@tiens.nguyen/gonext-local-worker 1.0.314 → 1.0.316

package/gonext_agent_chat.py

@@ -5052,8 +5052,27 @@ def run_agent_chat(cfg):
                         pass
                     break
             content = "".join(parts)
+            # Fallback OUTPUT-token count (task #112): some OpenAI-compatible servers (seen
+            # live: Moonshot/Kimi K3) don't emit a final usage chunk on every streamed call,
+            # and the runaway guard closes the stream BEFORE the trailing usage chunk — so a
+            # step that DID generate output reports completion_tokens=0 and contributes
+            # nothing to the per-turn "↓ out" total (~3x under-count on reasoning models,
+            # whose output is mostly reasoning-channel tokens). When the server reported no
+            # completion tokens, estimate from the streamed OUTPUT chars (visible content +
+            # reasoning, ~4 chars/token, matching the #83 input estimate) so no generating
+            # step silently counts 0. Authoritative server usage always wins when present
+            # (out_tok>0), so Ollama — which always reports usage — is unchanged; MLX, which
+            # reports none, now gets an estimate instead of a hard 0.
+            if out_tok <= 0:
+                approx_chars = len(content) + reasoning_len
+                if approx_chars > 0:
+                    out_tok = max(1, (approx_chars + 3) // 4)
+                    _log(f"streamed generate: server reported no completion tokens → "
+                         f"estimated out≈{out_tok} tokens from {approx_chars} streamed "
+                         "chars (content+reasoning) (#112)")
             if _runaway:
                 # Discard the garbage; "" makes _streamed_generate run the CODE-NOW retry.
+                # Keep out_tok (the estimate above) — those tokens WERE generated/billed.
                 _emit({"type": "step", "text": "Model output ran away — restarting this step…"})
                 _log(f"streamed generate: discarded {len(content)} runaway chars → empty content")
                 return "", role, in_tok, out_tok, reasoning_len

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.314",
+  "version": "1.0.316",
   "description": "Polls GoNext cloud API for async local LLM jobs and runs them against Ollama/OpenAI-compatible servers on this Mac",
   "type": "module",
   "license": "MIT",