CLAUDE.md

@@ -900,9 +900,18 @@ endpoint.
   a deploy-time decision, not a request-time one).
 - `hybrid → hybrid` allowed.
 - `hybrid → local` allowed (the inlined description replays as text).
-- `local → hybrid` rejected — the stored transcript has raw images in the
+- `local → hybrid` allowed. The stored transcript carries the raw photo in
-  first user message and OpenRouter providers don't accept that shape
+  its first user message, so each hybrid turn swaps it for a local vision
-  consistently. Regenerate the insight in hybrid mode instead.
+  description (cached per photo) in the copy sent to OpenRouter, then puts
+  the photo back before saving (`InsightChatService::inline_images_as_text`).
+  If the local model can't describe it, the turn fails. A stored model id is
+  only reused on the same backend (`continuation_model`).
+
+**Photos never go to OpenRouter.** `OpenRouterClient` refuses any request
+carrying image data (`IMAGES_STAY_LOCAL`) and its `describe_image` always
+errors — the client is the choke point, so a caller bug fails the turn
+instead of uploading a photo. Image work in hybrid mode goes through
+`ResolvedBackend::local()`.
 
 **Insight Chat Continuation:**
 
@@ -933,9 +942,9 @@ clients whether chat is available for a given insight.
 
 Backend routing rules (matches agentic-insight generation):
 - Stored `backend` on the insight row is authoritative by default.
-- `request.backend` may override per-turn. `local -> hybrid` is rejected in
+- `request.backend` may override per-turn, in either direction (see the
-  v1 (would require on-the-fly visual-description rewrite); `hybrid -> local`
+  cross-replay matrix above); `hybrid -> local` replays verbatim since the
-  replays verbatim since the description is already inlined as text.
+  description is already inlined as text.
 - `request.model` overrides the chat model (an Ollama id in local mode, an
   OpenRouter id in hybrid mode).
 

src/ai/insight_chat.rs

@@ -119,8 +119,16 @@ pub struct InsightChatService {
     generator: Arc<InsightGenerator>,
     insight_dao: Arc<Mutex<Box<dyn InsightDao>>>,
     chat_locks: ChatLockMap,
+    /// Local vision descriptions of transcript images, keyed by a hash of
+    /// the image data, so a local-generated chat continued on a text-only
+    /// (hybrid) model describes its photo once rather than every turn.
+    image_descriptions: Arc<Mutex<HashMap<u64, String>>>,
 }
 
+/// Cap on cached image descriptions; the cache is cleared when it fills.
+/// One entry per photo chatted about since the server started.
+const IMAGE_DESCRIPTION_CACHE_MAX: usize = 256;
+
 impl InsightChatService {
     pub fn new(
         generator: Arc<InsightGenerator>,
@@ -131,9 +139,62 @@ impl InsightChatService {
             generator,
             insight_dao,
             chat_locks,
+            image_descriptions: Arc::new(Mutex::new(HashMap::new())),
         }
     }
 
+    /// Prepare a replayed transcript for a backend that can't take images.
+    ///
+    /// A chat generated on the local backend carries the photo itself in its
+    /// first user message. Continuing it on hybrid would send that photo to
+    /// OpenRouter — which the OpenRouter client refuses — so each image is
+    /// described by the local vision model and the text takes its place in
+    /// the copy sent to the model. The returned stash puts the images back
+    /// before the transcript is saved, so the stored chat keeps its photo and
+    /// a later local turn still sees it.
+    ///
+    /// A no-op for backends that take images inline. If the local model
+    /// can't describe an image the turn fails: going ahead without the
+    /// photo would quietly change what the chat is about.
+    async fn inline_images_as_text(
+        &self,
+        backend: &ResolvedBackend,
+        messages: &mut [ChatMessage],
+    ) -> Result<ImageStash> {
+        if backend.images_inline {
+            return Ok(ImageStash::default());
+        }
+        let mut descriptions: HashMap<String, String> = HashMap::new();
+        for image in transcript_images(messages) {
+            let key = image_cache_key(&image);
+            let cached = self
+                .image_descriptions
+                .lock()
+                .ok()
+                .and_then(|c| c.get(&key).cloned());
+            let description = match cached {
+                Some(d) => d,
+                None => {
+                    let d = backend.local().describe_image(&image).await.map_err(|e| {
+                        anyhow!(
+                            "couldn't describe the photo with the local vision model, so this \
+                             chat can't continue on a text-only model: {e}"
+                        )
+                    })?;
+                    if let Ok(mut cache) = self.image_descriptions.lock() {
+                        if cache.len() >= IMAGE_DESCRIPTION_CACHE_MAX {
+                            cache.clear();
+                        }
+                        cache.insert(key, d.clone());
+                    }
+                    d
+                }
+            };
+            descriptions.insert(image, description);
+        }
+        Ok(apply_image_descriptions(messages, &descriptions))
+    }
+
     /// Accessor for the insight DAO (used by async job completion).
     pub fn insight_dao(&self) -> &Arc<Mutex<Box<dyn InsightDao>>> {
         &self.insight_dao
@@ -285,11 +346,12 @@ impl InsightChatService {
 
         let stored_model = insight.model_version.clone();
         let overrides = SamplingOverrides {
-            model: req
+            model: continuation_model(
-                .model
+                req.model.clone(),
-                .clone()
+                &stored_model,
-                .or_else(|| Some(stored_model.clone()))
+                &stored_backend,
-                .filter(|m| !m.is_empty()),
+                kind.as_str(),
+            ),
             num_ctx: req.num_ctx,
             temperature: req.temperature,
             top_p: req.top_p,
@@ -298,6 +360,9 @@ impl InsightChatService {
             enable_thinking: req.enable_thinking,
         };
         let backend = self.generator.resolve_backend(kind, &overrides).await?;
+        // Text-only backend (hybrid) on a transcript holding the photo:
+        // swap it for a local description. Restored before saving.
+        let image_stash = self.inline_images_as_text(&backend, &mut messages).await?;
         let model_used = backend.model().to_string();
         span.set_attribute("model", model_used.clone());
 
@@ -449,6 +514,8 @@ impl InsightChatService {
         // Drop the per-turn iteration-budget note from the system message
         // before we persist so it doesn't snowball on each subsequent turn.
         restore_system_content(&mut messages, original_system_content);
+        // Put the photo back so the stored transcript keeps it.
+        image_stash.restore(&mut messages);
 
         // Append mode: undo the per-turn system-prompt override so the
         // stored transcript keeps the original baked persona. Amend mode:
@@ -979,11 +1046,12 @@ impl InsightChatService {
 
         let stored_model = insight.model_version.clone();
         let overrides = SamplingOverrides {
-            model: req
+            model: continuation_model(
-                .model
+                req.model.clone(),
-                .clone()
+                &stored_model,
-                .or_else(|| Some(stored_model.clone()))
+                &stored_backend,
-                .filter(|m| !m.is_empty()),
+                kind.as_str(),
+            ),
             num_ctx: req.num_ctx,
             temperature: req.temperature,
             top_p: req.top_p,
@@ -992,6 +1060,9 @@ impl InsightChatService {
             enable_thinking: req.enable_thinking,
         };
         let backend = self.generator.resolve_backend(kind, &overrides).await?;
+        // Text-only backend (hybrid) on a transcript holding the photo:
+        // swap it for a local description. Restored before saving.
+        let image_stash = self.inline_images_as_text(&backend, &mut messages).await?;
         let model_used = backend.model().to_string();
         // llm.* attributes on the turn-level span (issue #130): clients can't
         // know local-vs-hybrid, but the turn knows its resolved backend.
@@ -1064,6 +1135,8 @@ impl InsightChatService {
         }
 
         restore_system_content(&mut messages, original_system_content);
+        // Put the photo back so the stored transcript keeps it.
+        image_stash.restore(&mut messages);
 
         if !req.amend {
             restore_system_prompt_override(&mut messages, override_stash);
@@ -1455,11 +1528,12 @@ impl InsightChatService {
 
         let stored_model = insight.model_version.clone();
         let overrides = SamplingOverrides {
-            model: req
+            model: continuation_model(
-                .model
+                req.model.clone(),
-                .clone()
+                &stored_model,
-                .or_else(|| Some(stored_model.clone()))
+                &stored_backend,
-                .filter(|m| !m.is_empty()),
+                kind.as_str(),
+            ),
             num_ctx: req.num_ctx,
             temperature: req.temperature,
             top_p: req.top_p,
@@ -1468,6 +1542,9 @@ impl InsightChatService {
             enable_thinking: req.enable_thinking,
         };
         let backend = self.generator.resolve_backend(kind, &overrides).await?;
+        // Text-only backend (hybrid) on a transcript holding the photo:
+        // swap it for a local description. Restored before saving.
+        let image_stash = self.inline_images_as_text(&backend, &mut messages).await?;
         let model_used = backend.model().to_string();
 
         // Tool set — images_inline mode + first user turn carries an image →
@@ -1533,6 +1610,8 @@ impl InsightChatService {
         // Drop the per-turn iteration-budget note before persisting so it
         // doesn't snowball on subsequent turns.
         restore_system_content(&mut messages, original_system_content);
+        // Put the photo back so the stored transcript keeps it.
+        image_stash.restore(&mut messages);
 
         // Append mode: undo the per-turn system-prompt override.
         // Amend mode: keep it — it becomes the new row's system message.
@@ -2071,34 +2150,122 @@ fn resolve_date_taken_for_context(
     })
 }
 
-/// Validate a stored→effective backend transition for a chat continuation.
+/// Validate the backend a chat continuation is switching to.
-/// Continuation runs against a transcript that was generated with a specific
+///
-/// backend; the only blocked transition is `local → hybrid`, because the
+/// Every transition is allowed now. `local → hybrid` used to be refused
-/// stored transcript has images embedded in the first user message and the
+/// because a local transcript carries the raw photo; the turn now replaces
-/// hybrid path (OpenRouter chat with describe-then-inline) can't replay
+/// it with a local vision description before anything reaches OpenRouter
-/// raw image bytes through OpenRouter consistently across providers.
+/// (`InsightChatService::inline_images_as_text`), and the OpenRouter client
-/// `hybrid → local` is allowed (the inlined description replays verbatim
+/// refuses image data outright as a backstop. `hybrid → local` replays the
-/// as text).
+/// inlined description as text. `stored` is kept for the call sites'
+/// symmetry and future rules.
 ///
 /// Whether "local" routes through Ollama or llama-swap is decided at
 /// startup by `LLM_BACKEND`; both share the same transcript shape from
 /// the chat-replay perspective.
-fn validate_cross_replay(stored: &str, effective: &str) -> Result<()> {
+fn validate_cross_replay(_stored: &str, effective: &str) -> Result<()> {
     if !matches!(effective, "local" | "hybrid") {
         bail!(
             "unknown backend '{}'; expected 'local' or 'hybrid'",
             effective
         );
     }
-    if stored == "local" && effective == "hybrid" {
-        bail!(
-            "switching from local to hybrid mid-chat isn't supported; \
-             regenerate the insight in hybrid mode if you want OpenRouter chat"
-        );
-    }
     Ok(())
 }
 
+/// The model to continue a chat with when the request names none: the one
+/// the insight was generated with, but only on the same backend. After a
+/// switch, the stored id belongs to the other backend (a llama-swap slot is
+/// not an OpenRouter model), so the new backend's default is used instead.
+fn continuation_model(
+    requested: Option<String>,
+    stored_model: &str,
+    stored_backend: &str,
+    effective_backend: &str,
+) -> Option<String> {
+    requested
+        .or_else(|| (stored_backend == effective_backend).then(|| stored_model.to_string()))
+        .filter(|m| !m.is_empty())
+}
+
+/// Every distinct image in a transcript, in order of first appearance.
+fn transcript_images(messages: &[ChatMessage]) -> Vec<String> {
+    let mut seen = std::collections::HashSet::new();
+    messages
+        .iter()
+        .flat_map(|m| m.images.iter().flatten())
+        .filter(|img| seen.insert((*img).clone()))
+        .cloned()
+        .collect()
+}
+
+fn image_cache_key(image: &str) -> u64 {
+    use std::hash::{Hash, Hasher};
+    let mut h = std::collections::hash_map::DefaultHasher::new();
+    image.hash(&mut h);
+    h.finish()
+}
+
+/// What `apply_image_descriptions` took out of a transcript, so it can be
+/// put back before the transcript is saved.
+#[derive(Debug, Default)]
+struct ImageStash {
+    entries: Vec<StashedImages>,
+}
+
+#[derive(Debug)]
+struct StashedImages {
+    index: usize,
+    original_content: String,
+    rewritten_content: String,
+    images: Option<Vec<String>>,
+}
+
+impl ImageStash {
+    /// Put the original content and images back. Each entry is matched by
+    /// index *and* by still carrying the rewritten text, so a turn that
+    /// reshaped the transcript can't get the wrong message restored.
+    fn restore(self, messages: &mut [ChatMessage]) {
+        for e in self.entries {
+            if let Some(m) = messages.get_mut(e.index)
+                && m.content == e.rewritten_content
+            {
+                m.content = e.original_content;
+                m.images = e.images;
+            }
+        }
+    }
+}
+
+/// Replace each image in `messages` with its description, appended to the
+/// message text. Returns what was taken out, for `ImageStash::restore`.
+fn apply_image_descriptions(
+    messages: &mut [ChatMessage],
+    descriptions: &HashMap<String, String>,
+) -> ImageStash {
+    let mut stash = ImageStash::default();
+    for (idx, m) in messages.iter_mut().enumerate() {
+        let Some(images) = m.images.as_ref().filter(|i| !i.is_empty()) else {
+            continue;
+        };
+        let mut content = m.content.clone();
+        for img in images {
+            if let Some(desc) = descriptions.get(img) {
+                content.push_str(&format!(
+                    "\n\nVisual description of the photo (from the local vision model):\n{desc}"
+                ));
+            }
+        }
+        stash.entries.push(StashedImages {
+            index: idx,
+            original_content: std::mem::replace(&mut m.content, content.clone()),
+            rewritten_content: content,
+            images: m.images.take(),
+        });
+    }
+    stash
+}
+
 /// Pick the backend label for bootstrap. Bootstrap has no stored insight
 /// to defer to (that's continuation's behaviour), so the default is
 /// `"local"`. Returns an error if the supplied label is non-empty but
@@ -3672,17 +3839,90 @@ mod tests {
     }
 
     #[test]
-    fn cross_replay_rejects_local_to_hybrid() {
+    fn cross_replay_allows_every_known_transition() {
-        let err = validate_cross_replay("local", "hybrid").unwrap_err();
-        assert!(format!("{}", err).contains("local to hybrid"));
-    }
-
-    #[test]
-    fn cross_replay_allows_supported_transitions() {
         assert!(validate_cross_replay("local", "local").is_ok());
         assert!(validate_cross_replay("hybrid", "hybrid").is_ok());
         // Hybrid → local replays the inlined description as plain text.
         assert!(validate_cross_replay("hybrid", "local").is_ok());
+        // Local → hybrid swaps the photo for a local description first.
+        assert!(validate_cross_replay("local", "hybrid").is_ok());
+    }
+
+    #[test]
+    fn continuation_model_drops_the_stored_model_across_a_switch() {
+        // Same backend: keep the model the insight was generated with.
+        assert_eq!(
+            continuation_model(None, "gemma4-31", "local", "local").as_deref(),
+            Some("gemma4-31")
+        );
+        // Switched: a llama-swap slot name is not an OpenRouter model id.
+        assert_eq!(
+            continuation_model(None, "gemma4-31", "local", "hybrid"),
+            None
+        );
+        // An explicit request always wins.
+        assert_eq!(
+            continuation_model(Some("x/y".into()), "gemma4-31", "local", "hybrid").as_deref(),
+            Some("x/y")
+        );
+        assert_eq!(continuation_model(None, "", "local", "local"), None);
+    }
+
+    fn with_image(text: &str, img: &str) -> ChatMessage {
+        let mut m = ChatMessage::user(text.to_string());
+        m.images = Some(vec![img.to_string()]);
+        m
+    }
+
+    #[test]
+    fn transcript_images_lists_each_image_once() {
+        let msgs = vec![
+            ChatMessage::system("sys"),
+            with_image("describe", "IMG_A"),
+            ChatMessage::user("more"),
+            with_image("again", "IMG_A"),
+        ];
+        assert_eq!(transcript_images(&msgs), vec!["IMG_A".to_string()]);
+    }
+
+    #[test]
+    fn photo_is_swapped_for_text_then_restored_exactly() {
+        let original = vec![
+            ChatMessage::system("sys"),
+            with_image("What's in this photo?", "IMG_A"),
+            assistant_text("A beach."),
+            ChatMessage::user("who's there?"),
+        ];
+        let mut msgs = original.clone();
+        let descriptions =
+            HashMap::from([("IMG_A".to_string(), "Two people on sand.".to_string())]);
+
+        let stash = apply_image_descriptions(&mut msgs, &descriptions);
+        // Nothing the text-only model receives carries image data.
+        assert!(msgs.iter().all(|m| m.images.is_none()));
+        assert!(msgs[1].content.starts_with("What's in this photo?"));
+        assert!(msgs[1].content.contains("Two people on sand."));
+
+        // The turn appends its exchange, then the stash is restored before
+        // saving: the stored transcript keeps the photo, unchanged.
+        msgs.push(assistant_text("Two friends."));
+        stash.restore(&mut msgs);
+        assert_eq!(
+            serde_json::to_string(&msgs[..original.len()]).unwrap(),
+            serde_json::to_string(&original).unwrap()
+        );
+        assert_eq!(msgs[1].images, Some(vec!["IMG_A".to_string()]));
+    }
+
+    #[test]
+    fn restore_leaves_a_message_the_turn_changed() {
+        let mut msgs = vec![with_image("q", "IMG_A")];
+        let descriptions = HashMap::from([("IMG_A".to_string(), "desc".to_string())]);
+        let stash = apply_image_descriptions(&mut msgs, &descriptions);
+        msgs[0].content = "something else entirely".into();
+        stash.restore(&mut msgs);
+        assert_eq!(msgs[0].content, "something else entirely");
+        assert!(msgs[0].images.is_none());
     }
 
     #[test]