Keep large PR reviews complete instead of degrading (#874)

actions/__init__.py

@@ -28,4 +28,4 @@
 #     ├── test_summarize_pr.py
 #     └── ...
 
-__version__ = "0.3.7"
+__version__ = "0.3.8"

actions/review_pr.py

@@ -536,12 +536,11 @@ def generate_pr_review(
         if file_contents:
             full_files_section = f"FULL FILE CONTENTS:\n{chr(10).join(file_contents)}\n\n"
 
-    # Calculate remaining budget for diff and check if truncation needed
-    diff_budget = max(
-        1000, REVIEW_PROMPT_CHARS - len(guidelines_section) - len(full_files_section) - len(history_section)
-    )
+    # Remaining budget for the diff: agent turns cache the prompt and page any remainder via read_diff, so the agent
+    # inlines ~5k diff lines while single-shot models keep the one-request ceiling
+    prompt_chars = REVIEW_PROMPT_CHARS * (2 if is_agent_review_model else 1)
+    diff_budget = max(1000, prompt_chars - len(guidelines_section) - len(full_files_section) - len(history_section))
     diff_truncated = len(augmented_diff) > diff_budget
-    is_large_pr = diff_truncated or len(file_list) > 30
     if is_agent_review_model:  # must match the get_agent_response fallback gate
         visibility_section = (  # function tools carry their own schema descriptions; only cross-tool rules belong here
             "EVIDENCE - every finding needs it:\n"
@@ -629,8 +628,8 @@ def generate_pr_review(
         '{"comments": [{"file": "exact/path", "line": N, "side": "RIGHT", "severity": "HIGH", "message": "..."}], "summary": "..."}\n\n'
         "JSON rules: exact paths (no ./), severity: CRITICAL|HIGH|MEDIUM|LOW|SUGGESTION\n"
         f"Files changed: {len(file_list)} ({', '.join(file_list[:30])}{'...' if len(file_list) > 30 else ''}), Lines: {lines_changed}\n"
-        f"{'Large or truncated PR: the diff below is incomplete. ' if is_large_pr else ''}"
-        f"{'Use list_changed_files and read_diff to inspect changed files not shown in the initial prompt. ' if is_large_pr and is_agent_review_model else ''}\n"
+        f"{'Large PR: the diff below is truncated. ' if diff_truncated else ''}"
+        f"{'Use list_changed_files and read_diff to inspect changed files not shown in the initial prompt. ' if diff_truncated and is_agent_review_model else ''}\n"
     )
 
     messages = [

actions/utils/openai_utils.py

@@ -424,7 +424,9 @@ def get_agent_response(
         "tools": tools,
         "parallel_tool_calls": True,  # batched tool calls share one turn, so the history is re-billed fewer times
         "prompt_cache_key": f"agent-run:{uuid4().hex}",
-        "context_management": [{"type": "compaction", "compact_threshold": 200_000}],
+        # Overflow guard only: compaction collapses the run to a few thousand tokens and drops the evidence gathered so
+        # far, so it only fires at the long-context billing boundary (2x input above 272k tokens on gpt-5.6)
+        "context_management": [{"type": "compaction", "compact_threshold": 272_000}],
     }
     if "gpt-5" in model:
         base_data["reasoning"] = {"effort": reasoning_effort or "medium"}
@@ -452,6 +454,8 @@ def get_agent_response(
         total_usage = _add_openai_usage(total_usage, response_json)
         previous_response_id = response_json.get("id")
         output_items = response_json.get("output", [])
+        if any(item.get("type") == "compaction" for item in output_items):
+            print("WARNING ⚠️ context compacted server-side; earlier tool results are summarized, not verbatim")
         turn_calls = _response_tool_calls(output_items)
         turn_cost = (
             _openai_usage_cost(response_json.get("usage") or {}, model)

tests/test_openai_utils.py

@@ -294,7 +294,7 @@ def test_get_agent_response_calls_function_tools(mock_post):
     assert first_payload["store"] is True
     assert first_payload["service_tier"] == "default"
     assert first_payload["reasoning"] == {"effort": "medium"}
-    assert first_payload["context_management"] == [{"type": "compaction", "compact_threshold": 200_000}]
+    assert first_payload["context_management"] == [{"type": "compaction", "compact_threshold": 272_000}]
     assert first_payload["prompt_cache_key"].startswith("agent-run:")
     assert "include" not in first_payload
     assert "previous_response_id" not in first_payload