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(vibey): scope Anthropic thinking fallback to Claude (#2928)

Call Whitef92922a4fc44
diff --git a/core/agent_core/ai/anthropic.py b/core/agent_core/ai/anthropic.py
index 47c8a1c44..97a903e1a 100644
--- a/core/agent_core/ai/anthropic.py
+++ b/core/agent_core/ai/anthropic.py
@@ -17,7 +17,7 @@ from typing import Any
import httpx
-from core.agent_core.ai.claude_thinking import uses_adaptive_thinking
+from core.agent_core.ai.claude_thinking import is_claude_model, uses_adaptive_thinking
from core.agent_core.ai.hosted import anthropic_hosted_block, citation, reported_counts
from core.agent_core.ai.image_limits import LoadedImage
from core.agent_core.ai._common import (
@@ -1045,15 +1045,16 @@ def build_messages_payload(
payload["tools"] = tools
effort = (request.reasoning_effort or "").lower()
adaptive = uses_adaptive_thinking(request.endpoint.model_id)
- if effort in {"none", "off", "disabled"} or (
- effort and not adaptive and _continues_unthought_tool_loop(payload_messages)
- ):
- # Under manual thinking, Anthropic requires the assistant message a tool loop continues to start with its
- # thinking block. One whose thinking cannot go back (another origin's, or a Session whose signatures the
- # provider refused) continues with thinking off, as Anthropic's refusal itself advises; the models that take
- # manual thinking all accept that. Adaptive thinking has no such rule, and several adaptive models refuse
- # ``disabled``, so it stays on.
+ if effort in {"none", "off", "disabled"}:
payload["thinking"] = {"type": "disabled"}
+ elif effort and not adaptive and _continues_unthought_tool_loop(payload_messages):
+ if is_claude_model(request.endpoint.model_id):
+ # Under manual thinking, Anthropic requires the assistant message a tool loop continues to start with its
+ # thinking block. One whose thinking cannot go back (another origin's, or a Session whose signatures the
+ # provider refused) continues with thinking off, as Anthropic's refusal itself advises; the models that
+ # take manual thinking all accept that. A non-Claude Anthropic-compatible model has no verified thinking
+ # shape here, so leave the field out rather than guessing.
+ payload["thinking"] = {"type": "disabled"}
elif effort:
if adaptive:
payload["thinking"] = {"type": "adaptive"}
diff --git a/core/agent_core/ai/claude_thinking.py b/core/agent_core/ai/claude_thinking.py
index 0e8fdc679..8482bc2a5 100644
--- a/core/agent_core/ai/claude_thinking.py
+++ b/core/agent_core/ai/claude_thinking.py
@@ -23,6 +23,13 @@ _CLAUDE_VERSION: Final = re.compile(r"claude-(?:[a-z]+-)?(\d+)(?:-(\d{1,2}))?(?!
_FIRST_ADAPTIVE_VERSION: Final = (4, 6)
+def is_claude_model(model_id: str) -> bool:
+ """Whether the model ID's final path segment names a Claude model."""
+
+ normalized = model_id.lower().replace(".", "-")
+ return normalized.rsplit("/", 1)[-1].startswith("claude-")
+
+
def uses_adaptive_thinking(model_id: str) -> bool:
"""Whether this Claude model takes adaptive thinking; False for a Claude version before 4.6 and for non-Claude."""
diff --git a/tests/agent_core/ai/test_claude_thinking.py b/tests/agent_core/ai/test_claude_thinking.py
index 800592182..6449f0c5d 100644
--- a/tests/agent_core/ai/test_claude_thinking.py
+++ b/tests/agent_core/ai/test_claude_thinking.py
@@ -4,7 +4,7 @@ from __future__ import annotations
import pytest
-from core.agent_core.ai.claude_thinking import uses_adaptive_thinking
+from core.agent_core.ai.claude_thinking import is_claude_model, uses_adaptive_thinking
@pytest.mark.parametrize(("model_id", "adaptive"), [
@@ -30,3 +30,10 @@ from core.agent_core.ai.claude_thinking import uses_adaptive_thinking
])
def test_only_a_claude_version_before_4_6_takes_a_token_budget(model_id, adaptive):
assert uses_adaptive_thinking(model_id) is adaptive
+
+
+def test_claude_model_detection_uses_the_model_name_not_the_protocol() -> None:
+ assert is_claude_model("claude-sonnet-4-5") is True
+ assert is_claude_model("anthropic/claude-sonnet-4-5") is True
+ assert is_claude_model("grok-4.7") is False
+ assert is_claude_model("relay-claude-opus-5-5") is False
diff --git a/tests/agent_core/ai/test_hosted_replay.py b/tests/agent_core/ai/test_hosted_replay.py
index 46409ba5f..40adcabc3 100644
--- a/tests/agent_core/ai/test_hosted_replay.py
+++ b/tests/agent_core/ai/test_hosted_replay.py
@@ -680,6 +680,7 @@ _READ_CALL = ToolCallBlock(id="toolu_1", name="read", arguments={"path": "a"})
# Sonnet 5.5, Fable and Mythos "reject `thinking: {type: "disabled"}`".
_ADAPTIVE_MODEL = "claude-opus-5-5"
_MANUAL_MODEL = "claude-sonnet-4-5"
+_NON_CLAUDE_MODEL = "grok-4.7"
@pytest.mark.parametrize(
@@ -698,16 +699,19 @@ _MANUAL_MODEL = "claude-sonnet-4-5"
(_MANUAL_MODEL, (_READ_CALL,), "result-then-user", "disabled"),
# ...while a new turn after an answer keeps it on.
(_MANUAL_MODEL, (TextBlock("done"),), "user", "enabled"),
+ # An Anthropic-compatible non-Claude model does not receive the Claude-specific fallback...
+ (_NON_CLAUDE_MODEL, (_READ_CALL,), "result", None),
+ # ...but its normal non-tool-loop reasoning request remains unchanged.
+ (_NON_CLAUDE_MODEL, (TextBlock("done"),), "user", "enabled"),
],
ids=["adaptive-thinking-as-text", "adaptive-call-first", "manual-signed-thinking-first", "manual-thinking-as-text",
- "manual-call-first", "manual-user-message-after-results", "manual-new-turn"],
+ "manual-call-first", "manual-user-message-after-results", "manual-new-turn", "non-claude-tool-loop",
+ "non-claude-new-turn"],
)
-async def test_a_tool_loop_whose_assistant_message_cannot_start_with_thinking_continues_with_thinking_off(
+async def test_anthropic_thinking_fallback_is_scoped_to_manual_claude_tool_loops(
model, continued, after, thinking
):
- """Under manual thinking a request whose continued assistant message cannot start with thinking (its thinking went
- as text) is sent with thinking off, as Anthropic's refusal advises; adaptive thinking has no such rule and is never
- turned off, since several adaptive models refuse ``disabled``; a new turn after an answer keeps thinking on."""
+ """Manual Claude uses the #2736 fallback; adaptive Claude stays on; other models omit it rather than guessing."""
endpoint = replace(_endpoint("anthropic", replay=False, native=False), hosted_tools={}, model_id=model)
result = ToolResultMessage("toolu_1", "read", (TextBlock("contents"),))
@@ -722,7 +726,10 @@ async def test_a_tool_loop_whose_assistant_message_cannot_start_with_thinking_co
wire = await build_wire_request(request, gateway=True)
- assert wire.body["thinking"]["type"] == thinking
+ if thinking is None:
+ assert "thinking" not in wire.body
+ else:
+ assert wire.body["thinking"]["type"] == thinking
def _thought_and_read() -> str:
diff --git a/tests/scenarios/model_hub/catalog.yaml b/tests/scenarios/model_hub/catalog.yaml
index 2927eaf8b..a74b56ded 100644
--- a/tests/scenarios/model_hub/catalog.yaml
+++ b/tests/scenarios/model_hub/catalog.yaml
@@ -439,7 +439,7 @@ scenarios:
- tests/agent_core/ai/test_hosted_replay.py::test_an_opaque_payload_refusal_ends_the_send_as_a_refusal_of_its_source_identity
- tests/agent_core/ai/test_hosted_replay.py::test_an_opaque_payload_another_account_refuses_is_sent_portably_from_then_on_in_the_session
- tests/agent_core/ai/test_hosted_replay.py::test_a_refusal_of_what_the_session_already_demoted_stands_as_the_error
- - tests/agent_core/ai/test_hosted_replay.py::test_a_tool_loop_whose_assistant_message_cannot_start_with_thinking_continues_with_thinking_off
+ - tests/agent_core/ai/test_hosted_replay.py::test_anthropic_thinking_fallback_is_scoped_to_manual_claude_tool_loops
- tests/agent_core/ai/test_hosted_replay.py::test_a_session_whose_signatures_were_refused_still_finishes_tool_loops_with_thinking_on
- tests/agent_core/ai/test_hosted_replay.py::test_a_decrypt_refusal_drops_every_opaque_payload_of_that_identity_for_the_rest_of_the_turn
- tests/agent_core/ai/test_hosted_replay.py::test_a_hosted_call_goes_back_verbatim_and_pinned_only_to_its_own_source_with_replay_evidence