feat: calculate context usage from history and add a final pre-request safety check after trimming.

Author: vinhnx

src/agent/runloop/unified/context_manager.rs

@@ -366,7 +366,18 @@ impl ContextManager {
         }
 
         let mut pruning_ledger = pruning_ledger;
-        let usage = self.token_budget.usage_ratio().await;
+        
+        // Use live calculation from history to ensure consistency with pre_request_check
+        // stored token_budget might be lagging if history was just modified
+        let estimated = self.estimate_request_tokens(history);
+        let max_tokens = self.trim_config.max_tokens;
+        
+        let usage = if max_tokens > 0 {
+             estimated as f64 / max_tokens as f64
+        } else {
+             0.0
+        };
+
         let mut outcome = ContextTrimOutcome::default();
 
         // Alert/compact: semantic prune first, then enforce window

src/agent/runloop/unified/turn/guards.rs

@@ -118,6 +118,12 @@ pub(crate) async fn run_proactive_guards(
         }
     }
 
+    // Final safety check: ensure we didn't exit the loop (e.g. max checks) while still in a dangerous state
+    let final_check = ctx.context_manager.pre_request_check(ctx.working_history);
+    if matches!(final_check, PreRequestAction::Block) {
+        anyhow::bail!("Context budget critical/overflow after trimming attempts. Conversation unsafe to proceed.");
+    }
+
     Ok(())
 }