impel-cli 0.20.46-beta.0 → 0.20.46

package/src/doctor.js

@@ -10,6 +10,7 @@ const MAX_ERROR_BODY_BYTES = 64 * 1024;
 const MAX_OUTPUT_CHARS = 16 * 1024;
 const MAX_PENDING_SSE_CHARS = 1024 * 1024;
 const MAX_STREAM_BYTES = 4 * 1024 * 1024;
+const CLAUDE_PROBE_MAX_TOKENS = 1_024;
 const GATEWAY_REQUEST_ID_RE = /^impel-[a-f0-9]{32}$/u;
 const POOL_STATES = new Set(["healthy", "exhausted", "rate_limited", "expired", "no_seat"]);
 
@@ -86,6 +87,12 @@ function consumeSseBlock(provider, block, state, startedAt, now) {
     if (!state.responseId && typeof event?.message?.id === "string") {
       state.responseId = event.message.id.slice(0, 512) || null;
     }
+    if (event?.type === "message_delta" && typeof event?.delta?.stop_reason === "string") {
+      state.stopReason = event.delta.stop_reason.slice(0, 100);
+      if (state.stopReason !== "end_turn") {
+        state.error ||= `provider stopped the response with reason ${redactSecretText(state.stopReason)}`;
+      }
+    }
     if (event?.type === "message_stop") state.terminalEvent = event.type;
     if (event?.type === "content_block_delta" && event?.delta?.type === "text_delta") {
       delta = typeof event.delta.text === "string" ? event.delta.text : "";
@@ -131,7 +138,11 @@ async function streamProbe({ provider, url, bearer, model, tenantId, attempt, ti
   const body = provider === "claude"
     ? {
         model,
-        max_tokens: 96,
+        // Current reasoning-capable Claude models may spend part of this budget
+        // before emitting the tiny acknowledgement. Ninety-six tokens made a
+        // healthy route look broken when reasoning exhausted the response
+        // budget before the first text delta.
+        max_tokens: CLAUDE_PROBE_MAX_TOKENS,
         stream: true,
         messages: [{ role: "user", content: prompt }],
       }
@@ -192,6 +203,7 @@ async function streamProbe({ provider, url, bearer, model, tenantId, attempt, ti
         exactAck: false,
         outputChars: 0,
         outputTruncated: false,
+        stopReason: null,
         terminalEvent: null,
         doneSentinel: false,
         receivedBytes: 0,
@@ -204,6 +216,7 @@ async function streamProbe({ provider, url, bearer, model, tenantId, attempt, ti
       responseId: null,
       ttftMs: null,
       outputTruncated: false,
+      stopReason: null,
       terminalEvent: null,
       doneSentinel: false,
       error: null,
@@ -262,6 +275,7 @@ async function streamProbe({ provider, url, bearer, model, tenantId, attempt, ti
       exactAck,
       outputChars: state.answer.length,
       outputTruncated: state.outputTruncated,
+      stopReason: state.stopReason,
       terminalEvent: state.terminalEvent,
       doneSentinel: state.doneSentinel,
       receivedBytes,
@@ -283,6 +297,7 @@ async function streamProbe({ provider, url, bearer, model, tenantId, attempt, ti
       exactAck: false,
       outputChars: 0,
       outputTruncated: false,
+      stopReason: null,
       terminalEvent: null,
       doneSentinel: false,
       receivedBytes: 0,
@@ -488,6 +503,7 @@ export async function probeTenant({
           exactAck: false,
           outputChars: 0,
           outputTruncated: false,
+          stopReason: null,
           terminalEvent: null,
           doneSentinel: false,
           receivedBytes: 0,

package/src/managedProfileVersion.js

@@ -1,4 +1,6 @@
 // Fleet generation shared by managed-profile writers and upstream clients.
 // Keep this isolated from apps.js so latency-sensitive transports do not load
 // desktop bundle machinery just to identify their managed config contract.
-export const CURRENT_CONFIG_VERSION = 37;
+// v38 replaces the generated desktop MCP Apps host with the authenticated web
+// board and removes the desktop-only Codex renderer flag.
+export const CURRENT_CONFIG_VERSION = 38;