@tiens.nguyen/gonext-local-worker 1.0.253 → 1.0.254

package/gonext_agent_chat.py

@@ -3117,11 +3117,17 @@ def run_agent_chat(cfg):
         [_active_ws] if _active_ws else _WS_ROOTS
     ) if _WS_AVAILABLE else []
     _ws_overview = _workspace_overview(_overview_roots) if _overview_roots else ""
-    _ws_tool_block = (
+    # VARIABLE workspace context (task #86): the active-folder line, workspace list and
+    # the on-disk folder overview change between turns (the agent creates/edits files),
+    # so they must NOT sit inside the static tool_hint — they'd bust the prefix cache
+    # mid-reference every turn. They join the variable tail of task_with_hint instead.
+    _ws_context_block = (
         _ws_current_line
         + f"  Other registered workspaces (use only if the user names them): {_ws_names}\n"
         f"  Current folder contents:\n"
         + "\n".join(f"    {line}" for line in _ws_overview.splitlines()) + "\n"
+    ) if _WS_AVAILABLE else ""
+    _ws_tool_block = (
         "    - list_dir(path) / read_text_file(path) — browse and read files.\n"
         "    - grep_repo(pattern, path='', glob='') — search code, returns file:line hits. "
         "Use this to find the EXACT lines to change.\n"
@@ -3225,19 +3231,18 @@ def run_agent_chat(cfg):
             f"(Permission denied publickey,password), STOP and tell the user to run "
             f"'ssh-copy-id {_du}@{_dh}' once, then retry. ***\n"
         )
+    # STATIC ONLY (task #86): everything in tool_hint must be byte-identical between
+    # turns so the model server's prefix KV-cache covers it (see task_with_hint below).
+    # Variable text — the deploy target and the current date/time — lives in the
+    # variable tail next to the task instead.
     tool_hint = (
-        # Deploy target FIRST (when a server is picked) — a small model buries context deep
-        # in a 20k-char hint, so the chosen server must lead, not trail (task #69 retest fix).
-        _deploy_block
-        + f"Solve the TASK step by step (up to {max_steps} steps). At EACH step write a "
+        f"Solve the TASK step by step (up to {max_steps} steps). At EACH step write a "
         "brief Thought, then ONE code block that calls a SINGLE tool. You will SEE that "
         "tool's result (Observation) before the next step — use it to decide what to do "
         "next (e.g. web_search to find a URL, THEN fetch_url to read it). When you have "
         "enough to answer, call final_answer(<your answer>) in a code block. Prefer "
         "FEWER steps, and never repeat a tool call that already succeeded.\n\n"
         "{{TOOL_LIST}}\n"  # filled in below, once from the REAL registered tool objects
-        f"(Current date/time: {now_str} — pass a timezone to get_current_datetime() only "
-        "if the task needs a DIFFERENT one.)\n"
         + _rag_tool_block + _ws_tool_block + _auto_test_block +
         "\n"
         "http_request RETURN FORMAT: 'HTTP 200\\n{body}' — first line is 'HTTP <code>', body follows.\n"
@@ -3322,14 +3327,30 @@ def run_agent_chat(cfg):
                 f"Previous task: {_ck.get('task', '')}\n" + "\n".join(_kept)
             )
 
-    # Lead with the TASK so the weak model anchors on what's actually being asked —
-    # not on the tool reference below. (Previously the hint led with the date, and the
-    # 3B model treated every message as a date question.)
+    # ORDER (task #86 — prefix-cache fix): STATIC first, VARIABLE last.
+    # The ~16k-char tool reference is byte-identical between turns; leading with it
+    # (right after the equally static system prompt) lets the model server's prefix
+    # KV-cache cover ~7k tokens across turns. Previously the variable task text came
+    # FIRST, so every new turn forced a cold re-eval of the whole reference (~2 min to
+    # first token on step 1 of each turn; steps 2+ were 10-30s because within a turn
+    # the prompt only grows at the end).
+    # The TASK now comes LAST, under a loud header: recency is the position a model
+    # attends to most, so anchoring is preserved (the historical "hint led with the
+    # date → 3B model answered everything as a date question" regression was about a
+    # misleading LEAD, not about where the task sits). The two VARIABLE blocks — the
+    # current date/time and the deploy target — live here in the tail, next to the
+    # task: the deploy target must stay near the task, not buried mid-hint (#69), and
+    # the timestamp changes every minute, which would bust the cache mid-reference.
     task_with_hint = (
-        "TASK (answer THIS, choose the tool that fits it):\n"
-        f"{task_text}\n\n"
-        "----- TOOL REFERENCE -----\n"
-        + tool_hint
+        "----- TOOL REFERENCE (how to work — your actual TASK follows below) -----\n"
+        + tool_hint +
+        "\n----- YOUR TASK -----\n"
+        f"(Current date/time: {now_str} — pass a timezone to get_current_datetime() only "
+        "if the task needs a DIFFERENT one.)\n"
+        + _deploy_block
+        + _ws_context_block +
+        "\nTASK (answer THIS, choose the tool that fits it):\n"
+        f"{task_text}\n"
     )
 
     # Track URLs that have already failed so we don't retry dead endpoints across steps.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.253",
+  "version": "1.0.254",
   "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",