test(cache): pin the prompt-cache scope isolation invariant for per-response session ids
Squash of the three commits on PR #96768 (net diff is tests-only: the mid-series production hunk in agent/transports/codex.py was reverted within the PR after review). Pins the cache-scope isolation invariant for hosts that mint one physical session per response, plus the system-prompt write-path lifecycle under a per-response session (#96570).
This commit is contained in:
joaomarcos
@@ -384,3 +384,188 @@ class TestAuxiliaryRuntimeThreading:
|
||||
assert aux._runtime_main_value("cache_scope") == ""
|
||||
finally:
|
||||
aux.reset_runtime_main(token)
|
||||
|
||||
|
||||
class TestPerResponseRunNonceIsolation:
|
||||
"""Issue #96570 — hosts that mint one physical session per RESPONSE.
|
||||
|
||||
Hermes Studio group chat builds ``gc_run_<room>_<profile>_<name>_<uuid4hex>``
|
||||
for every reply and destroys it when the reply completes
|
||||
(``groupRuntimeSessionId``), so every conversation-affinity hint Hermes
|
||||
derives from that id is re-keyed on every reply. What is demonstrated here
|
||||
is the routing/affinity mechanism moving per response; no provider cache
|
||||
telemetry or billing outcome is measured or claimed.
|
||||
|
||||
The normalizer cannot repair that from the id alone: a physical session id
|
||||
is an identity, and Hermes' public session API lets a client choose one
|
||||
freely (``POST /v1/sessions`` honors ``body["id"]``/``body["session_id"]``).
|
||||
These tests pin the isolation invariant that any future scope rule has to
|
||||
keep — collapsing a trailing token because it *looks* like per-run noise
|
||||
merges independent conversations, which is the failure class already
|
||||
recorded on #79017.
|
||||
"""
|
||||
|
||||
RUN = "gc_run_room42_default_Worker"
|
||||
RESPONSE_1 = f"{RUN}_5f2c1ab9d4e34f7a8b0c6d1e2f3a4b5c"
|
||||
RESPONSE_2 = f"{RUN}_9a7e3b1c05d24e6fb83a1c7d9e0f2a4b"
|
||||
|
||||
SYS = [
|
||||
{"role": "system", "content": "sys"},
|
||||
{"role": "user", "content": "hi"},
|
||||
]
|
||||
|
||||
@staticmethod
|
||||
def _prompt_cache_key(session_id: str) -> str:
|
||||
from agent.transports.codex import ResponsesApiTransport
|
||||
|
||||
return ResponsesApiTransport().build_kwargs(
|
||||
model="gpt-5.5",
|
||||
messages=[{"role": "system", "content": "same"}],
|
||||
tools=[],
|
||||
session_id=session_id,
|
||||
)["prompt_cache_key"]
|
||||
|
||||
def test_external_uuid_identity_remains_isolated(self):
|
||||
"""Two independent API-supplied ids may differ only in a trailing hex.
|
||||
|
||||
``POST /v1/sessions`` preserves a client-provided id verbatim, so a
|
||||
whole trailing 32-hex token is not per-run noise by construction.
|
||||
"""
|
||||
first = "customer_chat_11111111111141118111111111111111"
|
||||
second = "customer_chat_22222222222242228222222222222222"
|
||||
|
||||
assert _cache_scope_from_session_id(first) != _cache_scope_from_session_id(
|
||||
second
|
||||
)
|
||||
assert self._prompt_cache_key(first) != self._prompt_cache_key(second)
|
||||
|
||||
def test_studio_truncation_does_not_merge_members(self):
|
||||
"""Studio truncates the semantic prefix BEFORE appending the nonce.
|
||||
|
||||
``groupRuntimeSessionId`` slices its ``gc_run_<room>_<profile>_<name>``
|
||||
prefix to 96 characters and only then appends the per-response token,
|
||||
so two members of one room whose names diverge past that boundary are
|
||||
distinguished *only* by that token. Any rule that drops it merges two
|
||||
distinct agents onto one affinity key.
|
||||
"""
|
||||
room = "mabc1234qwerty"
|
||||
profile = "default"
|
||||
common_name = "A" * 70
|
||||
|
||||
first = (
|
||||
f"gc_run_{room}_{profile}_{common_name}Worker"[:96]
|
||||
+ "_11111111111141118111111111111111"
|
||||
)
|
||||
second = (
|
||||
f"gc_run_{room}_{profile}_{common_name}Reviewer"[:96]
|
||||
+ "_22222222222242228222222222222222"
|
||||
)
|
||||
|
||||
assert first != second
|
||||
assert _cache_scope_from_session_id(first) != _cache_scope_from_session_id(
|
||||
second
|
||||
)
|
||||
|
||||
def test_cron_normalization_stays_the_only_carve_out(self):
|
||||
"""The one accepted exception: cron's per-fire timestamp (#51395)."""
|
||||
assert (
|
||||
_cache_scope_from_session_id("cron_backup_20260814_120000")
|
||||
== "cron_backup"
|
||||
)
|
||||
assert (
|
||||
_cache_scope_from_session_id(self.RESPONSE_1) == self.RESPONSE_1
|
||||
)
|
||||
|
||||
def test_parentless_rows_resolve_to_their_own_scope(self, db):
|
||||
"""A row with no lineage is its own scope — permanently.
|
||||
|
||||
The Studio bridge creates the row with ``create_session(id, source,
|
||||
model)`` — no ``parent_session_id`` — so ``resolve_prompt_cache_scope``
|
||||
returns the physical id. This is the invariant, not a defect record:
|
||||
an owner Hermes was never told about must never be guessed. #96811
|
||||
closes the gap by having the host declare the logical conversation (a
|
||||
stable session id, or an explicit key); rows that still declare
|
||||
nothing keep resolving exactly like this.
|
||||
"""
|
||||
db.create_session(self.RESPONSE_1, source="studio")
|
||||
db.create_session(self.RESPONSE_2, source="studio")
|
||||
|
||||
assert resolve_prompt_cache_scope(_agent(self.RESPONSE_1, db)) == self.RESPONSE_1
|
||||
assert resolve_prompt_cache_scope(_agent(self.RESPONSE_2, db)) == self.RESPONSE_2
|
||||
|
||||
def test_distinct_ids_keep_distinct_affinity_keys(self):
|
||||
"""Every affinity surface isolates two distinct physical ids.
|
||||
|
||||
This is the isolation invariant restated at the wire layer, and it is
|
||||
also the reported symptom: Studio hands these two ids to the same
|
||||
logical conversation, so the routing/affinity key moves per response.
|
||||
|
||||
It stays true after #96811. The logical identity is supplied one layer
|
||||
up — ``cache_scope_id`` here, the ambient conversation context for the
|
||||
provider profiles — and these call sites pass neither, exactly as an
|
||||
undeclared conversation would.
|
||||
"""
|
||||
from agent.transports.chat_completions import _add_prompt_cache_key
|
||||
from agent.transports.codex import ResponsesApiTransport
|
||||
|
||||
transport = ResponsesApiTransport()
|
||||
base = dict(model="gpt-5.5", messages=self.SYS, tools=[])
|
||||
assert (
|
||||
transport.build_kwargs(**base, session_id=self.RESPONSE_1)[
|
||||
"prompt_cache_key"
|
||||
]
|
||||
!= transport.build_kwargs(**base, session_id=self.RESPONSE_2)[
|
||||
"prompt_cache_key"
|
||||
]
|
||||
)
|
||||
|
||||
xai = dict(model="grok-4", messages=self.SYS, tools=[], is_xai_responses=True)
|
||||
assert (
|
||||
transport.build_kwargs(**xai, session_id=self.RESPONSE_1)["extra_headers"][
|
||||
"x-grok-conv-id"
|
||||
]
|
||||
!= transport.build_kwargs(**xai, session_id=self.RESPONSE_2)[
|
||||
"extra_headers"
|
||||
]["x-grok-conv-id"]
|
||||
)
|
||||
|
||||
def chat_key(session_id):
|
||||
kwargs: dict = {}
|
||||
_add_prompt_cache_key(
|
||||
kwargs,
|
||||
messages=[{"role": "system", "content": "sys"}],
|
||||
tools=None,
|
||||
supports_prompt_cache_key=True,
|
||||
session_id=session_id,
|
||||
)
|
||||
return kwargs.get("prompt_cache_key")
|
||||
|
||||
assert chat_key(self.RESPONSE_1) != chat_key(self.RESPONSE_2)
|
||||
|
||||
@pytest.mark.parametrize("profile_name", ["openrouter", "nous"])
|
||||
def test_distinct_ids_keep_distinct_provider_sticky_keys(self, profile_name):
|
||||
"""OpenRouter/Nous route by this key, and it isolates distinct ids.
|
||||
|
||||
Same invariant as above on the sticky-routing surface: two ids of one
|
||||
Studio conversation re-key it on every reply, and a declared logical
|
||||
identity would arrive through the conversation contextvar, not from
|
||||
re-reading this id.
|
||||
"""
|
||||
from agent.portal_tags import (
|
||||
reset_conversation_context,
|
||||
set_conversation_context,
|
||||
)
|
||||
from providers import get_provider_profile
|
||||
|
||||
profile = get_provider_profile(profile_name)
|
||||
keys = []
|
||||
for session_id in (self.RESPONSE_1, self.RESPONSE_2):
|
||||
token = set_conversation_context(session_id)
|
||||
try:
|
||||
keys.append(
|
||||
profile.build_extra_body(session_id=session_id)["session_id"]
|
||||
)
|
||||
finally:
|
||||
reset_conversation_context(token)
|
||||
|
||||
assert keys[0] != keys[1]
|
||||
|
||||
@@ -379,5 +379,73 @@ class TestReconstructStaticPrefixMemoization:
|
||||
assert getattr(agent, "_static_rebuild_failed_for", None) is None
|
||||
|
||||
|
||||
class TestPerResponseSessionWritePath:
|
||||
"""The write path under an embedding host's per-response session (#96570).
|
||||
|
||||
Hermes Studio group chat pre-creates the SQLite row and pre-persists the
|
||||
user message BEFORE ``run_conversation()``, then runs one turn under a
|
||||
session id it destroys afterwards. The row therefore starts with a null
|
||||
system prompt and a non-empty history on its own genuine FIRST turn, which
|
||||
is what trips the "stored system prompt is null" warning — the warning is
|
||||
a first-turn artifact of that lifecycle, not evidence of a lost write.
|
||||
|
||||
This pins the write path against that exact lifecycle: the freshly built
|
||||
prompt must land in the pre-created row within the same run.
|
||||
"""
|
||||
|
||||
def _agent(self, db, session_id):
|
||||
agent = _make_agent(session_db=db, prebuilt_prompt="GROUP_PROMPT")
|
||||
agent.session_id = session_id
|
||||
return agent
|
||||
|
||||
def test_prepersisted_row_stores_the_freshly_built_prompt(self, tmp_path):
|
||||
from hermes_state import SessionDB
|
||||
|
||||
session_id = "gc_run_room42_default_Worker_5f2c1ab9d4e34f7a8b0c6d1e2f3a4b5c"
|
||||
with SessionDB(db_path=tmp_path / "state.db") as db:
|
||||
# What the bridge does before the turn starts.
|
||||
db.create_session(session_id, source="studio")
|
||||
db.append_message(session_id=session_id, role="user", content="hi")
|
||||
|
||||
_restore_or_build_system_prompt(
|
||||
self._agent(db, session_id),
|
||||
None,
|
||||
[{"role": "user", "content": "hi"}],
|
||||
)
|
||||
|
||||
assert db.get_session(session_id)["system_prompt"] == "GROUP_PROMPT"
|
||||
|
||||
def test_warning_is_a_first_turn_artifact_not_a_lost_write(
|
||||
self, tmp_path, caplog
|
||||
):
|
||||
"""Second turn of the SAME id restores — so nothing was dropped."""
|
||||
from hermes_state import SessionDB
|
||||
|
||||
session_id = "gc_run_room42_default_Worker_9a7e3b1c05d24e6fb83a1c7d9e0f2a4b"
|
||||
history = [{"role": "user", "content": "hi"}]
|
||||
with SessionDB(db_path=tmp_path / "state.db") as db:
|
||||
db.create_session(session_id, source="studio")
|
||||
db.append_message(session_id=session_id, role="user", content="hi")
|
||||
|
||||
with caplog.at_level(
|
||||
logging.WARNING, logger="agent.conversation_loop"
|
||||
):
|
||||
_restore_or_build_system_prompt(
|
||||
self._agent(db, session_id), None, history
|
||||
)
|
||||
assert "is null" in caplog.text
|
||||
|
||||
caplog.clear()
|
||||
second = self._agent(db, session_id)
|
||||
with caplog.at_level(
|
||||
logging.WARNING, logger="agent.conversation_loop"
|
||||
):
|
||||
_restore_or_build_system_prompt(second, None, history)
|
||||
|
||||
assert second._cached_system_prompt == "GROUP_PROMPT"
|
||||
second._build_system_prompt.assert_not_called()
|
||||
assert "is null" not in caplog.text
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
pytest.main([__file__, "-v"])
|
||||
|
||||
Reference in New Issue
Block a user