feat: Claude Sonnet 4 1M context; Cursor rules budget 10% (cap 100k)

README.md

@@ -181,6 +181,8 @@ alwaysApply: true
 
 The bot will automatically include these rules in its analysis, ensuring consistent code reviews that match your project's standards.
 
+By default, up to 10% of the model context (capped at 100k tokens) is allocated to Cursor rules. On 1M-context models, this allows up to 100k tokens of rules to be included.
+
 ## Usage
 
 ### Automatic Mode

pr_agent/algo/ai_handlers/litellm_ai_handler.py

@@ -1,7 +1,38 @@
 import os
-import litellm
 import openai
 import requests
+try:
+    # Shim: align Litellm's expected OpenAI type name with current OpenAI SDK
+    # Older Litellm versions import `ResponseTextConfig`, which was renamed to
+    # `ResponseFormatTextConfig` in newer OpenAI SDKs.
+    from openai.types.responses import (
+        response as _openai_response_mod,
+        response_create_params as _openai_response_params_mod,
+    )
+    if (
+        not hasattr(_openai_response_mod, "ResponseTextConfig")
+        and hasattr(_openai_response_mod, "ResponseFormatTextConfig")
+    ):
+        setattr(
+            _openai_response_mod,
+            "ResponseTextConfig",
+            _openai_response_mod.ResponseFormatTextConfig,
+        )
+    # Alias param type as well if needed
+    if (
+        not hasattr(_openai_response_params_mod, "ResponseTextConfigParam")
+        and hasattr(_openai_response_params_mod, "ResponseFormatTextConfigParam")
+    ):
+        setattr(
+            _openai_response_params_mod,
+            "ResponseTextConfigParam",
+            _openai_response_params_mod.ResponseFormatTextConfigParam,
+        )
+except Exception:
+    # Best effort compatibility; proceed even if aliasing fails
+    pass
+
+import litellm
 from litellm import acompletion
 from tenacity import retry, retry_if_exception_type, stop_after_attempt
 

pr_agent/git_providers/utils.py

@@ -178,8 +178,8 @@ def get_repository_rules_for_prompt():
             token_handler = TokenHandler()
             rules_tokens = token_handler.count_tokens(rules_content)
 
-            max_rules_tokens = int(get_settings().config.get('max_cursor_rules_tokens', 20000))
-            hard_cap_ratio = float(get_settings().config.get('cursor_rules_context_ratio', 0.25))
+            max_rules_tokens = int(get_settings().config.get('max_cursor_rules_tokens', 100000))
+            hard_cap_ratio = float(get_settings().config.get('cursor_rules_context_ratio', 0.10))
             model_ctx = get_max_tokens(model)
             hard_cap_tokens = max(2000, int(model_ctx * hard_cap_ratio))
             allowed_tokens = min(max_rules_tokens, hard_cap_tokens)

pr_agent/settings/configuration.toml

@@ -76,8 +76,8 @@ feedback_on_draft_pr=false # Set to true to enable processing of draft PRs
 use_cursor_rules=true # Set to true to enable reading official Cursor rules from repositories (.cursor/rules/*.mdc and .cursorrules)
 
 # Cursor rules budgeting
-max_cursor_rules_tokens = 20000
-cursor_rules_context_ratio = 0.25
+max_cursor_rules_tokens = 100000
+cursor_rules_context_ratio = 0.10
 
 # Default output budget when not using extended thinking
 default_max_output_tokens = 2048

requirements.txt

@@ -1,7 +1,7 @@
 # Core AI provider support
-# Bump LiteLLM for GPT-5 routing and latest providers
-litellm>=1.74.0
-openai>=1.55.3
+# Keep LiteLLM and OpenAI in compatible versions to avoid type import mismatches
+litellm==1.75.8
+openai==1.100.1
 anthropic>=0.48
 tiktoken==0.8.0