Merge origin/main into main

2026-10-07 01:47:47 +08:00 · 2026-10-05 12:10:23 +08:00
parent d4ed774423 0948f29da5
commit 84f0d5ce70
22 changed files with 2050 additions and 50 deletions
@@ -53,6 +53,75 @@ pub fn extract_provider_reasoning_effort_from_body(value: Option<&Value>) -> Opt
                .and_then(Value::as_str)
        })
        .and_then(normalize_provider_reasoning_effort)
+        .or_else(|| {
+            // Gemini also nests its payload one level down, so both the flat
+            // `generateContent` body and the `v1internal` envelope that carries it are read.
+            extract_gemini_reasoning_effort_from_body(object).or_else(|| {
+                object
+                    .get("request")
+                    .and_then(Value::as_object)
+                    .and_then(extract_gemini_reasoning_effort_from_body)
+            })
+        })
+}
+
+/// Gemini `generateContent` states its reasoning depth inside
+/// `generationConfig.thinkingConfig`, either as a symbolic `thinkingLevel` or as a token
+/// `thinkingBudget`. Both camelCase and snake_case spellings are read so that a captured client
+/// body and a converted provider body resolve to the same label.
+///
+/// `includeThoughts` alone is a visibility flag, not a depth, so it never produces a label.
+fn extract_gemini_reasoning_effort_from_body(
+    object: &serde_json::Map<String, Value>,
+) -> Option<String> {
+    let generation_config = object
+        .get("generationConfig")
+        .or_else(|| object.get("generation_config"))
+        .and_then(Value::as_object)?;
+    let thinking_config = generation_config
+        .get("thinkingConfig")
+        .or_else(|| generation_config.get("thinking_config"))
+        .and_then(Value::as_object)?;
+
+    if let Some(level) = thinking_config
+        .get("thinkingLevel")
+        .or_else(|| thinking_config.get("thinking_level"))
+        .and_then(Value::as_str)
+        .and_then(normalize_gemini_thinking_level)
+    {
+        return Some(level);
+    }
+
+    thinking_config
+        .get("thinkingBudget")
+        .or_else(|| thinking_config.get("thinking_budget"))
+        .and_then(Value::as_u64)
+        .map(|budget| {
+            // `0` disables reasoning outright. The shared budget ladder collapses it into `low`,
+            // which would report an explicitly disabled request as a shallow one.
+            if budget == 0 {
+                "none".to_string()
+            } else {
+                aether_ai_formats::formats::openai::shared::map_thinking_budget_to_openai_reasoning_effort(budget)
+                    .to_string()
+            }
+        })
+}
+
+/// Gemini also emits the protobuf enum spelling (`THINKING_LEVEL_HIGH`); the level itself is what
+/// the badge vocabulary understands, so the enum prefix is stripped before normalizing.
+///
+/// `THINKING_LEVEL_UNSPECIFIED` is the enum's "no explicit level" member, not a depth. It is
+/// rejected rather than surfaced, otherwise the badge would read `unspecified`.
+fn normalize_gemini_thinking_level(value: &str) -> Option<String> {
+    let normalized = value.trim().to_ascii_lowercase();
+    let normalized = normalized
+        .strip_prefix("thinking_level_")
+        .unwrap_or(normalized.as_str());
+    if normalized == "unspecified" {
+        return None;
+    }
+    normalize_provider_reasoning_effort(normalized)
 }

 fn normalize_provider_reasoning_effort(value: &str) -> Option<String> {
@@ -3588,6 +3657,157 @@ mod tests {
        assert_eq!(usage.provider_service_tier(), None);
    }

+    #[test]
+    fn gemini_thinking_level_supplies_reasoning_effort_for_both_body_spellings() {
+        let mut usage = sample_usage();
+        usage.provider_request_body = Some(json!({
+            "generationConfig": {
+                "thinkingConfig": { "includeThoughts": true, "thinkingLevel": "HIGH" }
+            }
+        }));
+
+        assert_eq!(usage.provider_reasoning_effort().as_deref(), Some("high"));
+
+        // The converted provider body keeps snake_case keys, and the client body may carry the
+        // protobuf enum spelling. Both must land on the same badge vocabulary.
+        usage.provider_request_body = Some(json!({
+            "generation_config": {
+                "thinking_config": { "thinking_level": "thinking_level_medium" }
+            }
+        }));
+
+        assert_eq!(usage.provider_reasoning_effort().as_deref(), Some("medium"));
+
+        usage.provider_request_body = Some(json!({
+            "generationConfig": {
+                "thinkingConfig": { "thinkingLevel": "  low  " }
+            }
+        }));
+
+        assert_eq!(usage.provider_reasoning_effort().as_deref(), Some("low"));
+    }
+
+    #[test]
+    fn gemini_thinking_budget_supplies_reasoning_effort_without_collapsing_zero() {
+        let mut usage = sample_usage();
+        usage.provider_request_body = Some(json!({
+            "generationConfig": { "thinkingConfig": { "thinkingBudget": 8192 } }
+        }));
+
+        assert_eq!(usage.provider_reasoning_effort().as_deref(), Some("xhigh"));
+
+        // `0` disables reasoning. The shared budget ladder maps 0..=1664 to `low`, which would
+        // report an explicitly disabled request as shallow, so the Gemini path reports `none`.
+        usage.provider_request_body = Some(json!({
+            "generationConfig": { "thinkingConfig": { "thinkingBudget": 0 } }
+        }));
+
+        assert_eq!(usage.provider_reasoning_effort().as_deref(), Some("none"));
+
+        usage.provider_request_body = Some(json!({
+            "generation_config": { "thinking_config": { "thinking_budget": 1280 } }
+        }));
+
+        assert_eq!(usage.provider_reasoning_effort().as_deref(), Some("low"));
+    }
+
+    #[test]
+    fn gemini_thinking_config_without_level_or_budget_yields_no_reasoning_effort() {
+        let mut usage = sample_usage();
+        usage.provider_request_body = Some(json!({
+            "generationConfig": { "thinkingConfig": { "includeThoughts": true } }
+        }));
+
+        assert_eq!(usage.provider_reasoning_effort(), None);
+
+        // A level-less, budget-less config must not fall back to metadata either: the captured
+        // body is authoritative and it says nothing about depth.
+        usage.request_metadata = Some(json!({ "provider_reasoning_effort": "max" }));
+        assert_eq!(usage.provider_reasoning_effort(), None);
+
+        // `THINKING_LEVEL_UNSPECIFIED` is the enum's "no explicit level" member, not a depth.
+        usage.request_metadata = None;
+        usage.provider_request_body = Some(json!({
+            "generationConfig": {
+                "thinkingConfig": { "thinkingLevel": "THINKING_LEVEL_UNSPECIFIED" }
+            }
+        }));
+
+        assert_eq!(usage.provider_reasoning_effort(), None);
+    }
+
+    /// The v1internal envelope nests the real `generateContent` payload under `request`. This is
+    /// the shape the Antigravity/Gemini CLI transports actually send upstream, so the extraction
+    /// has to descend into it or every converted `openai:chat -> gemini` request loses its badge.
+    #[test]
+    fn gemini_thinking_config_is_read_from_the_v1internal_envelope() {
+        let mut usage = sample_usage();
+        usage.provider_request_body = Some(json!({
+            "model": "gemini-3.8-flash-tiered",
+            "project": "aicode-consumers",
+            "requestId": "req-1",
+            "requestType": "agent",
+            "userAgent": "vscode/1.X.X (Antigravity/4.3.0)",
+            "request": {
+                "contents": [{ "role": "user", "parts": [{ "text": "hi" }] }],
+                "generationConfig": {
+                    "maxOutputTokens": 65536,
+                    "thinkingConfig": { "includeThoughts": true, "thinkingLevel": "high" }
+                }
+            }
+        }));
+
+        assert_eq!(usage.provider_reasoning_effort().as_deref(), Some("high"));
+
+        usage.provider_request_body = Some(json!({
+            "model": "gemini-3.8-flash-tiered",
+            "request": {
+                "generation_config": {
+                    "thinking_config": { "include_thoughts": true, "thinking_budget": 32768 }
+                }
+            }
+        }));
+
+        assert_eq!(usage.provider_reasoning_effort().as_deref(), Some("xhigh"));
+
+        usage.provider_request_body = Some(json!({
+            "model": "gemini-3.8-flash-tiered",
+            "request": {
+                "generationConfig": { "thinkingConfig": { "thinkingBudget": 0 } }
+            }
+        }));
+
+        assert_eq!(usage.provider_reasoning_effort().as_deref(), Some("none"));
+    }
+
+    /// A converted request that carries only `maxOutputTokens` must stay badge-less rather than
+    /// picking up a depth from somewhere else in the envelope.
+    #[test]
+    fn v1internal_envelope_without_thinking_config_yields_no_reasoning_effort() {
+        let mut usage = sample_usage();
+        usage.provider_request_body = Some(json!({
+            "model": "gemini-3.8-flash-tiered",
+            "project": "aicode-consumers",
+            "request": {
+                "contents": [{ "role": "user", "parts": [{ "text": "hi" }] }],
+                "generationConfig": { "maxOutputTokens": 65536 }
+            }
+        }));
+
+        assert_eq!(usage.provider_reasoning_effort(), None);
+    }
+
+    #[test]
+    fn gemini_thinking_config_does_not_shadow_explicit_effort_fields() {
+        let mut usage = sample_usage();
+        usage.provider_request_body = Some(json!({
+            "reasoning_effort": "max",
+            "generationConfig": { "thinkingConfig": { "thinkingLevel": "low" } }
+        }));
+
+        assert_eq!(usage.provider_reasoning_effort().as_deref(), Some("max"));
+    }
+
    #[test]
    fn requested_and_provider_reasoning_efforts_remain_independent() {
        let mut usage = sample_usage();