0fce82164a
Move provider adapters (anthropic, bedrock, azure), platform adapters (telegram, slack, discord, feishu, dingtalk, matrix), and terminal backends (modal, daytona) out of core into plugins/ workspace members. Core references them via the plugin registries (get_provider_namespace / get_provider_service / get_tool_provider / get_credential_pool_hook) instead of direct imports. - Provider/platform/terminal adapters relocated under plugins/; pyproject extras reference workspace members; nix variants aggregate per-platform extras. - Anthropic credential discovery + OAuth-masquerade guard live in the plugin's credential_pool_hook; browser-open guarded by _can_open_graphical_browser. - Vercel AI Gateway + Vercel Sandbox removed (upstream deletion); get_bedrock_model_ids removed (replaced by bedrock_model_ids_or_none + discover_bedrock_models). - Terminal backends resolve ModalEnvironment / DaytonaEnvironment lazily from the plugin registry. - uv.lock regenerated against the pluginified workspace. Plugin test suites updated for the relocation: imports point at hermes_agent_<plat>.adapter, caplog logger-name filters and monkeypatch targets use the new module paths, and credential/rollback tests patch registries.get_provider_service rather than the removed agent.*_adapter modules. Verified: zero dead imports of relocated modules in core (import smoke test + rename-map grep); nix develop succeeds; targeted plugin suites green (bedrock, anthropic-auxiliary, matrix, dingtalk, feishu, credential_pool, switch_model_rollback). Remaining full-suite failures are pre-existing on the pre-merge tree (telegram setUpModule __code__) or environmental (voice/media/ PTY/network-dependent), not introduced here.
243 lines
9.9 KiB
Python
243 lines
9.9 KiB
Python
"""Regression guard: preserve thinking blocks on DeepSeek's /anthropic endpoint.
|
|
|
|
DeepSeek's ``api.deepseek.com/anthropic`` route speaks the Anthropic Messages
|
|
protocol but, when thinking mode is enabled, requires ``thinking`` blocks from
|
|
prior assistant turns to round-trip on subsequent requests. The generic
|
|
third-party path strips them (signatures are Anthropic-proprietary and other
|
|
proxies cannot validate them), so without a DeepSeek-specific carve-out the
|
|
next tool-call turn fails with HTTP 400::
|
|
|
|
The content[].thinking in the thinking mode must be passed back to the
|
|
API.
|
|
|
|
DeepSeek's compatibility matrix lists ``thinking`` as supported but
|
|
``redacted_thinking`` and ``cache_control`` on thinking blocks as not
|
|
supported. Handling is the same as Kimi's ``/coding`` endpoint: strip
|
|
Anthropic-signed blocks (DeepSeek can't validate them) but preserve unsigned
|
|
blocks that Hermes synthesises from ``reasoning_content``.
|
|
|
|
See hermes-agent#16748.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import pytest
|
|
|
|
|
|
class TestDeepSeekAnthropicPreservesThinking:
|
|
"""convert_messages_to_anthropic must replay DeepSeek thinking blocks."""
|
|
|
|
@pytest.mark.parametrize(
|
|
"base_url",
|
|
[
|
|
"https://api.deepseek.com/anthropic",
|
|
"https://api.deepseek.com/anthropic/",
|
|
"https://api.deepseek.com/anthropic/v1",
|
|
"https://API.DeepSeek.com/anthropic",
|
|
],
|
|
)
|
|
def test_unsigned_thinking_block_survives_replay(self, base_url: str) -> None:
|
|
"""Unsigned thinking (synthesised from reasoning_content) must be preserved."""
|
|
from agent.anthropic_format import convert_messages_to_anthropic
|
|
|
|
messages = [
|
|
{"role": "user", "content": "hi"},
|
|
{
|
|
"role": "assistant",
|
|
"reasoning_content": "planning the tool call",
|
|
"tool_calls": [
|
|
{
|
|
"id": "call_1",
|
|
"type": "function",
|
|
"function": {"name": "skill_view", "arguments": "{}"},
|
|
}
|
|
],
|
|
},
|
|
{"role": "tool", "tool_call_id": "call_1", "content": "ok"},
|
|
]
|
|
_system, converted = convert_messages_to_anthropic(
|
|
messages, base_url=base_url
|
|
)
|
|
|
|
assistant_msg = next(m for m in converted if m["role"] == "assistant")
|
|
thinking_blocks = [
|
|
b for b in assistant_msg["content"]
|
|
if isinstance(b, dict) and b.get("type") == "thinking"
|
|
]
|
|
assert len(thinking_blocks) == 1, (
|
|
f"DeepSeek /anthropic ({base_url}) must preserve unsigned thinking "
|
|
"blocks synthesised from reasoning_content — upstream rejects "
|
|
"replayed tool-call messages without them."
|
|
)
|
|
assert thinking_blocks[0]["thinking"] == "planning the tool call"
|
|
# Synthesised block — never has a signature
|
|
assert "signature" not in thinking_blocks[0]
|
|
|
|
def test_unsigned_thinking_preserved_on_non_latest_assistant_turn(self) -> None:
|
|
"""DeepSeek validates history across every prior assistant turn, not just last."""
|
|
from agent.anthropic_format import convert_messages_to_anthropic
|
|
|
|
messages = [
|
|
{"role": "user", "content": "q1"},
|
|
{
|
|
"role": "assistant",
|
|
"reasoning_content": "r1",
|
|
"tool_calls": [
|
|
{
|
|
"id": "call_1",
|
|
"type": "function",
|
|
"function": {"name": "f", "arguments": "{}"},
|
|
}
|
|
],
|
|
},
|
|
{"role": "tool", "tool_call_id": "call_1", "content": "ok"},
|
|
{"role": "user", "content": "q2"},
|
|
{
|
|
"role": "assistant",
|
|
"reasoning_content": "r2",
|
|
"tool_calls": [
|
|
{
|
|
"id": "call_2",
|
|
"type": "function",
|
|
"function": {"name": "f", "arguments": "{}"},
|
|
}
|
|
],
|
|
},
|
|
{"role": "tool", "tool_call_id": "call_2", "content": "ok"},
|
|
]
|
|
_system, converted = convert_messages_to_anthropic(
|
|
messages, base_url="https://api.deepseek.com/anthropic"
|
|
)
|
|
|
|
assistants = [m for m in converted if m["role"] == "assistant"]
|
|
assert len(assistants) == 2
|
|
for assistant, expected in zip(assistants, ("r1", "r2")):
|
|
thinking = [
|
|
b for b in assistant["content"]
|
|
if isinstance(b, dict) and b.get("type") == "thinking"
|
|
]
|
|
assert len(thinking) == 1
|
|
assert thinking[0]["thinking"] == expected
|
|
|
|
def test_signed_anthropic_thinking_block_is_stripped(self) -> None:
|
|
"""Anthropic-signed blocks (that leaked through) must still be stripped.
|
|
|
|
DeepSeek issues its own signatures and cannot validate Anthropic's —
|
|
the strip-signed / keep-unsigned split matches the Kimi policy.
|
|
"""
|
|
from agent.anthropic_format import convert_messages_to_anthropic
|
|
|
|
messages = [
|
|
{"role": "user", "content": "hi"},
|
|
{
|
|
"role": "assistant",
|
|
"content": [
|
|
{
|
|
"type": "thinking",
|
|
"thinking": "anthropic-signed payload",
|
|
"signature": "anthropic-sig-xyz",
|
|
},
|
|
{"type": "text", "text": "hello"},
|
|
],
|
|
},
|
|
{"role": "user", "content": "again"},
|
|
]
|
|
_system, converted = convert_messages_to_anthropic(
|
|
messages, base_url="https://api.deepseek.com/anthropic"
|
|
)
|
|
|
|
assistant_msg = next(m for m in converted if m["role"] == "assistant")
|
|
thinking_blocks = [
|
|
b for b in assistant_msg["content"]
|
|
if isinstance(b, dict) and b.get("type") == "thinking"
|
|
]
|
|
assert thinking_blocks == [], (
|
|
"Signed Anthropic thinking blocks must be stripped on DeepSeek — "
|
|
"DeepSeek cannot validate Anthropic-proprietary signatures."
|
|
)
|
|
|
|
def test_cache_control_stripped_from_thinking_block(self) -> None:
|
|
"""cache_control must still be stripped even when the block is preserved.
|
|
|
|
DeepSeek's compatibility matrix lists cache_control on thinking blocks
|
|
as ignored — cache markers interfere with signature validation on
|
|
upstreams that do check them, so Hermes strips them everywhere.
|
|
"""
|
|
from agent.anthropic_format import convert_messages_to_anthropic
|
|
|
|
messages = [
|
|
{"role": "user", "content": "hi"},
|
|
{
|
|
"role": "assistant",
|
|
"reasoning_content": "r1",
|
|
"tool_calls": [
|
|
{
|
|
"id": "call_1",
|
|
"type": "function",
|
|
"function": {"name": "f", "arguments": "{}"},
|
|
}
|
|
],
|
|
},
|
|
{"role": "tool", "tool_call_id": "call_1", "content": "ok"},
|
|
]
|
|
# Inject cache_control on the synthesised thinking block after-the-fact
|
|
# by running conversion once, mutating, then re-running would be
|
|
# indirect. Instead check the simpler invariant: no thinking block in
|
|
# the converted output carries cache_control.
|
|
_system, converted = convert_messages_to_anthropic(
|
|
messages, base_url="https://api.deepseek.com/anthropic"
|
|
)
|
|
for m in converted:
|
|
if not isinstance(m.get("content"), list):
|
|
continue
|
|
for b in m["content"]:
|
|
if isinstance(b, dict) and b.get("type") in {"thinking", "redacted_thinking"}:
|
|
assert "cache_control" not in b
|
|
|
|
def test_openai_compat_deepseek_base_is_not_matched(self) -> None:
|
|
"""The OpenAI-compatible ``api.deepseek.com`` base must NOT trigger the
|
|
DeepSeek /anthropic branch — it never reaches this adapter, but the
|
|
detector should still fail closed so an accidental misuse doesn't
|
|
quietly send signed Anthropic blocks to an OpenAI endpoint.
|
|
"""
|
|
from agent.anthropic_format import _is_deepseek_anthropic_endpoint
|
|
|
|
assert _is_deepseek_anthropic_endpoint("https://api.deepseek.com") is False
|
|
assert _is_deepseek_anthropic_endpoint("https://api.deepseek.com/v1") is False
|
|
assert _is_deepseek_anthropic_endpoint("https://api.deepseek.com/anthropic") is True
|
|
assert _is_deepseek_anthropic_endpoint("https://api.deepseek.com/anthropic/v1") is True
|
|
|
|
def test_non_deepseek_third_party_still_strips_all_thinking(self) -> None:
|
|
"""MiniMax and other third-party Anthropic endpoints must keep the
|
|
generic strip-all behaviour (they reject unsigned blocks outright).
|
|
"""
|
|
from agent.anthropic_format import convert_messages_to_anthropic
|
|
|
|
messages = [
|
|
{"role": "user", "content": "hi"},
|
|
{
|
|
"role": "assistant",
|
|
"reasoning_content": "r1",
|
|
"tool_calls": [
|
|
{
|
|
"id": "call_1",
|
|
"type": "function",
|
|
"function": {"name": "f", "arguments": "{}"},
|
|
}
|
|
],
|
|
},
|
|
{"role": "tool", "tool_call_id": "call_1", "content": "ok"},
|
|
]
|
|
_system, converted = convert_messages_to_anthropic(
|
|
messages, base_url="https://api.minimax.io/anthropic"
|
|
)
|
|
assistant_msg = next(m for m in converted if m["role"] == "assistant")
|
|
thinking_blocks = [
|
|
b for b in assistant_msg["content"]
|
|
if isinstance(b, dict) and b.get("type") == "thinking"
|
|
]
|
|
assert thinking_blocks == [], (
|
|
"Non-DeepSeek third-party endpoints must keep the generic "
|
|
"strip-all-thinking behaviour — unsigned blocks get rejected."
|
|
)
|