Struct ShellSessionRegistry 

pub struct ShellSessionRegistry {
    shells: Mutex<HashMap<String, ShellEntry>>,
    status_map: Mutex<HashMap<String, Arc<Mutex<ShellStatusInfo>>>>,
    status_order: Mutex<VecDeque<String>>,
}

Fields

shells: Mutex<HashMap<String, ShellEntry>>

status_map: Mutex<HashMap<String, Arc<Mutex<ShellStatusInfo>>>>

status_order: Mutex<VecDeque<String>>

Implementations

impl ShellSessionRegistry

pub fn new() -> Arc<Self>

pub(crate) fn register_full( &self, shell_id: String, stop_tx: Sender<()>, stdin_tx: Option<UnboundedSender<String>>, status: Arc<Mutex<ShellStatusInfo>>, ) -> Vec<String>

fn prune_exited_statuses(&self) -> Vec<String>

Evict oldest EXITED status entries beyond MAX_EXITED_STATUS. Running shells are always retained (their status is read live via ShellStatus). Also drops any order-queue ids whose status entry is already gone. Returns the shell_ids actually evicted (status entry removed) so the caller can purge their broker-side persisted history too.

pub fn stop(&self, shell_id: &str) -> bool

Request stop of a running shell. Returns false if unknown or already exited. Dropping the stdin_tx here closes the relay channel → ChildStdin drops → child sees EOF, unblocking any stdin-reading process before the kill.

fn remove(&self, shell_id: &str)

Called by the runner on natural exit (stdout/stderr closed). Removes the shells entry (dropping stdin_tx → child gets stdin EOF). Status entry is intentionally kept so ShellStatus can query the final exit code/line count.

pub fn stop_all(&self) -> usize

Stop every running shell. For srv-shutdown cleanup.

pub fn resolve_stdin( &self, shell_id: &str, ) -> Result<UnboundedSender<String>, ShellInputFailure>

Resolve where a ShellInput write should go, distinguishing two failure modes:

• Ok(tx) — running with captured stdin
• Err(StdinNotCaptured) — running but created without capture_stdin
• Err(NotRunning) — unknown id or already exited

pub fn get_status(&self, shell_id: &str) -> ShellStatusInfo

Return a snapshot of the shell’s status. Returns running: false if unknown. Existing, documented contract for POST /api/v1/shell/status (the MCP ShellStatus tool) — do not change: an agent calling this for an id it doesn’t recognize expects a plain “not running” answer, not an error.

pub(crate) fn get_status_if_known( &self, shell_id: &str, ) -> Option<ShellStatusInfo>

Like Self::get_status, but None when no entry exists at all — distinguishing “genuinely unknown” from “known, and not running.”

Needed by the shellstatus RPC command (unlike the MCP tool above, which is fine collapsing “unknown” into “not running”): register_full only runs after the runner task spawns the child process, but shell_node_create is published to the frontend BEFORE that task is even scheduled (handle_shell_create in server/mod.rs). A caller that checks status immediately after seeing shell_node_create can race ahead of registration — reagent P1 on PR #2770: collapsing that race window into running: false was indistinguishable from a genuinely already-exited shell, so a fast, still-registering, genuinely LIVE shell (e.g. a real task dev) could be shown as failed in the Activity Dock for its entire run.

pub fn list_active(&self) -> Vec<ShellSummary>

List every currently-RUNNING shell across all blocks — the Swarm pane’s data source for its per-agent long-running-process rows (SPEC_SWARM_LONG_RUNNING_PROCESS_ROWS_2026_07_20). Unfiltered, like subagent.ListActive/ListDispatches — the Swarm pane fetches once and groups by block_id client-side in buildTree(), rather than one RPC per tracked agent block. Exited shells are intentionally excluded: Phase 1’s scope is “what’s happening now,” not a shell history view.

Trait Implementations

impl Default for ShellSessionRegistry

fn default() -> ShellSessionRegistry

Returns the “default value” for a type.