From c1aba3d021047300ca8068c97e6fe1e028bb7f1c Mon Sep 17 00:00:00 2001
From: Graham King <grahamk@nvidia.com>
Date: Wed, 30 Sep 2026 12:38:56 -0400
Subject: [PATCH] fix(translation): preserve native Anthropic thinking settings

`AnthropicMessagesCodec::encode_request` (crates/switchyard-translation/src/codecs/anthropic/buffered.rs:294-308) now re-emits the caller's native thinking object on the reconstructed path, and only falls back to `{"type": "adaptive"}` when effort is set and no native thinking exists.

Previously, any `reasoning.effort` clobbered thinking with adaptive, so reconstruction disagreed with verbatim replay (a caller sending thinking: disabled + output_config.effort replayed as disabled but reconstructed as adaptive).

GLM 5.3's review:
> Real bug, correctly fixed. README claims match the code, including omit-before-merge ordering and reasoning_effort only applying to OpenAI backends.

Fixes: https://linear.app/nvidia/issue/SWITCH-1490/

Assisted-by: Pi:GPT 6 Astra medium
Reviewed-by: Pi:GLM 5.3 high
Signed-off-by: Graham King <grahamk@nvidia.com>
---
 crates/libsy-llm-client/README.md             |  6 ++++
 .../src/codecs/anthropic/buffered.rs          | 10 +++++-
 .../tests/request_translation.rs              | 32 +++++++++++++++++++
 3 files changed, 47 insertions(+), 1 deletion(-)

diff --git a/crates/libsy-llm-client/README.md b/crates/libsy-llm-client/README.md
index 34ec8a989..8b0ef430d 100644
--- a/crates/libsy-llm-client/README.md
+++ b/crates/libsy-llm-client/README.md
@@ -248,6 +248,12 @@ fn build_multi_format_client(
   and `anthropic-version`. Header names are case-insensitive.
 - Per-target top-level request defaults go in `HttpBackendConfig::extra_body`.
   The merge is shallow and fields already present in the request take precedence.
+- Anthropic reconstruction preserves the caller's native `thinking` settings, including
+  disabled thinking and manual budgets. `output_config.effort` is kept separately.
+  Effort implies adaptive thinking only when native `thinking` is absent.
+  To replace either top-level field for a target, list it in `omit_body_fields` and
+  supply its replacement in `extra_body`. Omission runs before defaults are merged.
+  Target `reasoning_effort` overrides apply to OpenAI backends.
 - `HttpBackendConfig::max_retries` controls additional attempts after retryable
   transport failures, timeouts, HTTP 408/429, and 5xx responses. Buffered body
   transport failures are retried; streaming body failures are not replayed after
diff --git a/crates/switchyard-translation/src/codecs/anthropic/buffered.rs b/crates/switchyard-translation/src/codecs/anthropic/buffered.rs
index c6cdcfba5..231d97b84 100644
--- a/crates/switchyard-translation/src/codecs/anthropic/buffered.rs
+++ b/crates/switchyard-translation/src/codecs/anthropic/buffered.rs
@@ -291,8 +291,16 @@ impl FormatCodec for AnthropicMessagesCodec {
         if request.stream {
             body.insert("stream".to_string(), Value::Bool(true));
         }
+        // Native thinking controls the mode independently of output effort.
+        // Other codecs store their own provider's reasoning object in `raw`.
+        if is_anthropic_request(request)
+            && let Some(thinking) = &request.reasoning.raw
+        {
+            body.insert("thinking".to_string(), thinking.clone());
+        }
         if let Some(effort) = &request.reasoning.effort {
-            body.insert("thinking".to_string(), json!({"type": "adaptive"}));
+            body.entry("thinking".to_string())
+                .or_insert_with(|| json!({"type": "adaptive"}));
             body.insert("output_config".to_string(), json!({"effort": effort}));
         }
         if let Some(response_format) = &request.output.response_format
diff --git a/crates/switchyard-translation/tests/request_translation.rs b/crates/switchyard-translation/tests/request_translation.rs
index 2af80f9d3..eceb54693 100644
--- a/crates/switchyard-translation/tests/request_translation.rs
+++ b/crates/switchyard-translation/tests/request_translation.rs
@@ -484,6 +484,38 @@ fn anthropic_thinking_to_responses_uses_normalized_effort() -> TestResult {
     Ok(())
 }
 
+#[test]
+fn anthropic_reconstruction_preserves_thinking() -> TestResult {
+    let engine = TranslationEngine::default();
+    let policy = normalized_policy();
+    for (thinking, effort) in [
+        (json!({"type": "disabled"}), Some("high")),
+        (json!({"type": "enabled", "budget_tokens": 2048}), None),
+        (json!({"type": "adaptive"}), Some("high")),
+    ] {
+        let mut body = json!({
+            "model": "caller", "max_tokens": 4096,
+            "messages": [{"role": "user", "content": "hi"}],
+            "thinking": thinking
+        });
+        if let Some(effort) = effort {
+            body["output_config"] = json!({"effort": effort});
+        }
+        let mut request = engine
+            .decode_request(WireFormat::AnthropicMessages, &body, &policy)?
+            .request;
+        prepare_request_for_target(&mut request, &"target/model".into(), Some("target prompt"));
+        let output = engine
+            .encode_request(WireFormat::AnthropicMessages, &request, &policy)?
+            .body;
+
+        body["model"] = json!("target/model");
+        body["system"] = json!("target prompt");
+        assert_eq!(output, body);
+    }
+    Ok(())
+}
+
 #[test]
 fn anthropic_target_prompt_preserves_native_request_fields() -> TestResult {
     let engine = TranslationEngine::default();