nexrall-code 0.5.101 → 0.5.102

package/dist/index.js

@@ -153717,6 +153717,41 @@ async function runTurnHeadless(messages, modelAlias, workDir, _abortSignal, env4
       resultText += text;
       emit({ type: "text", text });
     },
+    // ── Reasoning-phase liveness ──────────────────────────────────────
+    // These used to be UNWIRED here while runTurn() (interactive) wired all
+    // three, and the asymmetry was not cosmetic: on a reasoning model at high
+    // effort the thinking phase emits no text and no tool calls, so a headless
+    // consumer saw ZERO events for the entire phase and could not distinguish
+    // "actively reasoning" from "process wedged".
+    //
+    // Measured on a Terminal-Bench 4.0 trial (2026-08-30, deepseek-v4-pro
+    // --effort max): nex-output.jsonl sat at 4 lines for 65 MINUTES while
+    // tcpdump inside the container's netns showed ~25 packets/s still flowing
+    // and the turn ultimately reported 310,872 output tokens. Nothing was
+    // actually wrong — but every signal available to the harness (log line
+    // count, file mtime) said "hung", and the socket/CPU forensics needed to
+    // prove otherwise are not something a CI wrapper can do.
+    //
+    // `thinking_progress` is the load-bearing one: it carries the backend's
+    // cumulative output-token count (routes/code.js's sendProgress, already
+    // throttled to <=5/s server-side, so this cannot flood the log) and fires
+    // DURING the phase. That makes it a real heartbeat.
+    onThinkingProgress: (tokens) => {
+      emit({ type: "thinking_progress", tokens });
+    },
+    // Fires once per turn at message_complete with the full reasoning text.
+    // Typed `thinking` deliberately: the bench's ATIF converter (atif.py)
+    // already routes exactly this event type into the trajectory's reasoning
+    // buffer, so wiring it here also fills in reasoning that was previously
+    // dropped on the floor for every headless run.
+    onThinking: (text) => {
+      emit({ type: "thinking", text });
+    },
+    // onThinkingDelta is intentionally NOT wired. Its chunks concatenate to the
+    // same string `onThinking` emits in full above, so emitting both would
+    // duplicate the entire reasoning trace — on the 310k-token turn measured
+    // above that is megabytes of redundant NDJSON — while adding no liveness
+    // signal `thinking_progress` does not already provide.
     // System notice (mid-run auto-prune/auto-compact) — emit as its own event
     // type instead of falling through to onText, so a stream-json consumer
     // doesn't see compaction housekeeping text mixed into the model's `text`

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nexrall-code",
-  "version": "0.5.101",
+  "version": "0.5.102",
   "description": "Nexrall Code — AI coding assistant for your terminal (headless agent for scripts, CI and automation)",
   "keywords": [
     "ai",