feat(background): runner PTY B8 + boucle sink fermée + refactor point-2

Livre le lot backend B8 des tâches de fond :

- infrastructure : runner de commandes concret (CommandBackgroundRunner)
  sur le port BackgroundTaskRunner, tail borné (bounded_tail/BoundedTail),
  éclatement du module background_task en sous-modules (mod/runner/tail,
  sink extrait de l'ancien background_task.rs).
- application : nouveau module background exposant les cas d'usage
  SpawnBackgroundCommand, CancelBackgroundTask, RetryBackgroundTask et le
  port BackgroundCommandArchive.
- domain : refactor point-2 de l'arbitrage Architect — sortie du trait
  BackgroundCommandArchive de la couche domaine vers application.
- app-tauri : câblage runtime (commands, dto, state, lib) des commandes
  spawn/cancel/retry et de la boucle de complétion sink fermée en
  composition root.

Build workspace + tests application/infrastructure verts (QA).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
2026-07-03 14:48:46 +02:00
parent 5d88c952ec
commit 8cac1470ac
12 changed files with 1180 additions and 11 deletions

View File

@ -30,7 +30,8 @@ use domain::ports::PtyHandle;
use crate::dto::{
parse_agent_id, parse_close_terminal, parse_delete_profile, parse_layout_id, parse_memory_slug,
parse_node_id, parse_profile_id, parse_project_id, parse_session_id, parse_skill_id,
parse_template_id, parse_ticket_id, AgentDriftListDto, AgentDto, AgentListDto,
parse_task_id, parse_template_id, parse_ticket_id, AgentDriftListDto, AgentDto, AgentListDto,
BackgroundTaskDto,
AssignSkillRequestDto, AttachLiveAgentRequestDto, AttachLiveAgentResponseDto,
ChangeAgentProfileDto, ChangeAgentProfileRequestDto, ConfigureProfilesRequestDto,
ConversationDetailsDto, CreateAgentFromTemplateRequestDto, CreateAgentRequestDto,
@ -2555,3 +2556,160 @@ pub async fn resolve_memory_links(
.map(MemoryLinksDto::from)
.map_err(ErrorDto::from)
}
// ---------------------------------------------------------------------------
// Background tasks (B8 — first-class background command)
// ---------------------------------------------------------------------------
/// Maps a background-task port error to the frontend [`ErrorDto`] shape.
fn background_error(err: domain::ports::BackgroundTaskPortError) -> ErrorDto {
use domain::ports::BackgroundTaskPortError as E;
let code = match &err {
E::NotFound => "NOT_FOUND",
E::AlreadyExists | E::Invalid(_) => "INVALID",
E::Runner(_) => "PROCESS",
E::Store(_) => "STORE",
};
ErrorDto {
code: code.to_owned(),
message: err.to_string(),
}
}
/// `spawn_background_command` — start a command-backed first-class background
/// task whose completion is delivered to the owning agent's inbox.
///
/// # Errors
/// Returns an [`ErrorDto`] (`INVALID` for a malformed id/path, `PROCESS` if the
/// command cannot be spawned, `STORE` on persistence failure).
#[tauri::command]
pub async fn spawn_background_command(
request: crate::dto::SpawnBackgroundCommandRequestDto,
state: State<'_, AppState>,
) -> Result<BackgroundTaskDto, ErrorDto> {
let project_id = parse_project_id(&request.project_id)?;
let owner_agent_id = parse_agent_id(&request.owner_agent_id)?;
let cwd = domain::project::ProjectPath::new(request.cwd.clone()).map_err(|_| ErrorDto {
code: "INVALID".to_owned(),
message: format!("invalid working directory: {}", request.cwd),
})?;
let command = domain::ports::SpawnSpec {
command: request.command,
args: request.args,
cwd,
env: request.env,
context_plan: None,
sandbox: None,
};
let wake_policy = if request.record_only {
domain::BackgroundTaskWakePolicy::RecordOnly
} else {
domain::BackgroundTaskWakePolicy::WakeOwner
};
state
.spawn_background_command
.execute(application::SpawnBackgroundCommandInput {
project_id,
owner_agent_id,
label: request.label,
command,
wake_policy,
deadline_ms: request.deadline_ms,
})
.await
.map(|out| BackgroundTaskDto::from(out.task))
.map_err(ErrorDto::from)
}
/// `cancel_background_task` — request cancellation of a running background task.
///
/// The terminal `Cancelled` state is written by the completion sink; this returns
/// the task as currently persisted (absent if it no longer exists).
///
/// # Errors
/// Returns an [`ErrorDto`] (`INVALID` for a malformed id, `PROCESS`/`STORE` on
/// failure).
#[tauri::command]
pub async fn cancel_background_task(
task_id: String,
state: State<'_, AppState>,
) -> Result<Option<BackgroundTaskDto>, ErrorDto> {
let id = parse_task_id(&task_id)?;
state
.cancel_background_task
.execute(id)
.await
.map(|out| out.task.map(BackgroundTaskDto::from))
.map_err(ErrorDto::from)
}
/// `retry_background_task` — re-run a terminal command task under a **new** task
/// id (the original id is never reused).
///
/// # Errors
/// Returns an [`ErrorDto`] (`INVALID` for a malformed id or a non-retryable task,
/// `NOT_FOUND` if the task is unknown, `PROCESS`/`STORE` on failure).
#[tauri::command]
pub async fn retry_background_task(
task_id: String,
state: State<'_, AppState>,
) -> Result<BackgroundTaskDto, ErrorDto> {
let id = parse_task_id(&task_id)?;
state
.retry_background_task
.execute(id)
.await
.map(|out| BackgroundTaskDto::from(out.task))
.map_err(ErrorDto::from)
}
/// `list_background_tasks` — read the background-task read-model for a project,
/// optionally narrowed to one owning agent.
///
/// Returns the union of the agent's open tasks and undelivered completions,
/// filtered to `projectId`. Without `agentId`, only the project's undelivered
/// completions are returned (the store cannot enumerate open tasks project-wide).
///
/// # Errors
/// Returns an [`ErrorDto`] (`INVALID` for a malformed id, `STORE` on failure).
#[tauri::command]
pub async fn list_background_tasks(
project_id: String,
agent_id: Option<String>,
state: State<'_, AppState>,
) -> Result<Vec<BackgroundTaskDto>, ErrorDto> {
let project_id = parse_project_id(&project_id)?;
let agent_id = match &agent_id {
Some(raw) => Some(parse_agent_id(raw)?),
None => None,
};
let store = &state.background_task_store;
let mut by_id: std::collections::HashMap<domain::TaskId, domain::BackgroundTask> =
std::collections::HashMap::new();
if let Some(agent_id) = agent_id {
for task in store
.list_open_for_agent(agent_id)
.await
.map_err(background_error)?
{
by_id.insert(task.id, task);
}
}
for task in store
.list_undelivered_completions()
.await
.map_err(background_error)?
{
by_id.entry(task.id).or_insert(task);
}
let mut tasks: Vec<domain::BackgroundTask> = by_id
.into_values()
.filter(|task| task.project_id == project_id)
.filter(|task| agent_id.map_or(true, |a| task.owner_agent_id == a))
.collect();
tasks.sort_by_key(|task| (task.created_at_ms, task.id));
Ok(tasks.into_iter().map(BackgroundTaskDto::from).collect())
}

View File

@ -2866,3 +2866,167 @@ mod ls6_pagination_tests {
assert!(dto.turns.is_empty() && !dto.has_more && dto.next_anchor.is_none());
}
}
// ---------------------------------------------------------------------------
// Background tasks (B8 — first-class background command)
// ---------------------------------------------------------------------------
use domain::{
BackgroundTask, BackgroundTaskKind, BackgroundTaskResult, BackgroundTaskState, TaskId,
};
/// One first-class background task as seen by the frontend (camelCase wire shape).
///
/// Mirrors the persisted [`BackgroundTask`], flattening the terminal
/// [`BackgroundTaskResult`] into optional `exitCode` / `summary` / tails so the
/// UI has a single shape for every lifecycle state. Optional fields are omitted
/// from the wire when absent (`skip_serializing_if`).
#[derive(Debug, Clone, Serialize)]
#[serde(rename_all = "camelCase")]
pub struct BackgroundTaskDto {
/// Stable task id (UUID string).
pub task_id: String,
/// Owning agent id (UUID string).
pub owner_agent_id: String,
/// Owning project id (UUID string).
pub project_id: String,
/// Kind discriminant (`"command"`, `"headlessRendezvous"`, `"sessionResume"`,
/// `"maintenance"`).
pub kind: String,
/// Lifecycle state (`"queued"`, `"running"`, `"waiting"`, `"completed"`,
/// `"failed"`, `"cancelled"`, `"expired"`).
pub state: String,
/// Process exit code, when the terminal result carries one.
#[serde(skip_serializing_if = "Option::is_none")]
pub exit_code: Option<i32>,
/// Human-readable summary / error / reason of the terminal result.
#[serde(skip_serializing_if = "Option::is_none")]
pub summary: Option<String>,
/// Bounded stdout tail (merged output for PTY-backed commands).
#[serde(skip_serializing_if = "Option::is_none")]
pub stdout_tail: Option<String>,
/// Bounded stderr tail (unset for PTY-backed commands, which merge streams).
#[serde(skip_serializing_if = "Option::is_none")]
pub stderr_tail: Option<String>,
/// Creation timestamp, epoch milliseconds.
pub created_at_ms: u64,
/// Last update timestamp, epoch milliseconds.
pub updated_at_ms: u64,
}
/// Kind discriminant string for a [`BackgroundTaskKind`].
fn background_kind_label(kind: &BackgroundTaskKind) -> &'static str {
match kind {
BackgroundTaskKind::Command { .. } => "command",
BackgroundTaskKind::HeadlessRendezvous { .. } => "headlessRendezvous",
BackgroundTaskKind::SessionResume { .. } => "sessionResume",
BackgroundTaskKind::Maintenance { .. } => "maintenance",
}
}
/// Lifecycle state string for a [`BackgroundTaskState`].
fn background_state_label(state: BackgroundTaskState) -> &'static str {
match state {
BackgroundTaskState::Queued => "queued",
BackgroundTaskState::Running => "running",
BackgroundTaskState::Waiting => "waiting",
BackgroundTaskState::Completed => "completed",
BackgroundTaskState::Failed => "failed",
BackgroundTaskState::Cancelled => "cancelled",
BackgroundTaskState::Expired => "expired",
}
}
impl From<BackgroundTask> for BackgroundTaskDto {
fn from(task: BackgroundTask) -> Self {
let (exit_code, summary, stdout_tail, stderr_tail) = match &task.result {
Some(BackgroundTaskResult::Success {
exit_code,
summary,
stdout_tail,
stderr_tail,
..
}) => (
*exit_code,
Some(summary.clone()),
stdout_tail.clone(),
stderr_tail.clone(),
),
Some(BackgroundTaskResult::Failure {
exit_code,
error,
stdout_tail,
stderr_tail,
..
}) => (
*exit_code,
Some(error.clone()),
stdout_tail.clone(),
stderr_tail.clone(),
),
Some(BackgroundTaskResult::Cancelled { reason, .. })
| Some(BackgroundTaskResult::Expired { reason, .. }) => {
(None, Some(reason.clone()), None, None)
}
None => (None, None, None, None),
};
Self {
task_id: task.id.to_string(),
owner_agent_id: task.owner_agent_id.to_string(),
project_id: task.project_id.to_string(),
kind: background_kind_label(&task.kind).to_owned(),
state: background_state_label(task.state).to_owned(),
exit_code,
summary,
stdout_tail,
stderr_tail,
created_at_ms: task.created_at_ms,
updated_at_ms: task.updated_at_ms,
}
}
}
/// Parses a task-id string (UUID) coming from the frontend.
///
/// # Errors
/// Returns an [`ErrorDto`] with code `INVALID` if the string is not a UUID.
pub fn parse_task_id(raw: &str) -> Result<TaskId, ErrorDto> {
uuid::Uuid::parse_str(raw)
.map(TaskId::from_uuid)
.map_err(|_| ErrorDto {
code: "INVALID".to_owned(),
message: format!("invalid task id: {raw}"),
})
}
/// Request DTO for `spawn_background_command`.
///
/// Builds a command-backed [`BackgroundTask`]: the command line runs under the
/// project host's PTY and its completion is delivered to `ownerAgentId`.
#[derive(Debug, Clone, Deserialize)]
#[serde(rename_all = "camelCase")]
pub struct SpawnBackgroundCommandRequestDto {
/// Owning project id (UUID string).
pub project_id: String,
/// Agent that owns completion delivery (UUID string).
pub owner_agent_id: String,
/// Human-facing label.
pub label: String,
/// Executable to run.
pub command: String,
/// Arguments.
#[serde(default)]
pub args: Vec<String>,
/// Working directory (absolute path).
pub cwd: String,
/// Extra environment variables.
#[serde(default)]
pub env: Vec<(String, String)>,
/// When `true` (the default), wake the owner on completion; otherwise only
/// record it.
#[serde(default)]
pub record_only: bool,
/// Optional absolute deadline, epoch milliseconds.
#[serde(default, skip_serializing_if = "Option::is_none")]
pub deadline_ms: Option<u64>,
}

View File

@ -214,6 +214,10 @@ pub fn run() {
commands::recall_memory,
commands::resolve_memory_links,
commands::move_tab_to_new_window,
commands::spawn_background_command,
commands::cancel_background_task,
commands::retry_background_task,
commands::list_background_tasks,
])
.run(tauri::generate_context!())
.expect("error while running IdeA Tauri application");

View File

@ -13,7 +13,8 @@ use std::sync::{Arc, Mutex};
use application::{
AgentResumer, AgentWakeService, AppError, AssignSkillToAgent, AttachLiveAgent,
ChangeAgentProfile, CheckEmbedderSuggestion, CloseProject, CloseTab, CloseTerminal,
BackgroundCommandArchive, CancelBackgroundTask, ChangeAgentProfile, CheckEmbedderSuggestion,
CloseProject, CloseTab, CloseTerminal, RetryBackgroundTask, SpawnBackgroundCommand,
ConfigureProfiles, ContextGuardUseCases, CreateAgentFromScratch, CreateAgentFromTemplate,
CreateLayout, CreateMemory, CreateProject, CreateSkill, CreateTemplate, DeleteAgent,
DeleteEmbedderProfile, DeleteLayout, DeleteMemory, DeleteProfile, DeleteSkill, DeleteTemplate,
@ -39,7 +40,8 @@ use application::{
use async_trait::async_trait;
use domain::ports::{
AgentContextStore, AgentRuntime, AgentSession, AgentSessionFactory, AgentWakePort,
BackgroundTaskPortError, BackgroundTaskStore, Clock, Embedder, EmbedderEnvInspector,
BackgroundTaskPortError, BackgroundTaskRunner, BackgroundTaskStore,
Clock, Embedder, EmbedderEnvInspector,
EmbedderProfileStore, EmbedderPromptStore, EventBus, FileSystem, GitPort, IdGenerator,
MemoryRecall, MemoryStore, PermissionStore, ProcessSpawner, ProfileStore, ProjectStore,
PtyPort, ScheduledTask, Scheduler, SkillStore, TemplateStore, WakeError, WakeReason,
@ -57,8 +59,9 @@ use serde_json::{json, Map, Value};
use uuid::Uuid;
use infrastructure::{
embedder_from_profile, AdaptiveMemoryRecall, BackgroundTaskReadyToDeliver,
ClaudePermissionProjector, ClaudeTranscriptInspector, CliAgentRuntime,
embedder_from_profile, AdaptiveMemoryRecall, BackgroundCompletionSink,
BackgroundTaskReadyToDeliver, ClaudePermissionProjector, ClaudeTranscriptInspector,
CliAgentRuntime, CommandBackgroundRunner,
CodexPermissionProjector, EmbedderEnvProbe, FsBackgroundTaskStore, FsConversationLog,
FsEmbedderProfileStore, FsEmbedderPromptStore, FsHandoffStore, FsLiveStateStore, FsMemoryStore,
FsOrchestratorWatcher, FsPermissionStore, FsProfileStore, FsProjectStore,
@ -795,6 +798,15 @@ pub struct AppState {
// --- Windows (L10) ---
/// Detach a tab into a new OS window (persists the workspace topology).
pub move_tab: Arc<MoveTabToNewWindow>,
// --- Background tasks (B8) ---
/// Spawn a command-backed first-class background task.
pub spawn_background_command: Arc<SpawnBackgroundCommand>,
/// Cancel a running background task.
pub cancel_background_task: Arc<CancelBackgroundTask>,
/// Retry a terminal command task under a fresh task id.
pub retry_background_task: Arc<RetryBackgroundTask>,
/// Store handle used by `list_background_tasks` to read the task read-model.
pub background_task_store: Arc<dyn BackgroundTaskStore>,
// --- Templates & sync (L7) ---
/// Create a template in the global store.
pub create_template: Arc<CreateTemplate>,
@ -1541,6 +1553,49 @@ impl AppState {
let background_tasks_port = Arc::clone(&background_tasks) as Arc<dyn BackgroundTaskStore>;
let (background_ready_tx, mut background_ready_rx) =
tokio::sync::mpsc::unbounded_channel::<BackgroundTaskReadyToDeliver>();
// --- B8 : runner de commandes + fermeture de la boucle du sink ---
// Le runner concret exécute les tâches command-backed sur le `PtyPort` composé
// (local ici ; SSH/WSL via `RemoteHost` quand ils atterriront — Liskov). Il
// émet UNE complétion par tâche ; le sink (single-writer) persiste l'état
// terminal AVANT de signaler la livraison (persist-avant-signal), puis le pont
// ready→inbox ci-dessous réveille l'agent propriétaire.
let background_runner = Arc::new(CommandBackgroundRunner::new(
Arc::clone(&pty_port),
Arc::clone(&clock) as Arc<dyn Clock>,
));
let background_runner_port =
Arc::clone(&background_runner) as Arc<dyn BackgroundTaskRunner>;
{
let sink = Arc::new(BackgroundCompletionSink::new(
Arc::clone(&background_tasks_port),
background_ready_tx.clone(),
));
let runner = Arc::clone(&background_runner_port);
// `start_from_runner` exige un runtime Tokio ambiant (`Handle::current`) :
// on l'appelle donc DANS la tâche async. Le `JoinHandle` du drain est gardé
// vivant en l'attendant (il ne se termine qu'à la fermeture du flux de
// complétions), ce qui maintient sink + runner en vie pour la session.
tauri::async_runtime::spawn(async move {
let drain = sink.start_from_runner(runner);
let _ = drain.await;
});
}
let spawn_background_command = Arc::new(SpawnBackgroundCommand::new(
Arc::clone(&background_tasks_port),
Arc::clone(&background_runner_port),
Arc::clone(&clock) as Arc<dyn Clock>,
Arc::clone(&ids) as Arc<dyn IdGenerator>,
));
let cancel_background_task = Arc::new(CancelBackgroundTask::new(
Arc::clone(&background_tasks_port),
Arc::clone(&background_runner_port),
));
let retry_background_task = Arc::new(RetryBackgroundTask::new(
Arc::clone(&background_tasks_port),
Arc::clone(&background_runner) as Arc<dyn BackgroundCommandArchive>,
Arc::clone(&spawn_background_command),
));
let background_wake = Arc::new(AgentWakeService::new(
Arc::clone(&mediated_inbox) as Arc<dyn AgentInbox>,
Arc::clone(&input_mediator),
@ -1977,6 +2032,10 @@ impl AppState {
fs_port: Arc::clone(&fs_port),
home_dir,
move_tab,
spawn_background_command,
cancel_background_task,
retry_background_task,
background_task_store: Arc::clone(&background_tasks_port),
}
}

View File

@ -0,0 +1,300 @@
//! First-class background command use cases (B8 sub-task 3).
//!
//! These use cases own the *application* side of a command-backed background
//! task: they allocate the [`TaskId`], create the persisted task
//! ([`BackgroundTaskState::Queued`] → [`Running`](BackgroundTaskState::Running))
//! and hand a [`BackgroundTaskSpec`] to the runner. They talk **only** to ports:
//! [`BackgroundTaskStore`], [`BackgroundTaskRunner`], [`Clock`] and
//! [`IdGenerator`], plus the runtime-only [`BackgroundCommandArchive`] abstraction
//! defined below (see its doc for why retry recall is not a domain port).
//!
//! The runner never writes the store; the completion sink does. So beyond the
//! initial create/transition, these use cases do not persist terminal state —
//! except on a *spawn failure*, where they mark the just-created task `Failed` so
//! it never dangles in `Running` with no worker behind it.
use std::sync::Arc;
use domain::ports::{
BackgroundTaskPortError, BackgroundTaskRunner, BackgroundTaskSpec, BackgroundTaskStore, Clock,
IdGenerator, SpawnSpec,
};
use domain::{
AgentId, BackgroundTask, BackgroundTaskKind, BackgroundTaskResult, BackgroundTaskState,
BackgroundTaskWakePolicy, ProjectId, TaskId,
};
use crate::error::AppError;
/// Maps a background-task port error into the application error shape.
fn map_port_err(err: BackgroundTaskPortError) -> AppError {
match err {
BackgroundTaskPortError::NotFound => AppError::NotFound("background task".to_owned()),
BackgroundTaskPortError::AlreadyExists => {
AppError::Invalid("background task already exists".to_owned())
}
BackgroundTaskPortError::Invalid(msg) => AppError::Invalid(msg),
BackgroundTaskPortError::Runner(msg) => AppError::Process(msg),
BackgroundTaskPortError::Store(msg) => AppError::Store(msg),
}
}
/// Input for [`SpawnBackgroundCommand`].
#[derive(Debug, Clone)]
pub struct SpawnBackgroundCommandInput {
/// Owning project.
pub project_id: ProjectId,
/// Agent that owns completion delivery.
pub owner_agent_id: AgentId,
/// Human-facing label.
pub label: String,
/// Command invocation to run.
pub command: SpawnSpec,
/// Completion wake policy.
pub wake_policy: BackgroundTaskWakePolicy,
/// Optional absolute deadline, epoch milliseconds.
pub deadline_ms: Option<u64>,
}
/// Output of [`SpawnBackgroundCommand`] / [`RetryBackgroundTask`].
#[derive(Debug, Clone)]
pub struct SpawnBackgroundCommandOutput {
/// The created task, in its `Running` state.
pub task: BackgroundTask,
}
/// Creates a command-backed background task and hands it to the runner.
pub struct SpawnBackgroundCommand {
store: Arc<dyn BackgroundTaskStore>,
runner: Arc<dyn BackgroundTaskRunner>,
clock: Arc<dyn Clock>,
ids: Arc<dyn IdGenerator>,
}
impl SpawnBackgroundCommand {
/// Wires the use case to its ports.
#[must_use]
pub fn new(
store: Arc<dyn BackgroundTaskStore>,
runner: Arc<dyn BackgroundTaskRunner>,
clock: Arc<dyn Clock>,
ids: Arc<dyn IdGenerator>,
) -> Self {
Self {
store,
runner,
clock,
ids,
}
}
/// Allocates a task id, persists it (`Queued`→`Running`) and spawns it.
///
/// # Errors
/// [`AppError`] if a store/runner operation or a domain invariant fails.
pub async fn execute(
&self,
input: SpawnBackgroundCommandInput,
) -> Result<SpawnBackgroundCommandOutput, AppError> {
let task = self
.create_and_run(
input.project_id,
input.owner_agent_id,
input.label,
input.command,
input.wake_policy,
input.deadline_ms,
)
.await?;
Ok(SpawnBackgroundCommandOutput { task })
}
/// Shared create/transition/spawn path, reused by retry.
async fn create_and_run(
&self,
project_id: ProjectId,
owner_agent_id: AgentId,
label: String,
command: SpawnSpec,
wake_policy: BackgroundTaskWakePolicy,
deadline_ms: Option<u64>,
) -> Result<BackgroundTask, AppError> {
let task_id = TaskId::from_uuid(self.ids.new_uuid());
let now = u64::try_from(self.clock.now_millis().max(0)).unwrap_or(0);
let task = BackgroundTask::new(
task_id,
project_id,
owner_agent_id,
BackgroundTaskKind::Command { label },
wake_policy,
now,
deadline_ms,
)
.map_err(|e| AppError::Invalid(e.to_string()))?;
self.store.create(&task).await.map_err(map_port_err)?;
let running = task
.transition(BackgroundTaskState::Running, now)
.map_err(|e| AppError::Invalid(e.to_string()))?;
self.store.save(&running).await.map_err(map_port_err)?;
let spec = BackgroundTaskSpec {
task_id,
project_id,
owner_agent_id,
kind: running.kind.clone(),
wake_policy,
command: Some(command),
deadline_ms,
};
if let Err(err) = self.runner.spawn(spec).await {
// Never leave the task dangling in `Running` with no worker: mark it
// Failed so reconcile/UI see a terminal state. Best-effort persist.
if let Ok(failed) = running.complete(BackgroundTaskResult::Failure {
finished_at_ms: now,
exit_code: None,
error: format!("failed to spawn background command: {err}"),
stdout_tail: None,
stderr_tail: None,
}) {
let _ = self.store.save(&failed).await;
}
return Err(map_port_err(err));
}
Ok(running)
}
}
/// Output of [`CancelBackgroundTask`].
#[derive(Debug, Clone)]
pub struct CancelBackgroundTaskOutput {
/// The task as currently persisted, if it still exists. The terminal
/// `Cancelled` state is written by the completion sink, not here.
pub task: Option<BackgroundTask>,
}
/// Requests cancellation of a running background task.
pub struct CancelBackgroundTask {
store: Arc<dyn BackgroundTaskStore>,
runner: Arc<dyn BackgroundTaskRunner>,
}
impl CancelBackgroundTask {
/// Wires the use case to its ports.
#[must_use]
pub fn new(store: Arc<dyn BackgroundTaskStore>, runner: Arc<dyn BackgroundTaskRunner>) -> Self {
Self { store, runner }
}
/// Signals the runner to cancel `task_id` (idempotent).
///
/// # Errors
/// [`AppError`] if the runner or store fails.
pub async fn execute(
&self,
task_id: TaskId,
) -> Result<CancelBackgroundTaskOutput, AppError> {
self.runner.cancel(task_id).await.map_err(map_port_err)?;
let task = self.store.get(task_id).await.map_err(map_port_err)?;
Ok(CancelBackgroundTaskOutput { task })
}
}
/// Read-only recall of the command invocation a background task was spawned with.
///
/// This is a **runtime, in-memory** abstraction, *not* a domain port: the recall
/// lives wherever the runner keeps its live invocations for the current session
/// and is deliberately **never persisted** (a [`SpawnSpec`] can carry secrets and
/// must never land in `.ideai/background-tasks/*.json`, which travels with the
/// project). Retry is therefore **session-scoped** in V1 — a task whose runner no
/// longer remembers its spec (e.g. after a restart) simply cannot be retried.
///
/// The persisted [`BackgroundTask`] carries no [`SpawnSpec`] (the command line is
/// an execution detail, not domain state). Retry, however, must re-run the *same*
/// command under a fresh task id, so it recovers the original invocation from
/// whatever component still holds it — the runner. Keeping that recall behind this
/// application-level trait lets [`RetryBackgroundTask`] stay decoupled from the
/// concrete runner without dragging the concept into the domain.
pub trait BackgroundCommandArchive: Send + Sync {
/// Returns the [`SpawnSpec`] a command-backed task was spawned with, if the
/// runner still remembers it. `None` when the task is unknown, was not
/// command-backed, or is no longer in the current session's registry.
fn spec_for(&self, task_id: TaskId) -> Option<SpawnSpec>;
}
/// Re-runs a terminal command-backed task under a brand-new task id.
///
/// The original command is recovered from the [`BackgroundCommandArchive`]; the
/// deadline is intentionally dropped (an absolute past deadline would expire the
/// retry instantly).
pub struct RetryBackgroundTask {
store: Arc<dyn BackgroundTaskStore>,
archive: Arc<dyn BackgroundCommandArchive>,
spawn: Arc<SpawnBackgroundCommand>,
}
impl RetryBackgroundTask {
/// Wires the use case to its ports and the spawn use case.
#[must_use]
pub fn new(
store: Arc<dyn BackgroundTaskStore>,
archive: Arc<dyn BackgroundCommandArchive>,
spawn: Arc<SpawnBackgroundCommand>,
) -> Self {
Self {
store,
archive,
spawn,
}
}
/// Retries `task_id`, returning the freshly-created task.
///
/// # Errors
/// [`AppError::NotFound`] if the task is unknown, [`AppError::Invalid`] if it
/// is not a terminal command task or its command is no longer available.
pub async fn execute(
&self,
task_id: TaskId,
) -> Result<SpawnBackgroundCommandOutput, AppError> {
let old = self
.store
.get(task_id)
.await
.map_err(map_port_err)?
.ok_or_else(|| AppError::NotFound(format!("background task {task_id}")))?;
if !old.is_terminal() {
return Err(AppError::Invalid(
"cannot retry a task that is still running".to_owned(),
));
}
let label = match &old.kind {
BackgroundTaskKind::Command { label } => label.clone(),
_ => {
return Err(AppError::Invalid(
"only command-backed tasks can be retried".to_owned(),
))
}
};
let command = self.archive.spec_for(task_id).ok_or_else(|| {
AppError::Invalid("original command is no longer available for retry".to_owned())
})?;
let task = self
.spawn
.create_and_run(
old.project_id,
old.owner_agent_id,
label,
command,
old.wake_policy,
None,
)
.await?;
Ok(SpawnBackgroundCommandOutput { task })
}
}

View File

@ -12,6 +12,7 @@
#![warn(missing_docs)]
pub mod agent;
pub mod background;
pub mod conversation;
pub mod diag;
pub mod embedder;
@ -47,6 +48,10 @@ pub use agent::{
StructuredSessionDescriptor, TurnOutcome, UpdateAgentContext, UpdateAgentContextInput,
AGENT_MEMORY_RECALL_BUDGET, CODEX_SUBMIT_DELAY_MS, LIVE_STATE_INJECT_MAX, RESUME_PROMPT,
};
pub use background::{
BackgroundCommandArchive, CancelBackgroundTask, CancelBackgroundTaskOutput, RetryBackgroundTask,
SpawnBackgroundCommand, SpawnBackgroundCommandInput, SpawnBackgroundCommandOutput,
};
pub use conversation::{
ConversationArchiveProvider, ReadConversationPage, ReadConversationPageInput, RecordTurn,
RotateConversationLog, RotateConversationLogInput, TurnPage, TurnSource, TurnView,

View File

@ -167,9 +167,9 @@ pub use orchestrator::{
};
pub use ports::{
AgentContextStore, AgentRuntime, BackgroundCompletionStream, BackgroundTaskCompletion,
BackgroundTaskHandle, BackgroundTaskPortError, BackgroundTaskRunner, BackgroundTaskSpec,
BackgroundTaskStore, Clock, ContextInjectionPlan, DirEntry, Embedder, EmbedderEnvInspector,
AgentContextStore, AgentRuntime, BackgroundCompletionStream,
BackgroundTaskCompletion, BackgroundTaskHandle, BackgroundTaskPortError, BackgroundTaskRunner,
BackgroundTaskSpec, BackgroundTaskStore, Clock, ContextInjectionPlan, DirEntry, Embedder, EmbedderEnvInspector,
EmbedderEnvReport, EmbedderError, EmbedderProfileStore, EmbedderPromptDismissal,
EmbedderPromptStore, EventBus, EventStream, ExitStatus, FileSystem, FsError, GitCommitInfo,
GitError, GitFileStatus, GitPort, GraphCommit, IdGenerator, LiveStateStore, MemoryError,

View File

@ -0,0 +1,22 @@
//! First-class background task infrastructure.
//!
//! Three cohesive pieces close the B2→B8 loop:
//!
//! - [`sink`] — the durable [`BackgroundCompletionSink`], single writer between a
//! runner completion and the ready-to-deliver signal (persist-before-signal).
//! - [`tail`] — a UTF-8-safe bounded ring buffer for stdout/stderr tails.
//! - [`runner`] — [`CommandBackgroundRunner`], the concrete
//! [`BackgroundTaskRunner`](domain::ports::BackgroundTaskRunner) that spawns a
//! command-backed task over a composed [`PtyPort`](domain::ports::PtyPort) and
//! emits exactly one completion per task.
mod runner;
mod sink;
mod tail;
pub use runner::CommandBackgroundRunner;
pub use sink::{
start_background_ready_inbox_bridge, BackgroundCompletionSink, BackgroundCompletionSinkError,
BackgroundCompletionSinkOutcome, BackgroundReadyInboxBridgeHandle, BackgroundTaskReadyToDeliver,
};
pub use tail::{bounded_tail, tail_cap_bytes, BoundedTail};

View File

@ -0,0 +1,308 @@
//! [`CommandBackgroundRunner`] — the concrete command-backed background task
//! runner (B8 sub-task 2).
//!
//! # Role
//!
//! This is the `impl` of the frozen [`BackgroundTaskRunner`] port that the
//! composition root wires under the [`BackgroundCompletionSink`]. It spawns a
//! command-backed task over a **composed** [`PtyPort`] (resolved by the project's
//! `RemoteHost`, so local/SSH/WSL are transparent — Liskov) and, for each task,
//! emits **exactly one** [`BackgroundTaskCompletion`] on the stream returned by
//! [`subscribe_completions`](BackgroundTaskRunner::subscribe_completions).
//!
//! It writes **neither** the store **nor** the inbox: the sink is the single
//! writer that persists the terminal state before signalling delivery
//! (persist-before-signal). The runner's only outward effect is the completion
//! event.
//!
//! # Lifecycle of one task
//!
//! `spawn` opens the PTY and starts a detached worker. The worker watches the
//! PTY's output stream for EOF (natural process exit), a cancel notification, or
//! the deadline, whichever comes first. It then snapshots the bounded output tail
//! from the PTY scrollback, reaps the real exit status via
//! [`PtyPort::kill`](domain::ports::PtyPort::kill), maps that to a terminal
//! [`BackgroundTaskResult`], and sends the completion.
//!
//! # Output tail & the live tee
//!
//! A PTY **merges** stdout and stderr onto one stream, so the completion carries
//! the merged output in `stdout_tail` and leaves `stderr_tail` empty. Exit
//! detection reads the PTY's single-consumer output subscription; a *live tee*
//! that re-subscribes the same PTY (to render it in an xterm cell) would supersede
//! the runner's subscription and break EOF detection — teeing a running command
//! therefore requires a multi-consumer PTY broadcast, which is out of scope here.
use std::collections::HashMap;
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::mpsc::{self, Receiver, Sender};
use std::sync::{Arc, Mutex};
use async_trait::async_trait;
use application::BackgroundCommandArchive;
use domain::ports::{
BackgroundCompletionStream, BackgroundTaskCompletion, BackgroundTaskHandle,
BackgroundTaskPortError, BackgroundTaskRunner, BackgroundTaskSpec, Clock, PtyHandle, PtyPort,
SpawnSpec,
};
use domain::terminal::PtySize;
use domain::{BackgroundTaskResult, TaskId};
use super::tail::{bounded_tail, tail_cap_bytes};
/// Default PTY geometry for a headless background command (no visible cell).
const DEFAULT_SIZE: PtySize = PtySize { rows: 24, cols: 80 };
/// Why the worker's terminal wait resolved.
enum WaitOutcome {
/// The process exited on its own (stream hit EOF).
Exited,
/// A [`cancel`](BackgroundTaskRunner::cancel) was requested.
Cancelled,
/// The deadline elapsed before exit.
Expired,
}
/// Live control handles for one running task (kept in the registry).
#[derive(Clone)]
struct TaskControl {
pty_handle: PtyHandle,
cancel: Arc<tokio::sync::Notify>,
cancelled: Arc<AtomicBool>,
}
/// Command-backed [`BackgroundTaskRunner`] over a composed [`PtyPort`].
pub struct CommandBackgroundRunner {
pty: Arc<dyn PtyPort>,
clock: Arc<dyn Clock>,
completion_tx: Sender<BackgroundTaskCompletion>,
completion_rx: Mutex<Option<Receiver<BackgroundTaskCompletion>>>,
running: Arc<Mutex<HashMap<TaskId, TaskControl>>>,
/// Retained command invocations, so [`BackgroundCommandArchive::spec_for`]
/// can recover a task's command for retry.
specs: Arc<Mutex<HashMap<TaskId, SpawnSpec>>>,
}
impl CommandBackgroundRunner {
/// Builds a runner over a composed PTY port and clock.
#[must_use]
pub fn new(pty: Arc<dyn PtyPort>, clock: Arc<dyn Clock>) -> Self {
let (completion_tx, completion_rx) = mpsc::channel();
Self {
pty,
clock,
completion_tx,
completion_rx: Mutex::new(Some(completion_rx)),
running: Arc::new(Mutex::new(HashMap::new())),
specs: Arc::new(Mutex::new(HashMap::new())),
}
}
fn now_ms(&self) -> u64 {
u64::try_from(self.clock.now_millis().max(0)).unwrap_or(0)
}
/// Detached worker driving one command to its single completion.
#[allow(clippy::too_many_arguments)]
async fn run_to_completion(
pty: Arc<dyn PtyPort>,
clock: Arc<dyn Clock>,
completion_tx: Sender<BackgroundTaskCompletion>,
running: Arc<Mutex<HashMap<TaskId, TaskControl>>>,
task_id: TaskId,
control: TaskControl,
deadline_ms: Option<u64>,
) {
let outcome = Self::await_terminal(&pty, &control, deadline_ms, &clock).await;
// A cancel that raced a natural EOF still wins: the user asked to cancel.
let outcome = if control.cancelled.load(Ordering::SeqCst) {
WaitOutcome::Cancelled
} else {
outcome
};
// Snapshot the merged output tail *before* kill tears the session down.
let tail = pty
.scrollback(&control.pty_handle)
.ok()
.map(|bytes| bounded_tail(&bytes, tail_cap_bytes()))
.filter(|s| !s.is_empty());
// Reap the real exit status (works for a natural exit and a forced kill).
let exit_code = pty
.kill(&control.pty_handle)
.await
.ok()
.and_then(|status| status.code);
let finished_at_ms = u64::try_from(clock.now_millis().max(0)).unwrap_or(0);
let result = match outcome {
WaitOutcome::Cancelled => BackgroundTaskResult::Cancelled {
finished_at_ms,
reason: "cancelled by request".to_owned(),
},
WaitOutcome::Expired => BackgroundTaskResult::Expired {
finished_at_ms,
reason: "deadline elapsed before completion".to_owned(),
},
WaitOutcome::Exited => {
if exit_code == Some(0) {
BackgroundTaskResult::Success {
finished_at_ms,
exit_code,
summary: "command completed successfully".to_owned(),
stdout_tail: tail,
stderr_tail: None,
}
} else {
BackgroundTaskResult::Failure {
finished_at_ms,
exit_code,
error: match exit_code {
Some(code) => format!("command exited with code {code}"),
None => "command terminated without an exit code".to_owned(),
},
stdout_tail: tail,
stderr_tail: None,
}
}
}
};
running.lock().expect("runner registry poisoned").remove(&task_id);
let _ = completion_tx.send(BackgroundTaskCompletion { task_id, result });
}
/// Resolves when the process exits, is cancelled, or hits its deadline.
async fn await_terminal(
pty: &Arc<dyn PtyPort>,
control: &TaskControl,
deadline_ms: Option<u64>,
clock: &Arc<dyn Clock>,
) -> WaitOutcome {
let eof = Self::wait_for_eof(Arc::clone(pty), control.pty_handle.clone());
tokio::pin!(eof);
let sleep_ms = deadline_ms.map(|deadline| {
let now = u64::try_from(clock.now_millis().max(0)).unwrap_or(0);
deadline.saturating_sub(now)
});
match sleep_ms {
Some(ms) => tokio::select! {
() = &mut eof => WaitOutcome::Exited,
() = control.cancel.notified() => WaitOutcome::Cancelled,
() = tokio::time::sleep(std::time::Duration::from_millis(ms)) => WaitOutcome::Expired,
},
None => tokio::select! {
() = &mut eof => WaitOutcome::Exited,
() = control.cancel.notified() => WaitOutcome::Cancelled,
},
}
}
/// Resolves when the PTY output stream ends (the process has exited).
///
/// The domain `OutputStream` is a blocking iterator, so it is drained on a
/// blocking task and its bytes discarded — the persisted tail is read from the
/// PTY scrollback instead, which the reader thread fills regardless of
/// subscribers. If the handle is already unknown, resolves immediately so the
/// worker proceeds to reap.
async fn wait_for_eof(pty: Arc<dyn PtyPort>, handle: PtyHandle) {
let Ok(stream) = pty.subscribe_output(&handle) else {
return;
};
let _ = tokio::task::spawn_blocking(move || {
for _chunk in stream {}
})
.await;
}
}
#[async_trait]
impl BackgroundTaskRunner for CommandBackgroundRunner {
async fn spawn(
&self,
spec: BackgroundTaskSpec,
) -> Result<BackgroundTaskHandle, BackgroundTaskPortError> {
let command = spec.command.clone().ok_or_else(|| {
BackgroundTaskPortError::Runner(
"command-backed runner requires a command spec".to_owned(),
)
})?;
let pty_handle = self
.pty
.spawn(command.clone(), DEFAULT_SIZE)
.await
.map_err(|e| BackgroundTaskPortError::Runner(format!("pty spawn failed: {e}")))?;
self.specs
.lock()
.expect("runner spec registry poisoned")
.insert(spec.task_id, command);
let control = TaskControl {
pty_handle,
cancel: Arc::new(tokio::sync::Notify::new()),
cancelled: Arc::new(AtomicBool::new(false)),
};
self.running
.lock()
.expect("runner registry poisoned")
.insert(spec.task_id, control.clone());
tokio::spawn(Self::run_to_completion(
Arc::clone(&self.pty),
Arc::clone(&self.clock),
self.completion_tx.clone(),
Arc::clone(&self.running),
spec.task_id,
control,
spec.deadline_ms,
));
// Touch `now_ms` only to keep the clock wired for future scheduling; the
// authoritative timestamps come from the worker at completion.
let _ = self.now_ms();
Ok(BackgroundTaskHandle {
task_id: spec.task_id,
})
}
async fn cancel(&self, task_id: TaskId) -> Result<(), BackgroundTaskPortError> {
let control = self
.running
.lock()
.expect("runner registry poisoned")
.get(&task_id)
.cloned();
// Unknown task ⇒ already terminal or never ours: cancellation is idempotent.
if let Some(control) = control {
control.cancelled.store(true, Ordering::SeqCst);
control.cancel.notify_one();
}
Ok(())
}
fn subscribe_completions(&self) -> BackgroundCompletionStream {
let rx = self
.completion_rx
.lock()
.expect("runner completion receiver poisoned")
.take()
.expect("subscribe_completions must be called exactly once");
Box::new(rx.into_iter())
}
}
impl BackgroundCommandArchive for CommandBackgroundRunner {
fn spec_for(&self, task_id: TaskId) -> Option<SpawnSpec> {
self.specs
.lock()
.expect("runner spec registry poisoned")
.get(&task_id)
.cloned()
}
}

View File

@ -0,0 +1,149 @@
//! UTF-8-safe bounded output tail (B8 sub-task 1).
//!
//! A background command can print megabytes; the completion contract only keeps a
//! bounded *tail* (the most recent bytes). [`BoundedTail`] is a byte ring buffer
//! capped at construction; [`bounded_tail`] renders the retained bytes as a
//! `String`, trimming any partial UTF-8 sequence at the front so a multi-byte
//! character never gets sliced by the cap.
use std::collections::VecDeque;
use domain::background_task::BACKGROUND_TASK_OUTPUT_TAIL_MAX_BYTES;
/// Default tail cap when `IDEA_BG_TAIL_BYTES` is unset (8 KiB).
const DEFAULT_TAIL_BYTES: usize = 8 * 1024;
/// Floor so a hostile/typo'd env value can't disable the tail entirely.
const MIN_TAIL_BYTES: usize = 1024;
/// Resolves the tail cap in bytes from `IDEA_BG_TAIL_BYTES`.
///
/// Defaults to 8 KiB, floored at 1 KiB, and clamped to the domain maximum
/// ([`BACKGROUND_TASK_OUTPUT_TAIL_MAX_BYTES`], 16 KiB) so the persisted result
/// always validates.
#[must_use]
pub fn tail_cap_bytes() -> usize {
std::env::var("IDEA_BG_TAIL_BYTES")
.ok()
.and_then(|v| v.trim().parse::<usize>().ok())
.unwrap_or(DEFAULT_TAIL_BYTES)
.clamp(MIN_TAIL_BYTES, BACKGROUND_TASK_OUTPUT_TAIL_MAX_BYTES)
}
/// A byte ring buffer that retains only the last `cap` bytes pushed into it.
#[derive(Debug)]
pub struct BoundedTail {
buf: VecDeque<u8>,
cap: usize,
}
impl BoundedTail {
/// Builds a tail retaining at most `cap` bytes (a `cap` of `0` retains
/// nothing but never panics).
#[must_use]
pub fn with_cap(cap: usize) -> Self {
Self {
buf: VecDeque::new(),
cap,
}
}
/// Appends a chunk, dropping the oldest bytes past the cap.
pub fn push(&mut self, chunk: &[u8]) {
if self.cap == 0 {
return;
}
// If the incoming chunk alone exceeds the cap, only its tail can survive.
let start = chunk.len().saturating_sub(self.cap);
self.buf.extend(&chunk[start..]);
let overflow = self.buf.len().saturating_sub(self.cap);
if overflow > 0 {
self.buf.drain(0..overflow);
}
}
/// Returns whether any byte is retained.
#[must_use]
pub fn is_empty(&self) -> bool {
self.buf.is_empty()
}
/// Renders the retained bytes as a UTF-8 string, dropping a leading partial
/// multi-byte sequence left dangling by the cap. Returns `None` when empty.
#[must_use]
pub fn to_tail_string(&self) -> Option<String> {
if self.buf.is_empty() {
return None;
}
let bytes: Vec<u8> = self.buf.iter().copied().collect();
Some(bounded_tail(&bytes, self.cap))
}
}
/// Renders the last `cap` bytes of `bytes` as a lossless-at-the-boundary
/// `String`.
///
/// The slice is taken from the end, then any partial UTF-8 sequence at the *front*
/// (a continuation byte with no lead byte, because the cap cut mid-character) is
/// skipped so the result is valid UTF-8 without replacement characters at the
/// seam. Invalid bytes in the interior are still replaced lossily.
#[must_use]
pub fn bounded_tail(bytes: &[u8], cap: usize) -> String {
let start = bytes.len().saturating_sub(cap);
let mut slice = &bytes[start..];
// Only realign when we actually truncated the front (start > 0): otherwise a
// legitimately leading continuation byte is just invalid input, handled below.
if start > 0 {
let mut skip = 0;
while skip < slice.len() && is_utf8_continuation(slice[skip]) {
skip += 1;
}
slice = &slice[skip..];
}
String::from_utf8_lossy(slice).into_owned()
}
/// Whether `b` is a UTF-8 continuation byte (`10xxxxxx`).
const fn is_utf8_continuation(b: u8) -> bool {
b & 0b1100_0000 == 0b1000_0000
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn keeps_only_the_last_cap_bytes() {
let mut tail = BoundedTail::with_cap(4);
tail.push(b"abcdef");
assert_eq!(tail.to_tail_string().as_deref(), Some("cdef"));
}
#[test]
fn push_bigger_than_cap_keeps_tail() {
let mut tail = BoundedTail::with_cap(3);
tail.push(b"0123456789");
assert_eq!(tail.to_tail_string().as_deref(), Some("789"));
}
#[test]
fn empty_tail_is_none() {
let tail = BoundedTail::with_cap(8);
assert!(tail.is_empty());
assert_eq!(tail.to_tail_string(), None);
}
#[test]
fn cut_multibyte_char_is_trimmed_at_front() {
// "é" is 0xC3 0xA9. Cap of 1 would keep only the trailing 0xA9 (a lone
// continuation byte); it must be dropped, yielding an empty string.
let bytes = "é".as_bytes();
assert_eq!(bounded_tail(bytes, 1), "");
// Cap of 2 keeps the whole character.
assert_eq!(bounded_tail(bytes, 2), "é");
}
#[test]
fn interior_is_preserved() {
assert_eq!(bounded_tail("héllo".as_bytes(), 100), "héllo");
}
}

View File

@ -38,9 +38,9 @@ pub mod store;
pub mod timeparse;
pub use background_task::{
start_background_ready_inbox_bridge, BackgroundCompletionSink, BackgroundCompletionSinkError,
BackgroundCompletionSinkOutcome, BackgroundReadyInboxBridgeHandle,
BackgroundTaskReadyToDeliver,
bounded_tail, start_background_ready_inbox_bridge, tail_cap_bytes, BackgroundCompletionSink,
BackgroundCompletionSinkError, BackgroundCompletionSinkOutcome, BackgroundReadyInboxBridgeHandle,
BackgroundTaskReadyToDeliver, BoundedTail, CommandBackgroundRunner,
};
pub use clock::SystemClock;
pub use conversation::InMemoryConversationRegistry;