diff --git a/src/ai/handlers.rs b/src/ai/handlers.rs
index b07b157..9c664b6 100644
--- a/src/ai/handlers.rs
+++ b/src/ai/handlers.rs
@@ -706,6 +706,63 @@ pub struct RenderedHistoryMessage {
     pub is_initial: bool,
 }
 
+#[derive(Debug, Deserialize)]
+pub struct ChatRewindHttpRequest {
+    pub file_path: String,
+    #[serde(default)]
+    pub library: Option<String>,
+    /// 0-based index into the rendered transcript. The message at this
+    /// index, and everything after it, is discarded. Must be > 0 — the
+    /// initial user message is protected.
+    pub discard_from_rendered_index: usize,
+}
+
+/// POST /insights/chat/rewind — truncate the stored conversation so the
+/// rendered message at `discard_from_rendered_index` (and everything after)
+/// is removed. Use when a user wants to retry a turn with a different
+/// prompt without prior replies poisoning context.
+#[post("/insights/chat/rewind")]
+pub async fn chat_rewind_handler(
+    _claims: Claims,
+    request: web::Json<ChatRewindHttpRequest>,
+    app_state: web::Data<AppState>,
+) -> impl Responder {
+    let library = match libraries::resolve_library_param(&app_state, request.library.as_deref()) {
+        Ok(Some(lib)) => lib,
+        Ok(None) => app_state.primary_library(),
+        Err(e) => {
+            return HttpResponse::BadRequest().json(serde_json::json!({
+                "error": format!("invalid library: {}", e)
+            }));
+        }
+    };
+
+    match app_state
+        .insight_chat
+        .rewind_history(
+            library.id,
+            &request.file_path,
+            request.discard_from_rendered_index,
+        )
+        .await
+    {
+        Ok(()) => HttpResponse::Ok().json(serde_json::json!({ "success": true })),
+        Err(e) => {
+            let msg = format!("{}", e);
+            log::error!("Chat rewind failed: {}", msg);
+            if msg.contains("no insight found") {
+                HttpResponse::NotFound().json(serde_json::json!({ "error": msg }))
+            } else if msg.contains("no chat history") {
+                HttpResponse::Conflict().json(serde_json::json!({ "error": msg }))
+            } else if msg.contains("cannot discard the initial") || msg.contains("out of range") {
+                HttpResponse::BadRequest().json(serde_json::json!({ "error": msg }))
+            } else {
+                HttpResponse::InternalServerError().json(serde_json::json!({ "error": msg }))
+            }
+        }
+    }
+}
+
 /// GET /insights/chat/history — return the rendered transcript for a photo.
 #[get("/insights/chat/history")]
 pub async fn chat_history_handler(
diff --git a/src/ai/insight_chat.rs b/src/ai/insight_chat.rs
index 7e5422c..41411d3 100644
--- a/src/ai/insight_chat.rs
+++ b/src/ai/insight_chat.rs
@@ -479,6 +479,109 @@ impl InsightChatService {
             model_used,
         })
     }
+
+    /// Truncate the stored conversation so the rendered message at
+    /// `discard_from_rendered_index` (and everything after it — including
+    /// the tool-call scaffolding that produced a discarded assistant reply)
+    /// is removed. The initial user turn cannot be discarded; attempting to
+    /// do so returns an error.
+    ///
+    /// Holds the per-file chat mutex so it serialises with `chat_turn`.
+    pub async fn rewind_history(
+        &self,
+        library_id: i32,
+        file_path: &str,
+        discard_from_rendered_index: usize,
+    ) -> Result<()> {
+        if discard_from_rendered_index == 0 {
+            bail!("cannot discard the initial user message");
+        }
+        let normalized = normalize_path(file_path);
+
+        let lock_key = (library_id, normalized.clone());
+        let entry_lock = {
+            let mut locks = self.chat_locks.lock().await;
+            locks
+                .entry(lock_key.clone())
+                .or_insert_with(|| Arc::new(TokioMutex::new(())))
+                .clone()
+        };
+        let _guard = entry_lock.lock().await;
+
+        let insight = {
+            let cx = opentelemetry::Context::new();
+            let mut dao = self.insight_dao.lock().expect("Unable to lock InsightDao");
+            dao.get_insight(&cx, &normalized)
+                .map_err(|e| anyhow!("failed to load insight: {:?}", e))?
+                .ok_or_else(|| anyhow!("no insight found for path"))?
+        };
+        let raw_history = insight
+            .training_messages
+            .as_ref()
+            .ok_or_else(|| anyhow!("insight has no chat history"))?;
+        let messages: Vec<ChatMessage> = serde_json::from_str(raw_history)
+            .map_err(|e| anyhow!("failed to deserialize chat history: {}", e))?;
+
+        let cut_at = find_raw_cut(&messages, discard_from_rendered_index)
+            .ok_or_else(|| anyhow!("discard_from_rendered_index out of range"))?;
+
+        let truncated = &messages[..cut_at];
+        let json = serde_json::to_string(truncated)
+            .map_err(|e| anyhow!("failed to serialize truncated history: {}", e))?;
+
+        let cx = opentelemetry::Context::new();
+        let mut dao = self.insight_dao.lock().expect("Unable to lock InsightDao");
+        dao.update_training_messages(&cx, library_id, &normalized, &json)
+            .map_err(|e| anyhow!("failed to persist truncated history: {:?}", e))?;
+        Ok(())
+    }
+}
+
+/// Is this raw message visible in the rendered transcript? Must match
+/// `load_history`'s filter exactly — `find_raw_cut` depends on it to map
+/// rendered indices back to raw positions.
+fn is_rendered(m: &ChatMessage) -> bool {
+    match m.role.as_str() {
+        "user" => true,
+        "assistant" => {
+            let has_tool_calls = m
+                .tool_calls
+                .as_ref()
+                .map(|c| !c.is_empty())
+                .unwrap_or(false);
+            !(has_tool_calls && m.content.trim().is_empty())
+        }
+        _ => false,
+    }
+}
+
+/// Given a rendered index to start discarding from, find the raw index at
+/// which to truncate. The cut position is the raw length after all prior
+/// rendered messages — which also strips any tool-call scaffolding that
+/// immediately precedes the discarded rendered message. Returns `None` if
+/// `discard_from_rendered_index` is past the end of the rendered view.
+pub(crate) fn find_raw_cut(
+    messages: &[ChatMessage],
+    discard_from_rendered_index: usize,
+) -> Option<usize> {
+    let mut rendered_count = 0usize;
+    let mut last_kept_raw_end = 0usize;
+    for (i, m) in messages.iter().enumerate() {
+        if !is_rendered(m) {
+            continue;
+        }
+        if rendered_count == discard_from_rendered_index {
+            return Some(last_kept_raw_end);
+        }
+        rendered_count += 1;
+        last_kept_raw_end = i + 1;
+    }
+    if rendered_count == discard_from_rendered_index {
+        // Discarding past the last rendered message is a no-op, but we
+        // surface it as "nothing to cut" rather than silent success.
+        return None;
+    }
+    None
 }
 
 /// Read AGENTIC_CHAT_MAX_ITERATIONS once per call. Cheap; keeps the code
@@ -637,4 +740,46 @@ mod tests {
         let dropped = apply_context_budget(&mut msgs, 1);
         assert!(!dropped);
     }
+
+    #[test]
+    fn rewind_strips_assistant_and_tool_scaffolding() {
+        // Rendered: [user1, asst1, user2, asst2] → cut at rendered index 3
+        // (the final asst2) should drop the tool-call scaffolding + asst2,
+        // leaving raw up through user2.
+        let msgs = vec![
+            ChatMessage::system("sys"),
+            ChatMessage::user("q1"),
+            assistant_text("a1"),
+            ChatMessage::user("q2"),
+            assistant_with_tool_call("lookup"),
+            ChatMessage::tool_result("data"),
+            assistant_text("a2 final"),
+        ];
+        let cut = find_raw_cut(&msgs, 3).expect("cut found");
+        // raw[0..cut] should end at user("q2") — indices 0..=3.
+        assert_eq!(cut, 4);
+        assert_eq!(msgs[cut - 1].role, "user");
+        assert_eq!(msgs[cut - 1].content, "q2");
+    }
+
+    #[test]
+    fn rewind_at_second_rendered_cuts_after_first_user() {
+        // Rendered index 1 = the first assistant reply → dropping it should
+        // leave just the initial user message.
+        let msgs = vec![
+            ChatMessage::system("s"),
+            ChatMessage::user("q1"),
+            assistant_with_tool_call("tool"),
+            ChatMessage::tool_result("r"),
+            assistant_text("a1"),
+        ];
+        let cut = find_raw_cut(&msgs, 1).expect("cut found");
+        assert_eq!(cut, 2); // sys + user("q1")
+    }
+
+    #[test]
+    fn rewind_beyond_range_returns_none() {
+        let msgs = vec![ChatMessage::user("q1"), assistant_text("a1")];
+        assert!(find_raw_cut(&msgs, 5).is_none());
+    }
 }
diff --git a/src/ai/mod.rs b/src/ai/mod.rs
index 93735c0..3c58b2a 100644
--- a/src/ai/mod.rs
+++ b/src/ai/mod.rs
@@ -11,10 +11,10 @@ pub mod sms_client;
 #[allow(unused_imports)]
 pub use daily_summary_job::{generate_daily_summaries, strip_summary_boilerplate};
 pub use handlers::{
-    chat_history_handler, chat_turn_handler, delete_insight_handler, export_training_data_handler,
-    generate_agentic_insight_handler, generate_insight_handler, get_all_insights_handler,
-    get_available_models_handler, get_insight_handler, get_openrouter_models_handler,
-    rate_insight_handler,
+    chat_history_handler, chat_rewind_handler, chat_turn_handler, delete_insight_handler,
+    export_training_data_handler, generate_agentic_insight_handler, generate_insight_handler,
+    get_all_insights_handler, get_available_models_handler, get_insight_handler,
+    get_openrouter_models_handler, rate_insight_handler,
 };
 pub use insight_generator::InsightGenerator;
 #[allow(unused_imports)]
diff --git a/src/ai/ollama.rs b/src/ai/ollama.rs
index 2cc2cfa..8c487c5 100644
--- a/src/ai/ollama.rs
+++ b/src/ai/ollama.rs
@@ -691,6 +691,17 @@ Analyze the image and use specific details from both the visual content and the
             .await
             .with_context(|| "Failed to parse Ollama chat response")?;
 
+        // Log performance counters returned by Ollama. Durations are
+        // reported in nanoseconds; we render ms + tokens/sec for skim-ability
+        // in the server log. Missing fields are left off the line rather
+        // than printed as `None`.
+        log_chat_metrics(
+            chat_response.prompt_eval_count,
+            chat_response.prompt_eval_duration,
+            chat_response.eval_count,
+            chat_response.eval_duration,
+        );
+
         Ok((
             chat_response.message,
             chat_response.prompt_eval_count,
@@ -915,8 +926,14 @@ struct OllamaChatResponse {
     done_reason: String,
     #[serde(default)]
     prompt_eval_count: Option<i32>,
+    /// Nanoseconds spent evaluating the prompt (context ingestion).
+    #[serde(default)]
+    prompt_eval_duration: Option<u64>,
     #[serde(default)]
     eval_count: Option<i32>,
+    /// Nanoseconds spent generating the response tokens.
+    #[serde(default)]
+    eval_duration: Option<u64>,
 }
 
 #[derive(Deserialize)]
@@ -924,6 +941,52 @@ struct OllamaResponse {
     response: String,
 }
 
+fn log_chat_metrics(
+    prompt_eval_count: Option<i32>,
+    prompt_eval_duration_ns: Option<u64>,
+    eval_count: Option<i32>,
+    eval_duration_ns: Option<u64>,
+) {
+    // Compute tokens/sec when both count and duration are present.
+    fn tokens_per_sec(count: Option<i32>, duration_ns: Option<u64>) -> Option<f64> {
+        match (count, duration_ns) {
+            (Some(c), Some(d)) if c > 0 && d > 0 => Some((c as f64) * 1_000_000_000.0 / (d as f64)),
+            _ => None,
+        }
+    }
+    let prompt_ms = prompt_eval_duration_ns.map(|ns| ns as f64 / 1_000_000.0);
+    let eval_ms = eval_duration_ns.map(|ns| ns as f64 / 1_000_000.0);
+    let prompt_tps = tokens_per_sec(prompt_eval_count, prompt_eval_duration_ns);
+    let eval_tps = tokens_per_sec(eval_count, eval_duration_ns);
+
+    let mut parts: Vec<String> = Vec::new();
+    if let Some(c) = prompt_eval_count {
+        let mut s = format!("prompt={} tok", c);
+        if let Some(ms) = prompt_ms {
+            s.push_str(&format!(" ({:.0} ms", ms));
+            if let Some(tps) = prompt_tps {
+                s.push_str(&format!(", {:.1} tok/s", tps));
+            }
+            s.push(')');
+        }
+        parts.push(s);
+    }
+    if let Some(c) = eval_count {
+        let mut s = format!("gen={} tok", c);
+        if let Some(ms) = eval_ms {
+            s.push_str(&format!(" ({:.0} ms", ms));
+            if let Some(tps) = eval_tps {
+                s.push_str(&format!(", {:.1} tok/s", tps));
+            }
+            s.push(')');
+        }
+        parts.push(s);
+    }
+    if !parts.is_empty() {
+        log::info!("Ollama chat metrics — {}", parts.join(", "));
+    }
+}
+
 #[derive(Deserialize)]
 struct OllamaTagsResponse {
     models: Vec<OllamaModel>,
diff --git a/src/main.rs b/src/main.rs
index 53ab607..8be397f 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -1358,6 +1358,7 @@ fn main() -> std::io::Result<()> {
                 .service(ai::get_openrouter_models_handler)
                 .service(ai::chat_turn_handler)
                 .service(ai::chat_history_handler)
+                .service(ai::chat_rewind_handler)
                 .service(ai::rate_insight_handler)
                 .service(ai::export_training_data_handler)
                 .service(libraries::list_libraries)