fix: anchor compaction to provider input usage

server/src/model/llm_call.rs

@@ -1,6 +1,6 @@
 use serde::Serialize;
 
-use super::ProviderType;
+use super::{ProviderType, Usage};
 
 #[derive(Clone, Debug)]
 pub struct NewLlmCall {
@@ -22,6 +22,14 @@ pub struct NewLlmCall {
     pub detailed: bool,
 }
 
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(crate) struct LlmCallUsageAnchor {
+    pub request_type: ProviderType,
+    pub usage: Usage,
+    pub message_count: usize,
+    pub tool_count: usize,
+}
+
 #[derive(Clone, Debug, Serialize)]
 pub struct LlmCallSummary {
     pub call_id: String,

server/src/model/usage.rs

@@ -2,6 +2,8 @@ use std::ops::AddAssign;
 
 use serde::{Deserialize, Serialize};
 
+use super::ProviderType;
+
 #[derive(Clone, Copy, Debug, Default, Serialize, Deserialize, PartialEq, Eq)]
 pub struct Usage {
     pub input_tokens: Option<u64>,
@@ -12,6 +14,19 @@ pub struct Usage {
     pub reasoning_tokens: Option<u64>,
 }
 
+impl Usage {
+    /// Returns the provider-visible input context without counting cached tokens twice.
+    pub(crate) fn context_input_tokens(self, provider: ProviderType) -> Option<u64> {
+        let input = self.input_tokens?;
+        match provider {
+            ProviderType::OpenAiChat | ProviderType::OpenAiResponses => Some(input),
+            ProviderType::Anthropic => input
+                .checked_add(self.cache_read_tokens.unwrap_or_default())?
+                .checked_add(self.cache_write_tokens.unwrap_or_default()),
+        }
+    }
+}
+
 impl AddAssign for Usage {
     fn add_assign(&mut self, rhs: Self) {
         self.input_tokens = sum(self.input_tokens, rhs.input_tokens);
@@ -30,6 +45,41 @@ fn sum(left: Option<u64>, right: Option<u64>) -> Option<u64> {
 #[cfg(test)]
 mod tests {
     use super::Usage;
+    use crate::model::ProviderType;
+
+    #[test]
+    fn openai_context_input_does_not_double_count_cached_tokens() {
+        let usage = Usage {
+            input_tokens: Some(140_649),
+            cache_read_tokens: Some(120_000),
+            cache_write_tokens: Some(10_000),
+            ..Usage::default()
+        };
+
+        assert_eq!(
+            usage.context_input_tokens(ProviderType::OpenAiResponses),
+            Some(140_649)
+        );
+        assert_eq!(
+            usage.context_input_tokens(ProviderType::OpenAiChat),
+            Some(140_649)
+        );
+    }
+
+    #[test]
+    fn anthropic_context_input_includes_disjoint_cache_tokens() {
+        let usage = Usage {
+            input_tokens: Some(10_649),
+            cache_read_tokens: Some(120_000),
+            cache_write_tokens: Some(10_000),
+            ..Usage::default()
+        };
+
+        assert_eq!(
+            usage.context_input_tokens(ProviderType::Anthropic),
+            Some(140_649)
+        );
+    }
 
     #[test]
     fn turn_total_only_reports_fields_known_for_every_cycle() {