity probe for ONE of an agent's integrations and return the LIVE verdict \u2014 use when a user reports an integration \"isn't working\" but the console shows it connected, or to confirm a fix landed. Unlike the cached status the dashboard shows, this SSH-less SSM-invokes the agent's host to run the probe against its actual wired `.mcp.json` + credentials, so it can't disagree with reality. Returns { agent_id, code_name, slug, host, verdict, message, probed_at, ssm_status }. `verdict` is the live result: `ok` | `degraded` | `transient_error` (retryable) | `down` | `not_probeable` (no probe wired for this kind) \u2014 or a central degrade: `unreachable` (agent has no host), `not_installed` (integration not on this agent), `host_cli_too_old` (the host predates the probe \u2014 update its agt-cli), `probe_error`. Pass { agent_id, slug } where slug is the integration's definition code_name (e.g. `gmail`, `slack`, `here-now`). Fails closed for an unauthorized org; every call is audited as a cross-org host access.",

package/dist/{persistent-session-6CU3BIHP.js → persistent-session-Q65SZNJ6.js}

@@ -34,8 +34,8 @@ import {
   writeDirectChatSessionState,
   writeEgressAllowlist,
   writePersistentClaudeWrapper
-} from "./chunk-WNN5WT42.js";
-import "./chunk-5DYG42FL.js";
+} from "./chunk-GBE523G5.js";
+import "./chunk-YUTJO6FU.js";
 import "./chunk-XWVM4KPK.js";
 export {
   EGRESS_BASELINE_DOMAINS,
@@ -74,4 +74,4 @@ export {
   writeEgressAllowlist,
   writePersistentClaudeWrapper
 };
-//# sourceMappingURL=persistent-session-6CU3BIHP.js.map
+//# sourceMappingURL=persistent-session-Q65SZNJ6.js.map

package/dist/{responsiveness-probe-GLZZZ2OK.js → responsiveness-probe-FS3JNFI2.js}

@@ -1,7 +1,7 @@
 import {
   paneLogPath
-} from "./chunk-WNN5WT42.js";
-import "./chunk-5DYG42FL.js";
+} from "./chunk-GBE523G5.js";
+import "./chunk-YUTJO6FU.js";
 import "./chunk-XWVM4KPK.js";
 
 // src/lib/responsiveness-probe.ts
@@ -250,4 +250,4 @@ export {
   parkPendingInbound,
   readAndResetChannelDeflections
 };
-//# sourceMappingURL=responsiveness-probe-GLZZZ2OK.js.map
+//# sourceMappingURL=responsiveness-probe-FS3JNFI2.js.map

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@integrity-labs/agt-cli",
-  "version": "0.28.104",
+  "version": "0.28.105",
   "description": "Augmented Team CLI — agent provisioning and management",
   "type": "module",
   "engines": {