686f6c61
e6b4f3750b
Automatic preflight used the full durable transcript even when the Codex Responses request would prune around a native compaction checkpoint. That false-triggered a 600s local summary against history the main request never sent. Estimate the converted, checkpoint-pruned payload when native compaction is eligible, and keep the generic estimate as the conservative fallback.
101 lines
3.5 KiB
Python
101 lines
3.5 KiB
Python
"""Native Responses preflight must count the checkpoint-pruned wire (#96155)."""
|
|
|
|
from types import SimpleNamespace
|
|
|
|
from agent.codex_responses_adapter import estimate_native_responses_preflight_tokens
|
|
from agent.model_metadata import estimate_request_tokens_rough
|
|
from agent.turn_context import _preflight_request_tokens
|
|
|
|
|
|
def _codex_agent(**over):
|
|
agent = SimpleNamespace(
|
|
api_mode="codex_responses",
|
|
provider="openai-codex",
|
|
model="gpt-5.6",
|
|
base_url="https://chatgpt.com/backend-api/codex",
|
|
_base_url_hostname="chatgpt.com",
|
|
_base_url_lower="https://chatgpt.com/backend-api/codex",
|
|
codex_responses_native_compaction=True,
|
|
compression_enabled=True,
|
|
_codex_reasoning_replay_enabled=True,
|
|
context_compressor=SimpleNamespace(threshold_tokens=765_000),
|
|
tools=None,
|
|
)
|
|
for key, value in over.items():
|
|
setattr(agent, key, value)
|
|
return agent
|
|
|
|
|
|
def _history_with_checkpoint():
|
|
# Pre-checkpoint assistant/tool rows are dropped from the wire; user
|
|
# asks are retained. A durable estimate that counts those dropped rows
|
|
# is what falsely tripped local compression in #96155.
|
|
pre = []
|
|
for i in range(30):
|
|
pre.append({"role": "user", "content": f"ask {i}"})
|
|
pre.append({"role": "assistant", "content": "working " + ("tool output " * 400)})
|
|
pre.append(
|
|
{
|
|
"role": "tool",
|
|
"content": "result " + ("payload " * 400),
|
|
"tool_call_id": f"call-{i}",
|
|
}
|
|
)
|
|
checkpoint_turn = {
|
|
"role": "assistant",
|
|
"content": "checkpointed turn",
|
|
"codex_reasoning_items": [
|
|
{
|
|
"type": "compaction",
|
|
"encrypted_content": "blob",
|
|
"_issuer_kind": "codex_backend",
|
|
}
|
|
],
|
|
}
|
|
tail = [{"role": "user", "content": "follow-up after checkpoint"}]
|
|
return pre + [checkpoint_turn] + tail
|
|
|
|
|
|
def test_returns_none_when_api_mode_is_not_codex_responses():
|
|
agent = _codex_agent(api_mode="chat_completions")
|
|
assert estimate_native_responses_preflight_tokens(agent, _history_with_checkpoint()) is None
|
|
|
|
|
|
def test_returns_none_when_native_compaction_is_disabled():
|
|
agent = _codex_agent(codex_responses_native_compaction=False)
|
|
assert estimate_native_responses_preflight_tokens(agent, _history_with_checkpoint()) is None
|
|
|
|
|
|
def test_returns_none_when_model_is_outside_gpt56_family():
|
|
agent = _codex_agent(model="gpt-5.2")
|
|
assert estimate_native_responses_preflight_tokens(agent, _history_with_checkpoint()) is None
|
|
|
|
|
|
def test_pruned_estimate_is_far_below_durable_transcript():
|
|
agent = _codex_agent()
|
|
messages = _history_with_checkpoint()
|
|
generic = estimate_request_tokens_rough(messages)
|
|
native = estimate_native_responses_preflight_tokens(agent, messages)
|
|
|
|
assert native is not None
|
|
assert generic > native * 2
|
|
assert native < 8_000
|
|
|
|
|
|
def test_preflight_wrapper_uses_pruned_estimate_when_eligible():
|
|
agent = _codex_agent()
|
|
messages = _history_with_checkpoint()
|
|
native = estimate_native_responses_preflight_tokens(agent, messages)
|
|
wrapped = _preflight_request_tokens(agent, messages, "")
|
|
|
|
assert native is not None
|
|
assert wrapped == native
|
|
|
|
|
|
def test_preflight_wrapper_falls_back_to_generic_when_ineligible():
|
|
agent = _codex_agent(api_mode="chat_completions")
|
|
messages = _history_with_checkpoint()
|
|
generic = estimate_request_tokens_rough(messages)
|
|
|
|
assert _preflight_request_tokens(agent, messages, "") == generic
|