fix(anthropic): send thinking=disabled explicitly on third-party endpoints (#15700) by briandevans · Pull Request #15712 · NousResearch/hermes-agent

briandevans · 2026-04-25T16:12:30Z

Summary

Third-party Anthropic-compatible endpoints (e.g. DeepSeek /anthropic) default to thinking mode when the thinking parameter is absent; native Anthropic disables thinking by default
When reasoning_config.enabled = False, Hermes now sends thinking: {"type": "disabled"} for third-party endpoints so they don't default to thinking on
Kimi /coding and native Anthropic are unaffected

The bug

build_anthropic_kwargs sets thinking only when thinking is enabled. For native Anthropic that is correct — thinking is disabled by default. But DeepSeek's /anthropic endpoint treats the absent parameter as "thinking enabled", causing every request with enabled=False to fail with HTTP 400:

The content[].thinking in the thinking mode must be passed back to the API.

The fix

Added an elif branch inside the existing reasoning_config block at agent/anthropic_adapter.py:

elif reasoning_config.get("enabled") is False and _is_third_party_anthropic_endpoint(base_url):
    kwargs["thinking"] = {"type": "disabled"}

_is_third_party_anthropic_endpoint returns True for any non-anthropic.com base URL. Kimi /coding is already excluded by the outer not _is_kimi_coding guard.

Test plan

Before: 2 of 4 new tests fail (assert None == {'type': 'disabled'}) — regression guard confirmed
After: all 4 new tests pass
Regression guard: reverted fix → test_reasoning_disabled_third_party_sends_explicit_disabled and test_reasoning_disabled_third_party_any_provider both fail with AssertionError: assert None == {'type': 'disabled'}; restored → 0 failures
tests/agent/test_anthropic_adapter.py — 131 passed, 10 pre-existing baseline failures (unchanged)
tests/agent/test_kimi_coding_anthropic_thinking.py — 8 passed (unchanged)
Existing test_reasoning_disabled still passes (native Anthropic still omits the key)

New test cases:

test_reasoning_disabled_third_party_sends_explicit_disabled — DeepSeek endpoint gets {"type": "disabled"}
test_reasoning_disabled_third_party_any_provider — any non-Anthropic endpoint gets {"type": "disabled"}
test_reasoning_disabled_kimi_coding_omits_thinking — Kimi /coding gets neither enabled nor disabled
test_reasoning_none_third_party_omits_thinking — reasoning_config=None on third-party endpoint still omits thinking

Pull request overview

Ensures Anthropic-compatible third-party endpoints that default to “thinking enabled” when the thinking field is absent will instead receive an explicit thinking: {"type": "disabled"} when reasoning_config.enabled is set to False, while preserving native Anthropic and Kimi /coding behavior.

Changes:

Add a third-party-only branch to send thinking={"type":"disabled"} when reasoning is explicitly disabled.
Add regression tests covering native Anthropic omission, third-party explicit disable, Kimi /coding omission, and reasoning_config=None behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
`agent/anthropic_adapter.py`	Adds explicit `thinking` disable for third-party Anthropic-compatible endpoints when `reasoning_config.enabled` is `False`.
`tests/agent/test_anthropic_adapter.py`	Adds tests validating the new third-party explicit-disable behavior and guarding native/Kimi exceptions.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

briandevans · 2026-04-25T17:07:54Z

+        elif reasoning_config.get("enabled") is False and _is_third_party_anthropic_endpoint(base_url):
+            # Third-party Anthropic-compatible endpoints (e.g. DeepSeek /anthropic)
+            # default to thinking mode when the `thinking` parameter is absent.
+            # Native Anthropic disables thinking by default, so the key is omitted
+            # there — but third-party implementations must be told explicitly.
+            kwargs["thinking"] = {"type": "disabled"}


Valid catch — addressed in 3a4ab0f074f083f897814c7bae7eb0ec2a64b92f.

Proxy-hostname substring misclassification ✅ — Replaced the raw "anthropic.com" in normalized substring check with a proper hostname comparison using urlparse. The function now extracts the hostname and checks host == "anthropic.com" or host.endswith(".anthropic.com"), so https://anthropic.com.proxy.company.com/anthropic is correctly identified as a third-party endpoint and receives thinking={"type":"disabled"}.

Regression test added ✅ — test_reasoning_disabled_proxy_hostname_containing_anthropic_com covers exactly this URL pattern; confirmed it fails with the old substring code and passes with the fix.

briandevans · 2026-04-25T17:07:59Z

+            base_url="https://third-party.example.com/anthropic",
+        )
+        assert kwargs.get("thinking") == {"type": "disabled"}
+


Addressed in 3a4ab0f074f083f897814c7bae7eb0ec2a64b92f — see reply above on the companion thread.

…oints (NousResearch#15700) DeepSeek's Anthropic-compatible endpoint (`/anthropic`) defaults to thinking mode when the `thinking` parameter is absent. Native Anthropic disables thinking by default, so the key has always been omitted there — but third-party implementations that mirror the Anthropic protocol must be told explicitly to disable it. Without the explicit `{"type": "disabled"}`, DeepSeek rejects the request with HTTP 400 "The content[].thinking in the thinking mode must be passed back to the API." The fix adds an `elif` branch inside the existing `reasoning_config` block: when `enabled` is `False` and the base URL is a third-party endpoint (i.e. not anthropic.com), set `thinking={"type": "disabled"}`. Kimi /coding is already excluded by the outer `not _is_kimi_coding` guard and remains unaffected. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

… _is_third_party_anthropic_endpoint A proxy URL like `https://anthropic.com.proxy.company.com/anthropic` contains `anthropic.com` as a substring in its full URL string, causing the old `"anthropic.com" in normalized` check to misclassify it as the native Anthropic API. This would suppress the `thinking={"type":"disabled"}` parameter required for third-party DeepSeek-style endpoints. Fix: parse the URL and compare the hostname against `anthropic.com` / `*.anthropic.com` rather than doing a substring match on the raw string. Adds a regression test for the proxy-hostname edge case. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

briandevans · 2026-04-29T15:15:41Z

Rebased onto current origin/main (13c238327) — was 94h stale and the previous base predated the related Kimi thinking-suppression work (83c288da0 fix(anthropic): broaden Kimi thinking-suppression to custom endpoints (#17455)) and the dot-mangling narrowing (7141cda96).

Re-checked the interaction with that newer Kimi suppression — they're orthogonal:

#17455 short-circuits the entire if reasoning_config… block via _is_kimi_coding so Kimi /coding endpoints emit no thinking parameter at all (server-side enabled).
This PR adds an elif reasoning_config.get("enabled") is False and _is_third_party_anthropic_endpoint(base_url) branch that explicitly sends thinking={"type": "disabled"} for third-party Anthropic-compat endpoints (DeepSeek /anthropic, etc.) where the parameter defaults to enabled when absent.

Clean rebase, no conflicts. Re-ran focused tests on the rebased head (4d630ffd6):

tests/agent/test_anthropic_adapter.py -k "thinking or third_party or reasoning_disabled" — 20/20 pass (covers the new test_reasoning_disabled_third_party_sends_explicit_disabled, the existing native-Anthropic test_reasoning_disabled, the Kimi-coding test_reasoning_disabled_kimi_coding_omits_thinking, and the hostname-vs-substring guard test_reasoning_disabled_proxy_hostname_containing_anthropic_com).

The fix and its hostname-comparison hardening (substring anthropic.com would have false-matched proxies like anthropic.com.example.com) are unchanged from review.

briandevans · 2026-05-19T10:31:52Z

Closing — issue #15700 was resolved by @teknium1's #26648 (fix(deepseek): wire thinking-mode via DeepSeekProfile), which closes #15700/#17212/#17825 from the DeepSeekProfile path instead of the adapter-level send-disabled approach taken here. That's the right place for the fix. Thanks for the merge!

Copilot AI review requested due to automatic review settings April 25, 2026 16:12

Copilot started reviewing on behalf of briandevans April 25, 2026 16:12 View session

Copilot AI reviewed Apr 25, 2026

View reviewed changes

alt-glitch added type/bug Something isn't working P2 Medium — degraded but workaround exists comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint provider/anthropic Anthropic native Messages API labels Apr 25, 2026

briandevans and others added 2 commits April 29, 2026 08:14

briandevans force-pushed the fix/anthropic-deepseek-thinking-disabled-15700 branch from 3a4ab0f to 4d630ff Compare April 29, 2026 15:15

briandevans closed this May 19, 2026

Navigation Menu

Uh oh!

fix(anthropic): send thinking=disabled explicitly on third-party endpoints (#15700) - #15712

fix(anthropic): send thinking=disabled explicitly on third-party endpoints (#15700)#15712
briandevans wants to merge 2 commits into
NousResearch:mainfrom
briandevans:fix/anthropic-deepseek-thinking-disabled-15700

briandevans commented Apr 25, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

briandevans Apr 25, 2026

Uh oh!

briandevans Apr 25, 2026

Uh oh!

briandevans commented Apr 29, 2026

Uh oh!

briandevans commented May 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

briandevans commented Apr 25, 2026

Summary

The bug

The fix

Test plan

Related

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

briandevans Apr 25, 2026

Choose a reason for hiding this comment

Uh oh!

briandevans Apr 25, 2026

Choose a reason for hiding this comment

Uh oh!

briandevans commented Apr 29, 2026

Uh oh!

briandevans commented May 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants