trim_entries: take ContextBudget instead of recomputing

compact() already computes context_budget() — pass it to trim_entries
so it has access to all budget components without recomputing them.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>

src/agent/context.rs

@@ -45,10 +45,11 @@ fn context_budget_tokens() -> usize {
 /// 2. Trim: drop oldest entries until the conversation fits, snapping
 ///    to user message boundaries.
 pub fn trim_entries(
-    context: &ContextState,
     entries: &[ConversationEntry],
     tokenizer: &CoreBPE,
+    budget: &ContextBudget,
 ) -> Vec<ConversationEntry> {
+    let fixed_tokens = budget.system + budget.identity + budget.journal;
     // --- Phase 1: dedup memory entries by key (keep last) ---
     let mut seen_keys: std::collections::HashMap<&str, usize> = std::collections::HashMap::new();
     let mut drop_indices: std::collections::HashSet<usize> = std::collections::HashSet::new();
@@ -67,19 +68,13 @@ pub fn trim_entries(
         .collect();
 
     // --- Phase 2: trim to fit context budget ---
-    // Everything in the context window is a message. Count them all,
-    // trim entries until the total fits.
     let max_tokens = context_budget_tokens();
     let count_msg = |m: &Message| msg_token_count(tokenizer, m);
 
-    let fixed_cost = count_msg(&Message::system(&context.system_prompt))
-        + count_msg(&Message::user(context.render_context_message()))
-        + count_msg(&Message::user(render_journal(&context.journal)));
-
     let msg_costs: Vec<usize> = deduped.iter()
         .map(|e| if e.is_log() { 0 } else { count_msg(e.api_message()) }).collect();
     let entry_total: usize = msg_costs.iter().sum();
-    let total: usize = fixed_cost + entry_total;
+    let total: usize = fixed_tokens + entry_total;
 
     let mem_tokens: usize = deduped.iter().zip(&msg_costs)
         .filter(|(e, _)| e.is_memory())
@@ -87,7 +82,7 @@ pub fn trim_entries(
     let conv_tokens: usize = entry_total - mem_tokens;
 
     dbglog!("[trim] max_tokens={} fixed={} mem={} conv={} total={} entries={}",
-        max_tokens, fixed_cost, mem_tokens, conv_tokens, total, deduped.len());
+        max_tokens, fixed_tokens, mem_tokens, conv_tokens, total, deduped.len());
 
     // Phase 2a: evict all DMN entries first — they're ephemeral
     let mut drop = vec![false; deduped.len()];

src/agent/mod.rs

@@ -987,11 +987,12 @@ impl Agent {
         self.load_startup_journal();
 
         // Dedup memory, trim to budget
+        let budget = self.context_budget();
         let entries = self.context.entries.clone();
         self.context.entries = crate::agent::context::trim_entries(
-            &self.context,
             &entries,
             &self.tokenizer,
+            &budget,
         );
 
         let after = self.context.entries.len();