Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.
Python agentaianthropicclaude-codecompressioncontext-engineeringcontext-windowcursorfastapilangchainllmmcpopenaiprompt-engineeringproxypythonragtoken-optimizationtokenstypescript
fix: correct preserved-entry index mapping in Gemini content round-trip (#836)
Ashish0ffe2b6ea49e
diff --git a/headroom/proxy/handlers/gemini.py b/headroom/proxy/handlers/gemini.pyindex 26a545f8d..fc5efcfde 100644--- a/headroom/proxy/handlers/gemini.py+++ b/headroom/proxy/handlers/gemini.py@@ -73,6 +73,40 @@ class GeminiHandlerMixin:return Truereturn False+ def _rebuild_gemini_contents(+ self,+ original_contents: list[dict],+ preserved_indices: set[int],+ preserved_contents: dict[int, dict],+ optimized_contents: list[dict],+ ) -> list[dict]:+ """Interleave preserved (non-text) entries back into optimized_contents at their+ original positions.++ preserved_indices uses original contents[] indices, but optimized_contents uses+ a different (shorter) index space because entries with no text parts were excluded+ from the messages[] sent for compression. Using orig_idx directly to overwrite+ optimized_contents[orig_idx] corrupts or silently drops entries.++ This method walks original_contents in order, placing each position with either+ the preserved original (for non-text entries) or the next optimized text entry.+ """+ opt_iter = iter(optimized_contents)+ result: list[dict] = []+ for idx, content in enumerate(original_contents):+ had_text = any("text" in p for p in content.get("parts", []))+ if idx in preserved_indices:+ result.append(preserved_contents[idx])+ if had_text:+ # Entry also produced a message; consume but discard the optimized version+ next(opt_iter, None)+ else:+ opt_entry = next(opt_iter, None)+ if opt_entry is not None:+ result.append(opt_entry)+ # else: dropped by compression — omit+ return result+def _gemini_contents_to_messages(self, contents: list[dict], system_instruction: dict | None = None) -> tuple[list[dict], set[int]]:@@ -501,12 +535,9 @@ class GeminiHandlerMixin:optimized_contents, optimized_system = self._messages_to_gemini_contents(optimized_messages)-- # Restore preserved content entries that had non-text parts- for orig_idx, original_content in preserved_contents.items():- if orig_idx < len(optimized_contents):- optimized_contents[orig_idx] = original_content-+ optimized_contents = self._rebuild_gemini_contents(+ contents, preserved_indices, preserved_contents, optimized_contents+ )body["contents"] = optimized_contentsif optimized_system:body["systemInstruction"] = optimized_system@@ -784,9 +815,12 @@ class GeminiHandlerMixin:optimized_contents, optimized_system = self._messages_to_gemini_contents(optimized_messages)- for orig_idx, original_content in preserved_contents.items():- if orig_idx < len(optimized_contents):- optimized_contents[orig_idx] = original_content+ optimized_contents = self._rebuild_gemini_contents(+ contents if isinstance(contents, list) else [],+ preserved_indices,+ preserved_contents,+ optimized_contents,+ )request_payload["contents"] = optimized_contentsif not is_antigravity:if optimized_system:@@ -1028,12 +1062,9 @@ class GeminiHandlerMixin:optimized_contents, optimized_system = self._messages_to_gemini_contents(optimized_messages)-- # Restore preserved content entries that had non-text parts- for orig_idx, original_content in preserved_contents.items():- if orig_idx < len(optimized_contents):- optimized_contents[orig_idx] = original_content-+ optimized_contents = self._rebuild_gemini_contents(+ contents, preserved_indices, preserved_contents, optimized_contents+ )body["contents"] = optimized_contentsif optimized_system:body["systemInstruction"] = optimized_systemdiff --git a/tests/test_google_multimodal.py b/tests/test_google_multimodal.pyindex c3f265885..03e583a8e 100644--- a/tests/test_google_multimodal.py+++ b/tests/test_google_multimodal.py@@ -627,3 +627,80 @@ class TestParametrizedPreservation:"""Parametrized test for preserved indices."""_, preserved_indices = proxy._gemini_contents_to_messages(contents)assert preserved_indices == expected_indices+++# =============================================================================+# Tests for _rebuild_gemini_contents+# =============================================================================+++class TestRebuildGeminiContents:+ """_rebuild_gemini_contents must re-insert preserved entries at their original positions."""++ def _round_trip(self, proxy, contents):+ """Simulate the full compression round-trip for a given contents list.++ Mimics what the handler does: convert → strip system msg → convert back → rebuild.+ """+ messages, preserved_indices = proxy._gemini_contents_to_messages(contents)+ preserved_contents = {idx: contents[idx] for idx in preserved_indices}+ optimized_contents, _ = proxy._messages_to_gemini_contents(messages)+ return proxy._rebuild_gemini_contents(+ contents, preserved_indices, preserved_contents, optimized_contents+ )++ def test_text_only_unchanged(self, proxy):+ """Text-only round-trip should produce identical contents."""+ contents = [TEXT_ONLY_CONTENT, MODEL_TEXT_CONTENT]+ result = self._round_trip(proxy, contents)+ assert len(result) == 2+ assert result[0]["parts"][0]["text"] == "Hello, world!"+ assert result[1]["parts"][0]["text"] == "Hello! How can I help you today?"++ def test_function_call_sequence_preserved(self, proxy):+ """functionCall and functionResponse entries must survive and appear at correct positions."""+ contents = [+ TEXT_ONLY_CONTENT, # idx 0: text+ FUNCTION_CALL_CONTENT, # idx 1: functionCall only — no text → preserved+ FUNCTION_RESPONSE_CONTENT, # idx 2: functionResponse only — no text → preserved+ MODEL_TEXT_CONTENT, # idx 3: text+ ]+ result = self._round_trip(proxy, contents)++ assert len(result) == 4, f"Expected 4 entries, got {len(result)}: {result}"+ # Position 0: original text+ assert result[0]["parts"][0].get("text") == "Hello, world!"+ # Position 1: functionCall preserved exactly+ assert "functionCall" in result[1]["parts"][0], "functionCall missing at position 1"+ assert result[1]["parts"][0]["functionCall"]["name"] == "get_weather"+ # Position 2: functionResponse preserved exactly+ assert "functionResponse" in result[2]["parts"][0], "functionResponse missing at position 2"+ # Position 3: text preserved+ assert result[3]["parts"][0].get("text") == "Hello! How can I help you today?"++ def test_function_call_at_start(self, proxy):+ """Preserved entry at idx=0 must not overwrite idx=0 of optimized_contents."""+ contents = [+ FUNCTION_CALL_CONTENT, # idx 0: no text → preserved+ TEXT_ONLY_CONTENT, # idx 1: text+ ]+ result = self._round_trip(proxy, contents)++ assert len(result) == 2+ assert "functionCall" in result[0]["parts"][0]+ assert result[1]["parts"][0].get("text") == "Hello, world!"++ def test_hybrid_entry_uses_original(self, proxy):+ """Entry with both text and functionCall keeps the original (with functionCall intact)."""+ contents = [+ TEXT_ONLY_CONTENT,+ FUNCTION_CALL_WITH_TEXT_CONTENT, # idx 1: has both text and functionCall → preserved+ MODEL_TEXT_CONTENT,+ ]+ result = self._round_trip(proxy, contents)++ assert len(result) == 3+ # Hybrid entry must come back as the original (functionCall retained)+ hybrid = result[1]+ part_keys = {k for p in hybrid["parts"] for k in p}+ assert "functionCall" in part_keys, "functionCall lost from hybrid entry"