From 92288eded441c6ba99362134d9f1e2f68a95ace8 Mon Sep 17 00:00:00 2001
From: Bryan Helmkamp <bryan@brynary.com>
Date: Sun, 24 May 2026 18:21:22 -0400
Subject: [PATCH] Use Codex-safe OpenAI context windows

Align Fabro's built-in OpenAI GPT-5 catalog defaults with Codex's 272k context window so Codex-backed runs compact before provider rejection. Direct API users can still raise the window through catalog overrides.
---
 lib/crates/fabro-model/src/catalog.rs         | 36 ++++++++++++++++++-
 .../src/catalog/providers/openai.toml         | 10 +++---
 2 files changed, 40 insertions(+), 6 deletions(-)

diff --git a/lib/crates/fabro-model/src/catalog.rs b/lib/crates/fabro-model/src/catalog.rs
index b1c931651..32a907ddf 100644
--- a/lib/crates/fabro-model/src/catalog.rs
+++ b/lib/crates/fabro-model/src/catalog.rs
@@ -3883,7 +3883,7 @@ reasoning_effort = "levels"
             family: "gpt-5",
             display_name: "GPT-5.4",
             limits: ModelLimits {
-                context_window: 1047576,
+                context_window: 272000,
                 max_output: Some(
                     128000,
                 ),
@@ -4059,6 +4059,40 @@ reasoning_effort = "levels"
         );
     }
 
+    #[test]
+    fn openai_codex_default_context_windows_match_codex_catalog() {
+        let catalog = Catalog::builtin();
+
+        for model in [
+            "gpt-5.2",
+            "gpt-5.3-codex",
+            "gpt-5.4",
+            "gpt-5.4-mini",
+            "gpt-5.5",
+        ] {
+            assert_eq!(
+                catalog.get(model).unwrap().context_window(),
+                272_000,
+                "{model} should use the Codex-safe default context window"
+            );
+        }
+    }
+
+    #[test]
+    fn openai_context_window_can_be_overridden_for_direct_api_usage() {
+        let catalog = Catalog::from_builtin_with_overrides(&minimal_settings(
+            r#"
+[models."gpt-5.5".limits]
+context_window = 1050000
+"#,
+        ))
+        .expect("sparse built-in model limit override should build");
+
+        let model = catalog.get("gpt-5.5").unwrap();
+        assert_eq!(model.context_window(), 1_050_000);
+        assert_eq!(model.max_output(), Some(128_000));
+    }
+
     // ---- Closest model tests ----
 
     #[test]
diff --git a/lib/crates/fabro-model/src/catalog/providers/openai.toml b/lib/crates/fabro-model/src/catalog/providers/openai.toml
index d9530972f..5ed6bed64 100644
--- a/lib/crates/fabro-model/src/catalog/providers/openai.toml
+++ b/lib/crates/fabro-model/src/catalog/providers/openai.toml
@@ -19,7 +19,7 @@ estimated_output_tps = 65
 aliases = ["gpt5"]
 
 [models."gpt-5.2".limits]
-context_window = 1047576
+context_window = 272000
 max_output = 128000
 
 [models."gpt-5.2".features]
@@ -44,7 +44,7 @@ estimated_output_tps = 100
 aliases = ["codex"]
 
 [models."gpt-5.3-codex".limits]
-context_window = 1047576
+context_window = 272000
 max_output = 128000
 
 [models."gpt-5.3-codex".features]
@@ -90,7 +90,7 @@ estimated_output_tps = 70
 aliases = ["gpt54", "gpt-54"]
 
 [models."gpt-5.4".limits]
-context_window = 1047576
+context_window = 272000
 max_output = 128000
 
 [models."gpt-5.4".features]
@@ -115,7 +115,7 @@ estimated_output_tps = 70
 aliases = ["gpt55", "gpt-55"]
 
 [models."gpt-5.5".limits]
-context_window = 1050000
+context_window = 272000
 max_output = 128000
 
 [models."gpt-5.5".features]
@@ -192,7 +192,7 @@ probe = true
 small_default = true
 
 [models."gpt-5.4-mini".limits]
-context_window = 400000
+context_window = 272000
 max_output = 128000
 
 [models."gpt-5.4-mini".features]