Skip to content

armgit / avibe

Public Imported from GitHub
0

Make your computer a home for your AI partner. Avibe is the local-first Agent OS: Vibey, its built-in AI partner, works out of the box, and Claude Code, Codex & OpenCode run unchanged beside it — reach them from your browser, phone, or any chat app.

fix(model-hub): let a Vibey call after a failover take the new Source (#2860)

Call Whiteb159b019c0bb
diff --git a/core/handlers/model_hub/turn_gateway.py b/core/handlers/model_hub/turn_gateway.py
index c690ba75c..42dfdd88d 100644
--- a/core/handlers/model_hub/turn_gateway.py
+++ b/core/handlers/model_hub/turn_gateway.py
@@ -570,12 +570,26 @@ class ModelHubTurnGateway:
raise ModelHubError("mapping_target_unavailable", status=409)
if backend == "vibey" and turn_id:
# The caller may read EOF before the previous handler's owned
- # teardown exits. Drain already-settled responses, not live model
- # work; the registry still refuses any remaining overlap.
+ # teardown exits, and Vibey stops reading at the stream's terminal
+ # event, which this gateway forwards before the upstream's EOF and
+ # before it settles the call. Drain responses whose protocol terminal
+ # was forwarded or whose upstream stream has ended, not live model
+ # work: a next call that resolves to another Source (the last call
+ # failed over) may only replace the turn's route once the previous
+ # one is released. The wait is bounded; the registry still refuses
+ # any remaining overlap.
finishing = [
owned.done for owned in self._turn_requests.get(turn_id, ())
if owned.execution is not None
- and (owned.execution.settlement_recorded or owned.execution.terminal_fact_committed)
+ and (
+ owned.execution.completed_at is not None
+ or (
+ owned.execution.wire_state is not None
+ and owned.execution.wire_state.terminal_outcome is not None
+ )
+ or owned.execution.settlement_recorded
+ or owned.execution.terminal_fact_committed
+ )
]
if finishing:
await asyncio.wait(finishing, timeout=self._transport_timeout)
diff --git a/modules/agents/vibey/errors.py b/modules/agents/vibey/errors.py
index 1484a5764..744b95bb4 100644
--- a/modules/agents/vibey/errors.py
+++ b/modules/agents/vibey/errors.py
@@ -47,6 +47,9 @@ _KIND_KEYS: dict[str, str] = {
"stream_cleanup": "internal",
# A context row this build cannot read (a shape it no longer accepts): the session cannot continue.
"TranscriptError": "historyUnreadable",
+ # Model Hub could not route a later model call of the Turn. A refusal that explains itself (no Source, all cooling
+ # down) shows its own copy first (``_preflight``); this names the rest instead of calling it an internal error.
+ "ModelHubError": "modelHubRoute",
# The loop refuses a route that declares no tool support before any request.
"UnsupportedModelRoute": "toolsUnsupported",
# A provider answer the canonical transcript cannot hold.
diff --git a/tests/scenarios/model_hub/catalog.yaml b/tests/scenarios/model_hub/catalog.yaml
index 1833e0352..2e25b4aae 100644
--- a/tests/scenarios/model_hub/catalog.yaml
+++ b/tests/scenarios/model_hub/catalog.yaml
@@ -648,6 +648,12 @@ scenarios:
test: tests/test_model_hub_vibey_consumer.py::test_sequential_vibey_retry_replaces_route_and_retains_served_provenance
related_tests:
- tests/test_model_hub_vibey_consumer.py::test_vibey_route_replacement_refuses_overlap_without_poisoning_current_request
+ - id: MH-VIBEY-010
+ name: A Vibey call that failed over to another Source does not refuse the Turn's next call while it settles
+ status: covered
+ kind: recovery
+ layer: contract
+ test: tests/test_model_hub_vibey_consumer.py::test_the_next_vibey_call_takes_the_failed_over_source_while_the_last_call_settles
- id: MH-VIBEY-008
name: Every Vibey resolution names its Turn, so each Turn on its Session's service-lifetime scope keeps exact provenance
status: covered
diff --git a/tests/test_model_hub_vibey_consumer.py b/tests/test_model_hub_vibey_consumer.py
index a6b204c98..83d4bc9d6 100644
--- a/tests/test_model_hub_vibey_consumer.py
+++ b/tests/test_model_hub_vibey_consumer.py
@@ -649,6 +649,80 @@ async def test_a_finished_turns_request_is_never_credited_to_the_next_turn_on_it
await gateway.close()
+class _SettlingLiveHandle(LiveInvokeHandle):
+ """A stream whose terminal frame was already forwarded, while the call has not finished.
+
+ ``held`` names what is still outstanding: the upstream's EOF after the terminal frame, or
+ the call's settlement (its outcome) after EOF.
+ """
+
+ def __init__(self, outcome, chunks, *, held):
+ self.settle = asyncio.Event()
+ self._held = held
+ super().__init__(outcome, chunks)
+
+ async def _iterate(self, chunks):
+ for chunk in chunks:
+ yield chunk
+ if self._held == "eof":
+ await self.settle.wait()
+
+ async def outcome(self):
+ if self._held == "settlement":
+ await self.settle.wait()
+ return self._outcome
+
+
[email protected]("held", ["eof", "settlement"])
+async def test_the_next_vibey_call_takes_the_failed_over_source_while_the_last_call_settles(tmp_path, held):
+ """MH-VIBEY-010: a call that failed over to another Source does not refuse the turn's next call.
+
+ Vibey stops reading a stream at its terminal event and resolves its next call at once, while the
+ gateway may still await the upstream's EOF or settle the last call. When that call failed over (its first Source now cooling
+ down), the next resolution names the other Source and replaces the turn's route; that waits for the
+ finished call instead of refusing the turn as overlapping live work.
+ """
+ primary = _source("src_primary01", "Primary", vendor="openai", protocol="openai_responses", model_id="gpt-x")
+ fallback = _source("src_fallback01", "Fallback", vendor="openai", protocol="openai_responses", model_id="gpt-x")
+ answer = _body("openai_responses", True)
+ settling = _SettlingLiveHandle(
+ _outcome(RawOutcomeKind.SUCCESS, source_id=fallback.id, stream_started=True), (answer,), held=held,
+ )
+ service = _vibey_service(tmp_path, [primary, fallback], handles=[
+ InvokeHandle(_outcome(RawOutcomeKind.HTTP_ERROR, status=503, code="server_error", source_id=primary.id)),
+ settling,
+ ])
+ gateway = ModelHubTurnGateway(service)
+ router = ModelHubRuntimeRouter(service=service, turn_gateway=gateway)
+ turn_id = "turn-failed-over"
+ try:
+ first = await router.resolve_hop("menu-alias", process_scope="vibey:test", turn_id=turn_id)
+ assert first["source_id"] == primary.id
+ async with aiohttp.ClientSession() as client:
+ async with client.post(
+ f'{first["base_url"]}/responses',
+ headers={"Authorization": f'Bearer {first["token"]}', **first["request_headers"]},
+ json={"model": first["runtime_model"], "stream": True, "input": "hi"},
+ ) as response:
+ assert _served_hop(response)["source"] == fallback.id
+ received = b""
+ while b"response.completed" not in received:
+ received += await response.content.readany()
+ # The terminal event is in: the next call resolves while this one still settles.
+ following = asyncio.create_task(
+ router.resolve_hop("menu-alias", process_scope="vibey:test", turn_id=turn_id)
+ )
+ await asyncio.sleep(0.05)
+ settling.settle.set()
+ second = await asyncio.wait_for(following, timeout=5)
+ await response.read()
+ assert second["source_id"] == fallback.id
+ finally:
+ settling.settle.set()
+ await gateway.close()
+
+
@pytest.mark.asyncio
async def test_vibey_route_replacement_refuses_overlap_without_poisoning_current_request(tmp_path):
"""MH-VIBEY-006: a new launch cannot take an in-flight request's turn identity."""
diff --git a/tests/test_vibey_agent.py b/tests/test_vibey_agent.py
index ce33fe740..a40832668 100644
--- a/tests/test_vibey_agent.py
+++ b/tests/test_vibey_agent.py
@@ -1212,6 +1212,8 @@ def test_an_unclassified_provider_failure_and_an_unexpected_exception_each_say_w
assert error_key("unknown") == "vibeyAgent.error.unclassified"
# Any kind outside the named set is the class of an exception the loop did not expect.
assert error_key("HookStateError") == "vibeyAgent.error.internal"
+ # A Model Hub routing refusal with no copy of its own says it is one, not an internal error.
+ assert error_key("ModelHubError") == "vibeyAgent.error.modelHubRoute"
assert error_key(None, reason="error") == "vibeyAgent.error.generic"
diff --git a/vibe/i18n/en.json b/vibe/i18n/en.json
index 3c201dc1b..b59cf50aa 100644
--- a/vibe/i18n/en.json
+++ b/vibe/i18n/en.json
@@ -1801,6 +1801,7 @@
"unclassified": "The request to the model failed with an error {backend} could not classify. Try again, or choose another model.",
"outputLimit": "The model reached its output limit before producing a usable answer or tool call. Ask for smaller steps, or choose a model with a larger output limit.",
"internal": "{backend} stopped because of an internal error. Try again; if it happens again, the service log names the cause.",
+ "modelHubRoute": "Model Hub could not route this step of {backend} to a model. Try again; if it happens again, the service log names the cause.",
"historyUnreadable": "This session's history can't be loaded, so {backend} can't continue it. Start a new session with /new.",
"generic": "{backend}'s run failed."
}
diff --git a/vibe/i18n/zh.json b/vibe/i18n/zh.json
index 2cd264957..4ef3a96fb 100644
--- a/vibe/i18n/zh.json
+++ b/vibe/i18n/zh.json
@@ -1801,6 +1801,7 @@
"unclassified": "请求模型时出现了 {backend} 无法归类的错误。可以重试,或换一个模型。",
"outputLimit": "模型在生成可用的答案或工具调用前达到了输出上限。请让它分成更小的步骤,或换用输出上限更大的模型。",
"internal": "{backend} 因内部错误而停止。可以重试;如果再次出现,服务日志里记录了原因。",
+ "modelHubRoute": "Model Hub 没能为 {backend} 的这一步选定模型路由。可以重试;如果再次出现,服务日志里记录了原因。",
"historyUnreadable": "这个会话的历史记录无法加载,{backend} 无法继续它。请发送 /new 开始一个新会话。",
"generic": "{backend} 运行失败。"
}