fix(llm, cli): Stop OpenAI silently shortening summary requests

crates/jp_cli/src/cmd/conversation/summarize.rs

@@ -15,6 +15,7 @@ use jp_llm::{
     event_builder::EventBuilder,
     model::ModelDetails,
     provider,
+    query::Truncation,
     retry::{RetryConfig, collect_with_retry},
     window,
 };
@@ -184,6 +185,10 @@ async fn summarize_stream(
             },
             tools: vec![],
             tool_choice: jp_config::assistant::tool_choice::ToolChoice::default(),
+            // The summary is stored as standing for every turn in the range,
+            // so a provider that quietly dropped part of the request would
+            // hand back a summary covering less than it claims.
+            truncation: Truncation::Forbidden,
         };
 
         let llm_events = collect_with_retry(provider, model_details, query, &retry_config).await?;
@@ -241,6 +246,12 @@ enum StreamOutcome {
 /// truncated or declined response would otherwise be stored as the summary and
 /// replace the turns it was meant to stand in for, silently dropping whatever
 /// the model never got to.
+///
+/// This covers a response the model cut short.
+/// A request the *provider* cut short arrives here as an ordinary
+/// [`FinishReason::Completed`] and is indistinguishable from a complete one, so
+/// that half is prevented when the request is built, by asking for
+/// [`Truncation::Forbidden`].
 fn summarize_events(events: Vec<Event>) -> StreamOutcome {
     let mut builder = EventBuilder::new();
     let mut flushed = Vec::new();

crates/jp_cli/src/cmd/query/tool/inquiry.rs

@@ -29,7 +29,7 @@ use jp_llm::{
     Provider,
     event_builder::structured_data,
     model::ModelDetails,
-    query::ChatQuery,
+    query::{ChatQuery, Truncation},
     retry::{RetryConfig, collect_with_retry},
     tool::ToolDefinition,
     window,
@@ -310,6 +310,7 @@ impl InquiryBackend for LlmInquiryBackend {
             thread,
             tools: self.tools.clone(),
             tool_choice: ToolChoice::None,
+            truncation: Truncation::Allowed,
         };
 
         let retry_config =

crates/jp_cli/src/cmd/query/turn_loop.rs

@@ -35,7 +35,7 @@ use jp_llm::{
     event::{Event, EventPart, FinishReason, ToolCallPart},
     model::ModelDetails,
     provider::get_provider,
-    query::ChatQuery,
+    query::{ChatQuery, Truncation},
     tool::{InvocationContext, ToolDefinition, executor::Executor},
     with_idle_timeout, with_output_limit,
 };
@@ -310,6 +310,10 @@ pub(super) async fn run_turn_loop(
                     thread,
                     tools: tools.to_vec(),
                     tool_choice: tool_choice.clone(),
+                    // Nothing fits the stream to the window on this path, so a
+                    // conversation that outgrows it stays answerable only if
+                    // the provider is allowed to drop what it cannot hold.
+                    truncation: Truncation::Allowed,
                 };
 
                 // Claim the waiting region BEFORE the HTTP request. Dropping

crates/jp_llm/src/provider/anthropic.rs

@@ -1213,6 +1213,7 @@ fn create_request(
         thread,
         tools,
         mut tool_choice,
+        ..
     } = query;
 
     let mut builder = types::CreateMessagesRequestBuilder::default();

crates/jp_llm/src/provider/anthropic_tests.rs

@@ -15,7 +15,10 @@ use serde_json::Map;
 use test_log::test;
 
 use super::*;
-use crate::test::{TestRequest, run_test};
+use crate::{
+    query::Truncation,
+    test::{TestRequest, run_test},
+};
 
 const MAGIC_STRING: &str = "ANTHROPIC_MAGIC_STRING_TRIGGER_REDACTED_THINKING_46C9A13E193C177646C7398A98432ECCCE4C1253D5E2D82641AC0E52CC2876CB";
 
@@ -278,6 +281,7 @@ fn test_opus_4_6_request_uses_adaptive_thinking() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -334,6 +338,7 @@ fn test_opus_4_7_xhigh_effort_mapping() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -385,6 +390,7 @@ fn test_opus_4_6_xhigh_falls_back_to_high() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -429,6 +435,7 @@ fn test_opus_4_6_max_effort_mapping() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -734,6 +741,7 @@ fn test_unknown_model_requests_summarized_thinking() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -811,6 +819,7 @@ fn test_unknown_reasoning_infers_adaptive_thinking() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -848,6 +857,7 @@ fn tier_request(
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     create_request(&model, query, true, &BetaFeatures(vec![])).map(|(request, ..)| request)
@@ -953,6 +963,7 @@ fn test_off_on_unknown_model_attempts_disable() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1070,6 +1081,7 @@ fn test_fable_5_reasoning_off_omits_disabled_thinking() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1105,6 +1117,7 @@ fn test_opus_4_5_uses_budgetted_thinking() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1157,6 +1170,7 @@ fn test_structured_output_sets_format() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1231,6 +1245,7 @@ fn test_schema_ignored_when_last_event_is_not_chat_request() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1271,6 +1286,7 @@ fn test_adaptive_thinking_with_structured_output() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1333,6 +1349,7 @@ fn test_forced_tool_with_reasoning_returns_fallback() {
             parameters: json!({ "type": "object", "properties": {} }),
         }],
         tool_choice: ToolChoice::Function("my_tool".into()),
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1405,6 +1422,7 @@ fn test_forced_tool_thinking_always_on_uses_escalating_nudge() {
             parameters: json!({ "type": "object", "properties": {} }),
         }],
         tool_choice: ToolChoice::Function("my_tool".into()),
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1489,6 +1507,7 @@ fn test_forced_tool_thinking_always_on_reasoning_off_still_soft_forces() {
             parameters: json!({ "type": "object", "properties": {} }),
         }],
         tool_choice: ToolChoice::Function("my_tool".into()),
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1554,6 +1573,7 @@ fn test_forced_tool_function_multi_tool_preserves_name() {
             },
         ],
         tool_choice: ToolChoice::Function("commit".into()),
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1621,6 +1641,7 @@ fn test_forced_tool_without_reasoning_no_fallback() {
             parameters: json!({ "type": "object", "properties": {} }),
         }],
         tool_choice: ToolChoice::Required,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1660,6 +1681,7 @@ fn test_auto_tool_choice_with_reasoning_no_fallback() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1802,6 +1824,7 @@ fn test_continue_injected_when_prefill_unsupported() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1851,6 +1874,7 @@ fn test_prefill_preserved_for_supported_models() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1890,6 +1914,7 @@ fn test_no_injection_when_last_message_is_user() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1933,6 +1958,7 @@ fn test_create_request_resends_signed_thinking_as_native_block() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -1989,6 +2015,7 @@ fn test_create_request_resends_redacted_thinking_as_native_block() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -2041,6 +2068,7 @@ fn test_create_request_falls_back_to_think_tags_without_signature() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -2099,6 +2127,7 @@ fn test_create_request_drops_empty_reasoning_instead_of_empty_think_tags() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -2155,6 +2184,7 @@ fn test_create_request_downgrades_trailing_assistant_thinking() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);
@@ -2220,6 +2250,7 @@ fn test_create_request_drops_trailing_redacted_thinking() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let beta = BetaFeatures(vec![]);

crates/jp_llm/src/provider/cerebras.rs

@@ -446,6 +446,7 @@ fn create_request(model: &ModelDetails, query: ChatQuery) -> Result<(Value, bool
         thread,
         tools,
         tool_choice,
+        ..
     } = query;
 
     let structured_schema = thread.events.schema();

crates/jp_llm/src/provider/cerebras_tests.rs

@@ -4,7 +4,7 @@ use jp_conversation::{ConversationEvent, event::ToolCallRequest};
 use reqwest_eventsource::Error as SseError;
 
 use super::*;
-use crate::provider::openai_compat::StreamChunk;
+use crate::{provider::openai_compat::StreamChunk, query::Truncation};
 
 /// Regression: a model absent from the table must still request the parsed
 /// reasoning format.
@@ -25,6 +25,7 @@ fn test_unknown_model_requests_parsed_reasoning() {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     };
 
     let (body, _) = create_request(&model, query).unwrap();
@@ -51,6 +52,7 @@ fn reasoning_query(reasoning: jp_config::model::parameters::PartialReasoningConf
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     }
 }
 
@@ -70,6 +72,7 @@ fn tier_query(tier: Option<ServiceTier>) -> ChatQuery {
         },
         tools: vec![],
         tool_choice: ToolChoice::Auto,
+        truncation: Truncation::default(),
     }
 }
 

crates/jp_llm/src/provider/google.rs

@@ -191,6 +191,7 @@ fn create_request(
         thread,
         tools,
         tool_choice,
+        ..
     } = query;
 
     // Extract schema and config before into_parts() consumes the thread.