fix(agent-core): parse provider cache token usage into accounting

src-tauri/crates/agent-core/src/core/providers/openai_compat/client.rs

@@ -249,6 +249,14 @@ impl OpenAICompatClient {
                     api_usage.completion_tokens,
                 );
                 usage.insert(usage_key::TOTAL_TOKENS.to_string(), api_usage.total_tokens);
+                if let Some(ref details) = api_usage.prompt_tokens_details {
+                    if details.cached_tokens > 0 {
+                        usage.insert(
+                            usage_key::CACHE_READ_TOKENS.to_string(),
+                            details.cached_tokens,
+                        );
+                    }
+                }
             }
         }
 

src-tauri/crates/agent-core/src/core/providers/openai_compat/streaming/chat.rs

@@ -178,6 +178,11 @@ pub(super) async fn run_chat(
         usage.insert("prompt_tokens".to_string(), api_usage.prompt_tokens);
         usage.insert("completion_tokens".to_string(), api_usage.completion_tokens);
         usage.insert("total_tokens".to_string(), api_usage.total_tokens);
+        if let Some(ref details) = api_usage.prompt_tokens_details {
+            if details.cached_tokens > 0 {
+                usage.insert("cache_read_tokens".to_string(), details.cached_tokens);
+            }
+        }
     }
 
     let tool_calls = choice

src-tauri/crates/agent-core/src/core/providers/openai_compat/streaming/sse_stream.rs

@@ -521,6 +521,16 @@ pub(super) async fn run_chat_streaming(
                 final_usage.insert("prompt_tokens".to_string(), usage.prompt_tokens);
                 final_usage.insert("completion_tokens".to_string(), usage.completion_tokens);
                 final_usage.insert("total_tokens".to_string(), usage.total_tokens);
+                if let Some(ref details) = usage.prompt_tokens_details {
+                    if details.cached_tokens > 0 {
+                        debug!(
+                            "[streaming-usage] OpenAI cached_tokens={}",
+                            details.cached_tokens
+                        );
+                        final_usage
+                            .insert("cache_read_tokens".to_string(), details.cached_tokens);
+                    }
+                }
             }
         }
         if stream_done {

src-tauri/crates/agent-core/src/core/providers/openai_compat/types.rs

@@ -172,6 +172,17 @@ pub(super) struct Usage {
     pub completion_tokens: i64,
     #[serde(default)]
     pub total_tokens: i64,
+    /// OpenAI prompt caching: `prompt_tokens_details.cached_tokens` reports
+    /// how many prompt tokens were served from the provider prompt cache.
+    #[serde(default)]
+    pub prompt_tokens_details: Option<PromptTokensDetails>,
+}
+
+/// `usage.prompt_tokens_details` from OpenAI-compatible responses.
+#[derive(Debug, Deserialize)]
+pub(super) struct PromptTokensDetails {
+    #[serde(default)]
+    pub cached_tokens: i64,
 }
 
 /// Error response from the API.

src-tauri/crates/agent-core/src/core/providers/responses_common/parser.rs

@@ -132,6 +132,14 @@ pub fn parse_response(resp: ResponsesResponse) -> Result<LLMResponse, ProviderEr
         if let Some(total) = u.total_tokens {
             usage.insert(usage_key::TOTAL_TOKENS.to_string(), total);
         }
+        if let Some(ref details) = u.input_tokens_details {
+            if details.cached_tokens > 0 {
+                usage.insert(
+                    usage_key::CACHE_READ_TOKENS.to_string(),
+                    details.cached_tokens,
+                );
+            }
+        }
     }
 
     Ok(LLMResponse {
@@ -164,6 +172,7 @@ mod tests {
             })],
             usage: Some(ResponsesUsage {
                 input_tokens: Some(10),
+                input_tokens_details: None,
                 output_tokens: Some(5),
                 total_tokens: Some(15),
             }),

src-tauri/crates/agent-core/src/core/providers/responses_common/types.rs

@@ -114,6 +114,16 @@ pub struct ResponsesUsage {
     pub input_tokens: Option<i64>,
     pub output_tokens: Option<i64>,
     pub total_tokens: Option<i64>,
+    /// `usage.input_tokens_details.cached_tokens` — prompt-cache hits.
+    #[serde(default)]
+    pub input_tokens_details: Option<ResponsesInputTokensDetails>,
+}
+
+/// `usage.input_tokens_details` from the Responses API.
+#[derive(Debug, Deserialize)]
+pub struct ResponsesInputTokensDetails {
+    #[serde(default)]
+    pub cached_tokens: i64,
 }
 
 /// Error from the Responses API.