orch-cli 0.98.7

//! Unified task store — SQLite as single source of truth.
//!
//! `TaskStore` replaces the legacy `internal_tasks` table and GitHub labels as the
//! authoritative task state. External backends (GitHub, Linear) become sync adapters
//! that mirror state changes outward.
//!
//! Uses sqlx for async SQLite access with file-based migrations.
//!
//! All task state, metrics, KV, and rate limits flow through this module.
//!
//! # Safety Warning: Never Use `SELECT *`
//!
//! Using `SELECT *` with `sqlx-sqlite`'s prepared statement cache can cause **OOB
//! (out-of-bounds) panics** when the schema changes (e.g., new migration adds a column).
//!
//! The issue occurs because:
//! 1. SQLite caches prepared statements with column metadata at prepare time
//! 2. When a migration adds a new column, `sqlite3_column_count()` returns the updated count
//! 3. The cached statement's column metadata vector was built with the old count
//! 4. Accessing column index N when only N columns exist (0..N-1) causes a panic
//!
//! **Solution:** Always use explicit column lists via constants like `TASK_COLS`,
//! `CONTROL_MESSAGE_COLS`, `JOB_STATE_COLS`, etc. See `tasks.rs` and `control.rs`
//! for examples. CI will fail if `SELECT *` is detected in store files.

use anyhow::Context;
use sqlx::sqlite::{SqliteConnectOptions, SqliteJournalMode, SqlitePool, SqlitePoolOptions};
use std::path::Path;
use std::sync::{Arc, OnceLock};

pub(crate) mod control;
pub(crate) mod dependencies;
mod helpers;
pub(crate) mod jobs;
pub(crate) mod kv;
mod metrics;
mod pricing;
pub(crate) mod replica;
pub(crate) mod tasks;

pub use control::{ChatCostSummary, ControlMessage, MemoryEntry, parse_since_duration};
pub use dependencies::{
    BlockerSnapshot, DepSource, NewDependency, ObservedPullRequest, TaskDependency,
    WaitingOnDependency, blocker_failure_kind, blocker_is_satisfied, blocker_snapshot_key,
    dep_label, worktree_refresh_key,
};
pub use helpers::{
    StoreResult, get_cost_estimate, get_cost_estimate_result, get_recent_memory,
    get_recent_memory_result, get_task_field, get_task_field_direct, get_token_usage,
    get_token_usage_result, opt_store_get_task, opt_store_get_task_by_id, resolve_store_id,
    store_increment, store_increment_by_id, store_log_activity, store_reset_counters,
    store_reset_failure_counters, store_set, store_set_by_id, store_set_result,
    store_set_result_by_id, store_touch_updated_at,
};
pub use jobs::JobState;
pub use metrics::{
    AgentStat, CostByGroup, CostPeriod, CostSummary, ErrorStat, HighReviewCycleTask,
    InsertTaskMetric, MetricsSummary, SlowTaskInfo,
};
pub use pricing::{CostEstimate, ModelPricing, TokenUsage, pricing_for_model};
pub use tasks::{
    CliTaskRef, CliTaskSpec, CompleteRun, NewTask, RunTokenUsage, StartRun, StoreRoute, Task,
    TaskActivity, TaskRun, TaskStatus, UpsertExternal, parse_cli_task_spec,
};

/// Default database path: `~/.orch/orch.db`
pub async fn default_db_path() -> anyhow::Result<std::path::PathBuf> {
    crate::home::db_path().await
}

/// The unified task store backed by SQLite (via sqlx).
#[derive(Clone)]
pub struct TaskStore {
    pool: SqlitePool,
    /// Set once on the daemon. Without it, replicated writes apply directly, as in tests and the CLI
    raft: Arc<OnceLock<crate::cluster::Node>>,
    /// Set on the CLI when a daemon runs, so its writes go through the daemon's log
    daemon: Arc<OnceLock<crate::cluster::local::Endpoint>>,
}

impl std::fmt::Debug for TaskStore {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("TaskStore").finish_non_exhaustive()
    }
}

impl TaskStore {
    /// Open (or create) the store at the given database path.
    ///
    /// Runs file-based migrations from the `migrations/` directory.
    pub async fn open(db_path: &Path) -> anyhow::Result<Self> {
        let options = SqliteConnectOptions::new()
            .filename(db_path)
            .journal_mode(SqliteJournalMode::Wal)
            .create_if_missing(true)
            .busy_timeout(std::time::Duration::from_secs(5))
            .pragma("foreign_keys", "ON");

        let pool = SqlitePoolOptions::new()
            .max_connections(5)
            .connect_with(options)
            .await
            .with_context(|| format!("opening task store: {}", db_path.display()))?;

        let store = Self {
            pool,
            raft: Arc::default(),
            daemon: Arc::default(),
        };
        store.migrate().await?;
        Ok(store)
    }

    /// Open a store with a single connection.
    ///
    /// Uses `max_connections(1)` to avoid WAL visibility issues where a write
    /// on one pooled connection isn't visible to a read on another. Used by
    /// integration tests that need read-your-own-writes consistency.
    pub async fn open_single(db_path: &Path) -> anyhow::Result<Self> {
        let options = SqliteConnectOptions::new()
            .filename(db_path)
            .journal_mode(SqliteJournalMode::Wal)
            .create_if_missing(true)
            .busy_timeout(std::time::Duration::from_secs(5))
            .pragma("foreign_keys", "ON");

        let pool = SqlitePoolOptions::new()
            .max_connections(1)
            .connect_with(options)
            .await
            .with_context(|| format!("opening task store (single): {}", db_path.display()))?;

        let store = Self {
            pool,
            raft: Arc::default(),
            daemon: Arc::default(),
        };
        store.migrate().await?;
        Ok(store)
    }

    /// Open an in-memory store (for testing).
    #[cfg(test)]
    pub async fn open_memory() -> anyhow::Result<Self> {
        let options = SqliteConnectOptions::new()
            .filename(":memory:")
            .pragma("foreign_keys", "ON");
        let pool = SqlitePoolOptions::new()
            .max_connections(1)
            .connect_with(options)
            .await?;

        let store = Self {
            pool,
            raft: Arc::default(),
            daemon: Arc::default(),
        };
        store.migrate().await?;
        Ok(store)
    }

    /// Run migrations.
    async fn migrate(&self) -> anyhow::Result<()> {
        sqlx::migrate!("./migrations")
            .run(&self.pool)
            .await
            .context("running task store migrations")?;
        Ok(())
    }

    pub fn attach_node(&self, node: crate::cluster::Node) {
        if self.raft.set(node).is_err() {
            tracing::warn!("raft already attached to the task store");
        }
    }

    /// True on the node that decides: the Raft leader, or a store with no Raft, as in the CLI and tests.
    /// A single node always leads, so one machine runs every phase as before
    pub fn is_leader(&self) -> bool {
        self.raft.get().is_none_or(super::cluster::Node::is_leader)
    }

    /// Writes go through Raft or a daemon that runs it, so they reach the other machines
    pub(crate) fn replicates(&self) -> bool {
        self.raft.get().is_some() || self.daemon.get().is_some()
    }

    pub fn node(&self) -> Option<&crate::cluster::Node> {
        self.raft.get()
    }

    pub fn attach_daemon(&self, endpoint: crate::cluster::local::Endpoint) {
        let _ = self.daemon.set(endpoint);
    }

    /// Propose a replicated write through Raft, or apply it directly when no Raft is attached
    pub(crate) async fn write(
        &self,
        cmd: crate::cluster::Command,
    ) -> anyhow::Result<crate::cluster::Reply> {
        let Some(node) = self.raft.get() else {
            if let Some(endpoint) = self.daemon.get() {
                match crate::cluster::local::send(endpoint, &cmd).await {
                    Ok(reply) => return reply_result(reply),
                    // Nothing was sent, so applying here cannot apply the command twice
                    Err(e) if e.is_connect() => {
                        tracing::debug!(err = %e, "daemon not listening, writing directly");
                    }
                    Err(e) => return Err(e.into()),
                }
            }

            return self.apply_direct(cmd).await;
        };
        let req = crate::cluster::Request {
            now: crate::cluster::now_stamp(),
            cmd,
        };
        reply_result(node.write(req).await?)
    }

    pub(crate) async fn write_int(&self, cmd: crate::cluster::Command) -> anyhow::Result<i64> {
        match self.write(cmd).await? {
            crate::cluster::Reply::Int(n) => Ok(n),
            other => anyhow::bail!("expected an integer reply, got {other:?}"),
        }
    }

    pub(crate) async fn write_bool(&self, cmd: crate::cluster::Command) -> anyhow::Result<bool> {
        match self.write(cmd).await? {
            crate::cluster::Reply::Bool(b) => Ok(b),
            other => anyhow::bail!("expected a boolean reply, got {other:?}"),
        }
    }

    /// The same SQL apply runs, in one transaction on this node only
    pub(crate) async fn apply_direct(
        &self,
        cmd: crate::cluster::Command,
    ) -> anyhow::Result<crate::cluster::Reply> {
        let mut attempt = 0;
        loop {
            match self.apply_once(cmd.clone()).await {
                // A dropped transaction rolls back, so running the command again is safe
                Err(e) if attempt < WRITE_LOCK_RETRIES && is_write_lock_error(&e) => {
                    attempt += 1;
                    tracing::warn!(attempt, err = %e, "database is locked, retrying the write");
                    tokio::time::sleep(std::time::Duration::from_millis(100 << (attempt - 1)))
                        .await;
                }
                result => return result,
            }
        }
    }

    async fn apply_once(
        &self,
        cmd: crate::cluster::Command,
    ) -> anyhow::Result<crate::cluster::Reply> {
        let mut tx = self.begin_write().await?;
        let reply = cmd
            .apply(&mut tx, &crate::cluster::now_stamp(), false)
            .await?;
        tx.commit().await?;
        Ok(reply)
    }

    /// Start a transaction that holds the write lock from its first statement.
    ///
    /// A deferred `BEGIN` that reads and then writes fails at once with `database is locked` when
    /// another connection commits in between, and `busy_timeout` cannot wait that out. `BEGIN
    /// IMMEDIATE` waits for the lock up front, where `busy_timeout` applies
    pub(crate) async fn begin_write(
        &self,
    ) -> sqlx::Result<sqlx::Transaction<'static, sqlx::Sqlite>> {
        self.pool.begin_with("BEGIN IMMEDIATE").await
    }

    /// Get a reference to the underlying pool (for advanced queries).
    pub fn pool(&self) -> &SqlitePool {
        &self.pool
    }
}

/// Retries for a write that still finds the database locked after `busy_timeout`
const WRITE_LOCK_RETRIES: u32 = 4;

/// `SQLITE_BUSY` (5) or `SQLITE_BUSY_SNAPSHOT` (517) anywhere in the error chain
fn is_write_lock_error(err: &anyhow::Error) -> bool {
    err.chain().any(|cause| {
        cause
            .downcast_ref::<sqlx::Error>()
            .and_then(|e| e.as_database_error())
            .and_then(sqlx::error::DatabaseError::code)
            .is_some_and(|code| code == "5" || code == "517")
    })
}

fn reply_result(reply: crate::cluster::Reply) -> anyhow::Result<crate::cluster::Reply> {
    match reply {
        crate::cluster::Reply::Err(e) => Err(anyhow::anyhow!(e)),
        other => Ok(other),
    }
}

#[cfg(test)]
mod tests;