From f462d733474e4192c19eba8ca0024ae320a850ca Mon Sep 17 00:00:00 2001
From: Douwe Osinga <douwe@block.xyz>
Date: Wed, 11 Mar 2026 15:13:11 -0400
Subject: [PATCH] fix: return ContextLengthExceeded when prompt exceeds
 effective KV cache size (#7815)

Co-authored-by: Douwe Osinga <douwe@squareup.com>
---
 .../src/providers/local_inference/inference_engine.rs      | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/crates/goose/src/providers/local_inference/inference_engine.rs b/crates/goose/src/providers/local_inference/inference_engine.rs
index 06256fca90..90940f357e 100644
--- a/crates/goose/src/providers/local_inference/inference_engine.rs
+++ b/crates/goose/src/providers/local_inference/inference_engine.rs
@@ -226,6 +226,13 @@ pub(super) fn validate_and_compute_context(
             )));
         }
     }
+    if prompt_token_count >= effective_ctx {
+        return Err(ProviderError::ContextLengthExceeded(format!(
+            "Prompt ({} tokens) exceeds context limit ({} tokens). \
+             Try reducing conversation length.",
+            prompt_token_count, effective_ctx,
+        )));
+    }
     Ok((prompt_token_count, effective_ctx))
 }