feat: capture thinking text via MITM dual-call merge

The LS makes TWO separate Google API calls for thinking models:
  Call 1: response + thinking token count (no thinking text)
  Call 2: thinking summary text (no thinking tokens)

Each hits a different StreamingAccumulator, so we:
1. Capture response_text in StreamingAccumulator (non-thinking parts)
2. In MitmStore::record_usage, detect when Call 2 arrives for a
   cascade that already has thinking tokens from Call 1
3. Merge Call 2's response_text as thinking_text on Call 1's usage

Also injects includeThoughts into Google API requests via MITM
modify to ensure thinking text is available in SSE responses.
This commit is contained in:
Nikketryhard
2026-02-14 19:49:15 -06:00
parent 905d55beb5
commit 34b9553484
4 changed files with 92 additions and 3 deletions

View File

@@ -59,6 +59,9 @@ pub struct StreamingAccumulator {
pub thinking_tokens: u64, pub thinking_tokens: u64,
/// Accumulated thinking/reasoning text from the model. /// Accumulated thinking/reasoning text from the model.
pub thinking_text: String, pub thinking_text: String,
/// Accumulated response text (non-thinking parts).
/// Used to identify "thinking summary" calls in the v1internal API.
pub response_text: String,
pub model: Option<String>, pub model: Option<String>,
pub stop_reason: Option<String>, pub stop_reason: Option<String>,
pub is_complete: bool, pub is_complete: bool,
@@ -83,16 +86,24 @@ impl StreamingAccumulator {
if let Some(model) = response["modelVersion"].as_str() { if let Some(model) = response["modelVersion"].as_str() {
self.model = Some(model.to_string()); self.model = Some(model.to_string());
} }
// Extract thinking text from parts with thought: true
if let Some(candidates) = response.get("candidates").and_then(|c| c.as_array()) { if let Some(candidates) = response.get("candidates").and_then(|c| c.as_array()) {
for candidate in candidates { for candidate in candidates {
if let Some(parts) = candidate["content"]["parts"].as_array() { if let Some(parts) = candidate["content"]["parts"].as_array() {
for part in parts { for part in parts {
// Public Gemini API: explicit thought flag
if part["thought"].as_bool() == Some(true) { if part["thought"].as_bool() == Some(true) {
if let Some(text) = part["text"].as_str() { if let Some(text) = part["text"].as_str() {
self.thinking_text.push_str(text); self.thinking_text.push_str(text);
} }
} }
// Capture non-thinking response text (skip thoughtSignature parts)
else if part.get("thoughtSignature").is_none() {
if let Some(text) = part["text"].as_str() {
if !text.is_empty() {
self.response_text.push_str(text);
}
}
}
} }
} }
// Check for completion // Check for completion
@@ -172,6 +183,11 @@ impl StreamingAccumulator {
} else { } else {
Some(self.thinking_text) Some(self.thinking_text)
}; };
let response_text = if self.response_text.is_empty() {
None
} else {
Some(self.response_text)
};
ApiUsage { ApiUsage {
input_tokens: self.input_tokens, input_tokens: self.input_tokens,
output_tokens: self.output_tokens, output_tokens: self.output_tokens,
@@ -179,6 +195,7 @@ impl StreamingAccumulator {
cache_read_input_tokens: self.cache_read_input_tokens, cache_read_input_tokens: self.cache_read_input_tokens,
thinking_output_tokens: self.thinking_tokens, thinking_output_tokens: self.thinking_tokens,
thinking_text, thinking_text,
response_text,
response_output_tokens: 0, response_output_tokens: 0,
model: self.model, model: self.model,
stop_reason: self.stop_reason, stop_reason: self.stop_reason,
@@ -203,6 +220,7 @@ fn extract_usage_from_message(msg: &Value) -> Option<ApiUsage> {
cache_read_input_tokens: usage["cache_read_input_tokens"].as_u64().unwrap_or(0), cache_read_input_tokens: usage["cache_read_input_tokens"].as_u64().unwrap_or(0),
thinking_output_tokens: 0, thinking_output_tokens: 0,
thinking_text: None, thinking_text: None,
response_text: None,
response_output_tokens: 0, response_output_tokens: 0,
model: msg["model"].as_str().map(|s| s.to_string()), model: msg["model"].as_str().map(|s| s.to_string()),
stop_reason: msg["stop_reason"].as_str().map(|s| s.to_string()), stop_reason: msg["stop_reason"].as_str().map(|s| s.to_string()),

View File

@@ -152,6 +152,47 @@ pub fn modify_request(body: &[u8]) -> Option<Vec<u8>> {
} }
} }
// ── 4. Inject includeThoughts to capture thinking text ───────────────
// Without this flag, Google only reports thinking token counts
// but doesn't send the thinking text in SSE parts.
{
// Ensure request.generationConfig.thinkingConfig.includeThoughts = true
let request = json.get_mut("request").and_then(|v| v.as_object_mut());
if let Some(req) = request {
let gen_config = req
.entry("generationConfig")
.or_insert_with(|| serde_json::json!({}));
if let Some(gc) = gen_config.as_object_mut() {
let thinking_config = gc
.entry("thinkingConfig")
.or_insert_with(|| serde_json::json!({}));
if let Some(tc) = thinking_config.as_object_mut() {
if !tc.contains_key("includeThoughts") {
tc.insert("includeThoughts".to_string(), Value::Bool(true));
changes.push("inject includeThoughts".to_string());
}
}
}
} else {
// Not wrapped in request — try top-level (public API format)
let gen_config = json.as_object_mut().and_then(|o| {
Some(o.entry("generationConfig")
.or_insert_with(|| serde_json::json!({})))
});
if let Some(gc) = gen_config.and_then(|v| v.as_object_mut()) {
let thinking_config = gc
.entry("thinkingConfig")
.or_insert_with(|| serde_json::json!({}));
if let Some(tc) = thinking_config.as_object_mut() {
if !tc.contains_key("includeThoughts") {
tc.insert("includeThoughts".to_string(), Value::Bool(true));
changes.push("inject includeThoughts (top-level)".to_string());
}
}
}
}
}
if changes.is_empty() { if changes.is_empty() {
return None; // Nothing modified return None; // Nothing modified
} }

View File

@@ -80,6 +80,7 @@ impl GrpcUsage {
output_tokens: self.output_tokens, output_tokens: self.output_tokens,
thinking_output_tokens: self.thinking_output_tokens, thinking_output_tokens: self.thinking_output_tokens,
thinking_text: None, // gRPC proto doesn't carry thinking text thinking_text: None, // gRPC proto doesn't carry thinking text
response_text: None,
response_output_tokens: self.response_output_tokens, response_output_tokens: self.response_output_tokens,
cache_creation_input_tokens: self.cache_write_tokens, cache_creation_input_tokens: self.cache_write_tokens,
cache_read_input_tokens: self.cache_read_tokens, cache_read_input_tokens: self.cache_read_tokens,

View File

@@ -26,6 +26,9 @@ pub struct ApiUsage {
/// Captured from Google SSE parts with `thought: true` or Anthropic thinking blocks. /// Captured from Google SSE parts with `thought: true` or Anthropic thinking blocks.
#[serde(skip_serializing_if = "Option::is_none")] #[serde(skip_serializing_if = "Option::is_none")]
pub thinking_text: Option<String>, pub thinking_text: Option<String>,
/// The response text captured from SSE parts (for merge detection).
#[serde(skip)]
pub response_text: Option<String>,
/// Google-specific: response output tokens (non-thinking portion) /// Google-specific: response output tokens (non-thinking portion)
pub response_output_tokens: u64, pub response_output_tokens: u64,
@@ -122,10 +125,36 @@ impl MitmStore {
} }
} }
// Store latest usage for the cascade (if we can identify it) // Store latest usage for the cascade (if we can identify it).
//
// Merge logic for v1internal thinking summaries:
// The LS makes TWO Google API calls per thinking request:
// Call 1: response + thinking token count (thinking_output_tokens > 0, no thinking text)
// Call 2: thinking summary text (thinking_output_tokens == 0, response_text has the summary)
//
// When Call 2 arrives, we merge its response_text as thinking_text into Call 1's usage.
let key = cascade_id.map(|s| s.to_string()).unwrap_or_else(|| "_latest".to_string()); let key = cascade_id.map(|s| s.to_string()).unwrap_or_else(|| "_latest".to_string());
let mut latest = self.latest_usage.write().await; let mut latest = self.latest_usage.write().await;
latest.insert(key, usage);
if let Some(existing) = latest.get_mut(&key) {
if existing.thinking_output_tokens > 0
&& existing.thinking_text.is_none()
&& usage.thinking_output_tokens == 0
&& usage.response_text.is_some()
{
// Call 2: thinking summary — merge into existing Call 1 usage
existing.thinking_text = usage.response_text;
debug!(
thinking_text_len = existing.thinking_text.as_ref().map_or(0, |t| t.len()),
"MITM: merged thinking summary text into existing usage"
);
} else {
// Normal case: replace existing usage
latest.insert(key, usage);
}
} else {
latest.insert(key, usage);
}
// Evict old entries to prevent unbounded memory growth // Evict old entries to prevent unbounded memory growth
const MAX_ENTRIES: usize = 500; const MAX_ENTRIES: usize = 500;