fix(auxiliary): pass reasoning_config and extra_body through to auxiliary Anthropic calls

[dorokuma]

Summary

_AnthropicCompletionsAdapter.create() in agent/auxiliary_client.py silently discards caller-supplied reasoning_config and extra_body for every auxiliary task routed through the Anthropic-Messages protocol (anthropic, minimax, kimi-coding, z.ai, any custom /anthropic-suffixed endpoint). The main agent path (agent/transports/anthropic.py) already handles these correctly; this PR aligns the auxiliary adapter with the same pattern.

Two bugs, one root cause

Bug A — reasoning_config hardcoded to None (L1000)

anthropic_kwargs = build_anthropic_kwargs(
    ...
    reasoning_config=None,   # always None, regardless of caller
    ...
)

The main agent path already does this right:

return build_anthropic_kwargs(..., reasoning_config=params.get("reasoning_config"), ...)

Bug B — extra_body dropped between caller and SDK (L1012)

create(**kwargs) accepts the caller's OpenAI-style kwargs but only forwards a hand-picked subset to self._client.messages.create(**anthropic_kwargs). extra_body (and any other caller-supplied field) never reaches the wire. The codex/responses transport in the same file already merges extra_body; the Anthropic branch is the gap.

Fix

--- a/agent/auxiliary_client.py
+++ b/agent/auxiliary_client.py
@@ -997,7 +997,7 @@ class _AnthropicCompletionsAdapter:
             messages=messages,
             tools=tools,
             max_tokens=max_tokens,
-            reasoning_config=None,
+            reasoning_config=kwargs.get("reasoning_config"),
             tool_choice=normalized_tool_choice,
             is_oauth=self._is_oauth,
         )
@@ -1009,6 +1009,19 @@ class _AnthropicCompletionsAdapter:
             if not _forbids_sampling_params(model):
                 anthropic_kwargs["temperature"] = temperature

+        # Merge caller-supplied extra_body so providers behind
+        # Anthropic-compatible gateways can receive per-vendor request
+        # fields (e.g. thinking control, metadata, service_tier). Dict
+        # form is the documented Anthropic SDK passthrough for
+        # non-standard request body keys.
+        caller_extra_body = kwargs.get("extra_body")
+        if caller_extra_body and isinstance(caller_extra_body, dict):
+            existing = anthropic_kwargs.get("extra_body") or {}
+            if not isinstance(existing, dict):
+                existing = {}
+            anthropic_kwargs["extra_body"] = {**existing, **caller_extra_body}
+
         response = self._client.messages.create(**anthropic_kwargs)

Behavior

• Backward-compatible. Callers that don't pass reasoning_config or extra_body see no change.
• Unlocks caller extra_body for the auxiliary path. Callers can now pass extra_body={"thinking": {"type": "disabled"}} (or any vendor field) via the standard OpenAI-style kwarg and have it reach the Anthropic SDK.
• reasoning_config kwarg now flows into build_anthropic_kwargs like the main agent does, instead of being silently replaced with None.

Tests

Add a unit test under tests/agent/ mocking self._client.messages.create() and asserting:

1. reasoning_config={"enabled": True, "effort": "medium"} from the caller → build_anthropic_kwargs is invoked with that value.
2. extra_body={"thinking": {"type": "disabled"}} from the caller → kwargs to messages.create() include the merged dict under extra_body.
3. Neither passed → kwargs to messages.create() are byte-identical to today (no regression).

Related issues

Same family of issue, different surfaces — flagging for reviewers:

• Bug: get_auxiliary_extra_body() ignores auxiliary.<task>.extra_body from config.yaml #35566 — auxiliary extra_body ignored from config.yaml
• feat(config): support model.extra_body passthrough for CLI and gateway agent requests #7209 — extra_body passthrough for CLI/gateway main agent
• Z.AI / GLM via zai provider never returns reasoning_content — Hermes sends extra_body.reasoning (OpenRouter-style) but Z.AI expects extra_body.thinking={"type":"enabled"} #16533 — Z.AI/GLM reasoning_content never returned (root cause: extra_body shape)
• [Feature]: Per-auxiliary reasoning effort configuration #32813 — per-auxiliary reasoning effort configuration (becomes solvable once this lands)
• /reasoning max rejected by frontend even though Anthropic adapter already supports it end-to-end #29248 — /reasoning max rejected by frontend

[teknium1]

Thanks for identifying the auxiliary Anthropic parity gap. It still exists on current main: agent/auxiliary_client.py:1283 hardcodes reasoning_config=None, while the main transport forwards its caller value at agent/transports/anthropic.py:70.

Problems

• The patch predates the current dispatch refactor. Current main calls create_anthropic_message() at agent/auxiliary_client.py:1295, rather than the direct SDK call in the PR diff. The helper forwards the constructed kwargs to messages.stream()/messages.create() at agent/anthropic_adapter.py:2773-2790, so the extra_body merge needs to be placed before that helper call.
• No regression test is included. Existing adapter tests use a mocked create_anthropic_message() seam at tests/agent/test_auxiliary_client.py:1260-1276, which is suitable for verifying both passthroughs.

Suggested changes

• Rebase the implementation concept onto the current helper-based adapter: forward kwargs.get("reasoning_config") at agent/auxiliary_client.py:1283, merge dict extra_body into anthropic_kwargs, then dispatch through the helper.
• Add focused tests for forwarded reasoning configuration, forwarded/merged extra_body, and the no-extra-input baseline.

Automated hermes-sweeper review.

[teknium1]

Both halves of this PR are now on main — closing with full credit, @dorokuma. Your diagnosis was exactly right on both counts:

Bug A (reasoning_config hardcoded to None): fixed via #64597/#64631, which route reasoning through build_anthropic_kwargs with per-task and per-slot resolution.

Bug B (extra_body dropped between caller and SDK): your commit was cherry-picked onto current main with your authorship preserved (rebase merge, commit 771571a) via PR #64942.

One scoping change on top of your merge block, informed by live probes against api.anthropic.com: the Messages API strictly rejects unknown body keys (400 "Extra inputs are not permitted"), including the OpenAI-shaped reasoning dict — which Hermes' own per-task config now places in extra_body. The merged passthrough therefore excludes reasoning (it's translated natively into thinking) and _-prefixed private keys, and merges over adapter-emitted fields rather than clobbering. Everything else a caller configures reaches the wire exactly as you intended.

Thanks for the precise two-bug analysis — the PR body's line-level diagnosis made this salvage straightforward.

[dorokuma]

Appreciate the thorough follow-up and the credit — glad the analysis helped make the fix straightforward.