Say a transcript's size, name a model by something, and ask before a reload

Five things asked for on the phone, and one trap behind the first of them.

A llama prompt carries `<__media__>` where a picture was, and llama.cpp
pairs each marker with a decoded image when it tokenizes -- so a marker in
words nobody attached a picture to fails the turn, and then fails every
later one, since the conversation is folded out of a transcript that holds
it for good. A model saying the marker back is enough to do it. Every
message with words in it now goes through `without_marker`.

A model's `general.name` is filled in by whatever converted the file, and
`convert_hf_to_gguf.py` fills it from the directory it converted: Prism ML's
Bonsai publishes `general.name = "Hf"`, which is unique and so passed the
label cascade and told a reader nothing. A name is now used only where it
shares a word with the repo or the file it came from; one that does not
drops to the file name.

The model settings dialog and the server card said what a save would cost in
a paragraph under the control, read after the decision if at all. Both ask
instead, in the shape the rest of that screen already uses for Stop and
Delete -- and the dialog's question is asked over the edits, so Cancel comes
back to them.

A field's label was `labelMedium` in the variant colour while every setting
beside it was body text, which on one form read as two ranks of setting.

`GET /sessions/{id}`'s `transcriptFile` now carries the file's size, and the
settings screen draws it beside what this phone has cached.

Checked on the emulator against the sandbox: the labels line up, the row
says "14 kB · 14 kB cached", and both confirmations appear over a loaded
Qwen3-0.6B. `cargo test` 275 passed, clippy and fmt clean, lint clean.

AGENTS.md

@@ -523,6 +523,17 @@ written, and the fold uses that same predicate to decide a reply is settled.
   `generate` now fails the turn for both -- a reply that stops early is not a
   reply, and the transcript keeps whatever arrived before it.
 
+- **`<__media__>` in a llama prompt is a picture, wherever it came from.**
+  llama.cpp takes an image out of the request and leaves that marker in the
+  rendered text, then pairs each marker with a decoded image at tokenize time
+  -- so one in words nobody attached a picture to fails the turn with `number
+  of media markers in text (1) exceeds number of bitmaps (0)`, which reaches
+  the phone as "Failed to tokenize prompt". A model saying it back is enough,
+  and then *every* later message fails too, since the conversation is folded
+  out of a transcript that now holds it. `Message::new` and
+  `Message::from_user` take it out of anything that is text (`without_marker`),
+  which is every message with words in it.
+
 - **A cancel flag is only as prompt as the next place somebody looks.** A
   llama turn waits on three things that look nowhere at all: a permission
   question, a tool call `llama-server` is running (a shell command there runs

app/androidApp/src/main/kotlin/com/example/aiapp/Api.kt

@@ -251,8 +251,18 @@ data class SessionSummary(
     val transcriptFile: FileOnMachine?,
 )
 
-/** A file somewhere the explorer can be pointed at: which machine, and the path on it. */
-data class FileOnMachine(val machine: String, val machineName: String, val path: String)
+/**
+ * A file somewhere the explorer can be pointed at: which machine, and the path on it.
+ *
+ * [bytes] is how big it is, and null is "the server did not say" rather than zero: a file of no
+ * bytes and a file nobody could measure are different answers.
+ */
+data class FileOnMachine(
+    val machine: String,
+    val machineName: String,
+    val path: String,
+    val bytes: Long? = null,
+)
 
 private fun parseSession(session: JSONObject) =
     SessionSummary(
@@ -292,6 +302,7 @@ private fun parseSession(session: JSONObject) =
                     machine = file.getString("machine"),
                     machineName = file.getString("machineName"),
                     path = file.getString("path"),
+                    bytes = if (file.has("bytes")) file.getLong("bytes") else null,
                 )
             },
     )

app/androidApp/src/main/kotlin/com/example/aiapp/Field.kt

@@ -57,12 +57,10 @@ fun LabelledField(
     keyboardActions: KeyboardActions = KeyboardActions.Default,
 ) {
     Column(modifier.fillMaxWidth()) {
-        Text(
-            label,
-            style = MaterialTheme.typography.labelMedium,
-            color = MaterialTheme.colorScheme.onSurfaceVariant,
-            modifier = Modifier.padding(bottom = 2.dp),
-        )
+        // Body size in the ordinary text colour, which is what a setting's label is where the
+        // control beside it is a switch or a picker. Smaller and greyer on the ones that are
+        // fields reads as two ranks of setting where there is one.
+        Text(label, modifier = Modifier.padding(bottom = 2.dp))
         hint?.let {
             Text(
                 it,

app/androidApp/src/main/kotlin/com/example/aiapp/ProviderScreen.kt

@@ -180,12 +180,6 @@ fun ProviderScreen(
                     if (view.models.isNotEmpty() && view.modelParams.isNotEmpty()) {
                         item("models-heading") {
                             Text("Models", style = MaterialTheme.typography.titleSmall)
-                            Text(
-                                "How a model is loaded belongs to the machine, not to a session: " +
-                                    "one copy of it in memory answers every session using it.",
-                                style = MaterialTheme.typography.bodySmall,
-                                color = MaterialTheme.colorScheme.onSurfaceVariant,
-                            )
                             Spacer(Modifier.height(8.dp))
                         }
                     }
@@ -324,6 +318,7 @@ private fun ServerCard(
     // half-typed is visibly not saved rather than quietly either way.
     val saved = maxLoaded?.toString().orEmpty()
     var typed by remember(saved) { mutableStateOf(saved) }
+    var confirming by remember { mutableStateOf(false) }
     Card(Modifier.fillMaxWidth()) {
         Column(Modifier.padding(12.dp)) {
             Text("Model server", style = MaterialTheme.typography.titleSmall)
@@ -354,20 +349,55 @@ private fun ServerCard(
                 Spacer(Modifier.weight(1f))
                 TextButton(
                     enabled = enabled && typed != saved,
-                    onClick = { onMaxLoaded(typed.toIntOrNull()) },
+                    // Saving this while the server is up changes nothing until it comes down
+                    // again, which is asked rather than written underneath -- see [RestartDialog].
+                    onClick = {
+                        if (server.running) confirming = true else onMaxLoaded(typed.toIntOrNull())
+                    },
                 ) {
                     Text("Save")
                 }
             }
-            if (typed != saved) {
-                Text(
-                    "Read when this server next starts.",
-                    style = MaterialTheme.typography.bodySmall,
-                    color = MaterialTheme.colorScheme.onSurfaceVariant,
+        }
+    }
+    if (confirming) {
+        RestartDialog(
+            title = "Save for the next start?",
+            text =
+                "This server is running, and how many models it keeps loaded was decided when it " +
+                    "started. Saving now changes what it does the next time it starts -- stop it " +
+                    "here to have that be now.",
+            onConfirm = {
+                confirming = false
+                onMaxLoaded(typed.toIntOrNull())
+            },
+            onDismiss = { confirming = false },
         )
     }
 }
-    }
+
+/**
+ * What a setting read at load time costs, asked before it is written.
+ *
+ * One shape for both of the questions on this screen — a model's settings and the server's —
+ * because they are the same question: this is read when something starts again, and here is what
+ * starting again costs. A paragraph under the control said the same thing and was read after the
+ * decision, if at all.
+ */
+@Composable
+private fun RestartDialog(
+    title: String,
+    text: String,
+    onConfirm: () -> Unit,
+    onDismiss: () -> Unit,
+) {
+    AlertDialog(
+        onDismissRequest = onDismiss,
+        title = { Text(title) },
+        text = { Text(text) },
+        confirmButton = { TextButton(onClick = onConfirm) { Text("Save") } },
+        dismissButton = { TextButton(onClick = onDismiss) { Text("Cancel") } },
+    )
 }
 
 @Composable
@@ -452,6 +482,10 @@ private fun ModelSettingsDialog(
     onSave: (Map<String, String>) -> Unit,
 ) {
     var params by remember(model.id) { mutableStateOf(model.settings) }
+    // Asked over this dialog rather than instead of it, so Cancel comes back to the edits rather
+    // than throwing them away.
+    var confirming by remember(model.id) { mutableStateOf(false) }
+    val loaded = model.status == "loaded" || model.status == "sleeping"
     AlertDialog(
         onDismissRequest = onDismiss,
         // Every control here is a number, so the keyboard is up for most of this dialog's life --
@@ -463,29 +497,35 @@ private fun ModelSettingsDialog(
         title = { Text(model.label) },
         text = {
             Column(Modifier.verticalScroll(rememberScrollState())) {
-                Text(
-                    if (model.status == "loaded" || model.status == "sleeping") {
-                        "This model is loaded. Saving takes it out of memory, and the sessions " +
-                            "using it load it again with these settings on their next message."
-                    } else {
-                        "Read when this model is next loaded."
-                    },
-                    style = MaterialTheme.typography.bodySmall,
-                    color = MaterialTheme.colorScheme.onSurfaceVariant,
-                )
-                Spacer(Modifier.height(12.dp))
                 ProviderParamFields(
                     specs = specs,
                     values = params,
                     onChange = { params = it },
-                    // Every one of these is read at load time, and the sentence above already
-                    // says when that is -- marking each control "on restart" would repeat it six
-                    // times.
+                    // Every one of these is read at load time, which is what Save stops to say --
+                    // marking each control "on restart" would repeat it six times over.
                     warnAboutRestart = false,
                 )
             }
         },
-        confirmButton = { TextButton(onClick = { onSave(params) }) { Text("Save") } },
+        confirmButton = {
+            TextButton(onClick = { if (loaded) confirming = true else onSave(params) }) {
+                Text("Save")
+            }
+        },
         dismissButton = { TextButton(onClick = onDismiss) { Text("Cancel") } },
     )
+    if (confirming) {
+        RestartDialog(
+            title = "Unload ${model.label}?",
+            text =
+                "It is in memory now, and these are read when it is loaded. Saving takes it out " +
+                    "of memory; the sessions using it load it again with these settings on their " +
+                    "next message.",
+            onConfirm = {
+                confirming = false
+                onSave(params)
+            },
+            onDismiss = { confirming = false },
+        )
+    }
 }

app/androidApp/src/main/kotlin/com/example/aiapp/SessionScreen.kt

@@ -2506,6 +2506,7 @@ fun SessionScreen(
             params = params,
             onParamsChanged = { params = it },
             cachedBytes = cachedBytes,
+            transcriptBytes = summary.transcriptFile?.bytes,
             // The purge finishes before the epoch moves, because the relaunched opening effect
             // reads the same directory and would otherwise draw what is about to be deleted. The
             // epoch is what makes the rest a cold open.

app/androidApp/src/main/kotlin/com/example/aiapp/SessionSettingsScreen.kt

@@ -116,6 +116,11 @@ fun SessionSettingsScreen(
      * the Reload row below, which is what would discard it.
      */
     cachedBytes: Long?,
+    /**
+     * How big the record on the server is, or null where it did not say. The other half of the pair
+     * beside it: what the conversation costs there, against what this phone is holding of it.
+     */
+    transcriptBytes: Long?,
     onReload: () -> Unit,
     /**
      * Opens the transcript file itself in the explorer. Null from a server that does not say where
@@ -569,24 +574,39 @@ fun SessionSettingsScreen(
                     modifier = Modifier.fillMaxWidth(),
                 ) {
                     Text("Transcript", modifier = Modifier.weight(1f))
-                    // The size is what the button discards, and the unknown state is drawn rather
-                    // than guessed: a spinner while the directory is being measured, and words when
-                    // there is nothing there, because "nothing cached" and "0 B" read as different
-                    // claims.
+                    // What the conversation costs on the server, and then what Reload would
+                    // discard here -- one line, so the two sizes read as a pair. A server that
+                    // did not measure its file leaves its half out rather than saying zero.
+                    val onServer = transcriptBytes?.let { humanSize(it) ?: "0 B" }
+                    // The unknown state is drawn rather than guessed: a spinner while the cache
+                    // is being measured, and words when there is nothing in it, because "nothing
+                    // cached" and "0 B" read as different claims.
                     when {
-                        cachedBytes == null ->
+                        cachedBytes == null -> {
+                            onServer?.let {
+                                Text(
+                                    it,
+                                    style = MaterialTheme.typography.bodySmall,
+                                    color = MaterialTheme.colorScheme.onSurfaceVariant,
+                                )
+                                Spacer(Modifier.width(8.dp))
+                            }
                             CircularProgressIndicator(
                                 modifier = Modifier.width(16.dp).height(16.dp),
                                 strokeWidth = 2.dp,
                             )
-                        else ->
+                        }
+                        else -> {
+                            val cached =
+                                humanSize(cachedBytes)?.let { "$it cached" } ?: "nothing cached"
                             Text(
-                                humanSize(cachedBytes)?.let { "$it cached" } ?: "nothing cached",
+                                onServer?.let { "$it · $cached" } ?: cached,
                                 style = MaterialTheme.typography.bodySmall,
                                 color = MaterialTheme.colorScheme.onSurfaceVariant,
                             )
                         }
                     }
+                }
                 // Both on a line of their own under what they act on, rather than crowded against
                 // the size on the line above: two buttons and a measurement do not fit the width
                 // of a phone, and the one that would lose is the number.

server/src/models.rs

@@ -74,7 +74,10 @@ pub struct LocalModel {
 pub fn labels(models: &[LocalModel]) -> Vec<String> {
     let rungs = |model: &LocalModel| {
         [
-            model.name.clone(),
+            model
+                .name
+                .clone()
+                .filter(|name| names_the_file(name, model)),
             Some(model.file.trim_end_matches(".gguf").to_string()),
             Some(model.key.clone()),
         ]
@@ -106,6 +109,35 @@ pub fn labels(models: &[LocalModel]) -> Vec<String> {
         .collect()
 }
 
+/// Whether a model's own `general.name` is a name for *this* model, or the
+/// converter's working directory wearing the same field.
+///
+/// The field is filled in by whatever produced the file, and
+/// `convert_hf_to_gguf.py` fills it from the directory it converted -- so a
+/// model converted out of a folder called `hf` publishes `general.name = "hf"`,
+/// which is unique, passes the cascade above, and tells a reader nothing at
+/// all. Prism ML's Bonsai is the one here (reported 2026-09-21, drawn as "hf"
+/// in the model picker).
+///
+/// The test is corroboration rather than a list of words to distrust: a real
+/// name shares something with where the file came from, both being about the
+/// same model, and a directory name picked up in passing does not. One word in
+/// common is enough. A name that fails it drops to the next rung, which is the
+/// file name the reader downloaded.
+fn names_the_file(name: &str, model: &LocalModel) -> bool {
+    let words = |text: &str| -> Vec<String> {
+        text.split(|c: char| !c.is_ascii_alphanumeric())
+            .filter(|word| word.len() > 1)
+            .map(str::to_ascii_lowercase)
+            .collect()
+    };
+    let from = words(&model.repo);
+    let from_file = words(&model.file);
+    words(name)
+        .iter()
+        .any(|word| from.contains(word) || from_file.contains(word))
+}
+
 /// What a download is doing, or did.
 ///
 /// Flat rather than a tagged enum carrying its message, because the phone
@@ -702,3 +734,64 @@ pub fn urlencode(value: &str) -> String {
         })
         .collect()
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn model(repo: &str, file: &str, name: Option<&str>) -> LocalModel {
+        LocalModel {
+            key: format!("{repo}/{file}"),
+            repo: repo.to_string(),
+            file: file.to_string(),
+            bytes: 1,
+            name: name.map(str::to_string),
+        }
+    }
+
+    /// The cascade, and the one rung that is not simply "is it unique": a name
+    /// the file it came from says nothing about is the converter's working
+    /// directory rather than the model's name.
+    #[test]
+    fn a_model_is_labelled_by_the_first_thing_that_identifies_it() {
+        let models = [
+            // Its own name, which is what a reader recognises.
+            model(
+                "Qwen/Qwen3-0.6B-GGUF",
+                "Qwen3-0.6B-Q8_0.gguf",
+                Some("Qwen3-0.6B"),
+            ),
+            // Two quantisations carry one name, so both fall to the file.
+            model("org/Big-GGUF", "Big-Q4_K_M.gguf", Some("Big")),
+            model("org/Big-GGUF", "Big-Q8_0.gguf", Some("Big")),
+            // `convert_hf_to_gguf.py` naming the directory it converted.
+            model("PrismML/Bonsai-GGUF", "Bonsai-1B-TQ1_0.gguf", Some("hf")),
+            // Nothing to go on but where it came from.
+            model("org/Quiet-GGUF", "weights.gguf", None),
+        ];
+        assert_eq!(
+            labels(&models),
+            [
+                "Qwen3-0.6B",
+                "Big-Q4_K_M",
+                "Big-Q8_0",
+                "Bonsai-1B-TQ1_0",
+                "weights",
+            ],
+        );
+    }
+
+    /// Two repos holding a file of the same name: the file rung collides as
+    /// well, and the key is what always terminates the cascade.
+    #[test]
+    fn a_file_name_two_repos_share_falls_through_to_the_key() {
+        let models = [
+            model("one/GGUF", "model.gguf", None),
+            model("two/GGUF", "model.gguf", None),
+        ];
+        assert_eq!(
+            labels(&models),
+            ["one/GGUF/model.gguf", "two/GGUF/model.gguf"]
+        );
+    }
+}

server/src/routes.rs

@@ -39,8 +39,8 @@
 //! GET    /sessions                     list (id, provider, title, model, status, last activity)
 //! GET    /sessions/{id}                one session, for refetching after a change. Its
 //!                                     `transcriptFile` is where the record itself is --
-//!                                     {machine, machineName, path} -- so the explorer can be
-//!                                     pointed at it
+//!                                     {machine, machineName, path, bytes?} -- so the explorer
+//!                                     can be pointed at it and its size said
 //! POST   /sessions                     spawn {machine, provider, title?, model?, cwd?, params?}
 //! POST   /sessions/order               {sessions} -- the list in the order it is drawn in,
 //!                                     as the reader dragged it; ids left out keep their

server/src/session/llama/mod.rs

@@ -220,11 +220,46 @@ enum Content {
     Parts(Vec<Value>),
 }
 
+/// llama.cpp's placeholder for a picture, as it appears *inside the prompt
+/// text* once the server has taken the image out of the request.
+///
+/// It is an in-band signal in the band that also carries what everybody typed,
+/// so it has to be taken back out of anything that is merely words.
+/// `mtmd_tokenize` splits the rendered prompt on it and pairs each occurrence
+/// with a decoded image, so one nobody sent an image for fails the turn --
+/// `number of media markers in text (1) exceeds number of bitmaps (0)`, which
+/// reaches the phone as "Failed to tokenize prompt" and then fails every later
+/// message too, the words being in the conversation for good. A model that
+/// says the marker back is enough to do it.
+///
+/// llama.cpp's own default, which is what it uses unless started with
+/// `--media-marker`; nothing here passes that, and the router writes the preset
+/// its children are started from (see [`router`]).
+const MEDIA_MARKER: &str = "<__media__>";
+
+/// `text` with any [`MEDIA_MARKER`] removed, for content that is words rather
+/// than a picture.
+///
+/// Removed rather than escaped: there is nothing to escape it *to* -- the
+/// marker is matched literally, and the only form llama.cpp will not read as a
+/// picture is its absence.
+fn without_marker(text: impl Into<String>) -> String {
+    let text = text.into();
+    if text.contains(MEDIA_MARKER) {
+        return text.replace(MEDIA_MARKER, "");
+    }
+    text
+}
+
 impl Message {
+    /// Words, with anything llama.cpp would read as a picture taken out of
+    /// them -- see [`without_marker`]. Every message with text goes through
+    /// here, which is what makes that structural rather than a rule to
+    /// remember at each call.
     fn new(role: &str, content: impl Into<String>) -> Self {
         Self {
             role: role.to_string(),
-            content: Content::Text(content.into()),
+            content: Content::Text(without_marker(content)),
             tool_calls: Vec::new(),
             tool_call_id: None,
         }
@@ -239,7 +274,7 @@ impl Message {
     /// templates, and only the first has ever been tested by everything else
     /// here.
     fn from_user(text: impl Into<String>, images: Vec<Value>) -> Self {
-        let text = text.into();
+        let text = without_marker(text);
         if images.is_empty() {
             return Self::new("user", text);
         }
@@ -3252,6 +3287,30 @@ mod tests {
         );
     }
 
+    /// llama.cpp's own picture placeholder is not content, and words that
+    /// contain it are still words: left in, one of them fails the turn it is
+    /// sent in and every turn after it, because the conversation is folded out
+    /// of a transcript that now holds it for good.
+    #[test]
+    fn a_media_marker_in_words_never_reaches_the_model() {
+        assert_eq!(
+            Message::new("user", format!("what is {MEDIA_MARKER} this?")).content,
+            Content::Text("what is  this?".to_string()),
+        );
+        // A tool's output is the other text somebody else wrote.
+        assert_eq!(
+            Message::result_of("call_1", MEDIA_MARKER).content,
+            Content::Text(String::new()),
+        );
+        // Beside a real picture, which is the part that says there is one.
+        let image =
+            json!({"type": "image_url", "image_url": {"url": "data:image/png;base64,UE5H"}});
+        assert_eq!(
+            Message::from_user(MEDIA_MARKER, vec![image.clone()]).content,
+            Content::Parts(vec![image]),
+        );
+    }
+
     /// A file that cannot be read is left out rather than failing the turn:
     /// the send path has already said so, and a turn is a worse place to learn
     /// it.

server/src/session/mod.rs

@@ -342,6 +342,12 @@ pub struct TranscriptFile {
     pub machine: String,
     pub machine_name: String,
     pub path: String,
+    /// How big the record is, for a reader deciding whether to open it.
+    /// Absent where the file could not be measured, which is a session that
+    /// has not written one yet -- and is deliberately not a zero, since "no
+    /// transcript" and "we could not look" are different answers.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub bytes: Option<u64>,
 }
 
 /// Where a session directory keeps its transcript.
@@ -352,12 +358,12 @@ fn transcript_in(dir: &Path) -> PathBuf {
 /// Where `id`'s transcript is, said the way the explorer takes it.
 fn transcript_file(config: &Config, data_dir: &Path, id: &str) -> Option<TranscriptFile> {
     let machine = config.machine(crate::config::LOCAL_MACHINE_ID)?;
+    let path = transcript_in(&data_dir.join(id));
     Some(TranscriptFile {
         machine: machine.id.clone(),
         machine_name: machine.name.clone(),
-        path: transcript_in(&data_dir.join(id))
-            .to_string_lossy()
-            .into_owned(),
+        bytes: std::fs::metadata(&path).ok().map(|file| file.len()),
+        path: path.to_string_lossy().into_owned(),
     })
 }