Struct RuntimeStore 

pub struct RuntimeStore { /* private fields */ }

SQLite-backed durable runtime state.

Implementations

impl RuntimeStore

pub fn open_default() -> Result<RuntimeStore, Error>

pub fn open(path: impl AsRef<Path>) -> Result<RuntimeStore, Error>

pub fn path(&self) -> &Path

pub fn sessions(&self) -> SessionsRepo<'_>

pub fn messages(&self) -> MessagesRepo<'_>

pub fn tasks(&self) -> TasksRepo<'_>

pub fn tool_runs(&self) -> ToolRunsRepo<'_>

pub fn approvals(&self) -> ApprovalsRepo<'_>

pub fn processes(&self) -> ProcessesRepo<'_>

pub fn checkpoints(&self) -> CheckpointsRepo<'_>

pub fn compactions(&self) -> CompactionsRepo<'_>

pub fn plugins(&self) -> PluginsRepo<'_>

pub fn provider_probes(&self) -> ProviderProbesRepo<'_>

pub fn pairing_tokens(&self) -> PairingTokensRepo<'_>

pub fn outcomes(&self) -> OutcomesRepo<'_>

pub fn reconcile_after_restart(&self) -> Result<(usize, usize), Error>

Recover state stranded by a previous daemon’s crash/stop (#120, #118). A Running task’s worker died with the daemon, so it can never finish — mark it failed with an event. An approval left in the transient approving claim state (a replay that crashed mid-effect, #118) is reset to undecided so it reappears as pending and stays re-runnable. Call once on daemon startup, before serving. Returns (tasks_reset, claims_released).

F18 (RC-E): only daemon-owned running tasks are reset. The store is shared with interactive mermaid CLI runs; their tasks are created with a NULL owner_kind and are LEFT RUNNING here, so a live CLI session isn’t wrongly flipped to failed (with a spurious “interrupted” event) just because the daemon restarted. The daemon tags the tasks it runs in-process via NewTask::daemon_owned.

pub fn gc( &self, retention_days: i64, outcomes_retention_days: i64, ) -> Result<u64, Error>

Best-effort retention GC (#130, F22/RC-F): prune archived approvals/checkpoints, the events of long-finished tasks, terminal tasks, and the high-churn / old rows of the remaining tables, all older than retention_days. The append-only outcomes reward table — the self-improving-loop training corpus — is pruned on its own, longer outcomes_retention_days window so a large training history survives the shorter task/session window. Deletes only archived, finished, or terminal-and-old rows — active data is never touched (a running task, a still-open tool run, a live process, or a recently-updated session all survive). Returns the number of rows removed.