[feat] add reqwest and sha2 dependencies to support new features in transcript processing

Cargo.toml

@@ -10,6 +10,8 @@ serde = { version = "1.0.219", features = ["derive"] }
 serde_json = "1.0.142"
 toml = "0.8"
 chrono = { version = "0.4", features = ["clock"] }
+reqwest = { version = "0.12", features = ["blocking", "json"] }
+sha2 = "0.10"
 whisper-rs = { git = "https://github.com/tazz4843/whisper-rs", optional = true }
 
 [features]

src/main.rs

@@ -2,11 +2,17 @@ use std::fs::{File, create_dir_all};
 use std::io::{self, Read, Write};
 use std::path::{Path, PathBuf};
 use std::process::Command;
+use std::env;
+use std::collections::BTreeMap;
 
 use anyhow::{anyhow, Context, Result};
 use clap::Parser;
 use serde::{Deserialize, Serialize};
 use chrono::Local;
+use reqwest::blocking::Client;
+use reqwest::redirect::Policy;
+use sha2::{Digest, Sha256};
+use std::time::Duration;
 
 #[cfg(feature = "native-whisper")]
 use whisper_rs::{FullParams, SamplingStrategy, WhisperContext, WhisperContextParameters};
@@ -15,7 +21,6 @@ use whisper_rs::{FullParams, SamplingStrategy, WhisperContext, WhisperContextPar
 #[command(name = "merge_transcripts", version, about = "Merge multiple JSON transcripts into one or transcribe audio using native whisper")]
 struct Args {
     /// Input .json transcript files or audio files to merge/transcribe
-    #[arg(required = true)]
     inputs: Vec<String>,
 
     /// Output file path base (date prefix will be added); if omitted, writes JSON to stdout
@@ -25,6 +30,10 @@ struct Args {
     /// Language code to use for transcription (e.g., en, de). No auto-detection.
     #[arg(short, long, value_name = "LANG")]
     language: Option<String>,
+
+    /// Launch interactive model downloader (list HF models, multi-select and download)
+    #[arg(long)]
+    download_models: bool,
 }
 
 #[derive(Debug, Deserialize)]
@@ -85,6 +94,15 @@ fn render_srt(items: &[OutputEntry]) -> String {
     out
 }
 
+fn sanitize_speaker_name(raw: &str) -> String {
+    if let Some((prefix, rest)) = raw.split_once('-') {
+        if !prefix.is_empty() && prefix.chars().all(|c| c.is_ascii_digit()) {
+            return rest.to_string();
+        }
+    }
+    raw.to_string()
+}
+
 // --- Helpers for audio transcription ---
 fn is_json_file(path: &Path) -> bool {
     matches!(path.extension().and_then(|s| s.to_str()).map(|s| s.to_lowercase()), Some(ext) if ext == "json")
@@ -127,8 +145,20 @@ fn normalize_lang_code(input: &str) -> Option<String> {
 fn find_model_file() -> Result<PathBuf> {
     let models_dir = Path::new("models");
     if !models_dir.exists() {
-        return Err(anyhow!("No models directory found at {}", models_dir.display()));
+        create_dir_all(models_dir).with_context(|| format!("Failed to create models directory: {}", models_dir.display()))?;
     }
+
+    // If env var WHISPER_MODEL is set and valid, prefer it
+    if let Ok(env_model) = env::var("WHISPER_MODEL") {
+        let p = PathBuf::from(env_model);
+        if p.is_file() {
+            // persist selection
+            let _ = std::fs::write(models_dir.join(".last_model"), p.display().to_string());
+            return Ok(p);
+        }
+    }
+
+    // Enumerate local models
     let mut candidates: Vec<PathBuf> = Vec::new();
     let rd = std::fs::read_dir(models_dir)
         .with_context(|| format!("Failed to read models directory: {}", models_dir.display()))?;
@@ -143,13 +173,64 @@ fn find_model_file() -> Result<PathBuf> {
             }
         }
     }
+
     if candidates.is_empty() {
-        return Err(anyhow!("No Whisper model files (*.bin) found in {}", models_dir.display()));
+        eprintln!("No Whisper model files (*.bin) found in {}.", models_dir.display());
+        eprint!("Would you like to download models now? [Y/n]: ");
+        io::stderr().flush().ok();
+        let mut input = String::new();
+        io::stdin().read_line(&mut input).ok();
+        let ans = input.trim().to_lowercase();
+        if ans.is_empty() || ans == "y" || ans == "yes" {
+            if let Err(e) = run_interactive_model_downloader() {
+                eprintln!("Downloader failed: {:#}", e);
+            }
+            // Re-scan
+            candidates.clear();
+            let rd2 = std::fs::read_dir(models_dir)
+                .with_context(|| format!("Failed to read models directory: {}", models_dir.display()))?;
+            for entry in rd2 {
+                let entry = entry?;
+                let path = entry.path();
+                if path.is_file() {
+                    if let Some(ext) = path.extension().and_then(|s| s.to_str()).map(|s| s.to_lowercase()) {
+                        if ext == "bin" {
+                            candidates.push(path);
+                        }
+                    }
+                }
+            }
+        }
     }
+
+    if candidates.is_empty() {
+        return Err(anyhow!("No Whisper model files (*.bin) available in {}", models_dir.display()));
+    }
+
+    // If only one, persist and return it
     if candidates.len() == 1 {
-        return Ok(candidates.remove(0));
+        let only = candidates.remove(0);
+        let _ = std::fs::write(models_dir.join(".last_model"), only.display().to_string());
+        return Ok(only);
     }
-    // Multiple models: prompt user to choose
+
+    // If a previous selection exists and is still valid, use it
+    let last_file = models_dir.join(".last_model");
+    if let Ok(prev) = std::fs::read_to_string(&last_file) {
+        let prev = prev.trim();
+        if !prev.is_empty() {
+            let p = PathBuf::from(prev);
+            if p.is_file() {
+                // Also ensure it's one of the candidates (same dir)
+                if candidates.iter().any(|c| c == &p) {
+                    eprintln!("Using previously selected model: {}", p.display());
+                    return Ok(p);
+                }
+            }
+        }
+    }
+
+    // Multiple models and no previous selection: prompt user to choose, then persist
     eprintln!("Multiple Whisper models found in {}:", models_dir.display());
     for (i, p) in candidates.iter().enumerate() {
         eprintln!("  {}) {}", i + 1, p.display());
@@ -162,7 +243,9 @@ fn find_model_file() -> Result<PathBuf> {
     if sel == 0 || sel > candidates.len() {
         return Err(anyhow!("Selection out of range"));
     }
-    Ok(candidates.swap_remove(sel - 1))
+    let chosen = candidates.swap_remove(sel - 1);
+    let _ = std::fs::write(models_dir.join(".last_model"), chosen.display().to_string());
+    Ok(chosen)
 }
 
 #[cfg(feature = "native-whisper")]
@@ -221,7 +304,16 @@ fn transcribe_native(audio_path: &Path, speaker: &str, lang_opt: Option<&str>) -
         }
     }
     let model_str = model.to_str().ok_or_else(|| anyhow!("Model path not valid UTF-8: {}", model.display()))?;
-    let ctx = WhisperContext::new_with_params(model_str, WhisperContextParameters::default())
+
+    // Initialize Whisper with GPU preference
+    let cparams = WhisperContextParameters::default();
+    // Prefer GPU if available; default whisper.cpp already has use_gpu=true. If the wrapper exposes
+    // a gpu_device field in the future, we could set it here from WHISPER_GPU_DEVICE.
+    if let Ok(dev_str) = env::var("WHISPER_GPU_DEVICE") {
+        let _ = dev_str.trim().parse::<i32>().ok();
+    }
+    // Even if we can't set fields explicitly (due to API differences), whisper.cpp defaults to GPU.
+    let ctx = WhisperContext::new_with_params(model_str, cparams)
         .with_context(|| format!("Failed to load Whisper model at {}", model.display()))?;
     let mut state = ctx.create_state()
         .map_err(|e| anyhow!("Failed to create Whisper state: {:?}", e))?;
@@ -252,6 +344,16 @@ fn transcribe_native(audio_path: &Path, speaker: &str, lang_opt: Option<&str>) -
 fn main() -> Result<()> {
     let args = Args::parse();
 
+    // If requested, run the interactive model downloader first. If no inputs were provided, exit after downloading.
+    if args.download_models {
+        if let Err(e) = run_interactive_model_downloader() {
+            eprintln!("Model downloader failed: {:#}", e);
+        }
+        if args.inputs.is_empty() {
+            return Ok(());
+        }
+    }
+
     // Determine inputs and optional output path
     let mut inputs = args.inputs;
     let mut output_path = args.output;
@@ -282,11 +384,11 @@ fn main() -> Result<()> {
 
     for input_path in &inputs {
         let path = Path::new(input_path);
-        let speaker = path
-            .file_stem()
-            .and_then(|s| s.to_str())
-            .unwrap_or("speaker")
-            .to_string();
+        let speaker = sanitize_speaker_name(
+            path.file_stem()
+                .and_then(|s| s.to_str())
+                .unwrap_or("speaker")
+        );
 
         let mut buf = String::new();
         if is_audio_file(path) {
@@ -400,3 +502,483 @@ fn main() -> Result<()> {
     }
     Ok(())
 }
+
+// --- Model downloader: list & download ggml models from Hugging Face ---
+
+#[derive(Debug, Deserialize)]
+struct HFLfsMeta {
+    oid: Option<String>,
+    size: Option<u64>,
+    sha256: Option<String>,
+}
+
+#[derive(Debug, Deserialize)]
+struct HFSibling {
+    rfilename: String,
+    size: Option<u64>,
+    sha256: Option<String>,
+    lfs: Option<HFLfsMeta>,
+}
+
+#[derive(Debug, Deserialize)]
+struct HFRepoInfo {
+    // When using ?expand=files the field is named 'siblings'
+    siblings: Option<Vec<HFSibling>>,
+}
+
+#[derive(Debug, Deserialize)]
+struct HFTreeItem {
+    path: String,
+    size: Option<u64>,
+    sha256: Option<String>,
+    lfs: Option<HFLfsMeta>,
+}
+
+#[derive(Clone, Debug)]
+struct ModelEntry {
+    // e.g. "tiny.en-q5_1"
+    name: String,
+    base: String,
+    subtype: String,
+    size: u64,
+    sha256: Option<String>,
+    repo: &'static str, // e.g. "ggerganov/whisper.cpp"
+}
+
+fn split_model_name(model: &str) -> (String, String) {
+    let mut idx = None;
+    for (i, ch) in model.char_indices() {
+        if ch == '.' || ch == '-' {
+            idx = Some(i);
+            break;
+        }
+    }
+    if let Some(i) = idx {
+        (model[..i].to_string(), model[i + 1..].to_string())
+    } else {
+        (model.to_string(), String::new())
+    }
+}
+
+fn human_size(bytes: u64) -> String {
+    const KB: f64 = 1024.0;
+    const MB: f64 = KB * 1024.0;
+    const GB: f64 = MB * 1024.0;
+    let b = bytes as f64;
+    if b >= GB { format!("{:.2} GiB", b / GB) }
+    else if b >= MB { format!("{:.2} MiB", b / MB) }
+    else if b >= KB { format!("{:.2} KiB", b / KB) }
+    else { format!("{} B", bytes) }
+}
+
+fn to_hex_lower(bytes: &[u8]) -> String {
+    let mut s = String::with_capacity(bytes.len() * 2);
+    for b in bytes { s.push_str(&format!("{:02x}", b)); }
+    s
+}
+
+fn expected_sha_from_sibling(s: &HFSibling) -> Option<String> {
+    if let Some(h) = &s.sha256 { return Some(h.to_lowercase()); }
+    if let Some(lfs) = &s.lfs {
+        if let Some(h) = &lfs.sha256 { return Some(h.to_lowercase()); }
+        if let Some(oid) = &lfs.oid {
+            // e.g. "sha256:abcdef..."
+            if let Some(rest) = oid.strip_prefix("sha256:") {
+                return Some(rest.to_lowercase().to_string());
+            }
+        }
+    }
+    None
+}
+
+fn size_from_sibling(s: &HFSibling) -> Option<u64> {
+    if let Some(sz) = s.size { return Some(sz); }
+    if let Some(lfs) = &s.lfs { return lfs.size; }
+    None
+}
+
+fn expected_sha_from_tree(s: &HFTreeItem) -> Option<String> {
+    if let Some(h) = &s.sha256 { return Some(h.to_lowercase()); }
+    if let Some(lfs) = &s.lfs {
+        if let Some(h) = &lfs.sha256 { return Some(h.to_lowercase()); }
+        if let Some(oid) = &lfs.oid {
+            if let Some(rest) = oid.strip_prefix("sha256:") {
+                return Some(rest.to_lowercase().to_string());
+            }
+        }
+    }
+    None
+}
+
+fn size_from_tree(s: &HFTreeItem) -> Option<u64> {
+    if let Some(sz) = s.size { return Some(sz); }
+    if let Some(lfs) = &s.lfs { return lfs.size; }
+    None
+}
+
+fn fill_meta_via_head(repo: &'static str, name: &str) -> (Option<u64>, Option<String>) {
+    let head_client = match Client::builder()
+        .user_agent("dialogue_merger/0.1 (+https://github.com/)")
+        .redirect(Policy::none())
+        .timeout(Duration::from_secs(30))
+        .build() {
+        Ok(c) => c,
+        Err(_) => return (None, None),
+    };
+    let url = format!("https://huggingface.co/{}/resolve/main/ggml-{}.bin", repo, name);
+    let resp = match head_client.head(url).send().and_then(|r| r.error_for_status()) {
+        Ok(r) => r,
+        Err(_) => return (None, None),
+    };
+    let headers = resp.headers();
+    let size = headers
+        .get("x-linked-size")
+        .and_then(|v| v.to_str().ok())
+        .and_then(|s| s.parse::<u64>().ok());
+    let mut sha = headers
+        .get("x-linked-etag")
+        .and_then(|v| v.to_str().ok())
+        .map(|s| s.trim().trim_matches('"').to_string());
+    if let Some(h) = &mut sha {
+        h.make_ascii_lowercase();
+        if h.len() != 64 || !h.chars().all(|c| c.is_ascii_hexdigit()) {
+            sha = None;
+        }
+    }
+    // Fallback: try x-xet-hash header if x-linked-etag is missing/invalid
+    if sha.is_none() {
+        sha = headers
+            .get("x-xet-hash")
+            .and_then(|v| v.to_str().ok())
+            .map(|s| s.trim().trim_matches('"').to_string());
+        if let Some(h) = &mut sha {
+            h.make_ascii_lowercase();
+            if h.len() != 64 || !h.chars().all(|c| c.is_ascii_hexdigit()) {
+                sha = None;
+            }
+        }
+    }
+    (size, sha)
+}
+
+fn hf_fetch_repo_models(client: &Client, repo: &'static str) -> Result<Vec<ModelEntry>> {
+    eprintln!("Fetching online data: listing models from {}...", repo);
+    // Prefer the tree endpoint for reliable size/hash metadata, then fall back to model metadata
+    let tree_url = format!("https://huggingface.co/api/models/{}/tree/main?recursive=1", repo);
+    let mut out: Vec<ModelEntry> = Vec::new();
+
+    match client.get(tree_url).send().and_then(|r| r.error_for_status()) {
+        Ok(resp) => {
+            match resp.json::<Vec<HFTreeItem>>() {
+                Ok(items) => {
+                    for it in items {
+                        let path = it.path.clone();
+                        if !(path.starts_with("ggml-") && path.ends_with(".bin")) { continue; }
+                        let model_name = path.trim_start_matches("ggml-").trim_end_matches(".bin").to_string();
+                        let (base, subtype) = split_model_name(&model_name);
+                        let size = size_from_tree(&it).unwrap_or(0);
+                        let sha256 = expected_sha_from_tree(&it);
+                        out.push(ModelEntry { name: model_name, base, subtype, size, sha256, repo });
+                    }
+                }
+                Err(_) => { /* fall back below */ }
+            }
+        }
+        Err(_) => { /* fall back below */ }
+    }
+
+    if out.is_empty() {
+        let url = format!("https://huggingface.co/api/models/{}", repo);
+        let resp = client
+            .get(url)
+            .send()
+            .and_then(|r| r.error_for_status())
+            .context("Failed to query Hugging Face API")?;
+
+        let info: HFRepoInfo = resp.json().context("Failed to parse Hugging Face API response")?;
+
+        if let Some(files) = info.siblings {
+            for s in files {
+                let rf = s.rfilename.clone();
+                if !(rf.starts_with("ggml-") && rf.ends_with(".bin")) { continue; }
+                let model_name = rf.trim_start_matches("ggml-").trim_end_matches(".bin").to_string();
+                let (base, subtype) = split_model_name(&model_name);
+                let size = size_from_sibling(&s).unwrap_or(0);
+                let sha256 = expected_sha_from_sibling(&s);
+                out.push(ModelEntry { name: model_name, base, subtype, size, sha256, repo });
+            }
+        }
+    }
+
+    // Fill missing metadata (size/hash) via HEAD request if necessary
+    if out.iter().any(|m| m.size == 0 || m.sha256.is_none()) {
+        eprintln!("Fetching online data: completing metadata checks for models in {}...", repo);
+    }
+    for m in out.iter_mut() {
+        if m.size == 0 || m.sha256.is_none() {
+            let (sz, sha) = fill_meta_via_head(m.repo, &m.name);
+            if m.size == 0 {
+                if let Some(s) = sz { m.size = s; }
+            }
+            if m.sha256.is_none() {
+                m.sha256 = sha;
+            }
+        }
+    }
+
+    // Sort by base then subtype then name for stable listing
+    out.sort_by(|a, b| a.base.cmp(&b.base).then(a.subtype.cmp(&b.subtype)).then(a.name.cmp(&b.name)));
+    Ok(out)
+}
+
+
+fn fetch_all_models(client: &Client) -> Result<Vec<ModelEntry>> {
+    eprintln!("Fetching online data: aggregating available models from Hugging Face...");
+    let mut v1 = hf_fetch_repo_models(client, "ggerganov/whisper.cpp")?; // main repo must succeed
+
+    // Optional tinydiarize repo; ignore errors but log to stderr
+    let mut v2: Vec<ModelEntry> = match hf_fetch_repo_models(client, "akashmjn/tinydiarize-whisper.cpp") {
+        Ok(v) => v,
+        Err(e) => {
+            eprintln!("Warning: failed to fetch optional repo akashmjn/tinydiarize-whisper.cpp: {:#}", e);
+            Vec::new()
+        }
+    };
+
+    v1.append(&mut v2);
+
+    // Deduplicate by name preferring ggerganov repo if duplicates
+    let mut map: BTreeMap<String, ModelEntry> = BTreeMap::new();
+    for m in v1 {
+        map.entry(m.name.clone()).and_modify(|existing| {
+            if existing.repo != "ggerganov/whisper.cpp" && m.repo == "ggerganov/whisper.cpp" {
+                *existing = m.clone();
+            }
+        }).or_insert(m);
+    }
+
+    let mut list: Vec<ModelEntry> = map.into_values().collect();
+    list.sort_by(|a, b| a.base.cmp(&b.base).then(a.subtype.cmp(&b.subtype)).then(a.name.cmp(&b.name)));
+    Ok(list)
+}
+
+fn print_grouped_models(models: &[ModelEntry]) {
+    let mut current = "".to_string();
+    for m in models {
+        if m.base != current {
+            current = m.base.clone();
+            println!("\n{}:", current);
+        }
+        let short_hash = m
+            .sha256
+            .as_ref()
+            .map(|h| h.chars().take(8).collect::<String>())
+            .unwrap_or_else(|| "-".to_string());
+        println!("  - {}  [{} | {} | {}]", m.name, m.repo, human_size(m.size), short_hash);
+    }
+    println!("\nEnter selection by indices (e.g., 1 3 5-7), or 'all', '*' for all, 'q' to cancel.");
+}
+
+fn prompt_select_models(models: &[ModelEntry]) -> Result<Vec<ModelEntry>> {
+    // Build a flat list but show group headers; indices count only models
+    println!("Available ggml Whisper models:");
+    let mut idx = 1usize;
+    let mut current = "".to_string();
+    // We'll record mapping from index -> position in models
+    let mut index_map: Vec<usize> = Vec::with_capacity(models.len());
+    for (pos, m) in models.iter().enumerate() {
+        if m.base != current {
+            current = m.base.clone();
+            println!("\n{}:", current);
+        }
+        let short_hash = m
+            .sha256
+            .as_ref()
+            .map(|h| h.chars().take(8).collect::<String>())
+            .unwrap_or_else(|| "-".to_string());
+        println!("  {}) {}  [{} | {} | {}]", idx, m.name, m.repo, human_size(m.size), short_hash);
+        index_map.push(pos);
+        idx += 1;
+    }
+    println!("\nEnter selection by indices (e.g., 1 3 5-7), or 'all', '*' for all, 'q' to cancel.");
+    loop {
+        eprint!("Selection: ");
+        io::stderr().flush().ok();
+        let mut line = String::new();
+        io::stdin().read_line(&mut line).context("Failed to read selection")?;
+        let s = line.trim().to_lowercase();
+        if s == "q" || s == "quit" || s == "exit" { return Ok(Vec::new()); }
+        let mut selected: Vec<usize> = Vec::new();
+        if s == "all" || s == "*" {
+            selected = (1..idx).collect();
+        } else if !s.is_empty() {
+            for part in s.split(|c| c == ',' || c == ' ' || c == ';') {
+                let part = part.trim();
+                if part.is_empty() { continue; }
+                if let Some((a, b)) = part.split_once('-') {
+                    if let (Ok(ia), Ok(ib)) = (a.parse::<usize>(), b.parse::<usize>()) {
+                        if ia >= 1 && ib < idx && ia <= ib { selected.extend(ia..=ib); }
+                    }
+                } else if let Ok(i) = part.parse::<usize>() {
+                    if i >= 1 && i < idx { selected.push(i); }
+                }
+            }
+        }
+        selected.sort_unstable();
+        selected.dedup();
+        if selected.is_empty() {
+            eprintln!("No valid selection. Please try again or 'q' to cancel.");
+            continue;
+        }
+        let chosen: Vec<ModelEntry> = selected.into_iter().map(|i| models[index_map[i - 1]].clone()).collect();
+        return Ok(chosen);
+    }
+}
+
+fn compute_file_sha256_hex(path: &Path) -> Result<String> {
+    let file = File::open(path)
+        .with_context(|| format!("Failed to open for hashing: {}", path.display()))?;
+    let mut reader = std::io::BufReader::new(file);
+    let mut hasher = Sha256::new();
+    let mut buf = [0u8; 1024 * 128];
+    loop {
+        let n = reader.read(&mut buf).context("Read error during hashing")?;
+        if n == 0 { break; }
+        hasher.update(&buf[..n]);
+    }
+    Ok(to_hex_lower(&hasher.finalize()))
+}
+
+fn download_one_model(client: &Client, models_dir: &Path, entry: &ModelEntry) -> Result<()> {
+    let final_path = models_dir.join(format!("ggml-{}.bin", entry.name));
+
+    // If the model already exists, verify against online metadata
+    if final_path.exists() {
+        if let Some(expected) = &entry.sha256 {
+            match compute_file_sha256_hex(&final_path) {
+                Ok(local_hash) => {
+                    if local_hash.eq_ignore_ascii_case(expected) {
+                        eprintln!("Model {} is up-to-date (hash match).", final_path.display());
+                        return Ok(());
+                    } else {
+                        eprintln!(
+                            "Local model {} hash differs from online (local {}.., online {}..). Updating...",
+                            final_path.display(),
+                            &local_hash[..std::cmp::min(8, local_hash.len())],
+                            &expected[..std::cmp::min(8, expected.len())]
+                        );
+                    }
+                }
+                Err(e) => {
+                    eprintln!(
+                        "Warning: failed to hash existing {}: {}. Will re-download to ensure correctness.",
+                        final_path.display(), e
+                    );
+                }
+            }
+        } else if entry.size > 0 {
+            match std::fs::metadata(&final_path) {
+                Ok(md) => {
+                    if md.len() == entry.size {
+                        eprintln!(
+                            "Model {} appears up-to-date by size ({}).",
+                            final_path.display(), entry.size
+                        );
+                        return Ok(());
+                    } else {
+                        eprintln!(
+                            "Local model {} size ({}) differs from online ({}). Updating...",
+                            final_path.display(), md.len(), entry.size
+                        );
+                    }
+                }
+                Err(e) => {
+                    eprintln!(
+                        "Warning: failed to stat existing {}: {}. Will re-download to ensure correctness.",
+                        final_path.display(), e
+                    );
+                }
+            }
+        } else {
+            eprintln!(
+                "Model {} exists but remote hash/size not available; will download to verify contents.",
+                final_path.display()
+            );
+            // Fall through to download for content comparison
+        }
+    }
+
+    let url = format!("https://huggingface.co/{}/resolve/main/ggml-{}.bin", entry.repo, entry.name);
+    eprintln!("Downloading {} ({} | {})...", entry.name, human_size(entry.size), url);
+    let mut resp = client
+        .get(url)
+        .send()
+        .and_then(|r| r.error_for_status())
+        .context("Failed to download model")?;
+
+    let tmp_path = models_dir.join(format!("ggml-{}.bin.part", entry.name));
+    if tmp_path.exists() {
+        let _ = std::fs::remove_file(&tmp_path);
+    }
+    let mut file = std::io::BufWriter::new(
+        File::create(&tmp_path)
+            .with_context(|| format!("Failed to create {}", tmp_path.display()))?
+    );
+
+    let mut hasher = Sha256::new();
+    let mut buf = [0u8; 1024 * 128];
+    loop {
+        let n = resp.read(&mut buf).context("Network read error")?;
+        if n == 0 { break; }
+        hasher.update(&buf[..n]);
+        file.write_all(&buf[..n]).context("Write error")?;
+    }
+    file.flush().ok();
+
+    let got = to_hex_lower(&hasher.finalize());
+    if let Some(expected) = &entry.sha256 {
+        if got != expected.to_lowercase() {
+            let _ = std::fs::remove_file(&tmp_path);
+            return Err(anyhow!(
+                "SHA-256 mismatch for {}: expected {}, got {}",
+                entry.name, expected, got
+            ));
+        }
+    } else {
+        eprintln!("Warning: no SHA-256 available for {}. Skipping verification.", entry.name);
+    }
+    // Replace existing file safely
+    if final_path.exists() {
+        let _ = std::fs::remove_file(&final_path);
+    }
+    std::fs::rename(&tmp_path, &final_path)
+        .with_context(|| format!("Failed to move into place: {}", final_path.display()))?;
+    eprintln!("Saved: {}", final_path.display());
+    Ok(())
+}
+
+fn run_interactive_model_downloader() -> Result<()> {
+    let models_dir = Path::new("models");
+    if !models_dir.exists() { create_dir_all(models_dir).context("Failed to create models directory")?; }
+    let client = Client::builder()
+        .user_agent("dialogue_merger/0.1 (+https://github.com/)")
+        .timeout(std::time::Duration::from_secs(600))
+        .build()
+        .context("Failed to build HTTP client")?;
+
+    eprintln!("Fetching online data: contacting Hugging Face to retrieve available models (this may take a moment)...");
+    let models = fetch_all_models(&client)?;
+    if models.is_empty() {
+        eprintln!("No models found on Hugging Face listing. Please try again later.");
+        return Ok(());
+    }
+    let selected = prompt_select_models(&models)?;
+    if selected.is_empty() {
+        eprintln!("No selection. Aborting download.");
+        return Ok(());
+    }
+    for m in selected {
+        if let Err(e) = download_one_model(&client, models_dir, &m) { eprintln!("Error: {:#}", e); }
+    }
+    Ok(())
+}