patchpal 0.22.3__tar.gz → 0.22.5__tar.gz

{patchpal-0.22.3/patchpal.egg-info → patchpal-0.22.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: patchpal
-Version: 0.22.3
+Version: 0.22.5
 Summary: An agentic coding and automation assistant, supporting both local and cloud LLMs
 Author: PatchPal Contributors
 License-Expression: Apache-2.0

{patchpal-0.22.3 → patchpal-0.22.5}/patchpal/__init__.py

@@ -1,6 +1,6 @@
 """PatchPal - An open-source Claude Code clone implemented purely in Python."""
 
-__version__ = "0.22.3"
+__version__ = "0.22.5"
 
 from patchpal.agent import create_agent, create_react_agent
 from patchpal.cli.autopilot import autopilot_loop

{patchpal-0.22.3 → patchpal-0.22.5}/patchpal/agent/function_calling.py

@@ -741,6 +741,7 @@ It's currently empty (just the template). The file is automatically loaded at se
                 response = litellm.completion(
                     model=self.model_id,
                     messages=messages,
+                    max_tokens=32000,  # Explicit output token limit for predictable context usage
                     timeout=LLM_TIMEOUT,
                     **self.litellm_kwargs,
                 )
@@ -957,10 +958,8 @@ It's currently empty (just the template). The file is automatically loaded at se
 
         # Check for compaction BEFORE starting work
         # This ensures we never compact mid-execution and lose tool results
-        # Use last_prompt_tokens from previous API call for accurate check (includes cache operations)
-        if self.enable_auto_compact and self.context_manager.needs_compaction(
-            self.messages, actual_prompt_tokens=self.last_prompt_tokens
-        ):
+        # Always estimates current messages to avoid staleness issues (no actual_prompt_tokens)
+        if self.enable_auto_compact and self.context_manager.needs_compaction(self.messages):
             self._perform_auto_compaction()
 
         # Agent loop with interrupt handling
@@ -1080,6 +1079,7 @@ It's currently empty (just the template). The file is automatically loaded at se
                         "messages": messages,
                         "tools": tools,
                         "tool_choice": "auto",
+                        "max_tokens": 32000,  # Explicit output token limit for predictable context usage
                         "timeout": LLM_TIMEOUT,
                         "stream": stream,
                         **self.litellm_kwargs,

{patchpal-0.22.3 → patchpal-0.22.5}/patchpal/context.py

@@ -257,7 +257,10 @@ Be comprehensive but concise. The goal is to continue work seamlessly without lo
         self.system_prompt = system_prompt
         self.estimator = TokenEstimator(model_id)
         self.context_limit = self._get_context_limit()
-        self.output_reserve = 4_096  # Reserve tokens for model output
+        # Reserve 16% of context for output (min 4K, max 32K)
+        # This ensures older models like GPT-4 (8K) get 1.28K reserve
+        # while modern models get full 32K reserve
+        self.output_reserve = min(32_000, max(4_000, int(self.context_limit * 0.16)))
 
     def _get_context_limit(self) -> int:
         """Get context limit for model.
@@ -321,30 +324,31 @@ Be comprehensive but concise. The goal is to continue work seamlessly without lo
     ) -> bool:
         """Check if context window needs compaction.
 
-        Supports both reactive (preferred) and proactive (fallback) approaches:
-        - Reactive: Use actual_prompt_tokens from latest API response
-        - Proactive: Estimate tokens if actual_prompt_tokens not available
+        ALWAYS estimates current messages to avoid staleness issues when predicting
+        whether the NEXT API call will overflow. Using actual_prompt_tokens from a
+        previous call can cause false negatives when large messages are added between
+        the last API call and the compaction check.
 
-        The reactive approach is preferred as it uses actual token counts from the LLM,
-        avoiding the need for tiktoken or other estimation methods.
+        Example of staleness bug (fixed):
+        - Previous API call: 120K tokens (60% usage)
+        - User pastes huge changelog: +90K tokens
+        - Total: 210K tokens (exceeds 200K limit)
+        - Bug: If we used actual_prompt_tokens=120K, we'd think we're at 60%
+        - Fix: Always re-estimate to see the 210K total
+
+        The actual_prompt_tokens parameter is kept for API compatibility but ignored
+        for compaction decisions. Use get_usage_stats() for display purposes where
+        actual tokens are appropriate (staleness OK for showing recent stats).
 
         Args:
             messages: Current message history
-            actual_prompt_tokens: Optional actual prompt token count from latest API response
+            actual_prompt_tokens: IGNORED - kept for API compatibility only
 
         Returns:
             True if compaction is needed
         """
-        # Reactive approach (preferred): use actual token counts from API response
-        if actual_prompt_tokens is not None:
-            # Add output reserve to account for response tokens
-            total_tokens = actual_prompt_tokens + self.output_reserve
-            usage_ratio = total_tokens / self.context_limit
-            return usage_ratio >= self.COMPACT_THRESHOLD
-
-        # Proactive approach (fallback): estimate tokens when API data not available
-        # This uses character-based estimation (3 chars per token) which works
-        # reliably without requiring tiktoken or network access
+        # ALWAYS estimate current messages - never use stale actual_prompt_tokens
+        # This ensures we detect large message additions that happen between API calls
         # Note: Dynamic date/time message adds ~30 tokens on each LLM call
         system_tokens = self.estimator.estimate_tokens(self.system_prompt)
         datetime_tokens = 30  # Approximate size of dynamic date/time message

{patchpal-0.22.3 → patchpal-0.22.5/patchpal.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: patchpal
-Version: 0.22.3
+Version: 0.22.5
 Summary: An agentic coding and automation assistant, supporting both local and cloud LLMs
 Author: PatchPal Contributors
 License-Expression: Apache-2.0