fix(harness): always emit a final answer before a turn ends (#4386)

src/openhuman/agent/harness/session/turn/core.rs

@@ -957,10 +957,11 @@ impl Agent {
             // cycle. Fold the extra call's usage into the turn accounting.
             let base = self.tool_dispatcher.to_provider_messages(&self.history);
             let (summary, summary_usage) = self
-                .summarize_iteration_checkpoint(
+                .summarize_turn_wrapup(
                     &base,
                     effective_model,
                     outcome.model_calls as u32 + 1,
+                    super::super::turn_checkpoint::MAX_ITER_CHECKPOINT_INSTRUCTION,
                 )
                 .await;
             if let Some(u) = summary_usage {
@@ -991,6 +992,61 @@ impl Agent {
                     iteration: outcome.model_calls,
                 },
             ));
+        } else if outcome.text.trim().is_empty() {
+            // #4093: the loop ran tool calls (tool_calls > 0, so the branch
+            // above did not fire) and then yielded a terminating response with
+            // no final text — the turn did work but would otherwise end
+            // silently, leaving the user with nothing. Enforce the
+            // "must produce a final response" terminal step: re-prompt the
+            // model (tools disabled) for a closing summary of what it did,
+            // falling back to a deterministic summary of the tool calls so the
+            // synthesized message is never itself empty. Fold the extra call's
+            // usage into the turn accounting, exactly like the cap path above.
+            let base = self.tool_dispatcher.to_provider_messages(&self.history);
+            let (summary, summary_usage) = self
+                .summarize_turn_wrapup(
+                    &base,
+                    effective_model,
+                    outcome.model_calls as u32 + 1,
+                    super::super::turn_checkpoint::FINAL_ANSWER_INSTRUCTION,
+                )
+                .await;
+            if let Some(u) = summary_usage {
+                input_tokens += u.input_tokens;
+                output_tokens += u.output_tokens;
+                cached_input_tokens += u.cached_input_tokens;
+                charged_amount_usd += u.charged_amount_usd;
+            }
+            let final_answer = if summary.trim().is_empty() {
+                super::super::turn_checkpoint::build_deterministic_final_summary(
+                    &tool_records_from_conversation(&outcome.conversation, &outcome.tool_outcomes),
+                )
+            } else {
+                summary
+            };
+            log::info!(
+                "[agent_loop] turn produced no final text after {} tool call(s); synthesized a closing summary ({} chars) — #4093",
+                outcome.tool_calls,
+                final_answer.chars().count()
+            );
+            // The empty terminal assistant response was already folded into
+            // `self.history` via `outcome.conversation` above (an empty
+            // `Chat(assistant(""))` — see `messages_to_conversation`). Drop that
+            // blank turn before appending the synthesized answer so the
+            // transcript and the next prompt don't carry a dangling empty
+            // assistant message immediately before the real reply (Codex review).
+            if matches!(
+                self.history.last(),
+                Some(ConversationMessage::Chat(msg))
+                    if msg.role == "assistant" && msg.content.trim().is_empty()
+            ) {
+                self.history.pop();
+            }
+            self.history
+                .push(ConversationMessage::Chat(ChatMessage::assistant(
+                    final_answer.clone(),
+                )));
+            final_answer
         } else {
             outcome.text.clone()
         };

src/openhuman/agent/harness/session/turn/session_io.rs

@@ -1,7 +1,6 @@
 //! Session persistence: transcript loading, checkpointing, and background tasks.
 
 use super::super::transcript;
-use super::super::turn_checkpoint::MAX_ITER_CHECKPOINT_INSTRUCTION;
 use super::super::types::Agent;
 use crate::openhuman::agent::harness;
 use crate::openhuman::agent::progress::AgentProgress;
@@ -63,27 +62,30 @@ impl Agent {
         }
     }
 
-    /// Ask the provider for a resumable checkpoint summary when a turn
-    /// hits the tool-call iteration cap, with native tools **disabled** so
-    /// the model returns prose rather than another tool call. Streams text
-    /// deltas to the progress sink (when attached) so the checkpoint
+    /// Ask the provider for a short wrap-up message with native tools
+    /// **disabled** so the model returns prose rather than another tool call.
+    /// Streams text deltas to the progress sink (when attached) so the summary
     /// appears in the UI like any other reply.
     ///
+    /// `instruction` is the synthetic user turn that steers the wrap-up — the
+    /// tool-call-cap checkpoint (`MAX_ITER_CHECKPOINT_INSTRUCTION`) or the
+    /// no-final-answer close (`FINAL_ANSWER_INSTRUCTION`, issue #4093).
+    ///
     /// Returns the summary text (empty when the provider call fails or
-    /// yields nothing — the caller then falls back to
-    /// [`build_deterministic_checkpoint`] so the thread is never left on an
-    /// unterminated tool cycle, bug-report-2026-05-26 A1) **paired with the
-    /// provider usage** for this extra call, so the caller can fold it into
-    /// the turn's cumulative token/cost accounting instead of silently
-    /// dropping it.
-    pub(super) async fn summarize_iteration_checkpoint(
+    /// yields nothing — the caller then falls back to a deterministic builder
+    /// so the turn is never left without a well-formed assistant message,
+    /// bug-report-2026-05-26 A1 / issue #4093) **paired with the provider
+    /// usage** for this extra call, so the caller can fold it into the turn's
+    /// cumulative token/cost accounting instead of silently dropping it.
+    pub(super) async fn summarize_turn_wrapup(
         &self,
         base_messages: &[ChatMessage],
         effective_model: &str,
         iteration_for_stream: u32,
+        instruction: &str,
     ) -> (String, Option<UsageInfo>) {
         let mut messages = base_messages.to_vec();
-        messages.push(ChatMessage::user(MAX_ITER_CHECKPOINT_INSTRUCTION));
+        messages.push(ChatMessage::user(instruction));
 
         // Mirror the main loop's streaming sink so the checkpoint renders
         // incrementally. Only text deltas are relevant here (tools are

src/openhuman/agent/harness/session/turn_checkpoint.rs

@@ -74,3 +74,32 @@ pub(super) fn build_deterministic_checkpoint(
     );
     out
 }
+
+/// Instruction appended (as a synthetic user turn) when a turn finished its
+/// tool work but the model produced **no final answer** — it yielded a
+/// terminating response with empty text after running tools (issue #4093).
+/// Native tools are disabled for this call so the model wraps up in prose
+/// instead of requesting more tools.
+pub(super) const FINAL_ANSWER_INSTRUCTION: &str = "\
+You have finished using tools for this turn but have not yet written a reply to the user. \
+Do not call any more tools. Write a short, self-contained final message that summarises what you did and \
+what you found or accomplished, grounded in the tool results above. If nothing conclusive resulted, say so plainly.";
+
+/// Build a deterministic final answer from this turn's tool-call records.
+/// Used as the guaranteed non-empty fallback when a turn ran tools but the
+/// model produced no closing message and the re-prompt for one also came
+/// back empty — so a turn that did work can never end silently (issue #4093).
+/// Distinct from [`build_deterministic_checkpoint`]: the turn did NOT hit the
+/// iteration cap, so this reads as a completed summary, not a paused one.
+pub(super) fn build_deterministic_final_summary(records: &[ToolCallRecord]) -> String {
+    if records.is_empty() {
+        return "I finished this turn but produced no result to report.".to_string();
+    }
+    let mut out = String::from("Here's a summary of what I did this turn:\n\n");
+    for r in records {
+        let status = if r.success { "ok" } else { "failed" };
+        out.push_str(&format!("- `{}` — {}\n", r.name, status));
+    }
+    out.push_str("\nLet me know if you'd like me to go further.");
+    out
+}

src/openhuman/agent/harness/session/turn_tests.rs

@@ -1038,6 +1038,165 @@ async fn turn_checkpoint_falls_back_to_deterministic_summary_when_model_summary_
     );
 }
 
+#[tokio::test]
+async fn turn_synthesizes_final_answer_when_tool_turn_yields_no_text() {
+    // #4093: the model runs a tool and then yields a terminating response with
+    // NO text and NO further tool calls — the turn did work but would end
+    // silently. Because the cap was not hit, this is not a checkpoint case; the
+    // harness must enforce the "must produce a final response" terminal step by
+    // re-prompting the model (tools disabled) for a closing summary and
+    // returning that instead of a blank reply.
+    let provider: Arc<dyn Provider> = Arc::new(SequenceProvider {
+        responses: AsyncMutex::new(vec![
+            // Tool iteration (well under the cap).
+            Ok(ChatResponse {
+                text: Some("<tool_call>{\"name\":\"echo\",\"arguments\":{}}</tool_call>".into()),
+                tool_calls: vec![],
+                usage: None,
+                reasoning_content: None,
+            }),
+            // Terminal response with no text and no tool calls — the silent end.
+            Ok(ChatResponse {
+                text: Some(String::new()),
+                tool_calls: vec![],
+                usage: None,
+                reasoning_content: None,
+            }),
+            // The harness's forced final-answer re-prompt (tools disabled).
+            Ok(ChatResponse {
+                text: Some("All done — I ran echo and it succeeded.".into()),
+                tool_calls: vec![],
+                usage: None,
+                reasoning_content: None,
+            }),
+        ]),
+        requests: AsyncMutex::new(Vec::new()),
+    });
+    let mut agent = make_agent_with_builder(
+        provider,
+        vec![Box::new(EchoTool)],
+        Box::new(FixedMemoryLoader {
+            context: String::new(),
+        }),
+        vec![],
+        crate::openhuman::config::AgentConfig {
+            max_tool_iterations: 5,
+            ..crate::openhuman::config::AgentConfig::default()
+        },
+        crate::openhuman::config::ContextConfig::default(),
+    );
+
+    let reply = agent
+        .turn("hello")
+        .await
+        .expect("a tool-only turn with no final text should synthesize one, not error");
+    assert!(
+        !reply.trim().is_empty(),
+        "turn must never end with an empty final message (#4093), got: {reply:?}"
+    );
+    assert!(
+        reply.contains("I ran echo"),
+        "the synthesized final message should be the model's wrap-up, got: {reply}"
+    );
+    // The transcript must end on the assistant's final message, not a dangling
+    // tool cycle.
+    assert!(
+        matches!(
+            agent.history.last(),
+            Some(ConversationMessage::Chat(msg))
+                if msg.role == "assistant" && !msg.content.trim().is_empty()
+        ),
+        "history should end on a non-empty assistant message, got: {:?}",
+        agent.history.last()
+    );
+    // ...and the blank terminal assistant response (folded in from the turn
+    // outcome) must have been dropped, not left dangling before the synthesized
+    // answer (Codex review).
+    assert!(
+        !agent.history.iter().any(|m| matches!(
+            m,
+            ConversationMessage::Chat(msg)
+                if msg.role == "assistant" && msg.content.trim().is_empty()
+        )),
+        "no blank assistant turn should remain in history, got: {:?}",
+        agent.history
+    );
+}
+
+#[tokio::test]
+async fn turn_final_answer_falls_back_to_deterministic_summary_when_reprompt_empty() {
+    // #4093 safety net: the tool ran, the model yielded no final text, and the
+    // forced final-answer re-prompt ALSO came back empty. The harness must fall
+    // back to a deterministic summary of the tool calls so the turn is never
+    // blank — and, unlike the cap path, it must read as a completed summary
+    // rather than a paused "tool-call limit" checkpoint.
+    let provider: Arc<dyn Provider> = Arc::new(SequenceProvider {
+        responses: AsyncMutex::new(vec![
+            Ok(ChatResponse {
+                text: Some("<tool_call>{\"name\":\"echo\",\"arguments\":{}}</tool_call>".into()),
+                tool_calls: vec![],
+                usage: None,
+                reasoning_content: None,
+            }),
+            Ok(ChatResponse {
+                text: Some(String::new()),
+                tool_calls: vec![],
+                usage: None,
+                reasoning_content: None,
+            }),
+            // Re-prompt for a final answer also returns empty.
+            Ok(ChatResponse {
+                text: Some(String::new()),
+                tool_calls: vec![],
+                usage: None,
+                reasoning_content: None,
+            }),
+        ]),
+        requests: AsyncMutex::new(Vec::new()),
+    });
+    let mut agent = make_agent_with_builder(
+        provider,
+        vec![Box::new(EchoTool)],
+        Box::new(FixedMemoryLoader {
+            context: String::new(),
+        }),
+        vec![],
+        crate::openhuman::config::AgentConfig {
+            max_tool_iterations: 5,
+            ..crate::openhuman::config::AgentConfig::default()
+        },
+        crate::openhuman::config::ContextConfig::default(),
+    );
+
+    let reply = agent
+        .turn("hello")
+        .await
+        .expect("empty final re-prompt should fall back deterministically, not error");
+    assert!(
+        !reply.trim().is_empty(),
+        "deterministic fallback must be non-empty (#4093), got: {reply:?}"
+    );
+    assert!(
+        reply.contains("echo"),
+        "fallback should list the tool that ran, got: {reply}"
+    );
+    assert!(
+        !reply.contains("tool-call limit"),
+        "a non-capped turn must not claim it hit the tool-call limit, got: {reply}"
+    );
+    // The blank terminal assistant response must not linger before the
+    // deterministic summary (Codex review).
+    assert!(
+        !agent.history.iter().any(|m| matches!(
+            m,
+            ConversationMessage::Chat(msg)
+                if msg.role == "assistant" && msg.content.trim().is_empty()
+        )),
+        "no blank assistant turn should remain in history, got: {:?}",
+        agent.history
+    );
+}
+
 #[tokio::test]
 async fn turn_checkpoint_usage_is_folded_into_transcript_accounting() {
     // The extra checkpoint provider call costs tokens; those must land in