utim-cli 2.3.1__tar.gz → 2.3.2__tar.gz

{utim_cli-2.3.1 → utim_cli-2.3.2}/CHANGELOG.md

@@ -1,5 +1,10 @@
 # Changelog
 
+## [2.3.2] - 2026-08-20
+
+### 🚀 Performance & Fixes
+- Token usage optimization and bug fixes
+
 ## [2.3.1] - 2026-08-19
 
 ### 🛠️ Fixes & Enhancements

{utim_cli-2.3.1/utim_cli.egg-info → utim_cli-2.3.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: utim-cli
-Version: 2.3.1
+Version: 2.3.2
 Summary: UTIM – Universal Terminal Intelligence Manager. A powerful agentic AI coding assistant for your terminal.
 License: Emend AI Proprietary EULA
 Project-URL: Homepage, https://utim.dev

{utim_cli-2.3.1 → utim_cli-2.3.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "utim-cli"
-version = "2.3.1"
+version = "2.3.2"
 
 description = "UTIM – Universal Terminal Intelligence Manager. A powerful agentic AI coding assistant for your terminal."
 readme = "README.md"

{utim_cli-2.3.1 → utim_cli-2.3.2}/utim_cli/__init__.py

@@ -5,7 +5,7 @@ try:
     from utim_cli._version import VERSION as __version__
 except ImportError:
     # Fallback during first-install / editable-install edge cases
-    __version__ = "2.3.1"
+    __version__ = "2.3.2"
 
 
 import builtins, os, pathlib, urllib.parse, re

{utim_cli-2.3.1 → utim_cli-2.3.2}/utim_cli/_version.py

@@ -11,6 +11,6 @@
 # from here so every surface stays in sync automatically.
 # ─────────────────────────────────────────────────────────────────────────────
 
-VERSION = "2.3.1"
+VERSION = "2.3.2"
 
 

{utim_cli-2.3.1 → utim_cli-2.3.2}/utim_cli/context_pruner.py

@@ -102,40 +102,27 @@ def _strip_thinking_blocks(messages: List[Dict], keep_last: int = 1) -> List[Dic
     return result
 
 
-def _deduplicate_tool_outputs(messages: List[Dict], keep_last_tools: int = 5) -> List[Dict]:
-    """Compress repeated identical tool responses to a one-liner stub,
-    WHILE STRICTLY PRESERVING the most recent `keep_last_tools` tool outputs 100% verbatim.
+def _deduplicate_tool_outputs(messages: List[Dict]) -> List[Dict]:
+    """Compress repeated identical tool responses to a one-liner stub.
 
-    The last 5 tool outputs must NEVER be deduplicated, truncated, or stubbed out,
-    even if they are duplicates of earlier reads, because the model requires the exact
-    code lines and context to execute precise file edits and tool actions.
+    When the agent calls the same tool with the same arguments repeatedly and gets
+    the same output (e.g. re-reading an unchanged file multiple times), only the
+    first occurrence is kept verbatim. Subsequent duplicates are replaced with:
+      [Duplicate tool output — identical to earlier result, omitted to save tokens]
     """
-    # Identify indices of all tool messages
-    tool_msg_indices = [i for i, m in enumerate(messages) if m.get("role") == "tool"]
-    exempt_indices = set(tool_msg_indices[-keep_last_tools:]) if len(tool_msg_indices) >= 1 else set()
-
-    seen: dict = {}  # (tool_name, content_hash) -> True
+    # Map: (tool_call_id_of_preceding_assistant_call -> output_hash)
+    # We key on (tool_name, args_hash) -> first_output_hash seen
+    seen: dict = {}  # (tool_name_or_id, content_hash) -> True
     result = []
 
-    for i, msg in enumerate(messages):
+    for msg in messages:
         if msg.get("role") == "tool":
             content = msg.get("content") or ""
             tool_name = msg.get("name", "")
-
-            # CRITICAL: Never deduplicate or stub any of the last 5 tool calls!
-            if i in exempt_indices:
-                result.append(msg)
-                continue
-
-            # Never deduplicate tool error messages
-            if isinstance(content, str) and any(err in content.lower() for err in ["error", "exception", "failed", "failure", "traceback", "syntaxerror", "not found"]):
-                result.append(msg)
-                continue
-
             # Use first 300 chars of content for the hash to avoid hashing huge outputs
             content_key = (tool_name, hashlib.md5(content[:300].encode(errors="replace")).hexdigest())
             if content_key in seen and len(content) > 200:
-                # Replace older duplicate with stub
+                # Replace with stub
                 msg = dict(msg)
                 msg["content"] = (
                     f"[Duplicate {tool_name!r} output — identical to earlier result, "
@@ -333,13 +320,13 @@ def resolve_model_context_envelope(model_id: str, raw_context_window: int = 0) -
     }
 
 
-def sanitize_message_sequence(messages: List[Dict], keep_last_tools: int = 5) -> List[Dict]:
+def sanitize_message_sequence(messages: List[Dict]) -> List[Dict]:
     """
     Ensure the messages list is valid for LLM API calls.
     - Each tool message must have a preceding assistant message with the matching tool_call_id.
     - Each assistant message with tool_calls must have corresponding tool messages.
     - Strips <think> blocks from non-current assistant messages to save tokens.
-    - Deduplicates older repeated identical tool outputs (strictly preserving the last 5 tool calls verbatim).
+    - Deduplicates repeated identical tool outputs.
     """
     if not messages:
         return messages
@@ -347,8 +334,8 @@ def sanitize_message_sequence(messages: List[Dict], keep_last_tools: int = 5) ->
     # Strip thinking blocks from older assistant messages
     messages = _strip_thinking_blocks(messages)
 
-    # Compress older duplicate tool outputs while protecting last 5 tool calls
-    messages = _deduplicate_tool_outputs(messages, keep_last_tools=keep_last_tools)
+    # Compress duplicate tool outputs
+    messages = _deduplicate_tool_outputs(messages)
         
     # Step 1: Identify all tool call IDs present in tool messages
     present_tool_ids = {
@@ -400,7 +387,7 @@ def sanitize_message_sequence(messages: List[Dict], keep_last_tools: int = 5) ->
 def soft_prune_messages(
     messages: List[Dict],
     turn_msg_start: int = 1,
-    keep_last_tools: int = 5,
+    keep_last_tools: int = 4,
     max_char_threshold: int = 1200,
 ) -> List[Dict]:
     """
@@ -408,7 +395,7 @@ def soft_prune_messages(
 
     Balanced policy ('not too aggressive, not too light'):
     - Anchors: System prompt (idx 0) and the root User prompt (turn_msg_start) are strictly preserved.
-    - Active Working Tail: The most recent `keep_last_tools` (at least 5) tool results & their assistant calls are kept 100% verbatim.
+    - Active Working Tail: The most recent `keep_last_tools` tool results & their assistant calls are kept 100% verbatim.
     - Errors / Failures: Any tool output containing error keywords (e.g. 'error', 'exception', 'failed', 'traceback', 'syntaxerror') is preserved verbatim to prevent debugging amnesia.
     - Read / Search Outputs (older than tail): If longer than `max_char_threshold`, softly truncates the middle (keeps head 350 chars + tail 200 chars) with a clean diagnostic placeholder.
     - Successful Mutation / Command Outputs: If verbose, compacted to clean status indicators without losing file targets.
@@ -430,7 +417,7 @@ def soft_prune_messages(
 
     # If not enough tool messages have accumulated in this turn, just sanitize and return
     if len(tool_indices) <= keep_last_tools:
-        return sanitize_message_sequence(cleaned_messages, keep_last_tools=keep_last_tools)
+        return sanitize_message_sequence(cleaned_messages)
 
     # Older tool message indices that are eligible for soft pruning
     eligible_tool_indices = set(tool_indices[:-keep_last_tools])
@@ -536,7 +523,7 @@ def soft_prune_messages(
 
         result_messages.append(msg)
 
-    return sanitize_message_sequence(result_messages, keep_last_tools=keep_last_tools)
+    return sanitize_message_sequence(result_messages)
 
 
 def score_message_importance(message: Dict, llm_key: str = None) -> float: