Merge pull request '1.2: Sandboxing' (#7) from 1.2 into main

Reviewed-on: #7

.devcontainer/Dockerfile

@@ -1,4 +1,4 @@
-FROM debian:trixie
+FROM rust:1.98-trixie
 
 ARG USERNAME=dev
 ARG USER_UID=1000
@@ -18,11 +18,11 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     && rm -rf /var/lib/apt/lists/*
 
 RUN groupadd --gid ${USER_GID:-1000} $USERNAME \
-    && useradd --uid ${USER_UID:-1000} --gid ${USER_GID:-1000} -m $USERNAME
+    && useradd --uid ${USER_UID:-1000} --gid ${USER_GID:-1000} -m $USERNAME \
+    && rustup component add clippy \
+    && rustup component add rustfmt
+
+
 
 USER $USERNAME
 WORKDIR /home/$USERNAME
-
-ENV PATH="/home/${USERNAME}/.cargo/bin:${PATH}"
-
-RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain stable

.devcontainer/devcontainer.json

@@ -19,5 +19,6 @@
     },
     "workspaceMount": "source=${localWorkspaceFolder},target=/workspaces/herald,type=bind",
     "workspaceFolder": "/workspaces/herald",
+   	"runArgs": ["--userns=keep-id", "--security-opt", "label=disable"],
     "appPort": [3000]
 }

.dockerignore

@@ -1,4 +0,0 @@
-target/
-.env
-.devcontainer/
-docs/

.env.example

@@ -1,5 +1,4 @@
 HTTP_PORT=3000
-BOT_NAME=Herald
 
 WEBHOOK_SIG_HEADER_SECRET=
 
@@ -18,5 +17,8 @@ SENTRY_DSN=
 RUST_LOG=info
 RUST_BACKTRACE=1
 
-
 METRICS_BIND_ADDR=
+
+# Sandboxed tool execution
+# DOCKER_HOST=
+SANDBOX_MAX_ITERATIONS=8

.vscode/settings.json

@@ -1,3 +0,0 @@
-{
-    "rust-analyzer.check.command": "clippy"
-}

.woodpecker/tests.yml

@@ -3,13 +3,27 @@ when:
         - push
 
 steps:
+    - name: fmt
+      image: rust:1.98
+      commands:
+          - rustup component add rustfmt
+          - cargo fmt --all -- --check
+
     - name: clippy
-      image: rust:1.96
+      image: rust:1.98
       commands:
           - rustup component add clippy
-          - cargo clippy
+          - cargo clippy --workspace --all-targets --all-features -- -D warnings
 
     - name: test
-      image: rust:1.96
+      image: rust:1.98
       commands:
-          - cargo test
+          - cargo test --workspace --all-targets
+
+    - name: container-build
+      image: quay.io/buildah/stable
+      privileged: true
+      volumes:
+          - /data/woodpecker-builds:/data
+      commands:
+          - buildah bud -f Containerfile -t herald-ci .

.zed/settings.json

@@ -0,0 +1,22 @@
+{
+    "languages": {
+        "Rust": {
+            "format_on_save": "on",
+            "formatter": "language_server"
+        }
+    },
+    "lsp": {
+        "rust-analyzer": {
+            "initialization_options": {
+                "check": {
+                    "command": "clippy",
+                    "extraArgs": [
+                        "--",
+                        "-D",
+                        "warnings"
+                    ]
+                }
+            }
+        }
+    }
+}

Cargo.toml

@@ -1,32 +1,33 @@
-[package]
+[workspace]
-name = "herald"
+members = [
-version = "1.1.0"
+    "crates/herald-server",
-edition = "2024"
+    "crates/devcontainer-rs",
+]
+resolver = "3"
 
-[profile.release]
+[workspace.dependencies]
-debug = 1
+reqwest = { version = "0.13", default-features = false, features = ["json", "native-tls", "stream"] }
-
-[dependencies]
-reqwest = { version = "0.12", default-features = false, features = ["json", "rustls-tls"] }
 tokio = { version = "1.53", features = ["full"] }
 tokio-stream = "0.1"
 tokio-util = "0.7"
 futures-util = "0.3"
 serde_json = "1.0"
 serde = { version = "1.0", features = ["derive"] }
-sentry = { version = "0.48", features = ["tower-axum-matched-path"] }
+sentry = { version = "0.49", features = ["tower-axum-matched-path"] }
-sentry-anyhow = { version = "0.48", features = ["backtrace"] }
+sentry-anyhow = { version = "0.49", features = ["backtrace"] }
-openrouter-rs = "0.12"
 dotenvy = "0.15"
 tower = "0.5"
-tower-http = {version = "0.6", features = ["trace"] }
+tower-http = { version = "0.7", features = ["trace"] }
 tracing = "0.1"
-tracing-subscriber = { version = "0.3", features=["env-filter"] }
+tracing-subscriber = { version = "0.3", features = ["env-filter"] }
 axum = "0.8"
-anyhow = { version = "1.0", features = ["backtrace"] }
+anyhow = { version = "1", features = ["backtrace"] }
 thiserror = "2.0"
 ring = "0.17"
 hex = "0.4"
 bytes = "1.1"
 metrics = "0.24"
 metrics-exporter-prometheus = { version = "0.18", default-features = false, features = ["http-listener"] }
+
+[profile.release]
+debug = 1

Containerfile

@@ -1,12 +1,31 @@
-FROM rust:1.96 as builder
+FROM rust:1.98-trixie as builder
 
 WORKDIR /app
-COPY . .
+
-RUN cargo build --release
+COPY Cargo.toml Cargo.lock ./
+COPY crates/ crates/
+
+RUN cargo build --release --package herald-server
 
 
 FROM debian:trixie-slim
 
+# git clones the pull request. ca-certificates is what every HTTPS call needs
+# (Gitea, OpenRouter, git). The shared libraries are the non-base ones the binary
+# links against, as reported by `ldd target/release/herald-server`; libc, libm
+# and libgcc_s come from the base image.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    ca-certificates \
+    git \
+    libssl3t64 \
+    libzstd1 \
+    zlib1g \
+    && rm -rf /var/lib/apt/lists/*
+
+
 WORKDIR /app
-COPY --from=builder /app/target/release/herald .
+COPY --from=builder /app/target/release/herald-server ./herald-server
-CMD [ "./herald" ]
+
+# Exec form, so the binary is PID 1 and receives the SIGTERM it handles to shut
+# down gracefully.
+CMD ["./herald-server"]

README.md

@@ -26,7 +26,6 @@ Herald reads its configuration from environment variables (a `.env` file is supp
 | Variable | Description |
 |---|---|
 | `HTTP_PORT` | Port to listen on |
-| `BOT_NAME` | The bot's Gitea username (used to detect mentions) |
 | `WEBHOOK_SIG_HEADER_SECRET` | Gitea webhook secret for signature verification |
 | `OPEN_ROUTER_API_KEY` | OpenRouter API key |
 | `OPEN_ROUTER_MODEL` | Model to use (e.g. `deepseek/deepseek-v4-flash`) |
@@ -38,12 +37,69 @@ Herald reads its configuration from environment variables (a `.env` file is supp
 | `METRICS_BIND_ADDR` | *(optional)* Bind address for the Prometheus metrics endpoint (e.g. `0.0.0.0:9100`). If unset, the metrics exporter is disabled. |
 | `SENTRY_DSN` | *(optional)* Sentry DSN for error tracking |
 | `RUST_LOG` | *(optional)* Log level, defaults to `info` |
+| `SANDBOX_MAX_ITERATIONS` | *(optional)* Maximum number of tool-calling iterations per sandboxed review. Defaults to `8` |
+| `DOCKER_HOST` | *(optional)* Container daemon socket Herald drives, e.g. `unix:///run/user/1000/podman/podman.sock` for a rootless podman. Defaults to `unix:///var/run/docker.sock` |
+
+## Sandboxed reviews
+
+Herald reviews pull requests inside an ephemeral
+[Dev Container](https://containers.dev/). For each review it:
+
+1. clones the pull request head into a temporary directory,
+2. builds and starts the repository's devcontainer (`devcontainer-rs`),
+3. reads the pull request diff and file list from the Gitea API with
+   `GITEA_TOKEN` (so private repositories work), tells the model which files and
+   lines changed — additions and deletions, with the line numbers of the new and
+   old versions of the file respectively — then lets it explore the repository
+   with read-only tools (`ls`, `file_size`, `read_file`, `grep`, `find`) run inside
+   the container: the code itself is not sent, so the model reads it at those lines,
+4. posts the review, anchoring each comment on the added or removed line it
+   refers to, and removes the container and the temporary clone.
+
+Generated files are left out of the changes handed to the model: lockfiles
+(`Cargo.lock`, `package-lock.json`, `yarn.lock`, `go.sum`…) are machine-written
+dependency churn whose thousands of lines would drown the code under review, and
+they are never a place where a comment belongs.
+
+Each comment is tagged with a severity — `bug`, `security`, `performance` or
+`maintainability` — shown at the start of the comment, and the summary also lists
+what the pull request does well.
+
+Herald drives the container daemon through its socket: `DOCKER_HOST` (default
+`unix:///var/run/docker.sock`), which covers both docker and podman's
+Docker-compatible socket. The repository must contain a
+`.devcontainer/devcontainer.json`.
+
+Herald can therefore run inside a container with only that socket mounted (no
+shared workspace directory is required): the clone is streamed to the daemon over
+the socket, like the build context, instead of being bind-mounted from a host
+path the daemon would have to see. This is the setup the `Containerfile`
+produces, e.g.:
+
+```sh
+podman run --env-file=.env -p 3001:3001 \
+    -v /run/user/$(id -u)/podman/podman.sock:/var/run/docker.sock \
+    herald:latest
+```
+
+The `runArgs` of that file are read but deliberately **not** passed to the daemon:
+they come from an untrusted pull request, and one of them (`--network host`) would
+attach the container to another network and quietly
+defeat the network cut described below. A devcontainer that relies on them
+(`--gpus all`, `--cap-add`, `--shm-size`…) will not get them.
+
+Each sandbox is isolated: it gets its own image tag, container and network. The
+container starts with network access so the `postCreateCommand` /
+`postStartCommand` hooks can install dependencies (e.g. `npm install`); once the
+hooks have run, the container is disconnected from the network for the rest of
+the review. Every container command is bounded by a timeout, and the container,
+network and image are removed when the review ends (including on failure).
 
 ## Development
 
 The easiest way to get started is with the provided [Dev Container](https://containers.dev/) (VS Code or Zed with the dev container extension).
 
-Open the project and reopen it in the container — the Rust toolchain is pre-installed.
+Open the project and reopen it in the container — the Rust toolchain is pre-installed, along with rootless podman, so sandboxed reviews can be exercised locally: start its API socket with `podman system service --time=0 &` and set `DOCKER_HOST=unix:///run/user/$(id -u)/podman/podman.sock`.
 
 **Without Dev Container**, you just need a Rust toolchain:
 

crates/devcontainer-rs/Cargo.toml

@@ -0,0 +1,19 @@
+[package]
+name = "devcontainer-rs"
+version = "0.1.0"
+edition = "2024"
+
+[dependencies]
+bollard = "0.21"
+bytes = { workspace = true }
+futures-util = { workspace = true }
+tar = "0.4"
+tokio = { workspace = true }
+tokio-stream = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+anyhow = { workspace = true }
+thiserror = { workspace = true }
+
+[dev-dependencies]
+tempfile = "3"

crates/devcontainer-rs/src/consts.rs

@@ -0,0 +1,24 @@
+//! Constantes de la crate : endpoint du daemon, timeouts et découpage du
+//! contexte de build.
+
+use std::time::Duration;
+
+/// Endpoint affiché dans les logs quand `DOCKER_HOST` n'est pas défini.
+pub(crate) const DEFAULT_ENDPOINT: &str = "unix:///var/run/docker.sock";
+
+/// Timeout appliqué aux opérations de build/run/stop/remove.
+pub(crate) const DEFAULT_COMMAND_TIMEOUT: Duration = Duration::from_secs(600);
+
+/// Timeout appliqué aux commandes exécutées dans un container en cours d'exécution.
+pub(crate) const DEFAULT_EXEC_TIMEOUT: Duration = Duration::from_secs(60);
+
+/// Taille des morceaux du contexte de build envoyés au daemon.
+pub(crate) const CONTEXT_CHUNK_SIZE: usize = 64 * 1024;
+
+/// Nombre de morceaux pouvant attendre dans le canal : c'est ce qui borne la
+/// mémoire occupée par le contexte, quelle que soit sa taille.
+pub(crate) const CONTEXT_CHUNKS: usize = 4;
+
+/// Lectures successives du code de sortie d'un exec, et attente entre elles.
+pub(crate) const EXIT_CODE_ATTEMPTS: usize = 10;
+pub(crate) const EXIT_CODE_DELAY: Duration = Duration::from_millis(20);

crates/devcontainer-rs/src/container.rs

@@ -0,0 +1,93 @@
+//! Un devcontainer en cours d'exécution : inspection et exécution de commandes.
+
+use std::time::Duration;
+
+use crate::{
+    consts::DEFAULT_EXEC_TIMEOUT, errors::ContainerError, exec::ExecOutput,
+    runtime::ContainerRuntime,
+};
+
+/// Un devcontainer en cours d'exécution.
+#[derive(Debug, Clone)]
+pub struct Container {
+    runtime: ContainerRuntime,
+    name: String,
+    workspace_folder: String,
+    remote_user: Option<String>,
+    network: Option<String>,
+    image: Option<String>,
+}
+
+impl Container {
+    /// Assemble un container démarré, avec son réseau et son image à nettoyer.
+    pub(crate) fn new(
+        runtime: ContainerRuntime,
+        name: String,
+        workspace_folder: String,
+        remote_user: Option<String>,
+        network: Option<String>,
+        image: Option<String>,
+    ) -> Self {
+        Self {
+            runtime,
+            name,
+            workspace_folder,
+            remote_user,
+            network,
+            image,
+        }
+    }
+
+    pub fn name(&self) -> &str {
+        &self.name
+    }
+
+    pub fn workspace_folder(&self) -> &str {
+        &self.workspace_folder
+    }
+
+    /// Exécute une commande dans le container et renvoie sa sortie.
+    ///
+    /// La commande est transmise sous forme de vecteur d'arguments (argv, sans shell),
+    /// donc aucun échappement ni interpolation n'est effectué.
+    pub async fn exec(&self, cmd: &[&str]) -> Result<ExecOutput, ContainerError> {
+        self.exec_with_timeout(cmd, DEFAULT_EXEC_TIMEOUT).await
+    }
+
+    pub(crate) async fn exec_with_timeout(
+        &self,
+        cmd: &[&str],
+        timeout: Duration,
+    ) -> Result<ExecOutput, ContainerError> {
+        self.runtime
+            .exec(&self.name, cmd, self.remote_user.as_deref(), timeout)
+            .await
+    }
+
+    /// Exécute un script shell dans le container via `sh -c`.
+    pub async fn exec_shell(&self, script: &str) -> Result<ExecOutput, ContainerError> {
+        self.exec(&["sh", "-c", script]).await
+    }
+
+    /// Arrête le container.
+    pub async fn stop(&self) -> Result<(), ContainerError> {
+        self.runtime.stop_container(&self.name).await
+    }
+
+    /// Supprime le container, son réseau et son image.
+    ///
+    /// La suppression du réseau et de l'image est best-effort : ils peuvent déjà être absents.
+    pub async fn remove(&self) -> Result<(), ContainerError> {
+        self.runtime.remove_container(&self.name).await?;
+
+        if let Some(network) = &self.network {
+            let _ = self.runtime.remove_network(network).await;
+        }
+
+        if let Some(image) = &self.image {
+            let _ = self.runtime.remove_image(image).await;
+        }
+
+        Ok(())
+    }
+}

crates/devcontainer-rs/src/context.rs

@@ -0,0 +1,166 @@
+//! Empaquetage en flux du contexte de build envoyé au daemon.
+
+use std::{
+    io::{BufWriter, Write},
+    path::Path,
+};
+
+use bytes::Bytes;
+use futures_util::Stream;
+use tokio::sync::mpsc;
+use tokio_stream::wrappers::ReceiverStream;
+
+use crate::consts::{CONTEXT_CHUNK_SIZE, CONTEXT_CHUNKS};
+
+/// Empaquette le contexte de build dans un tar, **en flux**.
+///
+/// Le CLI `docker build` empaquette le contexte puis l'envoie ; le faire d'un coup
+/// chargerait tout le dossier en mémoire, ce qui devient intenable dès que le
+/// `.devcontainer` embarque des binaires. Ici l'archive est écrite par une tâche
+/// bloquante et poussée morceau par morceau, avec la contre-pression du canal :
+/// si le daemon lit lentement, l'empaquetage ralentit.
+pub(crate) fn tar_directory_stream(
+    dir: &Path,
+) -> impl Stream<Item = Result<Bytes, std::io::Error>> + Send + 'static {
+    let (sender, receiver) = mpsc::channel(CONTEXT_CHUNKS);
+    let failures = sender.clone();
+    let dir = dir.to_path_buf();
+
+    tokio::task::spawn_blocking(move || {
+        let writer = BufWriter::with_capacity(CONTEXT_CHUNK_SIZE, ChannelWriter { sender });
+        let mut builder = tar::Builder::new(writer);
+
+        let result = builder
+            .append_dir_all(".", &dir)
+            .and_then(|()| builder.finish())
+            // `finish` écrit les blocs de fin d'archive ; il reste à vider le tampon
+            // pour que le daemon reçoive tout, y compris une archive vide.
+            .and_then(|()| builder.get_mut().flush());
+
+        if let Err(error) = result {
+            // L'échec est remonté au daemon par le flux : sinon il ne verrait qu'un
+            // tar tronqué et signalerait une erreur incompréhensible.
+            let _ = failures.blocking_send(Err(error));
+        }
+    });
+
+    ReceiverStream::new(receiver)
+}
+
+/// Écrit les morceaux du tar dans un canal, en attendant qu'il se vide.
+///
+/// Utilisé depuis une tâche bloquante, seul contexte où `blocking_send` est autorisé.
+struct ChannelWriter {
+    sender: mpsc::Sender<Result<Bytes, std::io::Error>>,
+}
+
+impl Write for ChannelWriter {
+    fn write(&mut self, buf: &[u8]) -> std::io::Result<usize> {
+        self.sender
+            .blocking_send(Ok(Bytes::copy_from_slice(buf)))
+            .map_err(|_| {
+                std::io::Error::new(
+                    std::io::ErrorKind::BrokenPipe,
+                    "the daemon dropped the build context",
+                )
+            })?;
+
+        Ok(buf.len())
+    }
+
+    fn flush(&mut self) -> std::io::Result<()> {
+        Ok(())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::fs;
+
+    use futures_util::StreamExt;
+
+    /// Rassemble le flux du contexte en un tar complet, pour l'inspecter.
+    async fn context_tar(dir: &Path) -> Vec<u8> {
+        let mut tar = Vec::new();
+        let mut chunks = tar_directory_stream(dir);
+
+        while let Some(chunk) = chunks.next().await {
+            tar.extend_from_slice(&chunk.unwrap());
+        }
+
+        tar
+    }
+
+    #[tokio::test]
+    async fn tar_context_holds_the_devcontainer_files() {
+        let dir = tempfile::tempdir().unwrap();
+        let devcontainer_dir = dir.path().join(".devcontainer");
+        fs::create_dir(&devcontainer_dir).unwrap();
+        fs::write(devcontainer_dir.join("Dockerfile"), "FROM alpine\n").unwrap();
+        fs::write(devcontainer_dir.join("devcontainer.json"), "{}").unwrap();
+
+        let tar = context_tar(&devcontainer_dir).await;
+
+        let mut archive = tar::Archive::new(tar.as_slice());
+        let names = archive
+            .entries()
+            .unwrap()
+            .map(|entry| entry.unwrap().path().unwrap().display().to_string())
+            .collect::<Vec<_>>();
+
+        assert!(names.iter().any(|name| name.ends_with("Dockerfile")));
+        assert!(names.iter().any(|name| name.ends_with("devcontainer.json")));
+    }
+
+    #[tokio::test]
+    async fn tar_context_keeps_the_executable_bit() {
+        let dir = tempfile::tempdir().unwrap();
+        let script = dir.path().join("setup.sh");
+        fs::write(&script, "#!/bin/sh\n").unwrap();
+
+        #[cfg(unix)]
+        {
+            use std::os::unix::fs::PermissionsExt;
+            fs::set_permissions(&script, fs::Permissions::from_mode(0o755)).unwrap();
+        }
+
+        let tar = context_tar(dir.path()).await;
+
+        let mut archive = tar::Archive::new(tar.as_slice());
+        let mode = archive
+            .entries()
+            .unwrap()
+            .map(|entry| entry.unwrap())
+            .find(|entry| entry.path().unwrap().ends_with("setup.sh"))
+            .unwrap()
+            .header()
+            .mode()
+            .unwrap();
+
+        assert_eq!(mode & 0o111, 0o111);
+    }
+
+    #[tokio::test]
+    async fn tar_context_is_sent_in_chunks() {
+        let dir = tempfile::tempdir().unwrap();
+        fs::write(
+            dir.path().join("big.bin"),
+            vec![0_u8; CONTEXT_CHUNK_SIZE * 3],
+        )
+        .unwrap();
+
+        let mut chunks = tar_directory_stream(dir.path());
+        let mut count = 0;
+
+        while let Some(chunk) = chunks.next().await {
+            assert!(chunk.unwrap().len() <= CONTEXT_CHUNK_SIZE);
+            count += 1;
+        }
+
+        assert!(
+            count > 1,
+            "the context should be streamed, not buffered in one piece"
+        );
+    }
+}

crates/devcontainer-rs/src/devcontainer.rs

@@ -0,0 +1,282 @@
+//! Représentation analysée d'un `devcontainer.json` et son cycle de vie :
+//! nommage, build, démarrage et hooks.
+
+use std::{
+    collections::HashMap,
+    path::{Path, PathBuf},
+    time::{Duration, SystemTime, UNIX_EPOCH},
+};
+
+use bollard::models::{ContainerCreateBody, HostConfig};
+
+use crate::{container::Container, errors::ContainerError, runtime::ContainerRuntime};
+
+#[derive(Debug)]
+pub struct DevContainer {
+    pub container_file_path: PathBuf,
+    pub name: Option<String>,
+    pub build_args: HashMap<String, String>,
+    pub container_env: HashMap<String, String>,
+    pub workspace_folder: Option<String>,
+    pub post_create_command: Option<String>,
+    pub post_start_command: Option<String>,
+    pub remote_user: Option<String>,
+}
+
+impl DevContainer {
+    /// Nom de l'image de base dérivé du nom du devcontainer.
+    pub fn image_name(&self) -> String {
+        let base = self.name.as_deref().unwrap_or("devcontainer");
+        format!("devcontainer-rs/{}", sanitize(base))
+    }
+
+    /// Tag d'image unique pour une exécution de sandbox donnée.
+    ///
+    /// L'unicité est importante : deux sandboxes concurrentes (éventuellement pour des
+    /// dépôts différents partageant un nom de devcontainer) ne doivent pas se disputer le même tag.
+    pub fn image_tag(&self) -> String {
+        format!("{}:{}", self.image_name(), unique_suffix())
+    }
+
+    /// Dossier de workspace dans le container, par défaut `/workspaces/workspace`.
+    pub fn workspace_folder(&self) -> String {
+        self.workspace_folder
+            .clone()
+            .unwrap_or_else(|| "/workspaces/workspace".to_string())
+    }
+
+    /// Nom de container unique pour cette exécution.
+    pub fn container_name(&self) -> String {
+        let base = self.name.as_deref().unwrap_or("devcontainer");
+        format!("devcontainer-rs-{}-{}", sanitize(base), unique_suffix())
+    }
+
+    /// Construit l'image devcontainer sous `image_tag`.
+    pub async fn build(
+        &self,
+        runtime: &ContainerRuntime,
+        image_tag: &str,
+    ) -> Result<(), ContainerError> {
+        let context = self.container_file_path.parent().ok_or_else(|| {
+            ContainerError::Unexpected(String::from(
+                "the devcontainer file path has no parent directory",
+            ))
+        })?;
+
+        let dockerfile = self
+            .container_file_path
+            .file_name()
+            .and_then(|name| name.to_str())
+            .ok_or_else(|| {
+                ContainerError::Unexpected(String::from(
+                    "the devcontainer file name is not valid UTF-8",
+                ))
+            })?;
+
+        runtime
+            .build_image(context, dockerfile, image_tag, &self.build_args)
+            .await
+    }
+
+    /// Construit l'image, démarre le container, exécute les hooks puis coupe le réseau.
+    ///
+    /// En cas d'échec, le container, le réseau et l'image sont nettoyés avant de
+    /// renvoyer l'erreur, afin qu'aucune ressource ne soit laissée en place.
+    pub async fn up(
+        &self,
+        runtime: &ContainerRuntime,
+        workspace_dir: &Path,
+    ) -> Result<Container, ContainerError> {
+        let image_tag = self.image_tag();
+        self.build(runtime, &image_tag).await?;
+
+        let name = self.container_name();
+        let network = format!("{name}-net");
+
+        // Réseau dédié afin de pouvoir couper la connectivité après les hooks.
+        if let Err(err) = runtime.create_network(&network).await {
+            let _ = runtime.remove_image(&image_tag).await;
+            return Err(err);
+        }
+
+        let workspace_folder = self.workspace_folder();
+        let body = ContainerCreateBody {
+            image: Some(image_tag.clone()),
+            cmd: Some(vec![String::from("sleep"), String::from("infinity")]),
+            user: self.remote_user.clone(),
+            env: Some(
+                self.container_env
+                    .iter()
+                    .map(|(key, value)| format!("{key}={value}"))
+                    .collect(),
+            ),
+            working_dir: Some(workspace_folder.clone()),
+            host_config: Some(HostConfig {
+                network_mode: Some(network.clone()),
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+
+        if let Err(err) = runtime.create_container(&name, body).await {
+            let _ = runtime.remove_network(&network).await;
+            let _ = runtime.remove_image(&image_tag).await;
+            return Err(err);
+        }
+
+        if let Err(err) = runtime.start_container(&name).await {
+            let _ = runtime.remove_container(&name).await;
+            let _ = runtime.remove_network(&network).await;
+            let _ = runtime.remove_image(&image_tag).await;
+            return Err(err);
+        }
+
+        // Le clone est copié dans le container par le socket, pas monté depuis un
+        // chemin de l'hôte : le daemon n'a pas besoin de voir le clone pour le rendre
+        // visible dans la sandbox, ce qui permet à Herald de tourner dans un container
+        // (avec le socket monté) sans partager de dossier avec l'hôte.
+        if let Err(err) = runtime
+            .upload_directory(&name, &workspace_folder, workspace_dir)
+            .await
+        {
+            let _ = runtime.remove_container(&name).await;
+            let _ = runtime.remove_network(&network).await;
+            let _ = runtime.remove_image(&image_tag).await;
+            return Err(err);
+        }
+
+        let container = Container::new(
+            runtime.clone(),
+            name,
+            self.workspace_folder(),
+            self.remote_user.clone(),
+            Some(network.clone()),
+            Some(image_tag),
+        );
+
+        // Les hooks s'exécutent avec accès au réseau (installation de dépendances, etc.).
+        if let Err(err) = self.run_hooks(&container, runtime.timeout()).await {
+            let _ = container.remove().await;
+            return Err(err);
+        }
+
+        // Coupe l'accès réseau pour le reste de la durée de vie de la sandbox.
+        if let Err(err) = runtime.disconnect_network(&network, container.name()).await {
+            let _ = container.remove().await;
+            return Err(err);
+        }
+
+        Ok(container)
+    }
+
+    async fn run_hooks(
+        &self,
+        container: &Container,
+        timeout: Duration,
+    ) -> Result<(), ContainerError> {
+        for command in [&self.post_create_command, &self.post_start_command]
+            .into_iter()
+            .flatten()
+        {
+            let output = container
+                .exec_with_timeout(&["sh", "-c", command], timeout)
+                .await?;
+
+            if !output.success() {
+                return Err(ContainerError::Unexpected(format!(
+                    "hook `{command}` failed with status {}: {}",
+                    output.status,
+                    output.stderr.trim()
+                )));
+            }
+        }
+
+        Ok(())
+    }
+}
+
+/// Nettoie une chaîne pour qu'elle puisse servir de nom d'image/container docker.
+fn sanitize(input: &str) -> String {
+    let sanitized: String = input
+        .chars()
+        .map(|c| {
+            if c.is_ascii_alphanumeric() || c == '-' || c == '_' || c == '.' {
+                c.to_ascii_lowercase()
+            } else {
+                '-'
+            }
+        })
+        .collect();
+
+    let trimmed = sanitized.trim_matches(|c| c == '-' || c == '.' || c == '_');
+    if trimmed.is_empty() {
+        "devcontainer".to_string()
+    } else {
+        trimmed.to_string()
+    }
+}
+
+/// Suffixe unique à une exécution de sandbox, combinant l'identifiant de processus et un horodatage.
+fn unique_suffix() -> String {
+    let nanos = SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|duration| duration.as_nanos())
+        .unwrap_or(0);
+
+    format!("{}-{}", std::process::id(), nanos)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::fs;
+
+    fn devcontainer(dir: &Path) -> DevContainer {
+        let devcontainer_path = dir.join("devcontainer.json");
+        let dockerfile_path = dir.join("Dockerfile");
+
+        fs::write(&dockerfile_path, "FROM alpine\n").unwrap();
+        fs::write(
+            &devcontainer_path,
+            r#"{
+                "name": "My Project",
+                "build": {
+                    "dockerfile": "Dockerfile",
+                    "args": { "VERSION": "1" }
+                },
+                "workspaceFolder": "/workspaces/my-project",
+                "containerEnv": { "RUST_LOG": "debug" },
+                "remoteUser": "dev"
+            }"#,
+        )
+        .unwrap();
+
+        let runtime = tokio::runtime::Runtime::new().unwrap();
+        runtime.block_on(crate::parse(&devcontainer_path)).unwrap()
+    }
+
+    #[test]
+    fn image_name_is_sanitized() {
+        let dir = tempfile::tempdir().unwrap();
+        let dc = devcontainer(dir.path());
+        assert_eq!(dc.image_name(), "devcontainer-rs/my-project");
+    }
+
+    #[test]
+    fn image_tags_are_unique() {
+        let dir = tempfile::tempdir().unwrap();
+        let dc = devcontainer(dir.path());
+        assert_ne!(dc.image_tag(), dc.image_tag());
+    }
+
+    #[test]
+    fn container_name_is_sanitized() {
+        let dir = tempfile::tempdir().unwrap();
+        let dc = devcontainer(dir.path());
+
+        assert!(
+            dc.container_name()
+                .starts_with("devcontainer-rs-my-project-")
+        );
+    }
+}

crates/devcontainer-rs/src/errors.rs

@@ -0,0 +1,52 @@
+//! Erreurs de la crate : échecs du daemon de containers et échecs d'analyse du
+//! `devcontainer.json`.
+
+use std::{path::PathBuf, time::Duration};
+
+use bollard::errors::Error as BollardError;
+
+#[derive(Debug, thiserror::Error)]
+pub enum ContainerError {
+    /// Le daemon est injoignable, ou a refusé une requête.
+    #[error("container daemon request failed: {source}")]
+    Request {
+        #[from]
+        source: BollardError,
+    },
+
+    /// Une opération a dépassé son timeout.
+    #[error("`{operation}` timed out after {timeout:?}")]
+    Timeout {
+        operation: String,
+        timeout: Duration,
+    },
+
+    /// La construction de l'image a échoué.
+    #[error("image build failed: {message}")]
+    Build { message: String },
+
+    /// Le daemon a répondu autre chose que ce qui était attendu.
+    #[error("{0}")]
+    Unexpected(String),
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum ParseError {
+    #[error("failed to read devcontainer file `{path}`: {source}")]
+    Read {
+        path: PathBuf,
+        source: std::io::Error,
+    },
+
+    #[error("invalid devcontainer JSON in `{path}`: {source}")]
+    Json {
+        path: PathBuf,
+        source: serde_json::Error,
+    },
+
+    #[error("container file `{0}` does not exist or is not a regular file")]
+    ContainerFileNotFound(PathBuf),
+
+    #[error("the devcontainer file path has no parent directory: `{0}`")]
+    InvalidDevContainerPath(PathBuf),
+}

crates/devcontainer-rs/src/exec.rs

@@ -0,0 +1,16 @@
+//! Résultat d'une commande exécutée dans un container.
+
+/// Résultat d'une commande exécutée dans un container.
+#[derive(Debug, Clone)]
+pub struct ExecOutput {
+    /// Code de sortie, ou `-1` si le processus a été terminé par un signal.
+    pub status: i32,
+    pub stdout: String,
+    pub stderr: String,
+}
+
+impl ExecOutput {
+    pub fn success(&self) -> bool {
+        self.status == 0
+    }
+}

crates/devcontainer-rs/src/lib.rs

@@ -0,0 +1,43 @@
+//! Primitives de cycle de vie de container pour un [`DevContainer`] analysé.
+//!
+//! Cette crate pilote l'API du daemon de containers pour construire l'image
+//! devcontainer, démarrer un container et y copier le workspace, exécuter les
+//! hooks `postCreateCommand` / `postStartCommand` et lancer des commandes à
+//! l'intérieur du container en cours d'exécution.
+//!
+//! Le daemon est joint par son **socket** (celui de l'hôte, monté dans le
+//! container de Herald) : [`ContainerRuntime::connect`] suit `DOCKER_HOST` comme
+//! le fait le CLI `docker`, et retombe sur le socket local. Aucun binaire de
+//! runtime n'est donc requis dans l'image, et `podman` fonctionne de la même
+//! façon dès lors qu'il expose son socket compatible Docker.
+//!
+//! # Isolation
+//!
+//! Chaque sandbox dispose de son propre tag d'image, de son propre container et de
+//! son propre réseau. Le container démarre attaché à ce réseau afin que les hooks
+//! `postCreateCommand` / `postStartCommand` puissent récupérer des dépendances
+//! (par ex. `npm install`) ; une fois les hooks exécutés, le container est déconnecté
+//! du réseau pour le reste de sa durée de vie. Chaque commande est bornée par un timeout.
+//!
+//! Les `runArgs` du `devcontainer.json` sont lus mais pas transmis au daemon : ils
+//! viennent du dépôt, donc d'une pull request non fiable, et pourraient rattacher le
+//! container à un autre réseau ou lui donner des privilèges qui annuleraient cette
+//! isolation.
+
+mod consts;
+mod container;
+mod context;
+mod devcontainer;
+mod errors;
+mod exec;
+mod path;
+mod runtime;
+mod schema;
+
+pub use container::Container;
+pub use devcontainer::DevContainer;
+pub use errors::{ContainerError, ParseError};
+pub use exec::ExecOutput;
+pub use path::normalize;
+pub use runtime::ContainerRuntime;
+pub use schema::{DevContainerBuildSchema, DevContainerSchema, parse};

crates/devcontainer-rs/src/path.rs

@@ -0,0 +1,40 @@
+//! Normalisation lexicale de chemins, sans accès au système de fichiers.
+
+use std::path::{Path, PathBuf};
+
+/// Normalise lexicalement un chemin, en résolvant `.` et `..` sans toucher au
+/// système de fichiers. Renvoie `None` si le chemin sort de sa racine.
+pub fn normalize(path: &Path) -> Option<PathBuf> {
+    use std::path::Component;
+
+    let mut out = PathBuf::new();
+    for component in path.components() {
+        match component {
+            Component::RootDir => out.push("/"),
+            Component::CurDir => {}
+            Component::ParentDir => {
+                if !out.pop() {
+                    return None;
+                }
+            }
+            Component::Normal(part) => out.push(part),
+            Component::Prefix(_) => return None,
+        }
+    }
+
+    Some(out)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn normalize_rejects_escaping_paths() {
+        assert_eq!(
+            normalize(Path::new("/workspaces/project/src/../main.rs")),
+            Some(PathBuf::from("/workspaces/project/main.rs"))
+        );
+        assert_eq!(normalize(Path::new("/workspaces/../../etc/passwd")), None);
+    }
+}

crates/devcontainer-rs/src/runtime.rs

@@ -0,0 +1,376 @@
+//! Client de l'API du daemon de containers et opérations de cycle de vie.
+
+use std::{collections::HashMap, path::Path, time::Duration};
+
+use bollard::{
+    Docker, body_try_stream,
+    container::LogOutput,
+    errors::Error as BollardError,
+    exec::{CreateExecOptions, StartExecOptions, StartExecResults},
+    models::{BuildInfo, ContainerCreateBody, NetworkCreateRequest, NetworkDisconnectRequest},
+    query_parameters::{
+        BuildImageOptions, CreateContainerOptions, RemoveContainerOptions, RemoveImageOptions,
+        StartContainerOptions, StopContainerOptions, UploadToContainerOptions,
+    },
+};
+use futures_util::StreamExt;
+
+use crate::{
+    consts::{DEFAULT_COMMAND_TIMEOUT, DEFAULT_ENDPOINT, EXIT_CODE_ATTEMPTS, EXIT_CODE_DELAY},
+    context::tar_directory_stream,
+    errors::ContainerError,
+    exec::ExecOutput,
+};
+
+/// Un client de l'API du daemon de containers.
+#[derive(Debug, Clone)]
+pub struct ContainerRuntime {
+    docker: Docker,
+    endpoint: String,
+    timeout: Duration,
+}
+
+impl ContainerRuntime {
+    /// Se connecte au daemon désigné par `DOCKER_HOST`, ou au socket local par défaut.
+    pub fn connect() -> Result<Self, ContainerError> {
+        let endpoint =
+            std::env::var("DOCKER_HOST").unwrap_or_else(|_| String::from(DEFAULT_ENDPOINT));
+
+        Ok(Self {
+            docker: Docker::connect_with_defaults()?,
+            endpoint,
+            timeout: DEFAULT_COMMAND_TIMEOUT,
+        })
+    }
+
+    /// Remplace le timeout appliqué aux opérations de build/run/stop/remove.
+    pub fn with_timeout(mut self, timeout: Duration) -> Self {
+        self.timeout = timeout;
+        self
+    }
+
+    /// Endpoint du daemon, pour les logs.
+    pub fn endpoint(&self) -> &str {
+        &self.endpoint
+    }
+
+    pub fn timeout(&self) -> Duration {
+        self.timeout
+    }
+
+    /// Vérifie que le daemon est joignable et répond.
+    pub async fn available(&self) -> bool {
+        self.docker.ping().await.is_ok()
+    }
+
+    /// Exécute une requête du daemon en appliquant le timeout des commandes.
+    async fn request<T>(
+        &self,
+        operation: &str,
+        request: impl Future<Output = Result<T, BollardError>>,
+    ) -> Result<T, ContainerError> {
+        match tokio::time::timeout(self.timeout, request).await {
+            Ok(Ok(value)) => Ok(value),
+            Ok(Err(source)) => Err(ContainerError::Request { source }),
+            Err(_) => Err(ContainerError::Timeout {
+                operation: String::from(operation),
+                timeout: self.timeout,
+            }),
+        }
+    }
+
+    /// Construit l'image `tag` depuis le contexte `context_dir`.
+    pub(crate) async fn build_image(
+        &self,
+        context_dir: &Path,
+        dockerfile: &str,
+        tag: &str,
+        build_args: &HashMap<String, String>,
+    ) -> Result<(), ContainerError> {
+        let options = BuildImageOptions {
+            dockerfile: String::from(dockerfile),
+            t: Some(String::from(tag)),
+            buildargs: Some(build_args.clone()),
+            rm: true,
+            ..Default::default()
+        };
+
+        let context = tar_directory_stream(context_dir);
+        let mut stream = self
+            .docker
+            .build_image(options, None, Some(body_try_stream(context)));
+
+        // Le flux porte la progression du build : la dernière erreur signalée par le
+        // daemon fait échouer l'opération.
+        let consume = async {
+            let mut failure = None;
+
+            while let Some(info) = stream.next().await {
+                let info: BuildInfo = info?;
+
+                if let Some(message) = info.error_detail.and_then(|detail| detail.message) {
+                    failure = Some(message);
+                }
+            }
+
+            Ok::<_, BollardError>(failure)
+        };
+
+        let failure = match tokio::time::timeout(self.timeout, consume).await {
+            Ok(Ok(failure)) => failure,
+            Ok(Err(source)) => return Err(ContainerError::Request { source }),
+            Err(_) => {
+                return Err(ContainerError::Timeout {
+                    operation: format!("build image `{tag}`"),
+                    timeout: self.timeout,
+                });
+            }
+        };
+
+        match failure {
+            Some(message) => Err(ContainerError::Build { message }),
+            None => Ok(()),
+        }
+    }
+
+    /// Crée un container, sans le démarrer.
+    pub(crate) async fn create_container(
+        &self,
+        name: &str,
+        body: ContainerCreateBody,
+    ) -> Result<(), ContainerError> {
+        let options = CreateContainerOptions {
+            name: Some(String::from(name)),
+            ..Default::default()
+        };
+
+        self.request(
+            "create container",
+            self.docker.create_container(Some(options), body),
+        )
+        .await?;
+
+        Ok(())
+    }
+
+    pub(crate) async fn upload_directory(
+        &self,
+        container: &str,
+        destination: &str,
+        directory: &Path,
+    ) -> Result<(), ContainerError> {
+        let options = UploadToContainerOptions {
+            path: String::from(destination),
+            ..Default::default()
+        };
+
+        self.request(
+            "upload workspace",
+            self.docker.upload_to_container(
+                container,
+                Some(options),
+                body_try_stream(tar_directory_stream(directory)),
+            ),
+        )
+        .await?;
+
+        Ok(())
+    }
+
+    pub(crate) async fn start_container(&self, name: &str) -> Result<(), ContainerError> {
+        self.request(
+            "start container",
+            self.docker
+                .start_container(name, None::<StartContainerOptions>),
+        )
+        .await?;
+
+        Ok(())
+    }
+
+    /// Exécute une commande dans un container et renvoie sa sortie.
+    pub(crate) async fn exec(
+        &self,
+        container: &str,
+        cmd: &[&str],
+        user: Option<&str>,
+        timeout: Duration,
+    ) -> Result<ExecOutput, ContainerError> {
+        let config = CreateExecOptions {
+            cmd: Some(cmd.iter().map(|arg| String::from(*arg)).collect()),
+            user: user.map(String::from),
+            attach_stdout: Some(true),
+            attach_stderr: Some(true),
+            ..Default::default()
+        };
+
+        let created = self
+            .request("create exec", self.docker.create_exec(container, config))
+            .await?;
+
+        let started = self
+            .request(
+                "start exec",
+                self.docker
+                    .start_exec(&created.id, None::<StartExecOptions>),
+            )
+            .await?;
+
+        let StartExecResults::Attached { mut output, .. } = started else {
+            return Err(ContainerError::Unexpected(String::from(
+                "the daemon detached an exec that was not requested as detached",
+            )));
+        };
+
+        let mut stdout = String::new();
+        let mut stderr = String::new();
+
+        // Le timeout couvre l'exécution de la commande elle-même, pas seulement sa
+        // mise en place.
+        let collect = async {
+            while let Some(message) = output.next().await {
+                match message? {
+                    LogOutput::StdOut { message } | LogOutput::Console { message } => {
+                        stdout.push_str(&String::from_utf8_lossy(&message));
+                    }
+                    LogOutput::StdErr { message } => {
+                        stderr.push_str(&String::from_utf8_lossy(&message));
+                    }
+                    LogOutput::StdIn { .. } => {}
+                }
+            }
+
+            Ok::<_, BollardError>(())
+        };
+
+        match tokio::time::timeout(timeout, collect).await {
+            Ok(Ok(())) => {}
+            Ok(Err(source)) => return Err(ContainerError::Request { source }),
+            Err(_) => {
+                // Le processus continue de tourner dans le container : sans arrêt, il
+                // consommerait CPU et mémoire et pourrait encore modifier le workspace
+                // pendant toute la durée de vie de la sandbox. L'API n'offre pas de
+                // moyen de tuer un exec, on arrête donc le container qui le porte.
+                let _ = self.stop_container(container).await;
+
+                return Err(ContainerError::Timeout {
+                    operation: format!("exec {}", cmd.join(" ")),
+                    timeout,
+                });
+            }
+        }
+
+        // Le code de sortie n'est pas forcément publié au moment où le flux se ferme :
+        // on laisse au daemon le temps de le renseigner, sans bloquer indéfiniment.
+        // Sans cela, une commande réussie peut être rapportée en échec (code -1) et
+        // l'outil renvoie une erreur au modèle à la place du contenu.
+        let mut inspected = self
+            .request("inspect exec", self.docker.inspect_exec(&created.id))
+            .await?;
+
+        for _ in 0..EXIT_CODE_ATTEMPTS {
+            if !inspected.running.unwrap_or(false) {
+                break;
+            }
+
+            tokio::time::sleep(EXIT_CODE_DELAY).await;
+
+            inspected = self
+                .request("inspect exec", self.docker.inspect_exec(&created.id))
+                .await?;
+        }
+
+        Ok(ExecOutput {
+            status: inspected.exit_code.unwrap_or(-1) as i32,
+            stdout,
+            stderr,
+        })
+    }
+
+    pub(crate) async fn stop_container(&self, name: &str) -> Result<(), ContainerError> {
+        let options = StopContainerOptions {
+            t: Some(1),
+            ..Default::default()
+        };
+
+        self.request(
+            "stop container",
+            self.docker.stop_container(name, Some(options)),
+        )
+        .await?;
+
+        Ok(())
+    }
+
+    /// Supprime un container, ses volumes anonymes et le processus qui y tourne.
+    pub(crate) async fn remove_container(&self, name: &str) -> Result<(), ContainerError> {
+        let options = RemoveContainerOptions {
+            force: true,
+            v: true,
+            ..Default::default()
+        };
+
+        self.request(
+            "remove container",
+            self.docker.remove_container(name, Some(options)),
+        )
+        .await?;
+
+        Ok(())
+    }
+
+    pub(crate) async fn remove_image(&self, tag: &str) -> Result<(), ContainerError> {
+        let options = RemoveImageOptions {
+            force: true,
+            ..Default::default()
+        };
+
+        self.request(
+            "remove image",
+            self.docker.remove_image(tag, Some(options), None),
+        )
+        .await?;
+
+        Ok(())
+    }
+
+    /// Crée le réseau dédié d'une sandbox.
+    pub(crate) async fn create_network(&self, name: &str) -> Result<(), ContainerError> {
+        let request = NetworkCreateRequest {
+            name: String::from(name),
+            ..Default::default()
+        };
+
+        self.request("create network", self.docker.create_network(request))
+            .await?;
+
+        Ok(())
+    }
+
+    /// Détache un container d'un réseau, ce qui coupe sa connectivité.
+    pub(crate) async fn disconnect_network(
+        &self,
+        network: &str,
+        container: &str,
+    ) -> Result<(), ContainerError> {
+        let request = NetworkDisconnectRequest {
+            container: String::from(container),
+            ..Default::default()
+        };
+
+        self.request(
+            "disconnect network",
+            self.docker.disconnect_network(network, request),
+        )
+        .await?;
+
+        Ok(())
+    }
+
+    pub(crate) async fn remove_network(&self, name: &str) -> Result<(), ContainerError> {
+        self.request("remove network", self.docker.remove_network(name))
+            .await?;
+
+        Ok(())
+    }
+}

crates/devcontainer-rs/src/schema.rs

@@ -0,0 +1,230 @@
+//! Schéma du `devcontainer.json`, analyse du fichier et résolution des
+//! références `${localEnv:…}`.
+
+use std::{
+    collections::HashMap,
+    path::{Path, PathBuf},
+};
+
+use serde::Deserialize;
+
+use crate::{devcontainer::DevContainer, errors::ParseError};
+
+#[derive(Debug, Deserialize)]
+pub struct DevContainerBuildSchema {
+    #[serde(default)]
+    pub dockerfile: String,
+    #[serde(default)]
+    pub args: HashMap<String, String>,
+}
+
+#[derive(Debug, Deserialize)]
+pub struct DevContainerSchema {
+    #[serde(default)]
+    pub name: Option<String>,
+    pub build: DevContainerBuildSchema,
+
+    #[serde(rename = "workspaceFolder", default)]
+    pub workspace_folder: Option<String>,
+
+    #[serde(rename = "containerEnv", default)]
+    pub container_env: HashMap<String, String>,
+
+    #[serde(rename = "postCreateCommand", default)]
+    pub post_create_command: Option<String>,
+
+    #[serde(rename = "postStartCommand", default)]
+    pub post_start_command: Option<String>,
+
+    #[serde(rename = "remoteUser", default)]
+    pub remote_user: Option<String>,
+
+    /// Arguments passés à `docker run`.
+    ///
+    /// Lus pour rester fidèle au format `devcontainer.json`, mais
+    /// **délibérément pas transmis** au runtime : ils viennent d'une pull
+    /// request non fiable et pourraient casser l'isolation de la sandbox, décrite
+    /// dans le doc de la crate.
+    #[serde(rename = "runArgs", default)]
+    pub run_args: Vec<String>,
+}
+
+impl TryFrom<(DevContainerSchema, PathBuf)> for DevContainer {
+    type Error = ParseError;
+
+    fn try_from(
+        (schema, devcontainer_path): (DevContainerSchema, PathBuf),
+    ) -> Result<Self, Self::Error> {
+        let base_dir = devcontainer_path
+            .parent()
+            .ok_or_else(|| ParseError::InvalidDevContainerPath(devcontainer_path.clone()))?;
+
+        let container_file_path = base_dir.join(schema.build.dockerfile);
+
+        if !container_file_path.is_file() {
+            return Err(ParseError::ContainerFileNotFound(container_file_path));
+        }
+
+        let build_args = schema
+            .build
+            .args
+            .into_iter()
+            .map(|(k, v)| (k, substitute_local_env(&v)))
+            .collect();
+
+        let container_env = schema
+            .container_env
+            .into_iter()
+            .map(|(k, v)| (k, substitute_local_env(&v)))
+            .collect();
+
+        Ok(Self {
+            container_file_path,
+            name: schema.name,
+            build_args,
+            container_env,
+            workspace_folder: schema.workspace_folder,
+            post_create_command: schema.post_create_command,
+            post_start_command: schema.post_start_command,
+            remote_user: schema.remote_user,
+        })
+    }
+}
+
+/// Résout les références `${localEnv:VAR}` et `${localEnv:VAR:default}` d'un
+/// `devcontainer.json`.
+///
+/// La spécification veut que `${localEnv:VAR}` soit lu dans l'environnement du
+/// client ; ici le `devcontainer.json` vient d'une **pull request**, donc de code
+/// non fiable : lire l'environnement du processus lui permettrait de récupérer
+/// `GITEA_TOKEN`, `OPEN_ROUTER_API_KEY` ou n'importe quel autre secret de Herald et
+/// de l'exfiltrer depuis un `build.args`, un `containerEnv`, son Dockerfile ou un
+/// hook. L'environnement n'est donc **jamais** consulté : seule la valeur par
+/// défaut est utilisée, et une référence sans défaut devient une chaîne vide.
+fn substitute_local_env(input: &str) -> String {
+    const PREFIX: &str = "${localEnv:";
+
+    let mut out = String::with_capacity(input.len());
+    let mut rest = input;
+
+    while let Some(start) = rest.find(PREFIX) {
+        out.push_str(&rest[..start]);
+        let after = &rest[start + PREFIX.len()..];
+
+        match after.find('}') {
+            Some(end) => {
+                let inner = &after[..end];
+                // La clé est lue pour délimiter la référence, pas pour la résoudre.
+                let (_key, default) = match inner.split_once(':') {
+                    Some((key, default)) => (key, Some(default)),
+                    None => (inner, None),
+                };
+
+                out.push_str(default.unwrap_or(""));
+
+                rest = &after[end + 1..];
+            }
+            None => {
+                out.push_str(PREFIX);
+                rest = after;
+            }
+        }
+    }
+
+    out.push_str(rest);
+    out
+}
+
+pub async fn parse(path: impl AsRef<Path>) -> Result<DevContainer, ParseError> {
+    let path = path.as_ref().to_path_buf();
+    let contents = tokio::fs::read_to_string(&path)
+        .await
+        .map_err(|source| ParseError::Read {
+            path: path.clone(),
+            source,
+        })?;
+
+    let schema = serde_json::from_str::<DevContainerSchema>(&contents).map_err(|source| {
+        ParseError::Json {
+            path: path.clone(),
+            source,
+        }
+    })?;
+
+    DevContainer::try_from((schema, path))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::fs;
+
+    #[tokio::test]
+    async fn parses_devcontainer_file() {
+        let dir = tempfile::tempdir().unwrap();
+        let devcontainer_path = dir.path().join("devcontainer.json");
+        let dockerfile_path = dir.path().join("Dockerfile");
+
+        fs::write(&dockerfile_path, "FROM alpine\n").unwrap();
+        fs::write(
+            &devcontainer_path,
+            r#"{
+                "name": "test",
+                "build": {
+                    "dockerfile": "Dockerfile",
+                    "args": {
+                        "VERSION": "1"
+                    }
+                },
+                "workspaceFolder": "/workspace",
+                "containerEnv": {
+                    "RUST_LOG": "debug"
+                },
+                "remoteUser": "dev"
+            }"#,
+        )
+        .unwrap();
+
+        let config = parse(&devcontainer_path).await.unwrap();
+
+        assert_eq!(config.name.as_deref(), Some("test"));
+        assert_eq!(config.container_file_path, dockerfile_path);
+        assert_eq!(config.build_args.get("VERSION").unwrap(), "1");
+        assert_eq!(config.container_env.get("RUST_LOG").unwrap(), "debug");
+        assert_eq!(config.workspace_folder.as_deref(), Some("/workspace"));
+        assert_eq!(config.remote_user.as_deref(), Some("dev"));
+    }
+
+    #[test]
+    fn local_env_references_are_never_read_from_the_environment() {
+        unsafe { std::env::set_var("DEVCONTAINER_TEST_SECRET", "s3cret") };
+
+        // Le `devcontainer.json` vient d'une pull request : résoudre la référence
+        // depuis l'environnement de Herald y exposerait ses jetons.
+        assert_eq!(
+            substitute_local_env("${localEnv:DEVCONTAINER_TEST_SECRET}"),
+            ""
+        );
+        assert_eq!(
+            substitute_local_env("token=${localEnv:DEVCONTAINER_TEST_SECRET}"),
+            "token="
+        );
+    }
+
+    #[test]
+    fn local_env_defaults_are_used() {
+        assert_eq!(
+            substitute_local_env("${localEnv:DEVCONTAINER_TEST_MISSING:fallback}"),
+            "fallback"
+        );
+        assert_eq!(
+            substitute_local_env("${localEnv:DEVCONTAINER_TEST_MISSING}"),
+            ""
+        );
+        assert_eq!(substitute_local_env("uid=${localEnv:UID:1000}"), "uid=1000");
+        assert_eq!(
+            substitute_local_env("no variables here"),
+            "no variables here"
+        );
+    }
+}

crates/herald-server/Cargo.toml

@@ -0,0 +1,28 @@
+[package]
+name = "herald-server"
+version = "1.2.0"
+edition = "2024"
+
+[dependencies]
+reqwest = { workspace = true }
+tokio = { workspace = true }
+tokio-util = { workspace = true }
+futures-util = { workspace = true }
+serde_json = { workspace = true }
+serde = { workspace = true }
+sentry = { workspace = true }
+sentry-anyhow = { workspace = true }
+dotenvy = { workspace = true }
+tower = { workspace = true }
+tower-http = { workspace = true }
+tracing = { workspace = true }
+tracing-subscriber = { workspace = true }
+axum = { workspace = true }
+anyhow = { workspace = true }
+thiserror = { workspace = true }
+ring = { workspace = true }
+hex = { workspace = true }
+metrics = { workspace = true }
+metrics-exporter-prometheus = { workspace = true }
+devcontainer-rs = { path = "../devcontainer-rs" }
+tempfile = "3"

crates/herald-server/src/bot.rs

@@ -1,7 +1,10 @@
 use crate::{
-    gitea::{GiteaAPI, WebhookType}, metrics, open_router::OpenRouterClient,
+    gitea::{GiteaAPI, WebhookType},
+    metrics,
+    open_router::OpenRouterClient,
+    sandbox::SandboxConfig,
 };
-use serde::Deserialize;
+use serde::{Deserialize, Deserializer};
 use std::{collections::HashSet, sync::Arc};
 use tokio::sync::Mutex;
 use tokio_util::sync::CancellationToken;
@@ -18,18 +21,111 @@ pub struct ReviewResult {
 pub struct ReviewItem {
     pub filename: String,
     pub line: Option<u64>,
-    pub code: String,
+    #[serde(default, deserialize_with = "deserialize_side")]
+    pub side: Option<ReviewSide>,
+    #[serde(default, deserialize_with = "deserialize_severity")]
+    pub severity: Option<ReviewSeverity>,
     pub message: String,
 }
 
+/// Which version of the file a review comment is anchored on.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum ReviewSide {
+    /// The line was added by the pull request: `line` is a line number of the
+    /// new version of the file.
+    Added,
+    /// The line was removed by the pull request: `line` is a line number of the
+    /// old version of the file.
+    Removed,
+}
+
+impl ReviewSide {
+    /// Reads the side the model asked for, tolerating casing and synonyms.
+    fn parse(raw: &str) -> Option<Self> {
+        match raw.trim().to_ascii_lowercase().as_str() {
+            "added" | "add" | "new" | "right" => Some(Self::Added),
+            "removed" | "remove" | "deleted" | "delete" | "old" | "left" => Some(Self::Removed),
+            _ => None,
+        }
+    }
+}
+
+/// What kind of problem a review reports.
+///
+/// The categories are the ones the prompt asks for; a review whose severity is
+/// missing or unreadable falls back to [`ReviewSeverity::Maintainability`], the
+/// least alarming bucket, instead of failing the whole review.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum ReviewSeverity {
+    /// Wrong behaviour.
+    Bug,
+    /// A vulnerability.
+    Security,
+    /// A resource problem.
+    Performance,
+    /// Everything else: readability, structure, naming, tests.
+    Maintainability,
+}
+
+impl ReviewSeverity {
+    /// Reads the severity the model asked for, tolerating casing and synonyms.
+    fn parse(raw: &str) -> Option<Self> {
+        match raw.trim().to_ascii_lowercase().as_str() {
+            "bug" | "bugs" | "correctness" | "error" => Some(Self::Bug),
+            "security" | "vulnerability" => Some(Self::Security),
+            "performance" | "perf" => Some(Self::Performance),
+            "maintainability" | "quality" | "style" => Some(Self::Maintainability),
+            _ => None,
+        }
+    }
+
+    /// Lower-case label used in the review markdown.
+    pub fn label(self) -> &'static str {
+        match self {
+            Self::Bug => "bug",
+            Self::Security => "security",
+            Self::Performance => "performance",
+            Self::Maintainability => "maintainability",
+        }
+    }
+}
+
+/// Reads the side the model asked for. An unreadable value is ignored instead of
+/// failing the whole review: the side is then resolved from the changed lines.
+fn deserialize_side<'de, D>(deserializer: D) -> Result<Option<ReviewSide>, D::Error>
+where
+    D: Deserializer<'de>,
+{
+    let raw = Option::<serde_json::Value>::deserialize(deserializer)?;
+
+    Ok(raw
+        .as_ref()
+        .and_then(serde_json::Value::as_str)
+        .and_then(ReviewSide::parse))
+}
+
+/// Reads the severity the model asked for, ignoring an unreadable value: the
+/// review is then reported as [`ReviewSeverity::Maintainability`].
+fn deserialize_severity<'de, D>(deserializer: D) -> Result<Option<ReviewSeverity>, D::Error>
+where
+    D: Deserializer<'de>,
+{
+    let raw = Option::<serde_json::Value>::deserialize(deserializer)?;
+
+    Ok(raw
+        .as_ref()
+        .and_then(serde_json::Value::as_str)
+        .and_then(ReviewSeverity::parse))
+}
+
 #[derive(Clone)]
 pub struct Bot {
     bot_name: String,
     gitea_api: GiteaAPI,
     open_router_client: OpenRouterClient,
-    http_client: reqwest::Client,
     max_concurrent: usize,
     open_router_model: String,
+    sandbox: SandboxConfig,
     actions_handled: Arc<Mutex<HashSet<u64>>>,
 }
 
@@ -38,17 +134,17 @@ impl Bot {
         bot_name: String,
         gitea_api: GiteaAPI,
         open_router_client: OpenRouterClient,
-        http_client: reqwest::Client,
         max_concurrent: usize,
         open_router_model: String,
+        sandbox: SandboxConfig,
     ) -> Self {
         Self {
             bot_name,
             gitea_api,
             open_router_client,
-            http_client,
             max_concurrent,
             open_router_model,
+            sandbox,
             actions_handled: Arc::new(Mutex::new(HashSet::new())),
         }
     }
@@ -112,12 +208,15 @@ impl Bot {
             }
         };
 
+        let tools = crate::sandbox::tools::for_webhook(&webhook);
+
         let exec_result = match webhook {
             WebhookType::Review(review_payload) => crate::bot_actions::review::exec_review(
                 &self.gitea_api,
                 &self.open_router_client,
-                &self.http_client,
                 &self.open_router_model,
+                &self.sandbox,
+                tools,
                 review_payload,
             ),
         }

crates/herald-server/src/consts.rs

@@ -0,0 +1,82 @@
+pub const GITEA_SIG_HEADER_NAME: &str = "x-gitea-signature";
+pub const GITEA_EVENT_TYPE_HEADER_NAME: &str = "x-gitea-event-type";
+
+pub const MAX_WEBHOOK_BODY_SIZE: usize = 1024 * 1024; // 1 MiB
+pub const MAX_DIFF_SIZE: usize = 1024 * 1024; // 1 MiB
+
+pub const BOT_PROCESS_MSG: &str = "
+    Review in progress with the model \"{model}\"...
+";
+
+pub const SANDBOX_SYSTEM_PROMPT: &str = "
+    You are a senior software engineer reviewing a pull request inside an
+    isolated sandbox.
+
+    The repository is checked out in your working directory. Explore it with the
+    provided read-only tools (ls, file_size, read_file, grep, find); paths are
+    relative to the repository root. Every line you read is prefixed with its
+    absolute line number: that is the number you must cite to anchor a comment.
+
+    Read before you assert: a review anchored on a line you did not read is
+    worthless, so never comment on code you have not seen. Gather enough context
+    to be confident, but do not re-read what you already have.
+
+    When you answer, send the requested JSON only: a raw JSON object, no markdown
+    code fence, nothing before or after it.
+";
+
+pub const REVIEW_PROMPT: &str = "
+    You are a senior software engineer reviewing a pull request.
+
+    Judge the changes for correctness, security, resource use and
+    maintainability. Report real problems; do not invent issues, and do not
+    report pure formatting a formatter would fix.
+
+    Be exhaustive: one review per distinct issue, and cover every changed file
+    that has something to report. Do not stop at the first few findings, and do
+    not merge several issues into a single review.
+
+    This is the pull request subject: \"{subject}\"
+
+    This is the user comment: \"{comment}\"
+    If the user asks about something precise, address that first.
+
+    The pull request changes these files and lines:
+
+    {changes}
+
+    The code is not provided. Read the files you need with the available tools
+    before answering.
+
+    `added` line numbers refer to the new version of the file, which is what your
+    working directory contains now. `removed` line numbers refer to the old
+    version, which is not in your working directory: you cannot read a removed
+    line, only the code around where it used to be.
+
+    Every review must anchor on one of the listed line numbers:
+    - filename: the full path exactly as listed above,
+    - line: one of the numbers listed for that file,
+    - side: \"added\" for a number from the `added` list, \"removed\" for one from
+      the `removed` list,
+    - severity: exactly one of \"bug\", \"security\", \"performance\" or
+      \"maintainability\" — a bug is wrong behaviour, security a vulnerability,
+      performance a resource problem, maintainability everything else.
+
+    Answer in french, with the raw json object only: no markdown code fence, no
+    text before or after it. All fields are mandatory. The `comment` field must
+    hold a short summary that lists the issues, and also what the pull request
+    does well: the author should get the compliments too.
+
+    {
+        \"reviews\": [
+            {
+                \"filename\": \"\",
+                \"line\": 0,
+                \"side\": \"\",
+                \"severity\": \"\",
+                \"message\": \"\"
+            }
+        ],
+        \"comment\": \"\"
+    }
+";

crates/herald-server/src/env.rs

@@ -12,6 +12,7 @@ pub struct EnvConfig {
     pub gitea_token: String,
     pub gitea_timeout: u64,
     pub metrics_bind_addr: Option<String>,
+    pub sandbox_max_iterations: usize,
 }
 
 pub fn load_config() -> anyhow::Result<EnvConfig> {
@@ -25,6 +26,10 @@ pub fn load_config() -> anyhow::Result<EnvConfig> {
     let gitea_token = try_get_env("GITEA_TOKEN")?;
     let gitea_timeout = try_get_env("GITEA_TIMEOUT")?.parse()?;
     let metrics_bind_addr = std::env::var("METRICS_BIND_ADDR").ok();
+    let sandbox_max_iterations = std::env::var("SANDBOX_MAX_ITERATIONS")
+        .ok()
+        .and_then(|value| value.parse().ok())
+        .unwrap_or(8);
 
     Ok(EnvConfig {
         http_port,
@@ -37,6 +42,7 @@ pub fn load_config() -> anyhow::Result<EnvConfig> {
         gitea_token,
         gitea_timeout,
         metrics_bind_addr,
+        sandbox_max_iterations,
     })
 }
 

crates/herald-server/src/gitea.rs

@@ -1,14 +1,28 @@
 use std::time::Duration;
 
+use futures_util::stream::TryStreamExt;
 use serde::Deserialize;
 use serde_json::{Value, json};
-use tracing::instrument;
+use tokio::io::AsyncReadExt;
+use tokio_util::io::StreamReader;
+use tracing::{instrument, warn};
 
-use crate::{bot::ReviewResult, errors::AppError};
+use crate::{
+    bot::{ReviewResult, ReviewSeverity, ReviewSide},
+    consts::MAX_DIFF_SIZE,
+    errors::AppError,
+};
+
+/// Page size requested when listing the files of a pull request.
+const FILE_PAGE_SIZE: u64 = 50;
+
+/// Maximum number of pages fetched for a pull request file list.
+const MAX_FILE_PAGES: u64 = 10;
 
 #[derive(Clone)]
 pub struct GiteaAPI {
     base_url: String,
+    token: String,
     client: reqwest::Client,
 }
 
@@ -22,6 +36,7 @@ impl GiteaAPI {
 
         Ok(Self {
             base_url: String::from(base_url),
+            token: String::from(token),
             client: reqwest::Client::builder()
                 .timeout(Duration::from_secs(timeout))
                 .default_headers(default_headers)
@@ -29,6 +44,11 @@ impl GiteaAPI {
         })
     }
 
+    /// API token used to authenticate against Gitea.
+    pub fn token(&self) -> &str {
+        &self.token
+    }
+
     #[instrument(skip(self))]
     pub async fn get_authorized_user(&self) -> anyhow::Result<User> {
         let url = format!("{}/api/v1/user", self.base_url);
@@ -119,6 +139,70 @@ impl GiteaAPI {
         Ok(())
     }
 
+    /// Raw unified diff of a pull request.
+    ///
+    /// The API endpoint is used rather than the `diff_url` carried by the
+    /// webhook: that one points at a web route, which is session authenticated
+    /// and therefore does not serve private repositories to an API token.
+    #[instrument(skip(self))]
+    pub async fn pull_request_diff(&self, full_name: &str, index: u64) -> anyhow::Result<String> {
+        let url = format!(
+            "{}/api/v1/repos/{}/pulls/{}.diff",
+            self.base_url, full_name, index
+        );
+
+        let res = self.client.get(url).send().await?;
+
+        if !res.status().is_success() {
+            return Err(anyhow::anyhow!(
+                "Failed to download pull request diff: {}",
+                res.status()
+            ));
+        }
+
+        read_capped(res).await
+    }
+
+    /// Files changed by a pull request, with their exact path and status.
+    #[instrument(skip(self))]
+    pub async fn pull_request_files(
+        &self,
+        full_name: &str,
+        index: u64,
+    ) -> anyhow::Result<Vec<PullRequestFile>> {
+        let mut files = Vec::new();
+
+        for page in 1..=MAX_FILE_PAGES {
+            let url = format!(
+                "{}/api/v1/repos/{}/pulls/{}/files?limit={FILE_PAGE_SIZE}&page={page}",
+                self.base_url, full_name, index
+            );
+
+            let res = self.client.get(url).send().await?;
+
+            if !res.status().is_success() {
+                return Err(anyhow::anyhow!(
+                    "Failed to list pull request files: {}",
+                    res.status()
+                ));
+            }
+
+            let page_files = res.json::<Vec<PullRequestFile>>().await?;
+
+            // The instance may cap the page size below the requested one, so a
+            // short page is not the end of the list: an empty one is.
+            if page_files.is_empty() {
+                return Ok(files);
+            }
+
+            files.extend(page_files);
+        }
+
+        warn!(files = files.len(), "Pull request file list was truncated");
+
+        Ok(files)
+    }
+
     #[instrument(skip(self, review_result))]
     pub async fn post_pull_request_review(
         &self,
@@ -132,19 +216,31 @@ impl GiteaAPI {
             self.base_url, full_name, index
         );
 
-        let comments = &review_result
+        let comments = review_result
             .reviews
             .iter()
-            .filter(|r| r.line.is_some())
+            .filter_map(|review| {
-            .map(|r| {
+                let line = review.line?;
-                let path = r.filename.clone();
+                let path = review.filename.clone();
-                let line = r.line.unwrap_or(0);
+                let severity = review
-                let body = r.message.clone();
+                    .severity
+                    .unwrap_or(ReviewSeverity::Maintainability)
+                    .label();
+                let body = format!("**[{severity}]** {}", review.message);
 
-                json!({
+                // A line removed by the pull request only exists in the old
-                    "path": path,
+                // version of the file, so it is anchored with `old_position`.
-                    "new_position": line,
+                Some(match review.side {
-                    "body": body
+                    Some(ReviewSide::Removed) => json!({
+                        "path": path,
+                        "old_position": line,
+                        "body": body
+                    }),
+                    _ => json!({
+                        "path": path,
+                        "new_position": line,
+                        "body": body
+                    }),
                 })
             })
             .collect::<Vec<_>>();
@@ -168,6 +264,23 @@ impl GiteaAPI {
     }
 }
 
+/// Reads a response body, refusing to buffer more than [`MAX_DIFF_SIZE`].
+async fn read_capped(response: reqwest::Response) -> anyhow::Result<String> {
+    let stream = response.bytes_stream().map_err(std::io::Error::other);
+
+    let mut buf = Vec::with_capacity(MAX_DIFF_SIZE);
+    StreamReader::new(stream)
+        .take((MAX_DIFF_SIZE + 1) as u64)
+        .read_to_end(&mut buf)
+        .await?;
+
+    if buf.len() > MAX_DIFF_SIZE {
+        anyhow::bail!("Pull request diff exceeds the maximum allowed size of 1 MiB");
+    }
+
+    Ok(String::from_utf8_lossy(&buf).into_owned())
+}
+
 #[derive(Debug)]
 pub enum WebhookType {
     Review(ReviewPayload),
@@ -197,8 +310,6 @@ pub struct ReviewPayload {
 
 #[derive(Deserialize, Debug)]
 pub struct PullRequest {
-    pub id: u64,
-    pub diff_url: String,
     pub number: u64,
     pub title: String,
 }
@@ -207,18 +318,29 @@ pub struct PullRequest {
 pub struct Comment {
     pub id: u64,
     pub body: String,
-    pub user: User,
 }
 
 #[derive(Deserialize, Debug)]
 pub struct User {
-    pub id: u64,
     pub login: String,
 }
 
 #[derive(Deserialize, Debug)]
 pub struct Repository {
     pub full_name: String,
+    pub clone_url: String,
+}
+
+/// A file changed by a pull request, as reported by the API.
+#[derive(Deserialize, Debug)]
+pub struct PullRequestFile {
+    /// Path of the file in the new version of the repository.
+    pub filename: String,
+    /// Previous path, for a renamed file.
+    #[serde(default)]
+    pub previous_filename: Option<String>,
+    /// `added`, `modified`, `deleted`, `renamed`…
+    pub status: String,
 }
 
 impl WebhookType {
@@ -264,7 +386,8 @@ mod tests {
                 "title": "My PR"
             },
             "repository": {
-                "full_name": "owner/repo"
+                "full_name": "owner/repo",
+                "clone_url": "https://github.com/owner/repo.git"
             },
             "comment": {
                 "id": 7,
@@ -282,10 +405,8 @@ mod tests {
         match result.unwrap() {
             WebhookType::Review(payload) => {
                 assert_eq!(payload.action, "created");
-                assert_eq!(payload.pull_request.id, 42);
                 assert_eq!(payload.comment.id, 7);
                 assert_eq!(payload.comment.body, "@test_bot LGTM");
-                assert_eq!(payload.comment.user.id, 100);
             }
         }
     }
@@ -329,7 +450,8 @@ mod tests {
                 "title": "My PR"
             },
             "repository": {
-                "full_name": "owner/repo"
+                "full_name": "owner/repo",
+                "clone_url": "https://github.com/owner/repo.git"
             },
             "comment": {
                 "id": 1,
@@ -361,7 +483,8 @@ mod tests {
                 "title": "My PR"
             },
             "repository": {
-                "full_name": "owner/repo"
+                "full_name": "owner/repo",
+                "clone_url": "https://github.com/owner/repo.git"
             },
             "comment": {
                 "id": 12,
@@ -375,10 +498,12 @@ mod tests {
 
         let payload: ReviewPayload = serde_json::from_value(json).unwrap();
         assert_eq!(payload.action, "created");
-        assert_eq!(payload.pull_request.id, 99);
         assert_eq!(payload.comment.id, 12);
         assert_eq!(payload.comment.body, "Needs work");
-        assert_eq!(payload.comment.user.id, 200);
+        assert_eq!(
+            payload.repository.clone_url,
+            "https://github.com/owner/repo.git"
+        );
     }
 
     #[test]
@@ -399,7 +524,8 @@ mod tests {
                 "title": "My PR"
             },
             "repository": {
-                "full_name": "owner/repo"
+                "full_name": "owner/repo",
+                "clone_url": "https://github.com/owner/repo.git"
             },
             "comment": {
                 "id": 1,
@@ -426,7 +552,8 @@ mod tests {
                 "title": "My PR"
             },
             "repository": {
-                "full_name": "owner/repo"
+                "full_name": "owner/repo",
+                "clone_url": "https://github.com/owner/repo.git"
             },
             "comment": {
                 "id": 1,

crates/herald-server/src/main.rs

@@ -1,4 +1,12 @@
-use crate::{bot::Bot, gitea::{GiteaAPI, WebhookType}, open_router::OpenRouterClient, state::AppState};
+use std::time::Duration;
+
+use crate::{
+    bot::Bot,
+    gitea::{GiteaAPI, WebhookType},
+    open_router::OpenRouterClient,
+    sandbox::SandboxConfig,
+    state::AppState,
+};
 
 use dotenvy::dotenv;
 use tokio::signal::unix::{SignalKind, signal};
@@ -15,7 +23,15 @@ mod errors;
 mod gitea;
 mod metrics;
 mod open_router;
+mod sandbox;
 mod state;
+mod text;
+
+/// Délai laissé aux reviews en cours et aux serveurs pour s'arrêter proprement.
+///
+/// Sans lui, une review prise dans une sandbox récalcitrante garderait le processus
+/// en vie jusqu'à ce que le superviseur le tue.
+const SHUTDOWN_TIMEOUT: Duration = Duration::from_secs(60);
 
 fn main() -> anyhow::Result<()> {
     dotenv().ok();
@@ -31,14 +47,12 @@ fn main() -> anyhow::Result<()> {
     let _sentry_guard = if let Ok(sentry_dsn) = env::try_get_env("SENTRY_DSN") {
         info!("Initialize sentry");
 
-        Some(sentry::init((
+        Some(sentry::init(
-            sentry_dsn,
+            sentry::ClientOptions::new()
-            sentry::ClientOptions {
+                .dsn(&sentry_dsn)
-                release: sentry::release_name!(),
+                .maybe_release(sentry::release_name!())
-                send_default_pii: true,
+                .send_default_pii(true),
-                ..Default::default()
+        ))
-            },
-        )))
     } else {
         warn!("SENTRY_DSN not set, sentry will not be initialized");
         None
@@ -73,13 +87,25 @@ async fn run() -> anyhow::Result<()> {
 
     let shutdown = CancellationToken::new();
 
+    let sandbox = SandboxConfig {
+        runtime: devcontainer_rs::ContainerRuntime::connect()?,
+        max_iterations: config.sandbox_max_iterations,
+    };
+
+    if !sandbox.runtime.available().await {
+        warn!(
+            endpoint = sandbox.runtime.endpoint(),
+            "Container daemon is not reachable, every review will fail"
+        );
+    }
+
     let bot = Bot::new(
         gitea_user.login,
         gitea_api,
         open_router_client,
-        reqwest::Client::new(),
         config.bot_max_concurrent,
         config.open_router_model.clone(),
+        sandbox,
     );
 
     let (tx, rx) = tokio::sync::mpsc::channel::<WebhookType>(config.bot_max_concurrent * 2);
@@ -102,11 +128,28 @@ async fn run() -> anyhow::Result<()> {
         anyhow::Ok(())
     };
 
-    tokio::try_join!(
+    let shutdown_deadline = async {
-        bot.start(rx, shutdown.clone()),
+        shutdown.cancelled().await;
-        api::start(app_state, shutdown.clone()),
+        tokio::time::sleep(SHUTDOWN_TIMEOUT).await;
-        signal
+    };
-    )?;
+
+    tokio::select! {
+        result = async {
+            tokio::try_join!(
+                bot.start(rx, shutdown.clone()),
+                api::start(app_state, shutdown.clone()),
+                signal
+            )
+        } => {
+            result?;
+        }
+        () = shutdown_deadline => {
+            warn!(
+                timeout = ?SHUTDOWN_TIMEOUT,
+                "Shutdown did not finish in time, exiting anyway"
+            );
+        }
+    }
 
     info!("Shutdown complete");
 

crates/herald-server/src/metrics.rs

@@ -1,10 +1,9 @@
 use std::{net::SocketAddr, str::FromStr};
 
-use metrics::{Unit, describe_counter, describe_gauge, counter, gauge};
+use metrics::{Unit, counter, describe_counter, describe_gauge, gauge};
 
 pub fn webhook_received(event_type: &str) {
-    counter!("herald_webhooks_received_total", "event_type" => event_type.to_string())
+    counter!("herald_webhooks_received_total", "event_type" => event_type.to_string()).increment(1);
-        .increment(1);
 }
 
 pub fn webhook_duplicate(event_type: &str) {
@@ -31,8 +30,7 @@ pub fn task_completed(event_type: &str) {
 }
 
 pub fn task_failed(event_type: &str) {
-    counter!("herald_bot_tasks_failed_total", "event_type" => event_type.to_string())
+    counter!("herald_bot_tasks_failed_total", "event_type" => event_type.to_string()).increment(1);
-        .increment(1);
 }
 
 pub fn openrouter_cost_usd(cost: f64) {
@@ -40,13 +38,41 @@ pub fn openrouter_cost_usd(cost: f64) {
 }
 
 pub fn describe() {
-    describe_counter!("herald_webhooks_received_total", Unit::Count, "Total webhooks received");
+    describe_counter!(
-    describe_counter!("herald_webhooks_duplicate_total", Unit::Count, "Webhooks rejected as duplicates");
+        "herald_webhooks_received_total",
-    describe_counter!("herald_webhooks_channel_full_total", Unit::Count, "Webhooks dropped because the bot channel was full");
+        Unit::Count,
-    describe_gauge!("herald_bot_tasks_active", Unit::Count, "Bot tasks currently in progress");
+        "Total webhooks received"
-    describe_counter!("herald_bot_tasks_completed_total", Unit::Count, "Bot tasks completed successfully");
+    );
-    describe_counter!("herald_bot_tasks_failed_total", Unit::Count, "Bot tasks that failed");
+    describe_counter!(
-    describe_counter!("herald_openrouter_cost_cents_total", Unit::Count, "Total OpenRouter cost in cents (divide by 100 for USD)");
+        "herald_webhooks_duplicate_total",
+        Unit::Count,
+        "Webhooks rejected as duplicates"
+    );
+    describe_counter!(
+        "herald_webhooks_channel_full_total",
+        Unit::Count,
+        "Webhooks dropped because the bot channel was full"
+    );
+    describe_gauge!(
+        "herald_bot_tasks_active",
+        Unit::Count,
+        "Bot tasks currently in progress"
+    );
+    describe_counter!(
+        "herald_bot_tasks_completed_total",
+        Unit::Count,
+        "Bot tasks completed successfully"
+    );
+    describe_counter!(
+        "herald_bot_tasks_failed_total",
+        Unit::Count,
+        "Bot tasks that failed"
+    );
+    describe_counter!(
+        "herald_openrouter_cost_cents_total",
+        Unit::Count,
+        "Total OpenRouter cost in cents (divide by 100 for USD)"
+    );
 }
 
 pub fn install(bind_addr: &str) -> anyhow::Result<()> {

crates/herald-server/src/open_router.rs

@@ -0,0 +1,473 @@
+//! Minimal OpenRouter chat-completions client.
+//!
+//! Herald only needs a non-streaming `POST /chat/completions` with optional
+//! tool calling, so the wire types are implemented in-tree instead of pulling a
+//! third-party SDK (and its own `reqwest` version) into the workspace.
+//!
+//! Only the response fields Herald consumes are modelled: `content`,
+//! `tool_calls` and `usage.cost`. Unknown fields are ignored.
+
+use std::time::Duration;
+
+use anyhow::Context;
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+use tracing::instrument;
+
+/// OpenRouter API root, version prefix included.
+const BASE_URL: &str = "https://openrouter.ai/api/v1";
+
+/// What the model is allowed to do on a given turn.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum ToolChoice {
+    /// The model decides whether to call a tool.
+    Auto,
+    /// The model must answer, without calling any tool. The tools stay declared,
+    /// so the conversation remains the same shape as on the other turns.
+    None,
+}
+
+impl ToolChoice {
+    fn as_str(self) -> &'static str {
+        match self {
+            Self::Auto => "auto",
+            Self::None => "none",
+        }
+    }
+}
+
+/// Result of a completion that may contain tool calls.
+pub struct ToolChatResult {
+    pub message: Option<String>,
+    pub tool_calls: Vec<ToolCall>,
+    pub cost: Option<f64>,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone, Copy, PartialEq, Eq)]
+#[serde(rename_all = "lowercase")]
+pub enum Role {
+    System,
+    User,
+    Assistant,
+    Tool,
+}
+
+/// A single turn of the conversation.
+#[derive(Serialize, Deserialize, Debug, Clone)]
+pub struct Message {
+    pub role: Role,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub content: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub tool_calls: Option<Vec<ToolCall>>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub tool_call_id: Option<String>,
+}
+
+impl Message {
+    pub fn new(role: Role, content: impl Into<String>) -> Self {
+        Self {
+            role,
+            content: Some(content.into()),
+            tool_calls: None,
+            tool_call_id: None,
+        }
+    }
+
+    /// Assistant turn requesting tool calls. Models often answer with tool
+    /// calls but no text, in which case the content is sent as `null`.
+    pub fn assistant_with_tool_calls(
+        content: impl Into<String>,
+        tool_calls: Vec<ToolCall>,
+    ) -> Self {
+        let content = content.into();
+
+        Self {
+            role: Role::Assistant,
+            content: (!content.is_empty()).then_some(content),
+            tool_calls: Some(tool_calls),
+            tool_call_id: None,
+        }
+    }
+
+    /// Result of a tool call, linked to the request by `tool_call_id`.
+    pub fn tool_response(tool_call_id: impl Into<String>, content: impl Into<String>) -> Self {
+        Self {
+            role: Role::Tool,
+            content: Some(content.into()),
+            tool_calls: None,
+            tool_call_id: Some(tool_call_id.into()),
+        }
+    }
+}
+
+/// A tool the model may call.
+#[derive(Serialize, Deserialize, Debug, Clone)]
+pub struct Tool {
+    #[serde(rename = "type", default = "function_type")]
+    pub kind: String,
+    pub function: FunctionDefinition,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone)]
+pub struct FunctionDefinition {
+    pub name: String,
+    pub description: String,
+    /// JSON schema describing the accepted arguments.
+    pub parameters: Value,
+}
+
+impl Tool {
+    pub fn new(name: impl Into<String>, description: impl Into<String>, parameters: Value) -> Self {
+        Self {
+            kind: function_type(),
+            function: FunctionDefinition {
+                name: name.into(),
+                description: description.into(),
+                parameters,
+            },
+        }
+    }
+}
+
+/// A tool call requested by the model.
+#[derive(Serialize, Deserialize, Debug, Clone)]
+pub struct ToolCall {
+    pub id: String,
+    #[serde(rename = "type", default = "function_type")]
+    pub kind: String,
+    pub function: FunctionCall,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone)]
+pub struct FunctionCall {
+    pub name: String,
+    /// Arguments as a JSON-encoded string. Kept verbatim so that echoing the
+    /// call back into the conversation does not re-encode or corrupt it.
+    pub arguments: String,
+}
+
+impl ToolCall {
+    pub fn id(&self) -> &str {
+        &self.id
+    }
+
+    pub fn name(&self) -> &str {
+        &self.function.name
+    }
+
+    pub fn arguments_json(&self) -> &str {
+        &self.function.arguments
+    }
+}
+
+fn function_type() -> String {
+    String::from("function")
+}
+
+#[derive(Serialize)]
+struct ChatRequest<'a> {
+    model: &'a str,
+    messages: &'a [Message],
+    reasoning: Reasoning,
+    tools: &'a [Tool],
+    tool_choice: &'a str,
+}
+
+/// OpenRouter reasoning config; Herald always asks for reasoning.
+#[derive(Serialize)]
+struct Reasoning {
+    enabled: bool,
+}
+
+#[derive(Deserialize)]
+struct ChatResponse {
+    #[serde(default)]
+    choices: Vec<Choice>,
+    #[serde(default)]
+    usage: Option<Usage>,
+}
+
+#[derive(Deserialize)]
+struct Choice {
+    message: ResponseMessage,
+}
+
+#[derive(Deserialize)]
+struct ResponseMessage {
+    #[serde(default)]
+    content: Option<Content>,
+    #[serde(default)]
+    tool_calls: Option<Vec<ToolCall>>,
+}
+
+/// Message content is a plain string for most models and a list of typed parts
+/// for others; both are collapsed to text.
+#[derive(Deserialize)]
+#[serde(untagged)]
+enum Content {
+    Text(String),
+    Parts(Vec<ContentPart>),
+}
+
+#[derive(Deserialize)]
+struct ContentPart {
+    #[serde(default)]
+    text: Option<String>,
+}
+
+impl Content {
+    fn into_text(self) -> String {
+        match self {
+            Self::Text(text) => text,
+            Self::Parts(parts) => parts
+                .into_iter()
+                .filter_map(|part| part.text)
+                .collect::<Vec<_>>()
+                .join(""),
+        }
+    }
+}
+
+#[derive(Deserialize)]
+struct Usage {
+    #[serde(default)]
+    cost: Option<f64>,
+}
+
+/// Error payload returned by OpenRouter for a failed request.
+#[derive(Deserialize)]
+struct ErrorResponse {
+    error: ErrorDetail,
+}
+
+#[derive(Deserialize)]
+struct ErrorDetail {
+    message: String,
+}
+
+#[derive(Clone)]
+pub struct OpenRouterClient {
+    client: reqwest::Client,
+    api_key: String,
+    model: String,
+}
+
+impl OpenRouterClient {
+    pub fn new(token: &str, model: &str, timeout: u64) -> anyhow::Result<Self> {
+        Ok(Self {
+            client: reqwest::Client::builder()
+                .timeout(Duration::from_secs(timeout))
+                .build()?,
+            api_key: String::from(token),
+            model: String::from(model),
+        })
+    }
+
+    /// Sends a conversation with tool definitions and returns either a final
+    /// message or the tool calls requested by the model.
+    ///
+    /// The conversation is borrowed rather than taken by value: the caller keeps
+    /// appending to it between iterations, and a copy per iteration would be pure
+    /// waste.
+    #[instrument(skip(self, messages, tools), err)]
+    pub async fn chat_with_tools(
+        &self,
+        messages: &[Message],
+        tools: &[Tool],
+        tool_choice: ToolChoice,
+    ) -> anyhow::Result<ToolChatResult> {
+        let response = self.complete(messages, tools, tool_choice).await?;
+
+        let cost = response.usage.and_then(|usage| usage.cost);
+        let message = response
+            .choices
+            .into_iter()
+            .next()
+            .ok_or_else(|| anyhow::anyhow!("No choices in response"))?
+            .message;
+
+        Ok(ToolChatResult {
+            message: message.content.map(Content::into_text),
+            tool_calls: message.tool_calls.unwrap_or_default(),
+            cost,
+        })
+    }
+
+    async fn complete(
+        &self,
+        messages: &[Message],
+        tools: &[Tool],
+        tool_choice: ToolChoice,
+    ) -> anyhow::Result<ChatResponse> {
+        let request = ChatRequest {
+            model: &self.model,
+            messages,
+            reasoning: Reasoning { enabled: true },
+            tools,
+            tool_choice: tool_choice.as_str(),
+        };
+
+        let response = self
+            .client
+            .post(format!("{BASE_URL}/chat/completions"))
+            .bearer_auth(&self.api_key)
+            .json(&request)
+            .send()
+            .await
+            .context("failed to reach OpenRouter")?;
+
+        let status = response.status();
+        if !status.is_success() {
+            let body = response.text().await.unwrap_or_default();
+            anyhow::bail!("OpenRouter returned {status}: {}", error_message(&body));
+        }
+
+        response
+            .json::<ChatResponse>()
+            .await
+            .context("invalid OpenRouter response")
+    }
+}
+
+/// Extracts the message from an OpenRouter error body, falling back to the raw
+/// body when it is not the expected JSON shape.
+fn error_message(body: &str) -> String {
+    serde_json::from_str::<ErrorResponse>(body)
+        .map(|response| response.error.message)
+        .unwrap_or_else(|_| body.trim().to_string())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    /// Builds a tool call the way the API returns one, so the fixture also
+    /// covers deserialization.
+    fn tool_call(id: &str, name: &str, arguments: &str) -> ToolCall {
+        serde_json::from_value(json!({
+            "id": id,
+            "type": "function",
+            "function": { "name": name, "arguments": arguments }
+        }))
+        .unwrap()
+    }
+
+    #[test]
+    fn tool_call_arguments_are_kept_verbatim() {
+        let call = tool_call("call_1", "grep", r#"{"pattern":"fn main"}"#);
+        let message = Message::assistant_with_tool_calls("", vec![call]);
+
+        let serialized = serde_json::to_value(&message).unwrap();
+
+        assert_eq!(serialized["content"], Value::Null);
+        assert_eq!(serialized["tool_calls"][0]["type"], "function");
+        assert_eq!(serialized["tool_calls"][0]["function"]["name"], "grep");
+        assert_eq!(
+            serialized["tool_calls"][0]["function"]["arguments"],
+            r#"{"pattern":"fn main"}"#
+        );
+    }
+
+    #[test]
+    fn tool_response_carries_role_and_tool_call_id() {
+        let message = Message::tool_response("call_1", "src/main.rs");
+        let serialized = serde_json::to_value(&message).unwrap();
+
+        assert_eq!(serialized["role"], "tool");
+        assert_eq!(serialized["tool_call_id"], "call_1");
+        assert_eq!(serialized["content"], "src/main.rs");
+    }
+
+    #[test]
+    fn request_serializes_tool_choice_and_reasoning() {
+        let messages = [Message::new(Role::User, "hi")];
+        let tools = [Tool::new("ls", "List files", json!({"type": "object"}))];
+
+        let request = serde_json::to_value(ChatRequest {
+            model: "some/model",
+            messages: &messages,
+            reasoning: Reasoning { enabled: true },
+            tools: &tools,
+            tool_choice: ToolChoice::Auto.as_str(),
+        })
+        .unwrap();
+
+        assert_eq!(request["model"], "some/model");
+        assert_eq!(request["reasoning"]["enabled"], true);
+        assert_eq!(request["tool_choice"], "auto");
+        assert_eq!(request["tools"][0]["function"]["name"], "ls");
+        assert_eq!(request["messages"][0]["role"], "user");
+    }
+
+    #[test]
+    fn a_forced_answer_turn_disables_the_tools() {
+        assert_eq!(ToolChoice::None.as_str(), "none");
+        assert_eq!(ToolChoice::Auto.as_str(), "auto");
+    }
+
+    #[test]
+    fn response_parses_tool_calls_and_cost() {
+        let response: ChatResponse = serde_json::from_value(json!({
+            "choices": [{
+                "message": {
+                    "role": "assistant",
+                    "content": null,
+                    "tool_calls": [{
+                        "id": "call_1",
+                        "type": "function",
+                        "function": { "name": "ls", "arguments": "{\"path\":\"src\"}" }
+                    }]
+                }
+            }],
+            "usage": { "cost": 0.0021 }
+        }))
+        .unwrap();
+
+        assert_eq!(response.usage.and_then(|usage| usage.cost), Some(0.0021));
+
+        let message = response.choices.into_iter().next().unwrap().message;
+        assert!(message.content.is_none());
+
+        let calls = message.tool_calls.unwrap();
+        assert_eq!(calls[0].name(), "ls");
+        assert_eq!(calls[0].id(), "call_1");
+        assert_eq!(calls[0].arguments_json(), r#"{"path":"src"}"#);
+    }
+
+    #[test]
+    fn response_parses_parts_content() {
+        let response: ChatResponse = serde_json::from_value(json!({
+            "choices": [{
+                "message": {
+                    "role": "assistant",
+                    "content": [
+                        { "type": "text", "text": "hello " },
+                        { "type": "text", "text": "world" }
+                    ]
+                }
+            }]
+        }))
+        .unwrap();
+
+        assert!(response.usage.is_none());
+
+        let message = response.choices.into_iter().next().unwrap().message;
+        assert_eq!(message.content.unwrap().into_text(), "hello world");
+    }
+
+    #[test]
+    fn error_message_prefers_api_message() {
+        let body = r#"{"error":{"message":"No auth credentials found","code":401}}"#;
+        assert_eq!(error_message(body), "No auth credentials found");
+    }
+
+    #[test]
+    fn error_message_falls_back_to_raw_body() {
+        assert_eq!(
+            error_message("  <html>bad gateway</html>  "),
+            "<html>bad gateway</html>"
+        );
+    }
+}

crates/herald-server/src/sandbox/agent.rs

@@ -0,0 +1,316 @@
+//! Tool-calling loop driving the model against a [`Sandbox`].
+//!
+//! The loop sends the conversation and the available tools to OpenRouter,
+//! executes any requested tool call inside the sandbox and feeds the results
+//! back, until the model produces a final message or the iteration budget is
+//! exhausted.
+
+use std::str::FromStr;
+use std::time::Instant;
+
+use anyhow::Context;
+use serde_json::Value;
+use tracing::{info, warn};
+
+use crate::{
+    open_router::{Message, OpenRouterClient, Role, Tool, ToolCall, ToolChoice},
+    sandbox::{Sandbox, tools},
+    text::{excerpt, truncate_bytes},
+};
+
+/// Number of characters of the arguments of a tool call kept in the logs.
+const MAX_LOGGED_ARGUMENTS: usize = 200;
+
+/// Number of characters of a tool result kept in the logs.
+///
+/// Sizes alone make a broken tool invisible: a command that fails returns a short
+/// `error: …` instead of the expected content, which is exactly what this excerpt
+/// makes obvious.
+const MAX_LOGGED_OUTPUT: usize = 120;
+
+/// Maximum number of bytes of a tool result handed to the model.
+///
+/// A result stays in the conversation and is re-sent on every following
+/// iteration, so an unbounded one (a whole file, a match on every line) inflates
+/// the context and the cost of the whole rest of the run — and can overflow the
+/// model context window outright.
+const MAX_TOOL_OUTPUT_BYTES: usize = 32 * 1024;
+
+/// Asked of the model when its answer does not satisfy the caller's contract.
+const RETRY_PROMPT: &str = "
+    Your answer is not valid: {error}
+
+    Answer again with the requested format only: a raw json object, without any
+    markdown code fence and without any text before or after it.
+";
+
+/// Asked of the model on the last turn, in place of one more exploration turn.
+///
+/// Without it, a model still exploring on its last turn would ask for another
+/// tool, and the run would end on the iteration budget with no answer at all.
+const FINAL_PROMPT: &str = "
+    You are out of turns: this is your last one. Answer now, with the requested
+    format, from what you have already gathered, without calling any tool.
+";
+
+/// Final output of an agent run.
+pub struct AgentResult<A> {
+    /// Answer of the model, parsed into the type the caller asked for.
+    pub answer: A,
+    pub cost: Option<f64>,
+    pub iterations: usize,
+}
+
+/// Runs the tool-calling loop until the model answers or `max_iterations` is
+/// reached.
+///
+/// `tool_definitions` is the set of tools the model may call; it is selected by
+/// the caller based on the webhook action (see [`tools::for_webhook`]).
+///
+/// The answer of the model is parsed into `A`, which is how the caller states the
+/// format it expects: while [`FromStr`] rejects the answer, the error is sent back
+/// to the model so that it can correct itself, which costs an iteration. The
+/// parsing happens here because that is where the conversation and the remaining
+/// budget are at hand.
+pub async fn run<A>(
+    open_router: &OpenRouterClient,
+    sandbox: &Sandbox,
+    tool_definitions: Vec<Tool>,
+    system_prompt: &str,
+    user_prompt: &str,
+    max_iterations: usize,
+) -> anyhow::Result<AgentResult<A>>
+where
+    A: FromStr,
+    A::Err: std::fmt::Display,
+{
+    let mut messages = vec![
+        Message::new(Role::System, system_prompt),
+        Message::new(Role::User, user_prompt),
+    ];
+
+    info!(
+        tools = %tool_definitions
+            .iter()
+            .map(|tool| tool.function.name.as_str())
+            .collect::<Vec<_>>()
+            .join(", "),
+        max_iterations,
+        "Starting tool-calling loop"
+    );
+
+    let mut total_cost = 0.0_f64;
+    let mut has_cost = false;
+    let mut last_rejection: Option<String> = None;
+
+    for iteration in 1..=max_iterations {
+        // Le dernier tour n'est plus un tour d'exploration : le modèle doit rendre
+        // sa réponse, avec ce qu'il a déjà vu.
+        let last = iteration == max_iterations;
+
+        if last {
+            info!(iteration, "Last turn: asking for the final answer");
+            messages.push(Message::new(Role::User, FINAL_PROMPT));
+        }
+
+        let started = Instant::now();
+        let response = open_router
+            .chat_with_tools(
+                &messages,
+                &tool_definitions,
+                if last {
+                    ToolChoice::None
+                } else {
+                    ToolChoice::Auto
+                },
+            )
+            .await?;
+
+        if let Some(cost) = response.cost {
+            total_cost += cost;
+            has_cost = true;
+        }
+
+        info!(
+            iteration,
+            max_iterations,
+            last_turn = last,
+            tool_calls = response.tool_calls.len(),
+            answer_bytes = response.message.as_deref().map_or(0, str::len),
+            elapsed_ms = started.elapsed().as_millis() as u64,
+            "Model answered"
+        );
+
+        // Au dernier tour, les outils sont refusés : même si le fournisseur les
+        // renvoie malgré tout, on ne les exécute pas et on tente la réponse.
+        if last || response.tool_calls.is_empty() {
+            let answer = response.message.unwrap_or_default();
+
+            match answer.parse::<A>() {
+                Ok(parsed) => {
+                    return Ok(AgentResult {
+                        answer: parsed,
+                        cost: has_cost.then_some(total_cost),
+                        iterations: iteration,
+                    });
+                }
+                Err(err) => {
+                    warn!(iteration, %err, "Model answer was rejected, asking for another one");
+
+                    // Keep the rejected answer in the history, so that the model
+                    // sees what it has to fix.
+                    messages.push(Message::new(Role::Assistant, answer));
+                    messages.push(Message::new(
+                        Role::User,
+                        RETRY_PROMPT.replace("{error}", &err.to_string()),
+                    ));
+
+                    last_rejection = Some(err.to_string());
+                    continue;
+                }
+            }
+        }
+
+        messages.push(Message::assistant_with_tool_calls(
+            response.message.unwrap_or_default(),
+            response.tool_calls.clone(),
+        ));
+
+        for call in &response.tool_calls {
+            let started = Instant::now();
+            let (content, truncated) = execute(sandbox, call).await;
+
+            info!(
+                iteration,
+                tool = call.name(),
+                arguments = %excerpt(call.arguments_json(), MAX_LOGGED_ARGUMENTS),
+                output_bytes = content.len(),
+                truncated,
+                // Les sauts de ligne casseraient la lisibilité d'une ligne de log.
+                output = %excerpt(&content, MAX_LOGGED_OUTPUT).replace('\n', " "),
+                elapsed_ms = started.elapsed().as_millis() as u64,
+                "Tool call finished"
+            );
+
+            messages.push(Message::tool_response(call.id(), content));
+        }
+    }
+
+    warn!(
+        max_iterations,
+        "Agent used its last turn without a valid answer"
+    );
+
+    match last_rejection {
+        Some(error) => anyhow::bail!(
+            "agent exceeded the maximum number of iterations ({max_iterations}), \
+             last answer rejected: {error}"
+        ),
+        None => anyhow::bail!("agent exceeded the maximum number of iterations ({max_iterations})"),
+    }
+}
+
+/// Runs a tool call and returns its result for the model, along with whether the
+/// result had to be truncated.
+async fn execute(sandbox: &Sandbox, call: &ToolCall) -> (String, bool) {
+    let args = match parse_args(call) {
+        Ok(args) => args,
+        Err(err) => return cap_output(format!("error: {err}")),
+    };
+
+    match tools::dispatch(sandbox, call.name(), &args).await {
+        Ok(output) => cap_output(output),
+        Err(err) => cap_output(format!("error: {err}")),
+    }
+}
+
+/// Bounds [`MAX_TOOL_OUTPUT_BYTES`] of a tool result, telling the model what was
+/// dropped so that it can narrow its request.
+fn cap_output(mut output: String) -> (String, bool) {
+    let total = output.len();
+
+    if !truncate_bytes(&mut output, MAX_TOOL_OUTPUT_BYTES) {
+        return (output, false);
+    }
+
+    let kept = output.len();
+    output.push_str(&format!(
+        "\n… output truncated: {total} bytes in total, the first {kept} are shown. \
+         Narrow the request (path, pattern or line range) to see the rest."
+    ));
+
+    (output, true)
+}
+
+fn parse_args(call: &ToolCall) -> anyhow::Result<Value> {
+    let raw = call.arguments_json().trim();
+    if raw.is_empty() {
+        return Ok(Value::Object(serde_json::Map::new()));
+    }
+
+    serde_json::from_str(raw)
+        .with_context(|| format!("invalid arguments for tool `{}`", call.name()))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    /// Builds a tool call the way the API returns one, so the fixture also
+    /// covers deserialization.
+    fn tool_call(name: &str, arguments: &str) -> ToolCall {
+        serde_json::from_value(json!({
+            "id": "call_1",
+            "type": "function",
+            "function": { "name": name, "arguments": arguments }
+        }))
+        .unwrap()
+    }
+
+    #[test]
+    fn parse_args_accepts_empty_arguments() {
+        let call = tool_call("ls", "");
+        assert_eq!(
+            parse_args(&call).unwrap(),
+            Value::Object(serde_json::Map::new())
+        );
+    }
+
+    #[test]
+    fn parse_args_parses_json_object() {
+        let call = tool_call("ls", r#"{"path":"src"}"#);
+        assert_eq!(parse_args(&call).unwrap()["path"], "src");
+    }
+
+    #[test]
+    fn parse_args_rejects_invalid_json() {
+        let call = tool_call("ls", "not json");
+        assert!(parse_args(&call).is_err());
+    }
+
+    #[test]
+    fn a_small_tool_output_is_kept_as_is() {
+        let (content, truncated) = cap_output(String::from("src/main.rs"));
+
+        assert_eq!(content, "src/main.rs");
+        assert!(!truncated);
+    }
+
+    #[test]
+    fn a_large_tool_output_is_truncated_and_told_to_the_model() {
+        let (content, truncated) = cap_output("a".repeat(MAX_TOOL_OUTPUT_BYTES + 1));
+
+        assert!(truncated);
+        assert!(content.starts_with(&"a".repeat(MAX_TOOL_OUTPUT_BYTES)));
+        assert!(content.contains("output truncated"));
+    }
+
+    #[test]
+    fn a_truncated_tool_output_stays_valid_utf8() {
+        let (content, truncated) = cap_output("é".repeat(MAX_TOOL_OUTPUT_BYTES));
+
+        assert!(truncated);
+        assert!(content.starts_with('é'));
+    }
+}

crates/herald-server/src/sandbox/instance.rs

@@ -0,0 +1,288 @@
+//! Le [`Sandbox`] : clone d'une pull request exécuté dans un devcontainer
+//! éphémère, et sa configuration.
+
+use std::{
+    path::{Path, PathBuf},
+    process::Stdio,
+};
+
+use anyhow::Context;
+use devcontainer_rs::{Container, ContainerRuntime, ExecOutput};
+use tempfile::TempDir;
+use tracing::{info, instrument};
+
+/// Devcontainer locations recognized within a repository, in priority order.
+const DEVCONTAINER_PATHS: [&str; 2] = [".devcontainer/devcontainer.json", ".devcontainer.json"];
+
+/// Sandbox-related runtime configuration.
+#[derive(Clone)]
+pub struct SandboxConfig {
+    /// Client du daemon de containers qui exécute la sandbox.
+    pub runtime: ContainerRuntime,
+    /// Maximum number of tool-calling iterations per agent run.
+    pub max_iterations: usize,
+}
+
+/// A cloned repository running inside an ephemeral devcontainer.
+pub struct Sandbox {
+    // Owns the temporary directory; dropping it cleans up the clone.
+    _workspace: TempDir,
+    /// Path of the clone, as copied into the container.
+    repo_dir: PathBuf,
+    container: Container,
+}
+
+impl Sandbox {
+    /// Clones the pull request head, builds the devcontainer and starts it.
+    ///
+    /// The clone is PR-aware: it fetches `refs/pull/<number>/head`, which works
+    /// for both same-repository and forked pull requests.
+    #[instrument(skip(runtime, token), fields(pr = pull_request_number))]
+    pub async fn create(
+        runtime: &ContainerRuntime,
+        clone_url: &str,
+        token: &str,
+        pull_request_number: u64,
+    ) -> anyhow::Result<Self> {
+        let workspace = tempfile::tempdir().context("failed to create sandbox workspace")?;
+        let repo_dir = workspace.path().join("repo");
+
+        clone_pull_request(clone_url, token, pull_request_number, &repo_dir).await?;
+        make_readable(&repo_dir).await?;
+
+        let devcontainer_path = find_devcontainer(&repo_dir)
+            .with_context(|| format!("no devcontainer found in `{clone_url}`"))?;
+
+        let devcontainer = devcontainer_rs::parse(&devcontainer_path).await?;
+
+        info!(image = %devcontainer.image_tag(), "Building and starting sandbox container");
+        let container = devcontainer.up(runtime, &repo_dir).await?;
+
+        let sandbox = Self {
+            _workspace: workspace,
+            repo_dir,
+            container,
+        };
+
+        if let Err(err) = sandbox.check_workspace().await {
+            let _ = sandbox.container.remove().await;
+            return Err(err);
+        }
+
+        Ok(sandbox)
+    }
+
+    /// Vérifie que le clone est bien visible dans le container.
+    ///
+    /// Sans ce contrôle, un workspace vide — par exemple un daemon qui n'a pas pu
+    /// recevoir le clone — fait échouer chaque outil ; le modèle enchaîne alors les
+    /// appels ratés jusqu'au budget d'itérations, sans jamais pouvoir reviewer quoi
+    /// que ce soit.
+    async fn check_workspace(&self) -> anyhow::Result<()> {
+        let workspace_folder = self.workspace_folder();
+        let probe = self
+            .container
+            .exec(&["ls", "-A", "--", workspace_folder])
+            .await
+            .with_context(|| format!("failed to list `{workspace_folder}` in the sandbox"))?;
+
+        if !probe.success() {
+            let details = self.diagnose_workspace().await;
+
+            anyhow::bail!(
+                "the sandbox cannot list `{workspace_folder}`: {} ({details})",
+                probe.stderr.trim()
+            );
+        }
+
+        if probe.stdout.trim().is_empty() {
+            anyhow::bail!(
+                "the sandbox workspace `{workspace_folder}` is empty: the container daemon does not see the clone"
+            );
+        }
+
+        info!(
+            clone = %self.repo_dir.display(),
+            workspace = %workspace_folder,
+            entries = probe.stdout.lines().count(),
+            "Sandbox workspace is readable"
+        );
+
+        Ok(())
+    }
+
+    /// Rassemble de quoi expliquer un refus d'accès au workspace.
+    ///
+    /// Un `EACCES` a deux causes possibles, indistinguables dans le message de
+    /// `ls` : les permissions du clone, ou un confinement du noyau qui bloque
+    /// l'accès. L'identité de l'utilisateur d'exec et les permissions du point de
+    /// montage permettent de trancher.
+    async fn diagnose_workspace(&self) -> String {
+        let mut details = Vec::new();
+
+        if let Ok(output) = self.container.exec(&["id"]).await {
+            details.push(output.stdout.trim().to_string());
+        }
+
+        if let Ok(output) = self
+            .container
+            .exec(&["ls", "-ld", "--", self.workspace_folder()])
+            .await
+            && output.success()
+        {
+            details.push(output.stdout.trim().to_string());
+        }
+
+        details.join(", ")
+    }
+
+    /// Executes a command in the container as an argv vector (no shell).
+    pub async fn exec(&self, cmd: &[&str]) -> anyhow::Result<ExecOutput> {
+        Ok(self.container.exec(cmd).await?)
+    }
+
+    /// Path of the repository inside the container.
+    pub fn workspace_folder(&self) -> &str {
+        self.container.workspace_folder()
+    }
+
+    /// Stops and removes the container. The temporary clone is removed on drop.
+    pub async fn cleanup(self) -> anyhow::Result<()> {
+        self.container.remove().await?;
+        Ok(())
+    }
+}
+
+fn find_devcontainer(repo_dir: &Path) -> Option<PathBuf> {
+    DEVCONTAINER_PATHS
+        .iter()
+        .map(|relative| repo_dir.join(relative))
+        .find(|candidate| candidate.is_file())
+}
+
+async fn clone_pull_request(
+    repo_url: &str,
+    token: &str,
+    pull_request_number: u64,
+    dest: &Path,
+) -> anyhow::Result<()> {
+    let dest = dest.display().to_string();
+
+    run_git(
+        token,
+        &[
+            "clone".to_string(),
+            "--depth".to_string(),
+            "1".to_string(),
+            repo_url.to_string(),
+            dest.clone(),
+        ],
+    )
+    .await?;
+
+    run_git(
+        token,
+        &[
+            "-C".to_string(),
+            dest.clone(),
+            "fetch".to_string(),
+            "--depth".to_string(),
+            "1".to_string(),
+            "origin".to_string(),
+            format!("refs/pull/{pull_request_number}/head"),
+        ],
+    )
+    .await?;
+
+    run_git(
+        token,
+        &[
+            "-C".to_string(),
+            dest,
+            "checkout".to_string(),
+            "FETCH_HEAD".to_string(),
+        ],
+    )
+    .await?;
+
+    Ok(())
+}
+
+/// Rend le clone lisible par les utilisateurs du container sandbox.
+///
+/// Le container peut ne pas avoir les mêmes uid que Herald (podman rootless
+/// mappe les uid à travers des plages subuid), et l'umask de l'opérateur peut être
+/// restrictif : sans cela, les outils de la sandbox échouent en `Permission
+/// denied` sur des fichiers que Herald vient de cloner lui-même.
+async fn make_readable(repo_dir: &Path) -> anyhow::Result<()> {
+    let output = tokio::process::Command::new("chmod")
+        .args(["-R", "a+rX"])
+        .arg(repo_dir)
+        .stdin(Stdio::null())
+        .output()
+        .await
+        .context("failed to spawn chmod")?;
+
+    if !output.status.success() {
+        anyhow::bail!(
+            "chmod failed on `{}`: {}",
+            repo_dir.display(),
+            String::from_utf8_lossy(&output.stderr).trim()
+        );
+    }
+
+    Ok(())
+}
+
+/// Runs git with the token injected through `http.extraHeader`, keeping the
+/// secret out of the process arguments.
+async fn run_git(token: &str, args: &[String]) -> anyhow::Result<()> {
+    let output = tokio::process::Command::new("git")
+        .args(args)
+        .env("GIT_CONFIG_COUNT", "1")
+        .env("GIT_CONFIG_KEY_0", "http.extraHeader")
+        .env(
+            "GIT_CONFIG_VALUE_0",
+            format!("Authorization: token {token}"),
+        )
+        .env("GIT_TERMINAL_PROMPT", "0")
+        .stdin(Stdio::null())
+        .output()
+        .await
+        .context("failed to spawn git")?;
+
+    if !output.status.success() {
+        anyhow::bail!(
+            "git {} failed: {}",
+            args.join(" "),
+            String::from_utf8_lossy(&output.stderr).trim()
+        );
+    }
+
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn find_devcontainer_prefers_dot_devcontainer_dir() {
+        let dir = tempfile::tempdir().unwrap();
+        let nested = dir.path().join(".devcontainer");
+        std::fs::create_dir(&nested).unwrap();
+        std::fs::write(nested.join("devcontainer.json"), "{}").unwrap();
+        std::fs::write(dir.path().join(".devcontainer.json"), "{}").unwrap();
+
+        assert_eq!(
+            find_devcontainer(dir.path()),
+            Some(nested.join("devcontainer.json"))
+        );
+    }
+
+    #[test]
+    fn find_devcontainer_returns_none_when_absent() {
+        let dir = tempfile::tempdir().unwrap();
+        assert_eq!(find_devcontainer(dir.path()), None);
+    }
+}

crates/herald-server/src/sandbox/mod.rs

@@ -0,0 +1,13 @@
+//! Sandboxed tool execution for the AI bot.
+//!
+//! A [`Sandbox`] clones a pull request into a temporary directory, builds and
+//! starts its devcontainer (via `devcontainer-rs`) and exposes command
+//! execution inside the resulting container. The [`tools`] module maps model
+//! tool calls to commands run in that container, and [`agent`] drives the
+//! tool-calling loop against OpenRouter.
+
+pub mod agent;
+mod instance;
+pub mod tools;
+
+pub use instance::{Sandbox, SandboxConfig};

crates/herald-server/src/sandbox/tools.rs

@@ -0,0 +1,303 @@
+//! Tool definitions exposed to the model and their execution inside a
+//! [`Sandbox`].
+//!
+//! Every tool is read-only and confined to the repository workspace: paths are
+//! resolved relative to the container workspace folder and rejected if they
+//! escape it. Commands are executed as argv vectors (never through a shell),
+//! so tool arguments cannot be used for shell injection.
+
+use std::path::Path;
+
+use anyhow::{Context, bail};
+use devcontainer_rs::{ExecOutput, normalize};
+use serde_json::{Value, json};
+
+use super::Sandbox;
+use crate::{gitea::WebhookType, open_router::Tool};
+
+/// Tools available to the model for a given webhook action.
+///
+/// The match is exhaustive on [`WebhookType`], so adding a new action forces a
+/// decision here about which tools that action may use.
+pub fn for_webhook(webhook: &WebhookType) -> Vec<Tool> {
+    match webhook {
+        WebhookType::Review(_) => review_tools(),
+    }
+}
+
+/// Read-only tools used to explore a repository during a review.
+fn review_tools() -> Vec<Tool> {
+    vec![
+        Tool::new(
+            "ls",
+            "List the entries of a directory inside the repository.",
+            json!({
+                "type": "object",
+                "properties": {
+                    "path": {
+                        "type": "string",
+                        "description": "Directory path relative to the repository root. Defaults to the repository root."
+                    }
+                }
+            }),
+        ),
+        Tool::new(
+            "file_size",
+            "Get the size of a file inside the repository in bytes.",
+            json!({
+                "type": "object",
+                "properties": {
+                    "path": {
+                        "type": "string",
+                        "description": "File path relative to the repository root."
+                    }
+                }
+            }),
+        ),
+        Tool::new(
+            "read_file",
+            "Read the content of a text file inside the repository. Every line is \
+             prefixed with its absolute line number, even when only a range is read.",
+            json!({
+                "type": "object",
+                "properties": {
+                    "path": {
+                        "type": "string",
+                        "description": "File path relative to the repository root."
+                    },
+                    "start_line": {
+                        "type": "integer",
+                        "description": "First line to read (1-based, inclusive). Defaults to the first line."
+                    },
+                    "end_line": {
+                        "type": "integer",
+                        "description": "Last line to read (1-based, inclusive). Defaults to the last line."
+                    }
+                },
+                "required": ["path"]
+            }),
+        ),
+        Tool::new(
+            "grep",
+            "Search for a regular expression across the repository files.",
+            json!({
+                "type": "object",
+                "properties": {
+                    "pattern": {
+                        "type": "string",
+                        "description": "Extended regular expression to search for."
+                    },
+                    "path": {
+                        "type": "string",
+                        "description": "File or directory to search in, relative to the repository root. Defaults to the repository root."
+                    }
+                },
+                "required": ["pattern"]
+            }),
+        ),
+        Tool::new(
+            "find",
+            "Find files by name pattern inside the repository.",
+            json!({
+                "type": "object",
+                "properties": {
+                    "pattern": {
+                        "type": "string",
+                        "description": "Glob pattern matched against file names, e.g. `*.rs`."
+                    },
+                    "path": {
+                        "type": "string",
+                        "description": "Directory to search in, relative to the repository root. Defaults to the repository root."
+                    }
+                },
+                "required": ["pattern"]
+            }),
+        ),
+    ]
+}
+
+/// Executes a tool call and returns its textual result.
+///
+/// Errors are returned as `Err` so the caller can decide whether to surface
+/// them to the model or abort.
+pub async fn dispatch(sandbox: &Sandbox, name: &str, args: &Value) -> anyhow::Result<String> {
+    match name {
+        "ls" => ls(sandbox, args).await,
+        "read_file" => read_file(sandbox, args).await,
+        "file_size" => file_size(sandbox, args).await,
+        "grep" => grep(sandbox, args).await,
+        "find" => find(sandbox, args).await,
+        other => bail!("unknown tool `{other}`"),
+    }
+}
+
+async fn file_size(sandbox: &Sandbox, args: &Value) -> anyhow::Result<String> {
+    let path = resolve(sandbox, required_str(args, "path")?)?;
+    let size = sandbox.exec(&["du", "-b", "--", &path]).await?;
+    into_stdout(size)
+}
+
+async fn ls(sandbox: &Sandbox, args: &Value) -> anyhow::Result<String> {
+    let path = resolve(sandbox, optional_str(args, "path").unwrap_or("."))?;
+    let output = sandbox.exec(&["ls", "-la", "--", &path]).await?;
+    into_stdout(output)
+}
+
+async fn read_file(sandbox: &Sandbox, args: &Value) -> anyhow::Result<String> {
+    let path = resolve(sandbox, required_str(args, "path")?)?;
+
+    let start = args.get("start_line").and_then(Value::as_u64);
+    let end = args.get("end_line").and_then(Value::as_u64);
+
+    let (first_line, output) = if start.is_none() && end.is_none() {
+        (1, sandbox.exec(&["cat", "--", &path]).await?)
+    } else {
+        let start = start.unwrap_or(1);
+        let end = end
+            .map(|line| line.to_string())
+            .unwrap_or_else(|| "$".to_string());
+        let range = format!("{start},{end}p");
+
+        (
+            start,
+            sandbox.exec(&["sed", "-n", &range, "--", &path]).await?,
+        )
+    };
+
+    Ok(number_lines(&into_stdout(output)?, first_line))
+}
+
+/// Préfixe chaque ligne par son numéro.
+///
+/// Le modèle doit citer une ligne précise pour ancrer son commentaire : sans
+/// numéros, il les compte lui-même et se décale de quelques lignes, ce qui place le
+/// commentaire à côté du code visé.
+fn number_lines(content: &str, first_line: u64) -> String {
+    content
+        .lines()
+        .enumerate()
+        .map(|(offset, line)| format!("{}:{line}", first_line + offset as u64))
+        .collect::<Vec<_>>()
+        .join("\n")
+}
+
+async fn grep(sandbox: &Sandbox, args: &Value) -> anyhow::Result<String> {
+    let pattern = required_str(args, "pattern")?;
+    let path = resolve(sandbox, optional_str(args, "path").unwrap_or("."))?;
+
+    let output = sandbox
+        .exec(&["grep", "-rn", "-E", "--", pattern, &path])
+        .await?;
+
+    // grep exits with 1 when there is no match, which is not an error.
+    match output.status {
+        0 | 1 => Ok(output.stdout),
+        _ => bail!("grep failed: {}", output.stderr.trim()),
+    }
+}
+
+async fn find(sandbox: &Sandbox, args: &Value) -> anyhow::Result<String> {
+    let pattern = required_str(args, "pattern")?;
+    let path = resolve(sandbox, optional_str(args, "path").unwrap_or("."))?;
+
+    let output = sandbox
+        .exec(&["find", &path, "-type", "f", "-name", pattern])
+        .await?;
+
+    into_stdout(output)
+}
+
+/// Resolves a tool path against the container workspace folder, rejecting any
+/// path that escapes it.
+fn resolve(sandbox: &Sandbox, path: &str) -> anyhow::Result<String> {
+    let workspace = sandbox.workspace_folder();
+    let candidate = Path::new(workspace).join(path);
+
+    let normalized =
+        normalize(&candidate).with_context(|| format!("path `{path}` escapes the workspace"))?;
+
+    if !normalized.starts_with(workspace) {
+        bail!("path `{path}` escapes the workspace");
+    }
+
+    Ok(normalized.display().to_string())
+}
+
+fn required_str<'a>(args: &'a Value, key: &str) -> anyhow::Result<&'a str> {
+    args.get(key)
+        .and_then(Value::as_str)
+        .with_context(|| format!("`{key}` is required"))
+}
+
+fn optional_str<'a>(args: &'a Value, key: &str) -> Option<&'a str> {
+    args.get(key).and_then(Value::as_str)
+}
+
+fn into_stdout(output: ExecOutput) -> anyhow::Result<String> {
+    if output.success() {
+        Ok(output.stdout)
+    } else {
+        bail!(
+            "command failed (status {}): {}",
+            output.status,
+            output.stderr.trim()
+        )
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::gitea::{Comment, PullRequest, Repository, ReviewPayload};
+
+    fn review_webhook() -> WebhookType {
+        WebhookType::Review(ReviewPayload {
+            action: "created".to_string(),
+            pull_request: PullRequest {
+                number: 1,
+                title: "My PR".to_string(),
+            },
+            repository: Repository {
+                full_name: "owner/repo".to_string(),
+                clone_url: "https://github.com/owner/repo.git".to_string(),
+            },
+            comment: Comment {
+                id: 1,
+                body: "@bot review".to_string(),
+            },
+        })
+    }
+
+    #[test]
+    fn review_webhook_exposes_read_only_tools() {
+        let names: Vec<String> = for_webhook(&review_webhook())
+            .into_iter()
+            .map(|tool| tool.function.name)
+            .collect();
+
+        assert_eq!(names, vec!["ls", "file_size", "read_file", "grep", "find"]);
+    }
+
+    #[test]
+    fn required_str_reports_missing_key() {
+        let err = required_str(&json!({}), "path").unwrap_err();
+        assert!(err.to_string().contains("path"));
+    }
+
+    #[test]
+    fn lines_are_numbered_from_the_first_one() {
+        assert_eq!(number_lines("a\nb\n", 1), "1:a\n2:b");
+    }
+
+    #[test]
+    fn a_range_keeps_the_absolute_line_numbers() {
+        // Un extrait lu à partir de la ligne 12 doit garder la numérotation du
+        // fichier : sinon le modèle citerait des lignes décalées.
+        assert_eq!(number_lines("x\ny", 12), "12:x\n13:y");
+    }
+
+    #[test]
+    fn an_empty_read_stays_empty() {
+        assert_eq!(number_lines("", 1), "");
+    }
+}

crates/herald-server/src/text.rs

@@ -0,0 +1,81 @@
+//! Small text helpers shared by the modules that log what the model answered and
+//! bound what the tools return to it.
+
+/// First `limit` characters of `text`, for logs.
+///
+/// Cuts on character boundaries, so the excerpt stays valid UTF-8, and marks a
+/// truncation with an ellipsis.
+pub fn excerpt(text: &str, limit: usize) -> String {
+    let mut chars = text.chars();
+    let excerpt = chars.by_ref().take(limit).collect::<String>();
+
+    if chars.next().is_some() {
+        return format!("{excerpt}…");
+    }
+
+    excerpt
+}
+
+/// Truncates `text` in place to at most `limit` bytes, cutting on a character
+/// boundary so the result stays valid UTF-8.
+///
+/// Returns `true` when something was dropped.
+pub fn truncate_bytes(text: &mut String, limit: usize) -> bool {
+    if text.len() <= limit {
+        return false;
+    }
+
+    let mut end = limit;
+    while !text.is_char_boundary(end) {
+        end -= 1;
+    }
+
+    text.truncate(end);
+
+    true
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn a_short_text_is_kept_as_is() {
+        assert_eq!(excerpt("abc", 3), "abc");
+    }
+
+    #[test]
+    fn a_long_text_is_truncated_and_marked() {
+        assert_eq!(excerpt("abcdef", 3), "abc…");
+    }
+
+    #[test]
+    fn truncation_cuts_on_character_boundaries() {
+        assert_eq!(excerpt("ééé", 2), "éé…");
+    }
+
+    #[test]
+    fn a_short_text_is_not_truncated() {
+        let mut text = String::from("abc");
+
+        assert!(!truncate_bytes(&mut text, 3));
+        assert_eq!(text, "abc");
+    }
+
+    #[test]
+    fn a_long_text_is_truncated_within_the_limit() {
+        let mut text = String::from("abcdef");
+
+        assert!(truncate_bytes(&mut text, 4));
+        assert_eq!(text, "abcd");
+    }
+
+    #[test]
+    fn byte_truncation_never_splits_a_character() {
+        let mut text = String::from("ééé");
+
+        // The limit falls in the middle of the second `é`: it is dropped.
+        assert!(truncate_bytes(&mut text, 3));
+        assert_eq!(text, "é");
+    }
+}

src/bot_actions/review.rs

@@ -1,214 +0,0 @@
-use futures_util::stream::TryStreamExt;
-use tokio::io::AsyncReadExt;
-use tokio_util::io::StreamReader;
-use tracing::instrument;
-
-use crate::{
-    bot::ReviewResult, consts::{BOT_PROCESS_MSG, MAX_DIFF_SIZE, REVIEW_PROMPT}, gitea::{GiteaAPI, ReviewPayload}, metrics, open_router::OpenRouterClient,
-};
-
-#[instrument(skip(gitea_api, open_router_client, http_client, review_payload))]
-pub async fn exec_review(
-    gitea_api: &GiteaAPI,
-    open_router_client: &OpenRouterClient,
-    http_client: &reqwest::Client,
-    model: &str,
-    review_payload: ReviewPayload,
-) -> anyhow::Result<()> {
-    tracing::info!(
-        repo = %review_payload.repository.full_name,
-        pr = review_payload.pull_request.number,
-        action = %review_payload.action,
-        "Starting review"
-    );
-
-    let new_comment = gitea_api
-        .comment(
-            &BOT_PROCESS_MSG.replace("{model}", model),
-            &review_payload.repository.full_name,
-            review_payload.pull_request.number,
-        )
-        .await?;
-
-    let bot_result: Result<ReviewResult, anyhow::Error> = async {
-        let git_diff =
-            download_git_diff(http_client, &review_payload.pull_request.diff_url).await?;
-
-        let diff_for_llm = format_diff_for_review(&git_diff);
-
-        let bot_request = REVIEW_PROMPT
-            .replace("{subject}", &review_payload.pull_request.title)
-            .replace("{comment}", &review_payload.comment.body)
-            .replace("{diff}", &diff_for_llm);
-
-        let chat_result = open_router_client.chat(&bot_request).await?;
-        let mut review_result = serde_json::from_str::<ReviewResult>(&chat_result.message)?;
-
-        review_result.cost = chat_result.cost;
-        if let Some(cost) = review_result.cost {
-            metrics::openrouter_cost_usd(cost);
-        }
-
-        let final_review_markdown = review_result_to_markdown(&review_result);
-
-        gitea_api
-            .post_pull_request_review(
-                &review_result,
-                &final_review_markdown,
-                &review_payload.repository.full_name,
-                review_payload.pull_request.number,
-            )
-            .await?;
-
-        Ok(review_result)
-    }
-    .await;
-
-    match bot_result {
-        Ok(_) => {
-            gitea_api
-                .delete_comment(&review_payload.repository.full_name, new_comment.id)
-                .await
-        }
-        Err(e) => {
-            gitea_api
-                .edit_comment(
-                    &format!("Error while reviewing: {}", e),
-                    &review_payload.repository.full_name,
-                    new_comment.id,
-                )
-                .await
-        }
-    }
-}
-
-fn review_result_to_markdown(review_result: &ReviewResult) -> String {
-    if review_result.reviews.is_empty() {
-        return String::from("No issues found. ✅");
-    }
-
-    let mut md = String::from("## Review Feedback\n\n");
-
-    md.push_str(&format!(
-        "### {} issues found.\n\n",
-        review_result.reviews.len()
-    ));
-
-    if !review_result.comment.is_empty() {
-        md.push_str("\n---\n\n");
-        md.push_str("### Summary\n\n");
-        md.push_str(&review_result.comment);
-        md.push('\n');
-    }
-
-    if let Some(cost) = review_result.cost {
-        md.push_str("\n---\n\n");
-        md.push_str(&format!("### Cost: ${}", cost));
-        md.push('\n');
-    }
-
-    md
-}
-
-async fn download_git_diff(http_client: &reqwest::Client, url: &str) -> anyhow::Result<String> {
-    let response = http_client.get(url).send().await?;
-    let stream = response.bytes_stream().map_err(std::io::Error::other);
-
-    let mut buf = Vec::with_capacity(MAX_DIFF_SIZE);
-    StreamReader::new(stream)
-        .take((MAX_DIFF_SIZE + 1) as u64)
-        .read_to_end(&mut buf)
-        .await?;
-
-    if buf.len() > MAX_DIFF_SIZE {
-        anyhow::bail!("Git diff exceeds the maximum allowed size of 1 Mo");
-    }
-
-    Ok(String::from_utf8_lossy(&buf).into_owned())
-}
-
-fn format_diff_for_review(git_diff: &str) -> String {
-    let mut output = String::new();
-    let mut current_file: Option<&str> = None;
-    let mut new_line: u64 = 0;
-
-    for line in git_diff.lines() {
-        if let Some(rest) = line.strip_prefix("diff --git a/") {
-            if let Some(end) = rest.find(' ') {
-                current_file = Some(&rest[..end]);
-            }
-            new_line = 0;
-            continue;
-        }
-
-        if line.starts_with("---") || line.starts_with("+++") {
-            continue;
-        }
-
-        if line.starts_with("@@") && line.contains('+') {
-            if let Some(start) = parse_hunk_new_start(line) {
-                new_line = start;
-            }
-            continue;
-        }
-
-        let Some(filename) = current_file else {
-            continue;
-        };
-
-        if line.starts_with(' ') {
-            new_line += 1;
-            continue;
-        }
-
-        if let Some(code) = line.strip_prefix('+') {
-            use std::fmt::Write;
-            let _ = writeln!(output, "{filename}:{new_line}:{code}");
-            new_line += 1;
-        }
-    }
-
-    output
-}
-
-fn parse_hunk_new_start(hunk_header: &str) -> Option<u64> {
-    let plus_part = hunk_header.split('+').nth(1)?;
-    let num_str = plus_part.split(|c: char| !c.is_ascii_digit()).next()?;
-    num_str.parse::<u64>().ok()
-}
-
-#[cfg(test)]
-#[test]
-fn test_format_diff_for_review() {
-    let diff = concat!(
-        "diff --git a/src/foo.rs b/src/foo.rs\n",
-        "--- a/src/foo.rs\n",
-        "+++ b/src/foo.rs\n",
-        "@@ -1,3 +1,6 @@\n",
-        " fn main() {\n",
-        "+    let x = 1;\n",
-        "     println!(\"hello\");\n",
-        "+    let y = 2;\n",
-        "+    let z = 3;\n",
-        " }\n",
-        "diff --git a/src/bar.rs b/src/bar.rs\n",
-        "--- a/src/bar.rs\n",
-        "+++ b/src/bar.rs\n",
-        "@@ -10,4 +10,6 @@\n",
-        " old context\n",
-        "+    let a = 10;\n",
-        " more context\n",
-        "+    let b = 20;\n",
-    );
-
-    let result = format_diff_for_review(diff);
-    let expected = concat!(
-        "src/foo.rs:2:    let x = 1;\n",
-        "src/foo.rs:4:    let y = 2;\n",
-        "src/foo.rs:5:    let z = 3;\n",
-        "src/bar.rs:11:    let a = 10;\n",
-        "src/bar.rs:13:    let b = 20;\n",
-    );
-
-    assert_eq!(result, expected);
-}

src/consts.rs

@@ -1,44 +0,0 @@
-pub const GITEA_SIG_HEADER_NAME: &str = "x-gitea-signature";
-pub const GITEA_EVENT_TYPE_HEADER_NAME: &str = "x-gitea-event-type";
-
-pub const MAX_WEBHOOK_BODY_SIZE: usize = 1024 * 1024; // 1 MiB
-pub const MAX_DIFF_SIZE: usize = 1024 * 1024; // 1 MiB
-
-pub const BOT_PROCESS_MSG: &str = "
-    Review in progress with the model \"{model}\"...
-";
-
-pub const REVIEW_PROMPT: &str = "
-    You are a senior software engineer reviewing code changes.
-
-    Check good practices and code quality.
-
-    This is the pull request subject: \"{subject}\"
-
-    This is the user comment: \"{comment}\"
-
-    The code changes (only added lines, with line numbers):
-
-    {diff}
-
-    Please review the code changes and provide feedback.
-
-    IMPORTANT: the `line` field must be the line number shown before each line.
-    The provided code has the format: `filename:line:code`
-
-    Return your feedback, in french, with only this json format, reviews must contain each review
-    All fields are mandatory.
-    (filename field must contain the full path with extension) and comment must contain a final summary:
-
-    {
-        \"reviews\": [
-            {
-                \"filename\": \"\",
-                \"line\": ,
-                \"code\": \"\",
-                \"message\": \"\"
-            }
-        ],
-        \"comment\": \"\"
-    }
-";

src/open_router.rs

@@ -1,50 +0,0 @@
-use std::time::Duration;
-
-use openrouter_rs::{Message, api::chat::ChatCompletionRequest};
-use tracing::instrument;
-
-pub struct ChatResult {
-    pub message: String,
-    pub cost: Option<f64>,
-}
-
-#[derive(Clone)]
-pub struct OpenRouterClient {
-    client: openrouter_rs::OpenRouterClient,
-    model: String,
-}
-
-impl OpenRouterClient {
-    pub fn new(token: &str, model: &str, timeout: u64) -> anyhow::Result<Self> {
-        Ok(Self {
-            client: openrouter_rs::OpenRouterClient::builder()
-                .api_key(token)
-                .http_client(
-                    reqwest::Client::builder()
-                        .timeout(Duration::from_secs(timeout))
-                        .build()?,
-                )
-                .build()?,
-            model: String::from(model),
-        })
-    }
-
-    #[instrument(skip(self), err)]
-    pub async fn chat(&self, msg: &str) -> anyhow::Result<ChatResult> {
-        let request = ChatCompletionRequest::builder()
-            .model(&self.model)
-            .enable_reasoning()
-            .messages(vec![Message::new(openrouter_rs::types::Role::User, msg)])
-            .build()?;
-
-        let response = self.client.chat().create(&request).await?;
-
-        Ok(ChatResult {
-            message: response.choices[0]
-                .content()
-                .map(String::from)
-                .ok_or(anyhow::anyhow!("No content"))?,
-            cost: response.usage.and_then(|u| u.cost),
-        })
-    }
-}