Skip to content

Drop in a screenshot and convert it to clean code (HTML/Tailwind/React/Vue)

Python

add openai prompt cache keys

Abi Rajab9bdca7161a6
diff --git a/backend/agent/providers/openai.py b/backend/agent/providers/openai.py
index 6e511f1b..818033a9 100644
--- a/backend/agent/providers/openai.py
+++ b/backend/agent/providers/openai.py
@@ -1,5 +1,6 @@
# pyright: reportUnknownVariableType=false
import copy
+import hashlib
import json
import uuid
from dataclasses import dataclass, field
@@ -124,6 +125,28 @@ def serialize_openai_tools(
return serialized
+def _build_prompt_cache_key(
+ model: Llm,
+ input_items: List[Dict[str, Any]],
+ tools: List[Dict[str, Any]],
+) -> str:
+ cache_identity = {
+ "version": 1,
+ "model": get_openai_api_name(model),
+ "reasoning_effort": get_openai_reasoning_effort(model),
+ "input": input_items,
+ "tools": tools,
+ }
+ payload = json.dumps(
+ cache_identity,
+ sort_keys=True,
+ separators=(",", ":"),
+ ensure_ascii=False,
+ )
+ digest = hashlib.sha256(payload.encode("utf-8")).hexdigest()[:24]
+ return f"s2c-openai-session-v1-{digest}"
+
+
@dataclass
class OpenAIResponsesParseState:
assistant_text: str = ""
@@ -425,10 +448,13 @@ class OpenAIProviderSession(ProviderSession):
self._input_items: List[Dict[str, Any]] = [
_convert_message_to_responses_input(message) for message in prompt_messages
]
+ self._prompt_cache_key = _build_prompt_cache_key(
+ model=model,
+ input_items=self._input_items,
+ tools=self._tools,
+ )
async def stream_turn(self, on_event: EventSink) -> ProviderTurn:
- self._turn_input_logger.record_turn_input(self._input_items)
-
params: Dict[str, Any] = {
"model": get_openai_api_name(self._model),
"input": self._input_items,
@@ -436,11 +462,17 @@ class OpenAIProviderSession(ProviderSession):
"tool_choice": "auto",
"stream": True,
"max_output_tokens": 50000,
+ "prompt_cache_key": self._prompt_cache_key,
}
reasoning_effort = get_openai_reasoning_effort(self._model)
if reasoning_effort:
params["reasoning"] = {"effort": reasoning_effort, "summary": "auto"}
+ self._turn_input_logger.record_turn_input(
+ self._input_items,
+ request_payload=params,
+ )
+
state = OpenAIResponsesParseState()
stream = await self._client.responses.create(**params) # type: ignore
async for event in stream: # type: ignore
diff --git a/backend/fs_logging/openai_turn_inputs.py b/backend/fs_logging/openai_turn_inputs.py
index b09ade16..8902f30c 100644
--- a/backend/fs_logging/openai_turn_inputs.py
+++ b/backend/fs_logging/openai_turn_inputs.py
@@ -334,6 +334,7 @@ class OpenAITurnUsageSummary:
class OpenAITurnInputReport:
turn_index: int
items: list[OpenAITurnInputItem]
+ request_payload: Any | None = None
usage: OpenAITurnUsageSummary | None = None
@@ -345,7 +346,11 @@ class OpenAITurnInputLogger:
_turn_index: int = 0
_turns: list[OpenAITurnInputReport] = field(default_factory=list)
- def record_turn_input(self, input_items: Sequence[Any]) -> None:
+ def record_turn_input(
+ self,
+ input_items: Sequence[Any],
+ request_payload: Any | None = None,
+ ) -> None:
if not self.enabled:
return
@@ -365,6 +370,7 @@ class OpenAITurnInputLogger:
OpenAITurnInputReport(
turn_index=self._turn_index,
items=turn_items,
+ request_payload=_to_serializable(request_payload),
)
)
@@ -457,6 +463,24 @@ class OpenAITurnInputLogger:
html_parts.append(
f" <h2>Turn {turn.turn_index} (items={len(turn.items)})</h2>"
)
+ if turn.request_payload is not None:
+ request_payload_json = json.dumps(
+ turn.request_payload,
+ indent=2,
+ ensure_ascii=False,
+ )
+ html_parts.append(" <details class='payload-wrap' open>")
+ html_parts.append(" <summary>Request payload</summary>")
+ html_parts.append(" <div class='json-view'>")
+ html_parts.append(_render_json_node(turn.request_payload, "root"))
+ html_parts.append(" </div>")
+ html_parts.append(" <details>")
+ html_parts.append(" <summary>Raw JSON payload</summary>")
+ html_parts.append(
+ f" <pre>{escape(request_payload_json)}</pre>"
+ )
+ html_parts.append(" </details>")
+ html_parts.append(" </details>")
if turn.usage is not None:
cost_text = "n/a"
if isinstance(turn.usage.cost_usd, (float, int)):
diff --git a/backend/tests/test_openai_provider_session.py b/backend/tests/test_openai_provider_session.py
new file mode 100644
index 00000000..5ed5f08f
--- /dev/null
+++ b/backend/tests/test_openai_provider_session.py
@@ -0,0 +1,153 @@
+import copy
+from typing import Any
+
+import pytest
+
+from agent.providers.base import ExecutedToolCall, ProviderTurn
+from agent.providers.openai import OpenAIProviderSession
+from agent.tools import ToolCall, ToolExecutionResult
+from llm import Llm
+
+
+class _EmptyAsyncStream:
+ def __aiter__(self) -> "_EmptyAsyncStream":
+ return self
+
+ async def __anext__(self) -> object:
+ raise StopAsyncIteration
+
+
+class _FakeResponses:
+ def __init__(self) -> None:
+ self.calls: list[dict[str, Any]] = []
+
+ async def create(self, **kwargs: Any) -> _EmptyAsyncStream:
+ self.calls.append(copy.deepcopy(kwargs))
+ return _EmptyAsyncStream()
+
+
+class _FakeOpenAIClient:
+ def __init__(self) -> None:
+ self.responses = _FakeResponses()
+
+ async def close(self) -> None:
+ return None
+
+
+async def _noop_event_sink(_: Any) -> None:
+ return None
+
+
+def _test_tools() -> list[dict[str, Any]]:
+ return [
+ {
+ "type": "function",
+ "name": "edit_file",
+ "description": "Apply an edit.",
+ "parameters": {
+ "type": "object",
+ "properties": {
+ "path": {"type": "string"},
+ },
+ "required": ["path"],
+ },
+ "strict": True,
+ }
+ ]
+
+
+async def test_openai_provider_session_reuses_prompt_cache_key_across_turns() -> None:
+ client = _FakeOpenAIClient()
+ session = OpenAIProviderSession(
+ client=client, # type: ignore[arg-type]
+ model=Llm.GPT_5_2_CODEX_HIGH,
+ prompt_messages=[{"role": "user", "content": "Build a landing page."}],
+ tools=_test_tools(),
+ )
+
+ first_turn = await session.stream_turn(_noop_event_sink)
+ session.append_tool_results(
+ ProviderTurn(
+ assistant_text=first_turn.assistant_text,
+ tool_calls=[],
+ assistant_turn=[
+ {
+ "type": "function_call",
+ "call_id": "call-1",
+ "name": "edit_file",
+ "arguments": '{"path":"index.html"}',
+ }
+ ],
+ ),
+ [
+ ExecutedToolCall(
+ tool_call=ToolCall(
+ id="call-1",
+ name="edit_file",
+ arguments={"path": "index.html"},
+ ),
+ result=ToolExecutionResult(
+ ok=True,
+ result={
+ "content": "Successfully edited file at index.html.",
+ "details": {
+ "diff": "--- index.html\n+++ index.html\n@@ -1 +1 @@\n-a\n+b\n",
+ "firstChangedLine": 1,
+ },
+ },
+ summary={"content": "Successfully edited file at index.html."},
+ ),
+ )
+ ],
+ )
+ await session.stream_turn(_noop_event_sink)
+
+ first_call = client.responses.calls[0]
+ second_call = client.responses.calls[1]
+ first_input = first_call["input"]
+ second_input = second_call["input"]
+
+ assert first_call["prompt_cache_key"] == second_call["prompt_cache_key"]
+ assert isinstance(first_call["prompt_cache_key"], str)
+ assert str(first_call["prompt_cache_key"]).startswith("s2c-openai-session-v1-")
+ assert isinstance(first_input, list)
+ assert isinstance(second_input, list)
+ assert len(second_input) > len(first_input)
+
+
+async def test_openai_provider_session_prompt_cache_key_is_deterministic() -> None:
+ first_client = _FakeOpenAIClient()
+ second_client = _FakeOpenAIClient()
+ different_prompt_client = _FakeOpenAIClient()
+
+ first_session = OpenAIProviderSession(
+ client=first_client, # type: ignore[arg-type]
+ model=Llm.GPT_5_2_CODEX_HIGH,
+ prompt_messages=[{"role": "user", "content": "Build a landing page."}],
+ tools=_test_tools(),
+ )
+ second_session = OpenAIProviderSession(
+ client=second_client, # type: ignore[arg-type]
+ model=Llm.GPT_5_2_CODEX_HIGH,
+ prompt_messages=[{"role": "user", "content": "Build a landing page."}],
+ tools=_test_tools(),
+ )
+ different_prompt_session = OpenAIProviderSession(
+ client=different_prompt_client, # type: ignore[arg-type]
+ model=Llm.GPT_5_2_CODEX_HIGH,
+ prompt_messages=[{"role": "user", "content": "Build a dashboard."}],
+ tools=_test_tools(),
+ )
+
+ await first_session.stream_turn(_noop_event_sink)
+ await second_session.stream_turn(_noop_event_sink)
+ await different_prompt_session.stream_turn(_noop_event_sink)
+
+ first_key = first_client.responses.calls[0]["prompt_cache_key"]
+ second_key = second_client.responses.calls[0]["prompt_cache_key"]
+ different_prompt_key = different_prompt_client.responses.calls[0]["prompt_cache_key"]
+
+ assert first_key == second_key
+ assert first_key != different_prompt_key
diff --git a/backend/tests/test_openai_turn_input_logging.py b/backend/tests/test_openai_turn_input_logging.py
index e0e50a41..96840163 100644
--- a/backend/tests/test_openai_turn_input_logging.py
+++ b/backend/tests/test_openai_turn_input_logging.py
@@ -79,6 +79,35 @@ def test_openai_turn_input_logger_preserves_full_large_payloads(
assert "truncated 50 chars" not in html
+def test_openai_turn_input_logger_includes_request_payload(
+ tmp_path, monkeypatch
+) -> None:
+ monkeypatch.setenv("LOGS_PATH", str(tmp_path))
+
+ logger = OpenAITurnInputLogger(model=Llm.GPT_5_2_CODEX_HIGH, enabled=True)
+ logger.record_turn_input(
+ [
+ {
+ "role": "user",
+ "content": "Build this page",
+ }
+ ],
+ request_payload={
+ "model": "gpt-5.2-codex",
+ "prompt_cache_key": "s2c-openai-session-v1-abc123",
+ "input": [{"role": "user", "content": "Build this page"}],
+ },
+ )
+
+ report_path = logger.write_html_report()
+
+ assert report_path is not None
+ html = Path(report_path).read_text(encoding="utf-8")
+ assert "Request payload" in html
+ assert "prompt_cache_key" in html
+ assert "s2c-openai-session-v1-abc123" in html
+
+
def test_openai_turn_input_logger_disabled_writes_nothing(tmp_path, monkeypatch) -> None:
monkeypatch.setenv("LOGS_PATH", str(tmp_path))