Files
Loki/rust/loki-rs/src/chat/host.rs
T
Claude fea055fd52 Lot 6 : le tour branché sur le moteur, les outils et la base (chat::host)
GoTurnHost est le TurnHost de la façade : envoi par EngineHttp (clé de
l'endpoint, pannes classées comme friendlyLLMError), outils par l'API du Go
avec leurs suites (capture, aperçu, image relayée), résultats complets
confiés à l'écrivain des discussions (id « <discussion>.<aléa> »),
compaction par Compactor avec un résumeur qui interroge le moteur
(EngineSummarizer) et l'archive recall du Go (StoreRecall), images jointes
lues sur disque et ouvertes par le Go, en cache borné à 64 Mio.

RecallArchive devient asynchrone ; la boucle consulte la mémoire des
refus de compaction avant d'annoncer une compaction en cours de tour.

Écarts : pas de résumé en continuation (état du slot tenu par le Go), ni
conseil de cache, ni garde d'écriture du mode Code.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016tcGfJWefW5kUGoC2wL8cv
2026-10-15 10:34:15 +00:00

581 lines
19 KiB
Rust

//! Le tour branché sur le vrai monde : le [`TurnHost`] de la façade.
//!
//! - envoi au moteur par [`EngineHttp`] (clé de CET endpoint), pannes
//! classées et formulées comme `friendlyLLMError` ;
//! - outils exécutés par le Go (`/internal/v1/tool/exec`), avec ce qu'il tire
//! du résultat (capture de la bulle, aperçu du navigateur, image relayée) ;
//! - résultats complets (« voir plus ») confiés à l'écrivain des discussions ;
//! - compaction par [`Compactor`] : résumé demandé au moteur
//! ([`EngineSummarizer`]), blocs archivés pour recall par le Go
//! ([`StoreRecall`]) ;
//! - images jointes (`loki-img:…`) lues sur disque et ouvertes par le Go
//! (déchiffrées), gardées en cache comme `chatImageDataURL`.
//!
//! Écarts assumés : le résumé en continuation (COMPACT_CONTINUATION) demande
//! l'état du slot du moteur, tenu par le Go — `slot_holds` rend faux et la
//! compaction résume la transcription. Ni conseil de cache ni garde
//! d'écriture du mode Code (état du Go) : `cache_hint` et `preflight` absents.
use std::collections::{HashMap, VecDeque};
use std::path::PathBuf;
use std::sync::{Arc, Mutex};
use futures_util::StreamExt;
use serde_json::{Map, Value, json};
use super::compact::{
CompactCaps, CompactConfig, CompactMemory, CompactOpts, Compactor, ContinuationRequest,
RecallArchive, Summarizer, SummaryChoice, SummaryError, TranscriptRequest, compact_opts_mid,
parse_summary_response,
};
use super::engine::{EngineHttp, classify, friendly_error};
use super::persist::{Persister, StoreSink, ToolResJob};
use super::stream::{CacheHintFn, PreflightFn, RetryPatterns};
use super::turn::{CompactPhase, Reply, SendError, ToolOutcome, TurnHost};
use super::types::{AskEvent, DiffLine, Message, Tool};
use super::wire::{
ChatEndpoint, ChatPayloadOpts, EchoRefusals, EffortFallbacks, IMG_REF_SCHEME,
build_chat_payload, content_parts, part_image_url, wire_messages,
};
use crate::store::Store;
/// Taille au-delà de laquelle le cache d'images oublie les plus anciennes
/// (`imgCacheMax`).
pub const IMG_CACHE_MAX: usize = 64 << 20;
/// Corps d'un refus de résumé gardé au plus (`postSummary`).
const SUMMARY_ERR_MAX: usize = 500;
/// Base64 standard (avec remplissage), comme `base64.StdEncoding`.
pub fn b64(data: &[u8]) -> String {
const T: &[u8; 64] = b"ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
let mut out = String::with_capacity(data.len().div_ceil(3) * 4);
for c in data.chunks(3) {
let n = (c[0] as u32) << 16
| (*c.get(1).unwrap_or(&0) as u32) << 8
| *c.get(2).unwrap_or(&0) as u32;
out.push(T[(n >> 18) as usize & 63] as char);
out.push(T[(n >> 12) as usize & 63] as char);
out.push(if c.len() > 1 {
T[(n >> 6) as usize & 63] as char
} else {
'='
});
out.push(if c.len() > 2 {
T[n as usize & 63] as char
} else {
'='
});
}
out
}
/// Type MIME d'une image rangée, d'après son extension (`imgExtByMime`).
fn image_mime(name: &str) -> &'static str {
for (ext, mime) in [
(".png", "image/png"),
(".gif", "image/gif"),
(".webp", "image/webp"),
] {
if name.ends_with(ext) {
return mime;
}
}
"image/jpeg"
}
/// Nom d'une image rangée, si la référence est sûre (ni chemin ni `..`).
fn image_name(r: &str) -> Option<&str> {
let name = r.strip_prefix(IMG_REF_SCHEME)?;
(!name.is_empty() && !name.contains(['/', '\\']) && !name.contains("..")).then_some(name)
}
#[derive(Default)]
struct ImgCacheInner {
map: HashMap<String, String>,
order: VecDeque<String>,
bytes: usize,
}
/// Images jointes en data-URL, bornées à [`IMG_CACHE_MAX`] (la plus ancienne
/// part en premier, jamais la dernière).
pub struct ImageCache {
dir: PathBuf,
inner: Mutex<ImgCacheInner>,
}
impl ImageCache {
/// `dir` : `LOKI_HOME/chatimg`.
pub fn new(dir: PathBuf) -> Self {
ImageCache {
dir,
inner: Mutex::new(ImgCacheInner::default()),
}
}
pub fn get(&self, r: &str) -> Option<String> {
let name = image_name(r)?;
self.inner.lock().ok()?.map.get(name).cloned()
}
fn put(&self, name: &str, url: String) {
let Ok(mut g) = self.inner.lock() else { return };
if g.map.contains_key(name) {
return;
}
g.bytes += url.len();
g.map.insert(name.to_string(), url);
g.order.push_back(name.to_string());
while g.bytes > IMG_CACHE_MAX && g.order.len() > 1 {
if let Some(old) = g.order.pop_front()
&& let Some(u) = g.map.remove(&old)
{
g.bytes -= u.len();
}
}
}
/// Lit et ouvre (par le Go) les images référencées pas encore en cache.
/// Une image illisible reste absente : la requête dira « indisponible ».
pub async fn prefetch(&self, store: &Store, msgs: &[Message]) {
for m in msgs {
for p in content_parts(m.content.as_ref()) {
let Some(name) = image_name(part_image_url(p)) else {
continue;
};
if self.inner.lock().is_ok_and(|g| g.map.contains_key(name)) {
continue;
}
let Ok(raw) = tokio::fs::read(self.dir.join(name)).await else {
continue;
};
let Ok(b) = store.open(raw).await else {
continue;
};
self.put(
name,
format!("data:{};base64,{}", image_mime(name), b64(&b)),
);
}
}
}
}
/// Le moteur vu de la compaction (`postSummary`).
pub struct EngineSummarizer {
pub http: EngineHttp,
pub ep: ChatEndpoint,
pub cfg: HashMap<String, String>,
pub patterns: Arc<RetryPatterns>,
pub port: u16,
}
impl EngineSummarizer {
async fn post(&self, payload: Value) -> Result<SummaryChoice, SummaryError> {
let body = serde_json::to_vec(&payload).unwrap_or_default();
let resp = self
.http
.post(&self.ep.url, &self.ep.key, body)
.await
.map_err(|e| {
SummaryError::Other(friendly_error(classify(&e), &e.to_string(), self.port, ""))
})?;
let status = resp.status().as_u16();
let raw = resp
.bytes()
.await
.map_err(|e| SummaryError::Other(e.to_string()))?;
if status != 200 {
let b = &raw[..raw.len().min(SUMMARY_ERR_MAX)];
return Err(SummaryError::Http {
who: if self.ep.external {
"API externe"
} else {
"llama-server"
}
.into(),
status,
body: String::from_utf8_lossy(b).trim().to_string(),
});
}
parse_summary_response(&raw)
}
}
impl Summarizer for EngineSummarizer {
async fn summarize_transcript(
&self,
req: TranscriptRequest,
) -> Result<SummaryChoice, SummaryError> {
let mut p = json!({
"model": self.ep.model,
"messages": [
{"role": "system", "content": req.system},
{"role": "user", "content": req.transcript},
],
"stream": false,
"temperature": 0.2,
"max_tokens": req.max_tokens,
"chat_template_kwargs": {"enable_thinking": false},
});
// Propre à llama.cpp : une API distante stricte refuse l'argument.
if self.ep.external
&& let Some(o) = p.as_object_mut()
{
o.remove("chat_template_kwargs");
}
self.post(p).await
}
async fn summarize_continuation(
&self,
req: ContinuationRequest,
) -> Result<SummaryChoice, SummaryError> {
let mut p = build_chat_payload(
&self.ep,
&req.messages,
0.2,
&ChatPayloadOpts {
tools: &req.tools,
disable_tools: false,
tool_choice_none: true,
effort: &req.effort,
kwargs: Some(&req.kwargs),
no_sampling: true,
},
&self.cfg,
&|_| None,
);
if let Some(o) = p.as_object_mut() {
o.insert("stream".into(), Value::Bool(false));
o.remove("stream_options");
o.insert("max_tokens".into(), json!(req.max_tokens));
}
self.post(p).await
}
fn slot_holds(&self, _conv: &str) -> bool {
false
}
fn textual_tool_call(&self, content: &str) -> bool {
self.patterns.textual_tool_call(content)
}
}
/// Archive recall tenue par le Go.
pub struct StoreRecall(pub Store);
impl RecallArchive for StoreRecall {
async fn archive(&self, label: &str, role: &str, content: &str) -> Result<String, String> {
self.0.recall_archive(label, role, content).await
}
}
/// Ce qui survit d'un tour à l'autre (un exemplaire par processus).
pub struct ChatShared {
pub http: EngineHttp,
pub store: Store,
pub persister: Persister<StoreSink>,
pub effort: EffortFallbacks,
pub echo_refusals: EchoRefusals,
pub compact_memory: CompactMemory,
pub images: ImageCache,
pub patterns: Arc<RetryPatterns>,
}
/// Ce que la compaction sait du tour.
pub struct TurnCompact {
pub cfg: CompactConfig,
pub caps: CompactCaps,
/// Mode Code : critères rendus de la discussion.
pub criteria: String,
/// Discussion du tour (vide : tâche ou sous-agent).
pub main_conv: String,
/// Forme de la requête du tour, pour les options de la compaction en
/// cours de tour.
pub tools: Vec<Tool>,
pub effort: String,
pub kwargs: Map<String, Value>,
}
/// Le [`TurnHost`] de la façade.
pub struct GoTurnHost {
pub shared: Arc<ChatShared>,
/// Discussion dont relèvent les résultats complets.
pub conv_id: String,
/// Capacités du tour, telles que l'API d'outils les attend (champs Go).
pub caps: Value,
pub cfg: HashMap<String, String>,
pub port: u16,
pub summarizer: EngineSummarizer,
pub compact: TurnCompact,
/// Messages mis en file par l'utilisateur pendant le tour.
pub queued: Box<dyn Fn() -> Vec<Message> + Send + Sync>,
}
/// `toolResID` : « <discussion>.<24 hex> ».
fn tool_res_id(conv: &str) -> String {
let mut b = [0u8; 12];
let _ = getrandom::fill(&mut b);
let conv = match conv.trim() {
"" => "nosession",
c => c,
};
format!("{conv}.{}", hex::encode(b))
}
/// Réponse de `/internal/v1/tool/exec` → [`ToolOutcome`].
pub fn tool_outcome(v: &Value, label: &str) -> ToolOutcome {
let s = |k: &str| v.get(k).and_then(Value::as_str).unwrap_or("").to_string();
let n = |k: &str| v.get(k).and_then(Value::as_i64).unwrap_or(0);
let mut relay: Option<Message> = v
.get("relay")
.and_then(|r| serde_json::from_value(r.clone()).ok());
if relay.is_none()
&& let Some(img) = v.get("vision_img").filter(|i| i.is_object())
{
relay = Some(see_image_message(label, img.clone()));
}
ToolOutcome {
result: s("result"),
diff: v
.get("diff")
.and_then(|d| serde_json::from_value::<Vec<DiffLine>>(d.clone()).ok())
.unwrap_or_default(),
added: n("added"),
removed: n("removed"),
ask: v
.get("ask")
.and_then(|a| serde_json::from_value::<AskEvent>(a.clone()).ok()),
image: s("image"),
live_shot: s("shot"),
relay,
}
}
/// `seeImageMessage` : l'image relayée au modèle.
pub fn see_image_message(label: &str, img: Value) -> Message {
let text = if label.is_empty() {
"Image :".to_string()
} else {
format!("Image demandée ({label}) :")
};
Message {
role: "user".into(),
content: Some(json!([{"type": "text", "text": text}, img])),
..Default::default()
}
}
impl TurnHost for GoTurnHost {
async fn send(&self, ep: &ChatEndpoint, body: Vec<u8>) -> Result<Reply, SendError> {
match self.shared.http.post(&ep.url, &ep.key, body).await {
Ok(resp) => {
let status = resp.status().as_u16();
let body = resp.bytes_stream().map(|r| r.map_err(|e| e.to_string()));
Ok(Reply {
status,
body: Box::pin(body),
})
}
Err(e) => {
let kind = classify(&e);
Err(SendError {
kind,
message: friendly_error(kind, &e.to_string(), self.port, ""),
})
}
}
}
async fn exec_tool(&self, name: &str, args: &Map<String, Value>, label: &str) -> ToolOutcome {
match self
.shared
.store
.tool_exec(name, &Value::Object(args.clone()), &self.caps)
.await
{
Ok(v) => tool_outcome(&v, label),
Err(e) => ToolOutcome {
result: format!("[erreur] exécution de l'outil impossible : {e}"),
..Default::default()
},
}
}
fn save_tool_result(&self, result: &str) -> String {
let id = tool_res_id(&self.conv_id);
self.shared.persister.enqueue_tool_res(ToolResJob {
key: id.clone(),
plain: Arc::from(result),
});
id
}
async fn compact(&self, phase: CompactPhase, msgs: &[Message]) -> Option<Vec<Message>> {
let recall = StoreRecall(self.shared.store.clone());
let c = Compactor {
cfg: &self.compact.cfg,
summarizer: &self.summarizer,
recall: &recall,
memory: &self.shared.compact_memory,
criteria: &self.compact.criteria,
};
let (out, changed) = match phase {
CompactPhase::Reactive | CompactPhase::FullWindow => {
c.compact_messages(msgs, self.compact.caps).await
}
CompactPhase::Mid { used, rescue } => {
let conv = self.compact.main_conv.as_str();
let opt: CompactOpts = compact_opts_mid(
&self.compact.cfg,
&self.shared.compact_memory,
self.summarizer.slot_holds(conv),
self.summarizer.ep.external,
conv,
msgs,
used,
self.compact.tools.clone(),
&self.compact.effort,
self.compact.kwargs.clone(),
rescue,
|m| {
wire_messages(m, false, &self.compact.cfg.echo)
.0
.into_owned()
},
);
c.compact_messages_noted(msgs, self.compact.caps, &opt)
.await
}
};
changed.then_some(out)
}
fn compact_refused(&self, used: i64, msgs: &[Message]) -> bool {
self.shared.compact_memory.refused_skip(
&self.compact.cfg,
&self.compact.main_conv,
used,
msgs,
)
}
fn drain_queued(&self) -> Vec<Message> {
(self.queued)()
}
async fn prepare_images(&self, msgs: &[Message]) {
self.shared.images.prefetch(&self.shared.store, msgs).await;
}
fn image_data_url(&self, r: &str) -> Option<String> {
self.shared.images.get(r)
}
fn effort_remember(&self, want: &str, got: &str) {
self.shared.effort.remember(&self.cfg, want, got);
}
fn echo_refuse(&self, model: &str) {
self.shared.echo_refusals.refuse(model);
}
fn cache_hint(&self) -> Option<CacheHintFn> {
None
}
fn preflight(&self) -> Option<PreflightFn> {
None
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn base64_standard() {
assert_eq!(b64(b""), "");
assert_eq!(b64(b"f"), "Zg==");
assert_eq!(b64(b"fo"), "Zm8=");
assert_eq!(b64(b"foo"), "Zm9v");
assert_eq!(b64(b"foobar"), "Zm9vYmFy");
assert_eq!(b64(&[0xff, 0xfe, 0x00]), "//4A");
}
#[test]
fn references_d_images() {
assert_eq!(image_name("loki-img:ab.png"), Some("ab.png"));
for bad in [
"loki-img:",
"loki-img:../x",
"loki-img:a/b",
"loki-img:a\\b",
"data:x",
"ab.png",
] {
assert_eq!(image_name(bad), None, "{bad}");
}
assert_eq!(image_mime("a.png"), "image/png");
assert_eq!(image_mime("a.webp"), "image/webp");
assert_eq!(image_mime("a.jpg"), "image/jpeg");
assert_eq!(image_mime("a.img"), "image/jpeg");
}
#[test]
fn cache_d_images_borne() {
let c = ImageCache::new(PathBuf::from("/nulle-part"));
let big = "x".repeat(IMG_CACHE_MAX / 2 + 1);
c.put("a.png", big.clone());
c.put("b.png", big.clone());
assert_eq!(c.get("loki-img:a.png"), None, "la plus ancienne part");
assert_eq!(c.get("loki-img:b.png").map(|u| u.len()), Some(big.len()));
// Une seule image, même trop grosse, reste.
let c = ImageCache::new(PathBuf::from("/nulle-part"));
c.put("g.png", "y".repeat(IMG_CACHE_MAX + 1));
assert!(c.get("loki-img:g.png").is_some());
}
#[test]
fn reponse_de_l_api_d_outils() {
let v = json!({
"label": "ls", "result": "a\nb",
"diff": [{"op": "+", "text": "x"}], "added": 1, "removed": 2,
"ask": {"question": "On continue ?", "options": ["oui"]},
"image": "captures/a.png", "shot": "s1",
"relay": {"role": "user", "content": [{"type": "text", "text": "Voici la capture demandée."}]}
});
let o = tool_outcome(&v, "ls");
assert_eq!(o.result, "a\nb");
assert_eq!((o.added, o.removed, o.diff.len()), (1, 2, 1));
assert_eq!(o.ask.unwrap().question, "On continue ?");
assert_eq!(
(o.image.as_str(), o.live_shot.as_str()),
("captures/a.png", "s1")
);
assert_eq!(o.relay.unwrap().role, "user");
// vision_img seul : le message relayé est bâti ici (seeImageMessage).
let img = json!({"type": "image_url", "image_url": {"url": "data:image/png;base64,AA=="}});
let o = tool_outcome(&json!({"result": "ok", "vision_img": img}), "photo.png");
assert_eq!(
o.relay.unwrap().content,
Some(json!([{"type": "text", "text": "Image demandée (photo.png) :"}, img]))
);
assert_eq!(
see_image_message("", json!({})).content.unwrap()[0]["text"],
"Image :"
);
assert_eq!(tool_outcome(&json!({}), "").result, "");
}
#[test]
fn identifiant_d_un_resultat_complet() {
let id = tool_res_id("c42");
let (conv, rand) = id.split_once('.').unwrap();
assert_eq!((conv, rand.len()), ("c42", 24));
assert!(tool_res_id(" ").starts_with("nosession."));
assert_ne!(tool_res_id("c"), tool_res_id("c"));
}
}