@tiens.nguyen/gonext-local-worker 1.0.255 → 1.0.256

package/gonext_agent_chat.py

@@ -3890,22 +3890,49 @@ def run_agent_chat(cfg):
         except Exception:  # noqa: BLE001
             pass
 
-        # (B) Trim OLD, LARGE tool observations out of the CodeAgent's step memory so
-        # they aren't re-sent to the model on every subsequent step (the #63 O(n²) token
-        # blow-up: an uncapped read_file_lines dump cost ~5-6k tokens PER remaining step).
-        # Keep the last 2 steps' observations intact (the model just acted on them);
-        # replace older long ones with a short placeholder — the model can re-read.
+        # (B) Trim OLD, LARGE step content out of the CodeAgent's step memory so it isn't
+        # re-sent to the model on every subsequent step (the #63 O(n²) token blow-up: an
+        # uncapped dump cost ~5-6k tokens PER remaining step). Keep the last 2 steps intact
+        # (the model just acted on them); shrink older long content. write_memory_to_messages
+        # rebuilds the prompt from memory.steps on EVERY request (no cache), so mutating the
+        # step objects here takes effect on the next request; old steps are only rendered
+        # into history, never re-parsed for execution, so trimming them is safe.
+        #   - observations = the tool's OUTPUT (a file/dir dump). Trimmed since #63.
+        #   - model_output = the model's own Thought + <code> block (task #87). For an OLD
+        #     step the code ALREADY ran and its result is in the observation, so re-sending
+        #     the code verbatim every step is pure waste — keep the Thought, drop the code.
+        # Both trims are IDEMPOTENT (a re-trimmed field falls under the size threshold, so a
+        # later step_callback skips it — the counter below only logs NEW trims).
         try:
             _mem = getattr(_agent_ref.get("a"), "memory", None)
             _steps = getattr(_mem, "steps", None)
             if _steps and len(_steps) > 2:
+                _trim_n, _trim_saved = 0, 0
                 for _st in _steps[:-2]:
                     _obs = getattr(_st, "observations", None)
                     if isinstance(_obs, str) and len(_obs) > 900:
+                        _before = len(_obs)
                         _st.observations = (
                             _obs[:400].rstrip()
                             + "\n…[earlier observation trimmed to save context — "
                             "re-read the file/dir with a small range if you still need it]")
+                        _trim_n += 1
+                        _trim_saved += _before - len(_st.observations)
+                    _mo = getattr(_st, "model_output", None)
+                    if isinstance(_mo, str) and len(_mo) > 700:
+                        _before = len(_mo)
+                        # Keep the Thought (everything before the code block); the code is
+                        # redundant with the observation that already came back.
+                        _head = re.split(r"<code[\s>]|```", _mo, maxsplit=1)[0].rstrip()
+                        _kept = (_head[:500].rstrip() if _head else _mo[:300].rstrip())
+                        _new = _kept + "\n…[code trimmed to save context — its result is in the Observation]"
+                        if len(_new) < _before:
+                            _st.model_output = _new
+                            _trim_n += 1
+                            _trim_saved += _before - len(_new)
+                if _trim_n:
+                    _log(f"context trim: shrank {_trim_n} old step field(s), "
+                         f"saved ~{_trim_saved} chars (#63/#87)")
         except Exception as _e:  # noqa: BLE001
             _log(f"memory-trim skip: {_e}")
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.255",
+  "version": "1.0.256",
   "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",