Skip to content

YC (S26) | Open Computer History | Continuously record your company computer work, map your workflows, help you find work worth automating, and power your agents' context

[codex] fix Claude large write output limits (#5873)

Louis Beaumontb3fe7591918d
diff --git a/crates/screenpipe-core/src/agents/pi.rs b/crates/screenpipe-core/src/agents/pi.rs
index b211cc690..ddc827555 100644
--- a/crates/screenpipe-core/src/agents/pi.rs
+++ b/crates/screenpipe-core/src/agents/pi.rs
@@ -21,6 +21,7 @@ pub const PI_AI_PACKAGE: &str = "@earendil-works/[email protected]";
pub const PI_NAMESPACE_DIR: &str = "@earendil-works";
pub const SCREENPIPE_API_URL: &str = "https://api.screenpipe.com/v1";
const CUSTOM_PROVIDER_USER_AGENT: &str = "screenpipe";
+const DEFAULT_CLOUD_MAX_OUTPUT_TOKENS: u64 = 32_000;
/// Apply compatibility settings required by OpenAI-compatible custom endpoints.
///
@@ -263,7 +264,14 @@ fn gateway_models_to_pi_models(data: &[serde_json::Value]) -> Vec<serde_json::Va
let ctx = m
.get("context_window")
.and_then(|v| v.as_u64())
+ .filter(|value| *value > 0)
.unwrap_or(128000);
+ let max_tokens = m
+ .get("max_output_tokens")
+ .and_then(|v| v.as_u64())
+ .filter(|value| *value > 0)
+ .unwrap_or(DEFAULT_CLOUD_MAX_OUTPUT_TOKENS)
+ .min(ctx);
let intelligence = m
.get("intelligence")
.and_then(|v| v.as_str())
@@ -277,7 +285,7 @@ fn gateway_models_to_pi_models(data: &[serde_json::Value]) -> Vec<serde_json::Va
"input": ["text", "image"],
"cost": {"input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0},
"contextWindow": ctx,
- "maxTokens": 32000,
+ "maxTokens": max_tokens,
// Pi sends its stable agent session ID as x-session-affinity.
// The hosted gateway uses that plus the user-message ordinal to
// count one visible turn once across a multi-call tool loop.
@@ -299,7 +307,7 @@ fn selectable_gateway_models(data: &[serde_json::Value]) -> Option<Vec<serde_jso
/// Only auto — if the gateway is down, nothing works anyway.
fn fallback_cloud_models() -> serde_json::Value {
json!([
- {"id": "auto", "name": "Auto (recommended)", "reasoning": true, "input": ["text", "image"], "cost": {"input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0}, "contextWindow": 128000, "maxTokens": 32000, "compat": {"sendSessionAffinityHeaders": true}},
+ {"id": "auto", "name": "Auto (recommended)", "reasoning": true, "input": ["text", "image"], "cost": {"input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0}, "contextWindow": 128000, "maxTokens": DEFAULT_CLOUD_MAX_OUTPUT_TOKENS, "compat": {"sendSessionAffinityHeaders": true}},
])
}
@@ -4562,6 +4570,70 @@ mod tests {
.is_none());
}
+ #[test]
+ fn gateway_catalog_uses_advertised_output_budget_with_safe_fallback() {
+ let models = gateway_models_to_pi_models(&[
+ json!({
+ "id": "claude-sonnet-5",
+ "context_window": 1_000_000,
+ "max_output_tokens": 128_000,
+ }),
+ json!({
+ "id": "legacy-model-without-output-metadata",
+ "context_window": 128_000,
+ }),
+ json!({
+ "id": "invalid-model-budget",
+ "context_window": 64_000,
+ "max_output_tokens": 0,
+ }),
+ json!({
+ "id": "oversized-model-budget",
+ "context_window": 64_000,
+ "max_output_tokens": 128_000,
+ }),
+ ]);
+
+ assert_eq!(models[0].get("maxTokens"), Some(&json!(128_000)));
+ assert_eq!(models[1].get("maxTokens"), Some(&json!(32_000)));
+ assert_eq!(models[2].get("maxTokens"), Some(&json!(32_000)));
+ assert_eq!(models[3].get("maxTokens"), Some(&json!(64_000)));
+ }
+
+ #[tokio::test]
+ async fn gateway_output_budget_flows_from_http_to_pi_catalog() {
+ use wiremock::{
+ matchers::{method, path},
+ Mock, MockServer, ResponseTemplate,
+ };
+
+ let server = MockServer::start().await;
+ Mock::given(method("GET"))
+ .and(path("/models"))
+ .respond_with(ResponseTemplate::new(200).set_body_json(json!({
+ "data": [
+ {
+ "id": "claude-sonnet-5",
+ "name": "Claude Sonnet 5",
+ "context_window": 1_000_000,
+ "max_output_tokens": 128_000,
+ "intelligence": "highest",
+ },
+ {
+ "id": "legacy-model",
+ "name": "Legacy model",
+ "context_window": 128_000,
+ },
+ ]
+ })))
+ .mount(&server)
+ .await;
+
+ let models = screenpipe_cloud_models(&server.uri(), None).await;
+ assert_eq!(models.pointer("/0/maxTokens"), Some(&json!(128_000)));
+ assert_eq!(models.pointer("/1/maxTokens"), Some(&json!(32_000)));
+ }
+
#[tokio::test]
async fn all_locked_gateway_catalog_uses_auto_fallback() {
use wiremock::{
diff --git a/packages/ai-gateway/src/handlers/models.ts b/packages/ai-gateway/src/handlers/models.ts
index b4a1d1ecf..979a20927 100644
--- a/packages/ai-gateway/src/handlers/models.ts
+++ b/packages/ai-gateway/src/handlers/models.ts
@@ -20,6 +20,8 @@ interface ModelEntry {
tags: string[];
free: boolean;
context_window: number;
+ /** Maximum completion budget clients should request from this model. */
+ max_output_tokens?: number;
best_for: string[];
speed: 'fast' | 'medium' | 'slow';
intelligence: 'standard' | 'high' | 'highest';
@@ -244,6 +246,7 @@ const CURATED_MODELS: ModelEntry[] = [
tags: ['premium', 'balanced', 'agentic', 'new'],
free: false,
context_window: 1000000,
+ max_output_tokens: 128000,
best_for: ['agentic work', 'coding', 'computer use', 'complex analysis'],
speed: 'medium',
intelligence: 'highest',
@@ -260,6 +263,7 @@ const CURATED_MODELS: ModelEntry[] = [
tags: ['business', 'frontier', 'agentic', 'new'],
free: false,
context_window: 1000000,
+ max_output_tokens: 128000,
best_for: ['hard reasoning', 'agentic coding', 'complex analysis'],
speed: 'slow',
intelligence: 'highest',
@@ -276,6 +280,7 @@ const CURATED_MODELS: ModelEntry[] = [
tags: ['business', 'frontier', 'agentic', 'new'],
free: false,
context_window: 1000000,
+ max_output_tokens: 128000,
best_for: ['hard reasoning', 'agentic coding', 'complex analysis'],
speed: 'slow',
intelligence: 'highest',
diff --git a/packages/ai-gateway/src/test/openai-models.unit.test.ts b/packages/ai-gateway/src/test/openai-models.unit.test.ts
index 9827fa220..ac742b4e0 100644
--- a/packages/ai-gateway/src/test/openai-models.unit.test.ts
+++ b/packages/ai-gateway/src/test/openai-models.unit.test.ts
@@ -113,6 +113,13 @@ describe('OpenAI API model catalog', () => {
expect(ids).toContain('claude-fable-5');
});
+ it('publishes the full Claude 5 output budget for agent tool calls', async () => {
+ const models = await listedModels();
+ for (const id of ['claude-sonnet-5', 'claude-opus-5', 'claude-fable-5']) {
+ expect(models.find(model => model.id === id)?.max_output_tokens).toBe(128_000);
+ }
+ });
+
it('does not expose provider-secret gates in /v1/models responses', async () => {
const models = await listedModels();
const openAiModel = models.find(model => model.id === 'gpt-5.5');