| 1 | //! Agent sessions: a long-lived container per session running tmux plus an |
| 2 | //! agent CLI against one repository, attachable from the browser. |
| 3 | //! |
| 4 | //! Deliberately a sibling of `anvil-ci` rather than part of it. The CI runner |
| 5 | //! is a single task processing one job at a time; a session lives for hours and |
| 6 | //! must never sit in that queue. Sharing the runner image and the Docker |
| 7 | //! plumbing (`anvil_docker`) is the extent of the overlap. |
| 8 | //! |
| 9 | //! The shape: |
| 10 | //! |
| 11 | //! ```text |
| 12 | //! browser <--ws--> anvil-web <--broadcast--, |
| 13 | //! | |
| 14 | //! supervisor task --docker exec (tty)--> tmux client |
| 15 | //! | | |
| 16 | //! '--> transcript on disk the agent |
| 17 | //! ``` |
| 18 | //! |
| 19 | //! One supervisor per session owns the container and a broadcast channel. |
| 20 | //! Attached browsers are subscribers, and so is the transcript writer — which |
| 21 | //! is why output is durable whether or not anyone is watching. |
| 22 | |
| 23 | pub mod container; |
| 24 | mod supervisor; |
| 25 | |
| 26 | use anvil_core::{ |
| 27 | App, |
| 28 | agent::{ |
| 29 | self, |
| 30 | status, |
| 31 | }, |
| 32 | }; |
| 33 | pub use supervisor::{ |
| 34 | attach_stream, |
| 35 | detached, |
| 36 | }; |
| 37 | |
| 38 | /// Why a session could not be started. |
| 39 | #[derive(Debug)] |
| 40 | pub enum StartError { |
| 41 | /// `agent.enabled` is false. |
| 42 | Disabled, |
| 43 | /// `agent.max_concurrent` sessions are already running. |
| 44 | AtCapacity(usize), |
| 45 | /// Anything else, with detail. |
| 46 | Failed(String), |
| 47 | } |
| 48 | |
| 49 | impl std::fmt::Display for StartError { |
| 50 | fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { |
| 51 | match self { |
| 52 | Self::Disabled => write!(f, "agent sessions are disabled (set agent.enabled)"), |
| 53 | Self::AtCapacity(n) => { |
| 54 | write!(f, "already running {n} sessions (agent.max_concurrent)") |
| 55 | } |
| 56 | Self::Failed(e) => write!(f, "{e}"), |
| 57 | } |
| 58 | } |
| 59 | } |
| 60 | |
| 61 | /// Start a session against `repo_id` at `base_ref`, returning its id. |
| 62 | /// |
| 63 | /// `secret_names` are the user secrets (`anvil_core::secrets::list_for_user`) |
| 64 | /// this session opted into at start time — see `supervisor::launch` for how |
| 65 | /// they actually get into the container. Empty is a normal choice, not an |
| 66 | /// error: most sessions want none. |
| 67 | /// |
| 68 | /// Creates the row first so a failure part-way through is still visible in the |
| 69 | /// UI, then hands off to a supervisor task. Returns as soon as the container is |
| 70 | /// up — the caller redirects to the session page and attaches from there. |
| 71 | pub async fn start( |
| 72 | app: &App, |
| 73 | repo_id: i64, |
| 74 | user_id: i64, |
| 75 | kind: &str, |
| 76 | base_ref: &str, |
| 77 | prompt: &str, |
| 78 | secret_names: &[String], |
| 79 | ) -> Result<i64, StartError> { |
| 80 | let cfg = &app.config.agent; |
| 81 | if !cfg.enabled { |
| 82 | return Err(StartError::Disabled); |
| 83 | } |
| 84 | let live = app.sessions.len(); |
| 85 | if cfg.max_concurrent > 0 && live >= cfg.max_concurrent { |
| 86 | return Err(StartError::AtCapacity(live)); |
| 87 | } |
| 88 | |
| 89 | let (repo_path, base_commit) = supervisor::resolve_base(app, repo_id, base_ref) |
| 90 | .await |
| 91 | .map_err(StartError::Failed)?; |
| 92 | |
| 93 | let session = agent::create( |
| 94 | &app.db, |
| 95 | repo_id, |
| 96 | user_id, |
| 97 | kind, |
| 98 | base_ref, |
| 99 | &base_commit, |
| 100 | prompt, |
| 101 | &cfg.image, |
| 102 | &secret_names.join(","), |
| 103 | ) |
| 104 | .await |
| 105 | .map_err(|e| StartError::Failed(e.to_string()))?; |
| 106 | |
| 107 | let id = session.id; |
| 108 | match supervisor::launch(app.clone(), session, repo_path).await { |
| 109 | Ok(()) => Ok(id), |
| 110 | Err(e) => { |
| 111 | let _ = agent::finish(&app.db, id, status::FAILED, 0, &e).await; |
| 112 | Err(StartError::Failed(e)) |
| 113 | } |
| 114 | } |
| 115 | } |
| 116 | |
| 117 | /// Ask a session to wind up. Idempotent, and safe for a session this process |
| 118 | /// does not have a handle for (it just marks the row). |
| 119 | pub async fn stop(app: &App, session_id: i64) { |
| 120 | if let Some(handle) = app.sessions.get(session_id) { |
| 121 | let _ = handle.shutdown.send(true); |
| 122 | return; |
| 123 | } |
| 124 | // No live handle: either it already ended, or it belongs to a previous |
| 125 | // process. Either way the row should not claim to be running. |
| 126 | if let Ok(Some(session)) = agent::get(&app.db, session_id).await |
| 127 | && status::is_live(&session.status) |
| 128 | { |
| 129 | let _ = agent::finish(&app.db, session_id, status::REAPED, 0, "").await; |
| 130 | } |
| 131 | } |
| 132 | |
| 133 | /// Reconcile `agent_sessions` rows against the containers Docker still has. |
| 134 | /// |
| 135 | /// The registry is in-memory, so after a restart this process has a handle for |
| 136 | /// nothing: every labelled container is an orphan and every row claiming to be |
| 137 | /// live is stale. Both are reaped, which mirrors what `ci::requeue_interrupted` |
| 138 | /// does for interrupted runs — except a terminal session cannot be resumed, so |
| 139 | /// it ends rather than re-queues. |
| 140 | pub async fn reconcile(app: &App) { |
| 141 | let docker = match anvil_docker::connect() { |
| 142 | Ok(docker) => docker, |
| 143 | Err(e) => { |
| 144 | tracing::warn!("agent: skipping reconcile, {e}"); |
| 145 | return; |
| 146 | } |
| 147 | }; |
| 148 | |
| 149 | match container::list_sessions(&docker).await { |
| 150 | Ok(found) => { |
| 151 | for (container_id, session_id) in found { |
| 152 | tracing::info!("agent: reaping orphaned container for session {session_id}"); |
| 153 | container::remove(&docker, &container_id).await; |
| 154 | } |
| 155 | } |
| 156 | Err(e) => tracing::warn!("agent: listing session containers: {e}"), |
| 157 | } |
| 158 | |
| 159 | match agent::live(&app.db).await { |
| 160 | Ok(stale) => { |
| 161 | for session in stale { |
| 162 | let _ = agent::finish( |
| 163 | &app.db, |
| 164 | session.id, |
| 165 | status::REAPED, |
| 166 | 0, |
| 167 | "anvil restarted while this session was running", |
| 168 | ) |
| 169 | .await; |
| 170 | } |
| 171 | } |
| 172 | Err(e) => tracing::error!("agent: reconciling session rows: {e}"), |
| 173 | } |
| 174 | } |
| 175 | |
| 176 | /// Periodic job enforcing the idle and wall-clock caps. |
| 177 | /// |
| 178 | /// A session with an attached viewer is never idle, however quiet the agent is |
| 179 | /// — otherwise watching a long think would kill it. |
| 180 | pub struct SweepJob; |
| 181 | |
| 182 | #[async_trait::async_trait] |
| 183 | impl anvil_core::periodic::PeriodicJob for SweepJob { |
| 184 | fn name(&self) -> &str { |
| 185 | "agent session sweep" |
| 186 | } |
| 187 | |
| 188 | async fn run(&self, app: &App) -> anvil_core::Result<()> { |
| 189 | let cfg = &app.config.agent; |
| 190 | let now = anvil_core::agent::now_secs(); |
| 191 | |
| 192 | for session in agent::live(&app.db).await? { |
| 193 | let Some(handle) = app.sessions.get(session.id) else { |
| 194 | // Live row, no handle: a leftover this process cannot drive. |
| 195 | let _ = agent::finish(&app.db, session.id, status::REAPED, 0, "").await; |
| 196 | continue; |
| 197 | }; |
| 198 | |
| 199 | let lifetime = now - session.started_at.max(session.created_at); |
| 200 | if cfg.max_lifetime_secs > 0 && lifetime as u64 > cfg.max_lifetime_secs { |
| 201 | tracing::info!("agent: session {} hit the lifetime cap", session.id); |
| 202 | let _ = handle.shutdown.send(true); |
| 203 | continue; |
| 204 | } |
| 205 | |
| 206 | if cfg.idle_timeout_secs > 0 |
| 207 | && !handle.has_viewers() |
| 208 | && handle.idle_secs() as u64 > cfg.idle_timeout_secs |
| 209 | { |
| 210 | tracing::info!("agent: session {} idle, reaping", session.id); |
| 211 | let _ = handle.shutdown.send(true); |
| 212 | } |
| 213 | } |
| 214 | Ok(()) |
| 215 | } |
| 216 | } |