@ai-matrx/agents 0.22.0 → 0.23.0

package/CHANGELOG.md

@@ -1,5 +1,12 @@
 # Changelog
 
+## 0.23.0
+
+- **Resume never runs a live run twice — every client rejoins it** (`./matrx`). New `resumeOrRejoin(transport, resumeCall, { organizationId, onEnvelope, … })`: runs the feature's resume door and pipes its NDJSON into `onEnvelope`; on `409 run_in_progress` it follows `POST /runtime/operations/{live_request_id}/rejoin` into the SAME handler under the run's own organization; on `409 live_stream_unavailable` it follows the durable lifecycle (`followRuntimeOperationToEnd`) and says so (`kind: "followed"`); `409 resume_conflict` comes back as `kind: "resume_conflict"` (retry is host policy). Building blocks exported: `readLiveRunRequestId` (reads `live_request_id`, NEVER the envelope's own `request_id`), `readMatrxErrorCode`, `isResumeConflict`, `isLiveStreamUnavailable`, `runtimeOperationRejoinPath`, `openMatrxStream` (any NDJSON door by path), `withRunOrganization`. Guard: `matrx/rejoin.test.ts` (real envelopes with the `request_id` decoy; proven red on a mutant reading `request_id` and on one dropping the run org).
+- **`streamErrorText(event)`** — the text for a stream `error` event: `user_message` first (e.g. "OpenAI refused this request: the platform's OpenAI account is out of credit."), then `message`, else null. A host fallback must never replace a server message. `runAgentToCompletion` now uses it.
+- **`reportProviderSessionFailure(transport, failure)`** — browser-held provider sessions (OpenAI/xAI realtime voice, Cartesia TTS) report their failures to `POST /broker/provider-failures` and get `{ error_type, retryable, user_message }` back; the body is bounded exactly as the server model bounds it; never throws (null + warn when the report itself fails). Guard: `matrx/provider-failures.test.ts`.
+- Consumer action: replace hand-written `409 run_in_progress` → rejoin code with `resumeOrRejoin`; render stream `error` events with `streamErrorText`; call `reportProviderSessionFailure` from every realtime/Cartesia session error handler and show its `user_message`.
+
 ## 0.22.0
 
 - **A signed-out visitor never reads the agent catalogue** (live 2026-10-01: every guest page that mounted a picker — `/education/progress`, `/demos/spatial` — fired `agx_get_list_full`, which `anon` cannot execute, and captured `permission denied for function agx_get_list_full` plus `identity.requireUserId() threw`). `ensureLoaded` now answers the empty catalogue with no read when there is no user, and does not stamp freshness, so the first load after sign-in reads for real. The pickers (web + native) show `labels.emptySignedOut` ("Sign in to see your agents") instead of "No agents found". Guard: `catalog/__tests__/store.test.ts` › "signed out never reads the catalogue".

package/dist/content-transfer/index.cjs

@@ -1019,39 +1019,6 @@ async function requestStream(transport, path, options) {
   return response;
 }
 
-// matrx/run.ts
-async function streamCall(transport, path, body, options) {
-  const response = await requestStream(transport, path, {
-    method: "POST",
-    // This client IS the streaming path — `stream: true` always, last so a
-    // caller-supplied value can never flip the response off NDJSON.
-    body: { ...body, stream: true },
-    ...options.signal ? { signal: options.signal } : {}
-  });
-  return toRunHandle(response, options);
-}
-function startMandateRun(transport, mandateKey, request, options = {}) {
-  return streamCall(
-    transport,
-    `/ai/mandates/${encodePathSegment(mandateKey)}`,
-    request,
-    options
-  );
-}
-function cancelAgentRun(transport, requestId, options = {}) {
-  const query = buildQuery(
-    options.mode === "interrupt" ? { mode: "interrupt" } : {}
-  );
-  return requestJson(
-    transport,
-    `/ai/cancel/${encodePathSegment(requestId)}${query}`,
-    {
-      method: "POST",
-      ...options.signal ? { signal: options.signal } : {}
-    }
-  );
-}
-
 // stream/sse.ts
 var FRAME_SEPARATOR = /\r\n\r\n|\n\n|\r\r/;
 var LINE_SEPARATOR = /\r\n|\n|\r/;
@@ -1250,6 +1217,39 @@ async function rejoinRuntimeOperation(transport, requestId, options = {}) {
   return toRunHandle(response, options);
 }
 
+// matrx/run.ts
+async function streamCall(transport, path, body, options) {
+  const response = await requestStream(transport, path, {
+    method: "POST",
+    // This client IS the streaming path — `stream: true` always, last so a
+    // caller-supplied value can never flip the response off NDJSON.
+    body: { ...body, stream: true },
+    ...options.signal ? { signal: options.signal } : {}
+  });
+  return toRunHandle(response, options);
+}
+function startMandateRun(transport, mandateKey, request, options = {}) {
+  return streamCall(
+    transport,
+    `/ai/mandates/${encodePathSegment(mandateKey)}`,
+    request,
+    options
+  );
+}
+function cancelAgentRun(transport, requestId, options = {}) {
+  const query = buildQuery(
+    options.mode === "interrupt" ? { mode: "interrupt" } : {}
+  );
+  return requestJson(
+    transport,
+    `/ai/cancel/${encodePathSegment(requestId)}${query}`,
+    {
+      method: "POST",
+      ...options.signal ? { signal: options.signal } : {}
+    }
+  );
+}
+
 // content-transfer/destination-ports.ts
 var success = (message, target) => ({ status: "success", delivered: "action", mimeTypes: [], message, ...target ? { target } : {} });
 var active = (bindings, content, expected) => {