Skip to content

The agent that grows with you

fix(usage-window): do not /retry a turn that ran while the limit-offer switch resolved

Tekniumd43d16c98880
diff --git a/gateway/slash_commands_model.py b/gateway/slash_commands_model.py
index edda8974297..8f9479431e3 100644
--- a/gateway/slash_commands_model.py
+++ b/gateway/slash_commands_model.py
@@ -662,7 +662,8 @@ class GatewayModelCommandsMixin:
if generation is not None and self._usage_limit_offer_is_stale(session_key, generation):
# A newer turn ran since the refusal: /retry would re-run that one, not the refused one.
return t("platform.discord.picker.choice_expired")
- return await self._accept_usage_limit_switch(event, value)
+ return await self._accept_usage_limit_switch(
+ event, value, offer=None if generation is None else (session_key, generation))
choices = [{"value": str(s["model"]),
"label": t("gateway.usage_limit.switch_choice", model=str(s.get("label") or s["model"]),
@@ -685,8 +686,10 @@ class GatewayModelCommandsMixin:
except (AttributeError, KeyError, TypeError, ValueError):
return False
- async def _accept_usage_limit_switch(self, event: MessageEvent, model_id: str) -> str:
- """The tap: switch this session to *model_id* on Nous, then re-send the refused message."""
+ async def _accept_usage_limit_switch(self, event: MessageEvent, model_id: str, *,
+ offer: Optional[tuple[str, int]] = None) -> str:
+ """The tap: switch this session to *model_id* on Nous, then re-send the refused message.
+ ``offer`` is the refused turn's ``(session_key, run_generation)``."""
from gateway.run import _hermes_home
source = await asyncio.to_thread(self._normalize_source_for_session_key, event.source)
session_key = self._session_key_for_source(source)
@@ -717,13 +720,15 @@ class GatewayModelCommandsMixin:
# Only after the switch landed (an error reply leaves the override unchanged). Exactly a typed /retry from the same sender (rewinds the refused turn, runs the message
# again, and the adapter delivers the reply). Called directly, the re-run's reply would be
# returned to nobody.
- self._track_background_task(self._dispatch_usage_limit_retry(event))
+ self._track_background_task(self._dispatch_usage_limit_retry(event, offer))
return reply
- async def _dispatch_usage_limit_retry(self, event: MessageEvent) -> None:
+ async def _dispatch_usage_limit_retry(self, event: MessageEvent, offer: Optional[tuple[str, int]] = None) -> None:
from gateway.platforms.event import MessageType
adapter = self._delivery_adapter_for(event.source)
- if adapter is None:
+ # Re-checked here, with no await before the /retry: the switch awaited model resolution, and a
+ # turn that ran meanwhile is what /retry would rewind. The switch itself stands.
+ if adapter is None or (offer is not None and self._usage_limit_offer_is_stale(*offer)):
return
try:
await adapter.handle_message(MessageEvent(
diff --git a/tests/gateway/test_usage_limit_offer.py b/tests/gateway/test_usage_limit_offer.py
index 74a9faffded..4b8b52898ff 100644
--- a/tests/gateway/test_usage_limit_offer.py
+++ b/tests/gateway/test_usage_limit_offer.py
@@ -223,6 +223,44 @@ async def test_a_tap_after_a_newer_turn_changes_nothing():
assert await on_choice("c1", "example/model-free:free") == "switched"
+async def test_a_turn_that_ran_during_the_switch_is_not_retried(monkeypatch):
+ """The tap was current, but a newer turn ran while the switch awaited model resolution: the
+ /retry would rewind that turn, so it is not sent."""
+ from gateway import slash_commands_model as scm
+
+ adapter = _ButtonAdapter()
+ adapter.handle_message = AsyncMock()
+ runner = _runner(adapter)
+ runner.config = SimpleNamespace(multiplex_profiles=False)
+ runner._session_model_overrides = {}
+ runner._normalize_source_for_session_key = lambda source: source
+ state = runner._session_state("chat:c1")
+ state.persistent.run_generation = 7
+
+ async def _perform(ctx, target, provider, source):
+ state.persistent.run_generation = 8 # another message's turn ran during this await
+ return SimpleNamespace(new_model=target), None
+
+ async def _commit(result, ctx, *, source, picker):
+ runner._session_model_overrides[ctx.session_key] = {"model": result.new_model, "provider": "nous"}
+ return "✓ Model switched to Model Free"
+
+ tasks = []
+ runner._perform_model_switch = _perform
+ runner._commit_model_switch = _commit
+ runner._track_background_task = lambda coro: tasks.append(coro)
+ monkeypatch.setattr(scm._ModelSwitchContext, "read_config", lambda self: None)
+ monkeypatch.setattr(scm._ModelSwitchContext, "apply_override", lambda self, override: None)
+
+ await runner._offer_usage_limit_switch(_event(), LIMIT, 7)
+ on_choice = runner._try_send_choice_picker.await_args.kwargs["on_choice_selected"]
+ assert (await on_choice("c1", "example/model-free:free")).startswith("✓")
+ for task in tasks:
+ await task
+ adapter.handle_message.assert_not_awaited()
+
+
def test_a_messaging_refusal_points_at_the_portal_not_topup():
from agent.plan_limits import portal_wording
assert portal_wording("You've hit your daily limit · resets 15:56 · turn on Keep going in /topup") == (