voro-core 0.2.0

use std::collections::HashMap;
use std::path::{Path, PathBuf};

use rusqlite::{Connection, OptionalExtension, params};

use crate::error::{Error, Result};
use crate::model::{
    Dep, DepKind, DepRef, Doc, Event, LivenessSource, Priority, Project, RefineOutcome, Repo,
    RunningRow, Session, SessionOutcome, Task, TaskState, location_is_url,
};

const MIGRATIONS: &[&str] = &[
    include_str!("../migrations/0001_init.sql"),
    include_str!("../migrations/0002_rename_backlog_to_parked.sql"),
    include_str!("../migrations/0003_track_pr.sql"),
    include_str!("../migrations/0004_add_session_ref.sql"),
    include_str!("../migrations/0005_add_branch.sql"),
    include_str!("../migrations/0006_one_open_session_per_task.sql"),
    include_str!("../migrations/0007_add_human.sql"),
    include_str!("../migrations/0008_add_stalled_state.sql"),
    include_str!("../migrations/0009_add_review_action.sql"),
    include_str!("../migrations/0010_add_waiting_state.sql"),
    include_str!("../migrations/0011_add_archived.sql"),
    include_str!("../migrations/0012_repos.sql"),
    include_str!("../migrations/0013_add_deep.sql"),
    include_str!("../migrations/0014_docs.sql"),
    include_str!("../migrations/0015_dep_kind_in_key.sql"),
    include_str!("../migrations/0016_add_refining_state.sql"),
    include_str!("../migrations/0017_schema_migrations.sql"),
    include_str!("../migrations/0018_project_viewer.sql"),
    include_str!("../migrations/0019_session_liveness_source.sql"),
    include_str!("../migrations/0020_store_meta.sql"),
];

/// Whether a path lies inside a Cargo build directory — a `target` component
/// followed immediately by a profile. Covers `target/debug/voro`,
/// `target/release/voro`, and the `target/debug/deps/` binaries the test
/// harness runs, in a worktree or the primary checkout alike.
fn path_is_cargo_target(path: &Path) -> bool {
    let parts: Vec<_> = path
        .components()
        .map(|c| c.as_os_str().to_string_lossy().into_owned())
        .collect();
    parts
        .windows(2)
        .any(|pair| pair[0] == "target" && (pair[1] == "debug" || pair[1] == "release"))
}

/// Write the journal rows for a migration pass (§5). Migrations applied before
/// the journal existed are backfilled with a NULL `sql`; what this pass applies
/// is recorded verbatim, signed with the build that applied it and, on a
/// protected store, with the consent that let it (§5).
fn record_in_journal(tx: &Connection, from_version: usize, consent: Option<&str>) -> Result<()> {
    let found: i64 = tx.query_row(
        "SELECT COUNT(*) FROM sqlite_master WHERE type = 'table' AND name = 'schema_migrations'",
        [],
        |row| row.get(0),
    )?;
    if found == 0 {
        return Ok(());
    }
    for idx in 1..=from_version {
        tx.execute(
            "INSERT OR IGNORE INTO schema_migrations (idx, sql, applied_at, applied_by)
             VALUES (?1, NULL, datetime('now'), NULL)",
            params![idx as i64],
        )?;
    }
    let by = applied_by(consent);
    for idx in (from_version + 1)..=MIGRATIONS.len() {
        tx.execute(
            "INSERT OR REPLACE INTO schema_migrations (idx, sql, applied_at, applied_by)
             VALUES (?1, ?2, datetime('now'), ?3)",
            params![idx as i64, MIGRATIONS[idx - 1], by],
        )?;
    }
    Ok(())
}

/// How a build signs the journal: crate version and the running executable's
/// path, which is what identifies the build behind a divergence. A consented
/// migration of a protected store appends how the consent was given, so even
/// a `--yes` override leaves a trace.
fn applied_by(consent: Option<&str>) -> String {
    let exe = std::env::current_exe()
        .map(|p| p.display().to_string())
        .unwrap_or_else(|_| "an unknown executable".to_string());
    let via = consent.map(|c| format!(", {c}")).unwrap_or_default();
    format!("voro {} at {exe}{via}", env!("CARGO_PKG_VERSION"))
}

/// The way out of a database carrying a migration this build does not have:
/// the dev store is rebuilt, the operator's is restored from a snapshot.
fn remedy_for_divergence(path: Option<&Path>) -> String {
    if path.is_some_and(|p| p == Store::dev_db_path()) {
        format!(
            "It is the dev store, which is disposable — rebuild it at this build's schema with \
             `voro seed --force`, or delete {} and it will be reseeded on the next run.",
            Store::dev_db_path().display()
        )
    } else {
        format!(
            "Restore the snapshot taken before that migration from {}; failing that, run the \
             build named above, which is the only one whose schema matches this database.",
            Store::backup_dir_for(path.unwrap_or(&Store::production_db_path())).display()
        )
    }
}

/// The way out of a database whose schema is ahead of this build: the dev
/// store is rebuilt, the operator's is restored from a snapshot.
fn remedy_for_schema_ahead(path: Option<&Path>) -> String {
    if path.is_some_and(|p| p == Store::dev_db_path()) {
        format!(
            "It is the dev store, which is disposable — rebuild it at this build's schema with \
             `voro seed --force`, or delete {} and it will be reseeded on the next run.",
            Store::dev_db_path().display()
        )
    } else {
        format!(
            "Restore a pre-migration snapshot from {}; failing that, run the build that migrated \
             it — though if that build was never released, doing so entrenches a schema no other \
             build can open.",
            Store::backup_dir_for(path.unwrap_or(&Store::production_db_path())).display()
        )
    }
}

/// Owns the SQLite database. All writes go through this type; task state in
/// particular is only ever changed by the transition API in `transition.rs`.
pub struct Store {
    pub(crate) conn: Connection,
}

/// Drop every row, leaving the schema in place — the reset behind
/// `voro seed --force`. The CLI is what confines this to the dev store.
impl Store {
    pub fn truncate_all(&mut self) -> Result<()> {
        let tx = self.conn.transaction()?;
        tx.pragma_update(None, "foreign_keys", false)?;
        for table in [
            "task_docs",
            "docs",
            "deps",
            "events",
            "sessions",
            "tasks",
            "repos",
            "projects",
        ] {
            tx.execute(&format!("DELETE FROM {table}"), [])?;
        }
        tx.execute("DELETE FROM sqlite_sequence", []).ok();
        tx.commit()?;
        self.conn.pragma_update(None, "foreign_keys", true)?;
        Ok(())
    }
}

/// Initial state for a task created by a human. `proposed` is quick capture;
/// `parked`/`ready` mean the creator has already triaged their own task.
#[derive(Debug, Clone)]
pub struct NewTask {
    pub project_id: i64,
    /// The repo the task runs in; `None` resolves to the project's default.
    pub repo_id: Option<i64>,
    pub title: String,
    pub body: String,
    pub priority: Priority,
    pub state: TaskState,
    pub agent: Option<String>,
    pub human: bool,
    pub deep: bool,
}

/// Content edits. State is deliberately absent — use `Store::apply`.
#[derive(Debug, Clone)]
pub struct TaskEdit {
    pub title: String,
    pub body: String,
    pub priority: Priority,
    pub agent: Option<String>,
    pub human: bool,
    pub deep: bool,
}

impl Store {
    pub fn open(path: &Path) -> Result<Store> {
        Store::open_with_consent(path, None)
    }

    /// Open with consent to migrate a protected store (§5): the TUI's launch
    /// prompt and `voro migrate` call this after a human has answered, or with
    /// `--yes` standing in for one. `consent` says how the consent was given
    /// and is recorded in the journal's `applied_by`. On an unprotected store
    /// it changes nothing — migration there never needed asking.
    pub fn open_migrate(path: &Path, consent: &str) -> Result<Store> {
        Store::open_with_consent(path, Some(consent))
    }

    fn open_with_consent(path: &Path, consent: Option<&str>) -> Result<Store> {
        if let Some(dir) = path.parent() {
            std::fs::create_dir_all(dir)
                .map_err(|e| Error::Invalid(format!("cannot create {}: {e}", dir.display())))?;
        }
        Store::open_at(
            Connection::open(path)?,
            path,
            &Store::production_db_path(),
            consent,
        )
    }

    pub fn open_in_memory() -> Result<Store> {
        Store::from_connection_at(Connection::open_in_memory()?, None)
    }

    /// `$XDG_DATA_HOME/voro`, defaulting to `~/.local/share/voro`.
    pub fn data_dir() -> PathBuf {
        let data_home = std::env::var_os("XDG_DATA_HOME")
            .map(PathBuf::from)
            .filter(|p| p.is_absolute())
            .unwrap_or_else(|| {
                let home = std::env::var_os("HOME")
                    .map(PathBuf::from)
                    .unwrap_or_default();
                home.join(".local/share")
            });
        data_home.join("voro")
    }

    /// The operator's store (DESIGN.md §5), at a path that does not vary with
    /// how the running binary was built. Dispatch renders `--db` against it,
    /// and `voro seed` refuses it.
    pub fn production_db_path() -> PathBuf {
        Store::data_dir().join("voro.db")
    }

    /// The store a build out of a `target/` directory opens instead (DESIGN.md
    /// §5). Seeded on first open and disposable: `voro seed --force` rebuilds
    /// it, and deleting it costs nothing.
    pub fn dev_db_path() -> PathBuf {
        Store::data_dir().join("dev.db")
    }

    /// Where snapshots taken before a migration land: beside the database they
    /// protect, so a store opened with `--db` keeps its own history.
    pub fn backup_dir_for(path: &Path) -> PathBuf {
        path.parent()
            .filter(|p| !p.as_os_str().is_empty())
            .unwrap_or(Path::new("."))
            .join("backups")
    }

    /// True when this binary was run out of a Cargo `target/` directory rather
    /// than installed. It picks the default store and bounds nothing:
    /// `cargo install --path` builds a working checkout, unreleased migrations
    /// and all, into an ordinary install location, where this reads as an
    /// install. The journal and the counter (§5) are what protect the schema.
    pub fn is_dev_build() -> bool {
        std::env::current_exe().is_ok_and(|exe| path_is_cargo_target(&exe))
    }

    /// The store a bare `voro` opens: the dev one for a dev build, the
    /// operator's otherwise.
    pub fn default_db_path() -> PathBuf {
        if Store::is_dev_build() {
            Store::dev_db_path()
        } else {
            Store::production_db_path()
        }
    }

    #[cfg(test)]
    fn from_connection(conn: Connection) -> Result<Store> {
        Store::from_connection_at(conn, None)
    }

    fn from_connection_at(conn: Connection, path: Option<&Path>) -> Result<Store> {
        Store::open_at_opt(conn, path, &Store::production_db_path(), None)
    }

    fn open_at(
        conn: Connection,
        path: &Path,
        production: &Path,
        consent: Option<&str>,
    ) -> Result<Store> {
        Store::open_at_opt(conn, Some(path), production, consent)
    }

    fn open_at_opt(
        conn: Connection,
        path: Option<&Path>,
        production: &Path,
        consent: Option<&str>,
    ) -> Result<Store> {
        conn.pragma_update(None, "foreign_keys", true)?;
        let mut store = Store { conn };
        let version = store.schema_version()?;
        // Ahead of the version check, which cannot see a divergence.
        store.verify_journal(path)?;
        if version > MIGRATIONS.len() {
            return Err(Error::SchemaAhead {
                version,
                known: MIGRATIONS.len(),
                remedy: remedy_for_schema_ahead(path),
            });
        }
        if version < MIGRATIONS.len()
            && let Some(path) = path
        {
            // The consent gate (§5). A store with no schema at all is exempt —
            // a fresh install creates its database silently, and there is
            // nothing yet to protect.
            if version > 0 && consent.is_none() && store.is_protected(path, production)? {
                return Err(Error::MigrationsPending {
                    path: path.to_path_buf(),
                    pending: MIGRATIONS.len() - version,
                    version,
                    known: MIGRATIONS.len(),
                });
            }
            store.snapshot(path, version)?;
        }
        store.migrate(consent)?;
        if path == Some(production) {
            store.mark_protected()?;
        }
        Ok(store)
    }

    /// Whether this store is the operator's (§5): opened at the production
    /// path, or carrying the `protected` marker a past open there wrote — how
    /// the property survives a symlink, a moved data directory, or a restored
    /// copy. Runs before any migration, so it must read a store from before
    /// `store_meta` existed, where only the path can answer.
    fn is_protected(&self, path: &Path, production: &Path) -> Result<bool> {
        if path == production {
            return Ok(true);
        }
        let has_meta: i64 = self.conn.query_row(
            "SELECT COUNT(*) FROM sqlite_master WHERE type = 'table' AND name = 'store_meta'",
            [],
            |row| row.get(0),
        )?;
        if has_meta == 0 {
            return Ok(false);
        }
        let marked: Option<String> = self
            .conn
            .query_row(
                "SELECT value FROM store_meta WHERE key = 'protected'",
                [],
                |row| row.get(0),
            )
            .optional()?;
        Ok(marked.as_deref() == Some("1"))
    }

    /// `INSERT OR IGNORE` so an already-marked store takes no write at all:
    /// opening must not bump `data_version` for connections polling it.
    fn mark_protected(&self) -> Result<()> {
        self.conn.execute(
            "INSERT OR IGNORE INTO store_meta (key, value) VALUES ('protected', '1')",
            [],
        )?;
        Ok(())
    }

    /// Check the journal (§5) against the migrations this build carries. The
    /// counter reports a database that is *ahead*; this reports one that is
    /// *different*, which two branches numbering a migration alike produce.
    /// History predating the journal has a NULL `sql` and is skipped as
    /// unverifiable.
    fn verify_journal(&self, path: Option<&Path>) -> Result<()> {
        if !self.has_journal()? {
            return Ok(());
        }
        let mut stmt = self.conn.prepare(
            "SELECT idx, sql, applied_at, applied_by FROM schema_migrations
             WHERE sql IS NOT NULL ORDER BY idx",
        )?;
        let rows = stmt.query_map([], |row| {
            Ok((
                row.get::<_, i64>(0)? as usize,
                row.get::<_, String>(1)?,
                row.get::<_, String>(2)?,
                row.get::<_, Option<String>>(3)?,
            ))
        })?;
        for row in rows {
            let (idx, applied, applied_at, applied_by) = row?;
            // Indices beyond this build's list are the counter's to report.
            let Some(carried) = MIGRATIONS.get(idx - 1) else {
                continue;
            };
            if applied != *carried {
                return Err(Error::SchemaDiverged {
                    idx,
                    applied_at,
                    applied_by: applied_by.unwrap_or_else(|| "an unrecorded build".to_string()),
                    remedy: remedy_for_divergence(path),
                });
            }
        }
        Ok(())
    }

    fn has_journal(&self) -> Result<bool> {
        let found: i64 = self.conn.query_row(
            "SELECT COUNT(*) FROM sqlite_master WHERE type = 'table' AND name = 'schema_migrations'",
            [],
            |row| row.get(0),
        )?;
        Ok(found > 0)
    }

    /// The store's `user_version`. An open store always reads as the count of
    /// migrations its build carries; `voro migrate` reports it when there was
    /// nothing to apply.
    pub fn schema_version(&self) -> Result<usize> {
        Ok(self
            .conn
            .query_row("PRAGMA user_version", [], |row| row.get::<_, i64>(0))? as usize)
    }

    /// Copy the database beside itself before a migration touches it, since a
    /// migration that renames or drops a column is not reversible from the
    /// migrated file alone. A database with no schema yet is skipped, and a
    /// failure to write the copy is reported rather than fatal.
    fn snapshot(&self, path: &Path, version: usize) -> Result<()> {
        if version == 0 || !path.exists() {
            return Ok(());
        }
        let stamp: String =
            self.conn
                .query_row("SELECT strftime('%Y%m%d-%H%M%S', 'now')", [], |row| {
                    row.get(0)
                })?;
        let stem = path
            .file_stem()
            .map(|s| s.to_string_lossy().into_owned())
            .unwrap_or_else(|| "voro".to_string());
        let dir = Store::backup_dir_for(path);
        let target = dir.join(format!("{stem}-v{version}-{stamp}.db"));
        let copied = std::fs::create_dir_all(&dir).and_then(|()| std::fs::copy(path, &target));
        if let Err(e) = copied {
            eprintln!(
                "voro: could not snapshot {} before migrating to schema {}: {e}",
                path.display(),
                MIGRATIONS.len()
            );
        }
        Ok(())
    }

    /// SQLite's `PRAGMA data_version`, which increments whenever another
    /// connection commits a change to the database. The value is stable across
    /// commits made on this connection, so a caller can poll it to detect
    /// external writes without reacting to its own mutations.
    pub fn data_version(&self) -> Result<i64> {
        Ok(self
            .conn
            .query_row("PRAGMA data_version", [], |r| r.get(0))?)
    }

    /// Migrations may rebuild tables (SQLite cannot alter CHECK constraints),
    /// so foreign-key enforcement is suspended for the duration and integrity
    /// verified afterwards — the procedure SQLite documents for schema changes.
    fn migrate(&mut self, consent: Option<&str>) -> Result<()> {
        self.conn.pragma_update(None, "foreign_keys", false)?;
        let applied = self.apply_migrations(consent);
        let restored = self.conn.pragma_update(None, "foreign_keys", true);
        applied?;
        restored?;
        let violations: i64 =
            self.conn
                .query_row("SELECT COUNT(*) FROM pragma_foreign_key_check", [], |r| {
                    r.get(0)
                })?;
        if violations > 0 {
            return Err(Error::Invalid(format!(
                "{violations} foreign key violation(s) after migration"
            )));
        }
        Ok(())
    }

    fn apply_migrations(&mut self, consent: Option<&str>) -> Result<()> {
        let tx = self.conn.transaction()?;
        let version: usize =
            tx.query_row("PRAGMA user_version", [], |row| row.get::<_, i64>(0))? as usize;
        for (i, sql) in MIGRATIONS.iter().enumerate().skip(version) {
            tx.execute_batch(sql)?;
            tx.pragma_update(None, "user_version", (i + 1) as i64)?;
        }
        record_in_journal(&tx, version, consent)?;
        tx.commit()?;
        Ok(())
    }

    // --- projects ---

    /// Create a project and its default repo in one transaction, so a project
    /// with no checkout is never observable (DESIGN.md §5). The repo is named
    /// after the project; `voro repo` renames nothing, but `repo add` puts
    /// further checkouts beside it.
    pub fn create_project(&mut self, name: &str, path: &str) -> Result<Project> {
        let tx = self.conn.transaction()?;
        tx.execute("INSERT INTO projects (name) VALUES (?1)", params![name])?;
        let id = tx.last_insert_rowid();
        tx.execute(
            "INSERT INTO repos (project_id, name, path, is_default) VALUES (?1, ?2, ?3, 1)",
            params![id, name, path],
        )?;
        tx.commit()?;
        self.project(id)
    }

    pub fn project(&self, id: i64) -> Result<Project> {
        self.conn
            .query_row(
                &format!("SELECT {PROJECT_COLUMNS} FROM projects WHERE id = ?1"),
                [id],
                project_from_row,
            )
            .optional()?
            .ok_or(Error::ProjectNotFound(id))
    }

    pub fn projects(&self) -> Result<Vec<Project>> {
        let mut stmt = self.conn.prepare(&format!(
            "SELECT {PROJECT_COLUMNS} FROM projects ORDER BY name"
        ))?;
        let rows = stmt.query_map([], project_from_row)?;
        Ok(rows.collect::<rusqlite::Result<_>>()?)
    }

    pub fn set_weight(&mut self, project_id: i64, weight: i64) -> Result<()> {
        if !(0..=5).contains(&weight) {
            return Err(Error::Invalid(format!("weight {weight} out of range 0-5")));
        }
        let changed = self.conn.execute(
            "UPDATE projects SET weight = ?1 WHERE id = ?2",
            params![weight, project_id],
        )?;
        if changed == 0 {
            return Err(Error::ProjectNotFound(project_id));
        }
        Ok(())
    }

    /// Name the `voro.toml` viewer this project's local diffs open in
    /// (DESIGN.md §8/§11a). `None` stores NULL — no viewer named, so `open`
    /// falls back to the config's default viewer.
    pub fn set_viewer(&mut self, project_id: i64, viewer: Option<&str>) -> Result<Project> {
        let viewer = match viewer.map(str::trim) {
            Some("") => {
                return Err(Error::Invalid(
                    "viewer name is required — name no viewer to use the default one".into(),
                ));
            }
            named => named,
        };
        let changed = self.conn.execute(
            "UPDATE projects SET viewer = ?1 WHERE id = ?2",
            params![viewer, project_id],
        )?;
        if changed == 0 {
            return Err(Error::ProjectNotFound(project_id));
        }
        self.project(project_id)
    }

    /// Archive or unarchive a project (DESIGN.md §5). Archiving hides the
    /// project and all of its tasks from the cockpit views; the tasks
    /// themselves are not touched — no state change, no event — so unarchiving
    /// restores the pre-archive view exactly. Refuses a no-op so a typo'd
    /// second archive is heard rather than silently absorbed.
    pub fn set_archived(&mut self, project_id: i64, archived: bool) -> Result<Project> {
        let project = self.project(project_id)?;
        if project.archived == archived {
            return Err(Error::Invalid(format!(
                "project '{}' is {} archived",
                project.name,
                if archived { "already" } else { "not" }
            )));
        }
        self.conn.execute(
            "UPDATE projects SET archived = ?1 WHERE id = ?2",
            params![archived, project_id],
        )?;
        self.project(project_id)
    }

    /// Tasks reference a project by id, not name, so renaming is a pure
    /// label change — no task or dependency is touched.
    pub fn rename_project(&mut self, project_id: i64, name: &str) -> Result<Project> {
        let changed = self.conn.execute(
            "UPDATE projects SET name = ?1 WHERE id = ?2",
            params![name, project_id],
        )?;
        if changed == 0 {
            return Err(Error::ProjectNotFound(project_id));
        }
        self.project(project_id)
    }

    /// Re-point a project's *default* repo. This is what `voro project path`
    /// and the projects screen's path field edit — the single-repo spelling of
    /// `repo path`, kept because a one-repo project is still the common case.
    pub fn set_default_repo_path(&mut self, project_id: i64, path: &str) -> Result<Repo> {
        let repo = self.default_repo(project_id)?;
        self.set_repo_path(repo.id, path)
    }

    /// Delete a project outright — only safe when it has no tasks, since tasks
    /// reference their project by id and deleting would orphan history. A project
    /// with tasks in any state refuses; weight 0 snoozes without losing history.
    /// Its repos go with it: no task references them, so nothing is orphaned.
    pub fn delete_project(&mut self, project_id: i64) -> Result<()> {
        self.project(project_id)?;
        let task_count: i64 = self.conn.query_row(
            "SELECT COUNT(*) FROM tasks WHERE project_id = ?1",
            [project_id],
            |r| r.get(0),
        )?;
        if task_count > 0 {
            return Err(Error::ProjectHasTasks {
                id: project_id,
                count: task_count,
            });
        }
        let tx = self.conn.transaction()?;
        tx.execute("DELETE FROM repos WHERE project_id = ?1", [project_id])?;
        tx.execute("DELETE FROM projects WHERE id = ?1", [project_id])?;
        tx.commit()?;
        Ok(())
    }

    // --- repos ---
    //
    // A project owns at least one repo, exactly one of which is its default
    // (DESIGN.md §3/§5). The at-most-one-default half is schema-enforced by a
    // partial unique index; the rest — never zero repos, no deleting the
    // default or a referenced one — lives here, the same place the state
    // machine's invariants live, so no interface can bypass them.

    /// A project's repos, default first, then by name.
    pub fn repos(&self, project_id: i64) -> Result<Vec<Repo>> {
        let mut stmt = self.conn.prepare(&format!(
            "SELECT {REPO_COLUMNS} FROM repos WHERE project_id = ?1
             ORDER BY is_default DESC, name"
        ))?;
        let rows = stmt.query_map([project_id], repo_from_row)?;
        Ok(rows.collect::<rusqlite::Result<_>>()?)
    }

    pub fn repo(&self, id: i64) -> Result<Repo> {
        self.conn
            .query_row(
                &format!("SELECT {REPO_COLUMNS} FROM repos WHERE id = ?1"),
                [id],
                repo_from_row,
            )
            .optional()?
            .ok_or(Error::RepoIdNotFound(id))
    }

    /// A project's repo by name. An unknown name errors listing the project's
    /// repos, so a filing agent gets a correction rather than a wrong checkout.
    pub fn repo_by_name(&self, project_id: i64, name: &str) -> Result<Repo> {
        let repos = self.repos(project_id)?;
        repos
            .into_iter()
            .find(|r| r.name == name)
            .ok_or_else(|| Error::RepoNotFound {
                project: self.project(project_id).map(|p| p.name).unwrap_or_default(),
                name: name.to_string(),
                known: self.repo_names(project_id),
            })
    }

    pub fn default_repo(&self, project_id: i64) -> Result<Repo> {
        self.conn
            .query_row(
                &format!(
                    "SELECT {REPO_COLUMNS} FROM repos WHERE project_id = ?1 AND is_default = 1"
                ),
                [project_id],
                repo_from_row,
            )
            .optional()?
            .ok_or_else(|| match self.project(project_id) {
                // A project always has a default repo — `create_project` makes
                // it in the same transaction — so the only way here is an id
                // that names no project at all.
                Err(e) => e,
                Ok(_) => Error::Invalid(format!("project {project_id} has no default repo")),
            })
    }

    /// The checkout a task's work runs in (DESIGN.md §8): its own repo when it
    /// names one, else its project's default. The single resolution point —
    /// dispatch, `pr`/`open`, worktree cleanup, and `import` all come here
    /// rather than reading `repo_id` themselves.
    pub fn repo_for_task(&self, task: &Task) -> Result<Repo> {
        match task.repo_id {
            Some(id) => self.repo(id),
            None => self.default_repo(task.project_id),
        }
    }

    /// Add a repo to a project. The first repo of a project is made by
    /// `create_project`, so one added here is never the default; `set_default`
    /// promotes it.
    pub fn add_repo(&mut self, project_id: i64, name: &str, path: &str) -> Result<Repo> {
        self.project(project_id)?;
        if name.trim().is_empty() {
            return Err(Error::Invalid("a repo name is required".into()));
        }
        if self.repos(project_id)?.iter().any(|r| r.name == name) {
            return Err(Error::Invalid(format!(
                "project already has a repo named '{name}'"
            )));
        }
        self.conn.execute(
            "INSERT INTO repos (project_id, name, path, is_default) VALUES (?1, ?2, ?3, 0)",
            params![project_id, name, path],
        )?;
        self.repo(self.conn.last_insert_rowid())
    }

    pub fn set_repo_path(&mut self, repo_id: i64, path: &str) -> Result<Repo> {
        let changed = self.conn.execute(
            "UPDATE repos SET path = ?1 WHERE id = ?2",
            params![path, repo_id],
        )?;
        if changed == 0 {
            return Err(Error::RepoIdNotFound(repo_id));
        }
        self.repo(repo_id)
    }

    /// Make a repo its project's default. Clearing the old default and setting
    /// the new one share a transaction, because the partial unique index would
    /// otherwise refuse the intermediate state.
    pub fn set_default_repo(&mut self, repo_id: i64) -> Result<Repo> {
        let repo = self.repo(repo_id)?;
        let tx = self.conn.transaction()?;
        tx.execute(
            "UPDATE repos SET is_default = 0 WHERE project_id = ?1",
            [repo.project_id],
        )?;
        tx.execute("UPDATE repos SET is_default = 1 WHERE id = ?1", [repo_id])?;
        tx.commit()?;
        self.repo(repo_id)
    }

    /// Remove a repo, refusing the three ways it would leave the store
    /// inconsistent: the project's last repo (a project always has a
    /// checkout), its default while others remain (set a new one first), and
    /// one any task still names (re-point those tasks first).
    pub fn delete_repo(&mut self, repo_id: i64) -> Result<()> {
        let repo = self.repo(repo_id)?;
        let project = self.project(repo.project_id)?;
        if self.repos(repo.project_id)?.len() == 1 {
            return Err(Error::LastRepo {
                project: project.name,
                name: repo.name,
            });
        }
        if repo.is_default {
            return Err(Error::DefaultRepo {
                project: project.name,
                name: repo.name,
            });
        }
        let used: i64 = self.conn.query_row(
            "SELECT COUNT(*) FROM tasks WHERE repo_id = ?1",
            [repo_id],
            |r| r.get(0),
        )?;
        if used > 0 {
            return Err(Error::RepoInUse {
                name: repo.name,
                count: used,
            });
        }
        self.conn
            .execute("DELETE FROM repos WHERE id = ?1", [repo_id])?;
        Ok(())
    }

    /// Re-point a task at a repo of its own project, or back at the default
    /// with `None`. A repo belonging to another project is refused — a task's
    /// checkout is chosen from its project's repos, not from every repo.
    pub fn set_task_repo(&mut self, task_id: i64, repo_id: Option<i64>) -> Result<Task> {
        let task = self.task(task_id)?;
        if let Some(id) = repo_id {
            let repo = self.repo(id)?;
            if repo.project_id != task.project_id {
                return Err(Error::Invalid(format!(
                    "repo '{}' belongs to another project",
                    repo.name
                )));
            }
        }
        self.conn.execute(
            "UPDATE tasks SET repo_id = ?1 WHERE id = ?2",
            params![repo_id, task_id],
        )?;
        self.task(task_id)
    }

    fn repo_names(&self, project_id: i64) -> String {
        self.repos(project_id)
            .map(|repos| {
                repos
                    .into_iter()
                    .map(|r| r.name)
                    .collect::<Vec<_>>()
                    .join(", ")
            })
            .unwrap_or_default()
    }

    // --- docs ---
    //
    // A document is a plan a project's work derives from (DESIGN.md §3/§5).
    // It is *owned* by one project, which is where a relative location resolves
    // and where `doc list` shows it, but the task edge is deliberately not
    // constrained to that project: one strategy doc routinely spawns work
    // across several, and refusing the cross-project link would defeat the
    // "which tasks came from this plan?" query the table exists for.

    /// Register a document against a project. `location` is a checkout-relative
    /// path, an absolute path, or a URL; an absolute path that lies inside one
    /// of the project's checkouts is stored relative to it (DESIGN.md §5), so
    /// the link survives the checkout moving. `repo` names which checkout a
    /// relative path resolves against, `None` meaning the project's default.
    pub fn create_doc(
        &mut self,
        project_id: i64,
        repo_id: Option<i64>,
        location: &str,
        title: Option<&str>,
    ) -> Result<Doc> {
        self.project(project_id)?;
        let (location, repo_id) = self.normalise_location(project_id, repo_id, location)?;
        if self
            .docs(project_id)?
            .iter()
            .any(|d| d.location == location)
        {
            return Err(Error::Invalid(format!(
                "this project already has a document at '{location}'"
            )));
        }
        self.conn.execute(
            "INSERT INTO docs (project_id, repo_id, title, location, created_at)
             VALUES (?1, ?2, ?3, ?4, datetime('now'))",
            params![project_id, repo_id, title, location],
        )?;
        let doc = self.doc(self.conn.last_insert_rowid())?;
        log_global_event(&self.conn, "doc-added", Some(&doc.location))?;
        Ok(doc)
    }

    /// Reduce an operator-supplied location to what is stored: a URL verbatim
    /// (and never against a repo, since it resolves unaided), an absolute path
    /// relativised against the checkout that contains it, and anything else
    /// left as given. An explicit `repo_id` pins which checkout is meant, and
    /// an absolute path outside it is refused rather than silently stored whole.
    fn normalise_location(
        &self,
        project_id: i64,
        repo_id: Option<i64>,
        location: &str,
    ) -> Result<(String, Option<i64>)> {
        let location = location.trim();
        if location.is_empty() {
            return Err(Error::Invalid("a document path or URL is required".into()));
        }
        if let Some(id) = repo_id {
            let repo = self.repo(id)?;
            if repo.project_id != project_id {
                return Err(Error::Invalid(format!(
                    "repo '{}' belongs to another project",
                    repo.name
                )));
            }
        }
        if location_is_url(location) {
            if repo_id.is_some() {
                return Err(Error::Invalid(
                    "a URL resolves on its own — drop --repo, which only picks the checkout a \
                     relative path is read from"
                        .into(),
                ));
            }
            return Ok((location.to_string(), None));
        }
        if !Path::new(location).is_absolute() {
            return Ok((location.to_string(), repo_id));
        }
        // An absolute path: prefer the checkout that contains it, so the stored
        // location survives that checkout moving. The longest matching path
        // wins, for the case of a repo nested inside another.
        let mut repos = match repo_id {
            Some(id) => vec![self.repo(id)?],
            None => self.repos(project_id)?,
        };
        repos.sort_by_key(|r| std::cmp::Reverse(r.path.len()));
        for repo in &repos {
            if let Ok(rel) = Path::new(location).strip_prefix(&repo.path) {
                return Ok((rel.to_string_lossy().into_owned(), Some(repo.id)));
            }
        }
        match repo_id {
            // An explicit --repo said which checkout to read this path from, so
            // a path outside it is a mistake worth hearing rather than storing.
            Some(id) => Err(Error::Invalid(format!(
                "'{location}' is not inside repo '{}' ({})",
                self.repo(id)?.name,
                self.repo(id)?.path
            ))),
            // Outside every checkout: a legitimate external document, kept
            // absolute and resolving against no repo.
            None => Ok((location.to_string(), None)),
        }
    }

    pub fn doc(&self, id: i64) -> Result<Doc> {
        self.conn
            .query_row(
                &format!("SELECT {DOC_COLUMNS} FROM docs WHERE id = ?1"),
                [id],
                doc_from_row,
            )
            .optional()?
            .ok_or(Error::DocNotFound(id))
    }

    /// A project's documents, oldest first — registration order is the closest
    /// thing a plan library has to a meaningful one.
    pub fn docs(&self, project_id: i64) -> Result<Vec<Doc>> {
        let mut stmt = self.conn.prepare(&format!(
            "SELECT {DOC_COLUMNS} FROM docs WHERE project_id = ?1 ORDER BY id"
        ))?;
        let rows = stmt.query_map([project_id], doc_from_row)?;
        Ok(rows.collect::<rusqlite::Result<_>>()?)
    }

    pub fn all_docs(&self) -> Result<Vec<Doc>> {
        let mut stmt = self
            .conn
            .prepare(&format!("SELECT {DOC_COLUMNS} FROM docs ORDER BY id"))?;
        let rows = stmt.query_map([], doc_from_row)?;
        Ok(rows.collect::<rusqlite::Result<_>>()?)
    }

    /// Every document with the given location, across projects — what a `--doc`
    /// flag naming a path rather than an id matches. More than one match is
    /// returned rather than resolved, so the caller can say which ids collided.
    pub fn docs_at(&self, location: &str) -> Result<Vec<Doc>> {
        let location = location.trim();
        Ok(self
            .all_docs()?
            .into_iter()
            .filter(|d| d.location == location)
            .collect())
    }

    /// Where a document actually is: a URL or absolute path verbatim, and a
    /// relative one joined onto its checkout. The single resolution point —
    /// dispatch and every renderer come here rather than joining paths itself.
    pub fn resolve_doc(&self, doc: &Doc) -> Result<String> {
        if doc.is_url() || Path::new(&doc.location).is_absolute() {
            return Ok(doc.location.clone());
        }
        let repo = match doc.repo_id {
            Some(id) => self.repo(id)?,
            None => self.default_repo(doc.project_id)?,
        };
        Ok(Path::new(&repo.path)
            .join(&doc.location)
            .to_string_lossy()
            .into_owned())
    }

    /// Remove a document and every task link to it, in one transaction. Unlike
    /// a repo, a doc is navigational — nothing resolves to nothing when it goes
    /// — so this unlinks rather than refusing, and returns the tasks it freed
    /// so the caller can say how far the removal reached.
    pub fn delete_doc(&mut self, doc_id: i64) -> Result<Vec<i64>> {
        let doc = self.doc(doc_id)?;
        let linked = self.tasks_for_doc(doc_id)?;
        let tx = self.conn.transaction()?;
        for task in &linked {
            log_event(&tx, task.id, "doc-unlinked", Some(doc.label()))?;
        }
        tx.execute("DELETE FROM task_docs WHERE doc_id = ?1", [doc_id])?;
        tx.execute("DELETE FROM docs WHERE id = ?1", [doc_id])?;
        log_global_event(&tx, "doc-removed", Some(&doc.location))?;
        tx.commit()?;
        Ok(linked.into_iter().map(|t| t.id).collect())
    }

    /// Link a task to a document. Returns whether the edge was new, so a
    /// repeated link reads as a no-op rather than an error — and logs the link
    /// on the task's own event trail only when something changed.
    pub fn link_doc(&mut self, task_id: i64, doc_id: i64) -> Result<bool> {
        self.task(task_id)?;
        let doc = self.doc(doc_id)?;
        let changed = self.conn.execute(
            "INSERT OR IGNORE INTO task_docs (task_id, doc_id) VALUES (?1, ?2)",
            params![task_id, doc_id],
        )?;
        if changed > 0 {
            log_event(&self.conn, task_id, "doc-linked", Some(doc.label()))?;
        }
        Ok(changed > 0)
    }

    pub fn unlink_doc(&mut self, task_id: i64, doc_id: i64) -> Result<bool> {
        self.task(task_id)?;
        let doc = self.doc(doc_id)?;
        let changed = self.conn.execute(
            "DELETE FROM task_docs WHERE task_id = ?1 AND doc_id = ?2",
            params![task_id, doc_id],
        )?;
        if changed > 0 {
            log_event(&self.conn, task_id, "doc-unlinked", Some(doc.label()))?;
        }
        Ok(changed > 0)
    }

    /// Replace a task's whole document list — what `set --doc` writes, matching
    /// `--blocked-by`'s replace semantics so the flag can remove a link as well
    /// as add one. Each added and dropped edge is logged individually.
    pub fn set_task_docs(&mut self, task_id: i64, doc_ids: &[i64]) -> Result<Vec<Doc>> {
        self.task(task_id)?;
        let wanted: Vec<Doc> = doc_ids
            .iter()
            .map(|id| self.doc(*id))
            .collect::<Result<_>>()?;
        let current = self.docs_for_task(task_id)?;
        let tx = self.conn.transaction()?;
        for doc in &current {
            if !wanted.iter().any(|d| d.id == doc.id) {
                tx.execute(
                    "DELETE FROM task_docs WHERE task_id = ?1 AND doc_id = ?2",
                    params![task_id, doc.id],
                )?;
                log_event(&tx, task_id, "doc-unlinked", Some(doc.label()))?;
            }
        }
        for doc in &wanted {
            if !current.iter().any(|d| d.id == doc.id) {
                tx.execute(
                    "INSERT INTO task_docs (task_id, doc_id) VALUES (?1, ?2)",
                    params![task_id, doc.id],
                )?;
                log_event(&tx, task_id, "doc-linked", Some(doc.label()))?;
            }
        }
        tx.commit()?;
        self.docs_for_task(task_id)
    }

    /// The documents a task cites, in registration order.
    pub fn docs_for_task(&self, task_id: i64) -> Result<Vec<Doc>> {
        let mut stmt = self.conn.prepare(&format!(
            "SELECT {} FROM docs d JOIN task_docs td ON td.doc_id = d.id
             WHERE td.task_id = ?1 ORDER BY d.id",
            prefixed(DOC_COLUMNS, "d")
        ))?;
        let rows = stmt.query_map([task_id], doc_from_row)?;
        Ok(rows.collect::<rusqlite::Result<_>>()?)
    }

    /// Every document link keyed by task id, loaded whole — what the TUI reads
    /// once per refresh so the render path never queries the store, the same
    /// shape as the dependency maps.
    pub fn docs_by_task(&self) -> Result<HashMap<i64, Vec<Doc>>> {
        let mut stmt = self.conn.prepare(&format!(
            "SELECT td.task_id, {} FROM docs d JOIN task_docs td ON td.doc_id = d.id
             ORDER BY td.task_id, d.id",
            prefixed(DOC_COLUMNS, "d")
        ))?;
        let rows = stmt.query_map([], |row| {
            Ok((row.get::<_, i64>(0)?, doc_from_row_at(row, 1)?))
        })?;
        let mut map: HashMap<i64, Vec<Doc>> = HashMap::new();
        for row in rows {
            let (task_id, doc) = row?;
            map.entry(task_id).or_default().push(doc);
        }
        Ok(map)
    }

    /// The tasks derived from a document — the "which tasks came from this
    /// plan?" query, in id order so a plan's rollout reads chronologically.
    pub fn tasks_for_doc(&self, doc_id: i64) -> Result<Vec<Task>> {
        let mut stmt = self.conn.prepare(&format!(
            "SELECT {} FROM tasks t JOIN task_docs td ON td.task_id = t.id
             WHERE td.doc_id = ?1 ORDER BY t.id",
            prefixed(TASK_COLUMNS, "t")
        ))?;
        let rows = stmt.query_map([doc_id], task_from_row)?;
        Ok(rows.collect::<rusqlite::Result<_>>()?)
    }

    // --- tasks ---

    pub fn create_task(&mut self, new: NewTask) -> Result<Task> {
        if !matches!(
            new.state,
            TaskState::Proposed | TaskState::Parked | TaskState::Ready
        ) {
            return Err(Error::Invalid(format!(
                "a task cannot be created in state '{}'",
                new.state
            )));
        }
        if new.human && new.agent.is_some() {
            return Err(Error::Invalid(
                "a human-only task cannot carry an agent override — the override only \
                 selects a dispatch agent, and no agent can execute the task"
                    .into(),
            ));
        }
        if new.human && new.deep {
            return Err(Error::Invalid(
                "a human-only task cannot be deep — deep only selects a dispatch model, \
                 and no agent can execute the task"
                    .into(),
            ));
        }
        // An archived project accepts no new work through any door — `add`,
        // `propose`, and import all create through here (DESIGN.md §5).
        let project = self.project(new.project_id)?;
        if project.archived {
            return Err(Error::ProjectArchived { name: project.name });
        }
        // A task's checkout is chosen from its own project's repos; NULL means
        // the project default, which is what every task created without one gets.
        if let Some(repo_id) = new.repo_id {
            let repo = self.repo(repo_id)?;
            if repo.project_id != new.project_id {
                return Err(Error::Invalid(format!(
                    "repo '{}' belongs to another project",
                    repo.name
                )));
            }
        }
        let tx = self.conn.transaction()?;
        tx.execute(
            "INSERT INTO tasks (project_id, repo_id, title, body, priority, state, agent, human,
                                deep, state_since, created_at)
             VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, ?8, ?9, datetime('now'), datetime('now'))",
            params![
                new.project_id,
                new.repo_id,
                new.title,
                new.body,
                new.priority,
                new.state,
                new.agent,
                new.human,
                new.deep
            ],
        )?;
        let id = tx.last_insert_rowid();
        log_event(&tx, id, "created", Some(new.state.as_str()))?;
        tx.commit()?;
        self.task(id)
    }

    pub fn task(&self, id: i64) -> Result<Task> {
        get_task(&self.conn, id)?.ok_or(Error::TaskNotFound(id))
    }

    pub fn tasks(&self) -> Result<Vec<Task>> {
        let mut stmt = self
            .conn
            .prepare(&format!("SELECT {TASK_COLUMNS} FROM tasks ORDER BY id"))?;
        let rows = stmt.query_map([], task_from_row)?;
        Ok(rows.collect::<rusqlite::Result<_>>()?)
    }

    pub fn update_task(&mut self, id: i64, edit: TaskEdit) -> Result<Task> {
        let current = self.task(id)?;
        if edit.human && edit.agent.is_some() {
            return Err(Error::HumanTask {
                id,
                reason: "an agent override is meaningless on a task no agent can execute — \
                         clear one or the other"
                    .into(),
            });
        }
        if edit.human && edit.deep {
            return Err(Error::HumanTask {
                id,
                reason: "the deep flag is meaningless on a task no agent can execute — it \
                         only selects a dispatch model; clear one or the other"
                    .into(),
            });
        }
        // `needs-input`, `review`, and `stalled` are unreachable for human tasks
        // (§6), so a task sitting in one — or one an agent session is still open
        // on — is demonstrably agent-executed and cannot be flagged human.
        if edit.human && !current.human {
            if matches!(
                current.state,
                TaskState::NeedsInput | TaskState::Review | TaskState::Stalled
            ) {
                return Err(Error::HumanTask {
                    id,
                    reason: format!(
                        "a task in state '{}' was executed by an agent; resolve it first",
                        current.state
                    ),
                });
            }
            let open_sessions: i64 = self.conn.query_row(
                "SELECT COUNT(*) FROM sessions WHERE task_id = ?1 AND ended_at IS NULL",
                [id],
                |r| r.get(0),
            )?;
            if open_sessions > 0 {
                return Err(Error::HumanTask {
                    id,
                    reason: "an agent session is still open on it; complete or abort it first"
                        .into(),
                });
            }
        }
        let tx = self.conn.transaction()?;
        tx.execute(
            "UPDATE tasks SET title = ?1, body = ?2, priority = ?3, agent = ?4, human = ?5,
                              deep = ?6
             WHERE id = ?7",
            params![
                edit.title,
                edit.body,
                edit.priority,
                edit.agent,
                edit.human,
                edit.deep,
                id
            ],
        )?;
        // A body edit overwrites the task's whole brief in place, so the log
        // keeps the text it replaced (DESIGN.md §8) — the append-only audit
        // covering the one field whose loss cannot be reconstructed from state.
        // Only a real change is logged, since every `set` lands here.
        if edit.body != current.body && !current.body.is_empty() {
            log_event(&tx, id, "body", Some(&current.body))?;
        }
        tx.commit()?;
        self.task(id)
    }

    /// Re-prioritise a task in isolation (DESIGN.md §7). Unlike `update_task`
    /// this touches only `priority`, and it logs the change. Task state is left
    /// untouched.
    pub fn set_priority(&mut self, id: i64, priority: Priority) -> Result<Task> {
        let changed = self.conn.execute(
            "UPDATE tasks SET priority = ?1 WHERE id = ?2",
            params![priority, id],
        )?;
        if changed == 0 {
            return Err(Error::TaskNotFound(id));
        }
        log_event(&self.conn, id, "priority", Some(&priority.to_string()))?;
        self.task(id)
    }

    /// Flag a task as warranting the agent's strongest model, or clear the flag
    /// (DESIGN.md §8). Like [`set_priority`] this touches one field and logs the
    /// change, so the TUI's toggle needs no full edit; task state is untouched.
    /// Refused on a human task, which is never dispatched and so has no model.
    ///
    /// [`set_priority`]: Store::set_priority
    pub fn set_deep(&mut self, id: i64, deep: bool) -> Result<Task> {
        let task = self.task(id)?;
        if deep && task.human {
            return Err(Error::HumanTask {
                id,
                reason: "the deep flag only selects a dispatch model, and no agent can \
                         execute the task"
                    .into(),
            });
        }
        self.conn.execute(
            "UPDATE tasks SET deep = ?1 WHERE id = ?2",
            params![deep, id],
        )?;
        log_event(
            &self.conn,
            id,
            "deep",
            Some(if deep { "set" } else { "cleared" }),
        )?;
        self.task(id)
    }

    /// Track (or, with `None`, untrack) a GitHub PR on a task (DESIGN.md §11c).
    /// The URL is stored verbatim — validation is the caller's job — and the
    /// change is logged. Leaves task state untouched.
    pub fn set_pr(&mut self, id: i64, pr_url: Option<&str>) -> Result<Task> {
        let changed = self.conn.execute(
            "UPDATE tasks SET pr_url = ?1 WHERE id = ?2",
            params![pr_url, id],
        )?;
        if changed == 0 {
            return Err(Error::TaskNotFound(id));
        }
        log_event(&self.conn, id, "pr", pr_url.or(Some("cleared")))?;
        self.task(id)
    }

    /// Record (or, with `None`, clear) the git branch a task's work lives on —
    /// the intended name a human sets for dispatch to inject, or the name an
    /// agent reports through `voro done --branch`. Stored verbatim (Voro never
    /// runs git) and logged; task state is left untouched.
    pub fn set_branch(&mut self, id: i64, branch: Option<&str>) -> Result<Task> {
        let changed = self.conn.execute(
            "UPDATE tasks SET branch = ?1 WHERE id = ?2",
            params![branch, id],
        )?;
        if changed == 0 {
            return Err(Error::TaskNotFound(id));
        }
        log_event(&self.conn, id, "branch", branch.or(Some("cleared")))?;
        self.task(id)
    }

    /// Set or replace a task's completion summary outside `done` (DESIGN.md §8):
    /// append a `summary` event, which [`latest_summary`] supersedes with, so the
    /// PR body, detail view, and incomplete-report flag all pick up the newest.
    /// This amends a stale PR body or supplies a missing `[incomplete report]`
    /// summary without a `reject` → re-`done` round trip. Allowed only on a
    /// `running` or `review` task; it never touches `tasks.state`.
    ///
    /// [`latest_summary`]: Store::latest_summary
    pub fn set_summary(&mut self, id: i64, summary: &str) -> Result<Task> {
        if summary.trim().is_empty() {
            return Err(Error::Invalid("a summary is required".into()));
        }
        let task = self.task(id)?;
        if !matches!(task.state, TaskState::Running | TaskState::Review) {
            return Err(Error::Invalid(format!(
                "a summary can only be set on a running or review task; task {} is {}",
                id, task.state
            )));
        }
        log_event(&self.conn, id, "summary", Some(summary.trim()))?;
        self.task(id)
    }

    /// How the newest concluded refine round on a task ended (DESIGN.md §6),
    /// read off the `refine` event the `refining → proposed` transition logs.
    /// `None` for a task no round has ever concluded on. This is what the two
    /// row markers are derived from, so a proposal says which of "reworked" and
    /// "the rewrite died" it is rather than leaving the operator to notice an
    /// absence.
    pub fn latest_refine_outcome(&self, task_id: i64) -> Result<Option<RefineOutcome>> {
        let detail: Option<String> = self
            .conn
            .query_row(
                "SELECT detail FROM events WHERE task_id = ?1 AND kind = 'refine'
                 ORDER BY id DESC LIMIT 1",
                [task_id],
                |r| r.get::<_, Option<String>>(0),
            )
            .optional()?
            .flatten();
        detail.map(|d| RefineOutcome::parse(&d)).transpose()
    }

    /// Whether `task_id` is a `proposed` task whose last refine round rewrote
    /// its body (DESIGN.md §6) — what renders the `↻ refined` marker. Gated on
    /// `proposed`, so triaging the task clears it, and on the *concluded* round,
    /// so a proposal is only marked once the improved body exists. Derived
    /// fresh, never stored.
    pub fn refined_flag(&self, task_id: i64) -> Result<bool> {
        self.refine_marker(task_id, RefineOutcome::Applied)
    }

    /// The other half of [`refined_flag`]: a `proposed` task whose last refine
    /// round died without applying anything, which renders the `⚠ refine
    /// failed` marker. Same lifecycle — shown while `proposed`, cleared by
    /// triage — because a failed refine must be visibly different from a
    /// proposal nobody has refined.
    ///
    /// [`refined_flag`]: Store::refined_flag
    pub fn refine_failed_flag(&self, task_id: i64) -> Result<bool> {
        self.refine_marker(task_id, RefineOutcome::Failed)
    }

    /// Correct the outcome of a round that concluded `failed` but whose rewrite
    /// then arrived anyway (DESIGN.md §6): a body replacement landing on a
    /// `proposed` task whose last round failed says that round did apply
    /// something, however late, so the recorded outcome is superseded by
    /// `applied` and the row's marker flips from `⚠ refine failed` to `↻
    /// refined`. A rewritten body sitting under a failure marker is worse than
    /// no marker at all: it teaches the operator to disbelieve the one signal
    /// that exists to say a rewrite they asked for silently never happened.
    ///
    /// This corrects a *concluded* round and nothing else — the task neither
    /// re-enters `refining` nor transitions, and the round's session keeps the
    /// outcome the reconciler observed of its process. Returns whether anything
    /// was corrected, so a caller can say so; a no-op on any other state or any
    /// other last outcome, and idempotent, since the correction it appends is
    /// itself the newest outcome.
    pub fn correct_late_refine(&mut self, task_id: i64) -> Result<bool> {
        if !self.refine_failed_flag(task_id)? {
            return Ok(false);
        }
        log_event(
            &self.conn,
            task_id,
            "refine",
            Some(RefineOutcome::Applied.as_str()),
        )?;
        Ok(true)
    }

    fn refine_marker(&self, task_id: i64, wanted: RefineOutcome) -> Result<bool> {
        let state: Option<TaskState> = self
            .conn
            .query_row("SELECT state FROM tasks WHERE id = ?1", [task_id], |r| {
                r.get(0)
            })
            .optional()?;
        if state != Some(TaskState::Proposed) {
            return Ok(false);
        }
        Ok(self.latest_refine_outcome(task_id)? == Some(wanted))
    }

    /// The newest refine note recorded on a task — the note that rode the
    /// `proposed → refining` transition — for the seed context a refine agent is
    /// launched with and for the detail views.
    pub fn latest_refine_note(&self, task_id: i64) -> Result<Option<String>> {
        Ok(self
            .conn
            .query_row(
                "SELECT detail FROM events WHERE task_id = ?1 AND kind = 'refined'
                 ORDER BY id DESC LIMIT 1",
                [task_id],
                |r| r.get::<_, Option<String>>(0),
            )
            .optional()?
            .flatten())
    }

    // --- deps ---

    /// The task a proposal was discovered from (the `discovered-from` edge of
    /// §5), if any — the context a sloppy proposal is usually missing, which is
    /// what a refine session is seeded with. The newest edge wins if a task
    /// somehow carries several.
    pub fn discovered_from(&self, task_id: i64) -> Result<Option<Task>> {
        let parent: Option<i64> = self
            .conn
            .query_row(
                "SELECT depends_on FROM deps
                 WHERE task_id = ?1 AND kind = 'discovered-from'
                 ORDER BY depends_on DESC LIMIT 1",
                [task_id],
                |r| r.get(0),
            )
            .optional()?;
        parent.map(|id| self.task(id)).transpose()
    }

    pub fn add_dep(&mut self, task_id: i64, depends_on: i64, kind: DepKind) -> Result<()> {
        if kind != DepKind::Blocks && task_id == depends_on {
            return Err(Error::Invalid("a task cannot depend on itself".into()));
        }
        let tx = self.conn.transaction()?;
        if kind == DepKind::Blocks {
            crate::transition::reject_blocks_cycle(&tx, task_id, depends_on)?;
        }
        let inserted = tx.execute(
            "INSERT INTO deps (task_id, depends_on, kind) VALUES (?1, ?2, ?3)
             ON CONFLICT (task_id, depends_on, kind) DO NOTHING",
            params![task_id, depends_on, kind],
        )?;
        if inserted == 0 {
            return Err(Error::Invalid(format!(
                "#{task_id} already has a {kind} dependency on #{depends_on}"
            )));
        }
        if kind == DepKind::Blocks {
            crate::transition::reconcile_readiness(&tx, task_id)?;
        }
        tx.commit()?;
        Ok(())
    }

    /// Drop one edge. The kind is part of the identity of an edge — a pair may
    /// carry several — so removing a blocker must not take the
    /// `discovered-from` edge beside it with it.
    pub fn remove_dep(&mut self, task_id: i64, depends_on: i64, kind: DepKind) -> Result<()> {
        let tx = self.conn.transaction()?;
        let removed = tx.execute(
            "DELETE FROM deps WHERE task_id = ?1 AND depends_on = ?2 AND kind = ?3",
            params![task_id, depends_on, kind],
        )?;
        if removed == 0 {
            return Err(Error::Invalid(format!(
                "#{task_id} has no {kind} dependency on #{depends_on}"
            )));
        }
        if kind == DepKind::Blocks {
            crate::transition::reconcile_readiness(&tx, task_id)?;
        }
        tx.commit()?;
        Ok(())
    }

    /// Every dependency edge of every kind, keyed by the depending task and
    /// resolved to the dependency's current title and state — the forward
    /// direction a detail view renders as `blocked by #N`. One query feeds every
    /// pane, so the render path never issues a per-row lookup.
    pub fn deps_by_task(&self) -> Result<HashMap<i64, Vec<DepRef>>> {
        self.dep_refs(
            "SELECT d.task_id, t.id, t.title, t.state, d.kind
             FROM deps d JOIN tasks t ON t.id = d.depends_on
             ORDER BY d.task_id, t.id, d.kind",
        )
    }

    /// The reverse edges: every dependency keyed by the task depended *on*,
    /// resolved to the depending task — who a task blocks (or spawned).
    pub fn dependents_by_task(&self) -> Result<HashMap<i64, Vec<DepRef>>> {
        self.dep_refs(
            "SELECT d.depends_on, t.id, t.title, t.state, d.kind
             FROM deps d JOIN tasks t ON t.id = d.task_id
             ORDER BY d.depends_on, t.id, d.kind",
        )
    }

    fn dep_refs(&self, sql: &str) -> Result<HashMap<i64, Vec<DepRef>>> {
        let mut stmt = self.conn.prepare(sql)?;
        let rows = stmt.query_map([], |row| {
            let key: i64 = row.get(0)?;
            let dep = DepRef {
                id: row.get(1)?,
                title: row.get(2)?,
                state: row.get(3)?,
                kind: row.get(4)?,
            };
            Ok((key, dep))
        })?;
        let mut map: HashMap<i64, Vec<DepRef>> = HashMap::new();
        for row in rows {
            let (key, dep) = row?;
            map.entry(key).or_default().push(dep);
        }
        Ok(map)
    }

    pub fn deps_of(&self, task_id: i64) -> Result<Vec<Dep>> {
        let mut stmt = self.conn.prepare(
            "SELECT task_id, depends_on, kind FROM deps WHERE task_id = ?1
             ORDER BY depends_on, kind",
        )?;
        let rows = stmt.query_map([task_id], |row| {
            Ok(Dep {
                task_id: row.get(0)?,
                depends_on: row.get(1)?,
                kind: row.get(2)?,
            })
        })?;
        Ok(rows.collect::<rusqlite::Result<_>>()?)
    }

    // --- sessions ---

    /// Open a session for a running task, stamping `started_at`. `ended_at` and
    /// `outcome` stay NULL until [`end_session`](Store::end_session).
    /// `liveness_source` is which source reconciliation must read the session by
    /// (DESIGN.md §8), which only the caller that spawned the process knows.
    pub fn create_session(
        &mut self,
        task_id: i64,
        agent: &str,
        pid: Option<i64>,
        liveness_source: LivenessSource,
        log_path: Option<&str>,
    ) -> Result<Session> {
        let id = insert_session(&self.conn, task_id, agent, pid, liveness_source, log_path)?;
        self.session(id)
    }

    /// Record the agent's own reference for a session, captured
    /// after launch — the row necessarily exists before the reference does,
    /// so this is an update rather than a `create_session` parameter.
    pub fn set_session_ref(&mut self, id: i64, session_ref: &str) -> Result<Session> {
        let changed = self.conn.execute(
            "UPDATE sessions SET session_ref = ?1 WHERE id = ?2",
            params![session_ref, id],
        )?;
        if changed == 0 {
            return Err(Error::SessionNotFound(id));
        }
        self.session(id)
    }

    /// Record what a confirmed headless send did to a session (DESIGN.md §8):
    /// the process now carrying the turn, and — where the agent forked rather
    /// than resumed in place — the reference the conversation continues under.
    /// One statement, so a reconcile in another window never reads the new
    /// reference beside the old process or the reverse.
    pub fn record_session_send(
        &mut self,
        id: i64,
        session_ref: Option<&str>,
        pid: i64,
    ) -> Result<Session> {
        let changed = self.conn.execute(
            "UPDATE sessions SET pid = ?1, session_ref = COALESCE(?2, session_ref)
             WHERE id = ?3",
            params![pid, session_ref, id],
        )?;
        if changed == 0 {
            return Err(Error::SessionNotFound(id));
        }
        self.session(id)
    }

    /// Close a session with its outcome, stamping `ended_at`.
    pub fn end_session(&mut self, id: i64, outcome: SessionOutcome) -> Result<Session> {
        if set_session_outcome(&self.conn, id, outcome)? == 0 {
            return Err(Error::SessionNotFound(id));
        }
        self.session(id)
    }

    pub fn session(&self, id: i64) -> Result<Session> {
        self.conn
            .query_row(
                &format!("SELECT {SESSION_COLUMNS} FROM sessions WHERE id = ?1"),
                [id],
                session_from_row,
            )
            .optional()?
            .ok_or(Error::SessionNotFound(id))
    }

    pub fn sessions_for(&self, task_id: i64) -> Result<Vec<Session>> {
        let mut stmt = self.conn.prepare(&format!(
            "SELECT {SESSION_COLUMNS} FROM sessions WHERE task_id = ?1 ORDER BY id DESC"
        ))?;
        let rows = stmt.query_map([task_id], session_from_row)?;
        Ok(rows.collect::<rusqlite::Result<_>>()?)
    }

    /// Every task's newest session, keyed by task id, in one query — what the
    /// TUI loads per refresh to answer "what is/was this session doing?" without
    /// querying the store mid-draw. Session ids are monotonic, so `max(id)` is
    /// the latest.
    pub fn latest_sessions(&self) -> Result<std::collections::HashMap<i64, Session>> {
        let mut stmt = self.conn.prepare(&format!(
            "SELECT {SESSION_COLUMNS} FROM sessions s
             WHERE s.id = (SELECT max(id) FROM sessions WHERE task_id = s.task_id)"
        ))?;
        let rows = stmt.query_map([], session_from_row)?;
        rows.map(|r| r.map(|s| (s.task_id, s)))
            .collect::<rusqlite::Result<_>>()
            .map_err(Into::into)
    }

    /// Sessions that have not yet ended, newest first.
    pub fn live_sessions(&self) -> Result<Vec<Session>> {
        let mut stmt = self.conn.prepare(&format!(
            "SELECT {SESSION_COLUMNS} FROM sessions WHERE ended_at IS NULL ORDER BY id DESC"
        ))?;
        let rows = stmt.query_map([], session_from_row)?;
        Ok(rows.collect::<rusqlite::Result<_>>()?)
    }

    /// Whether `task_id` is a `review` task carrying a *half-written* completion
    /// report — a branch with no summary (DESIGN.md §8). A summary with no
    /// branch is not flagged: an investigation, triage or audit produces no code
    /// and its summary is the whole deliverable. Gated on `review`, derived
    /// fresh rather than stored.
    pub fn incomplete_report_flag(&self, task_id: i64) -> Result<bool> {
        let row: Option<(TaskState, Option<String>)> = self
            .conn
            .query_row(
                "SELECT state, branch FROM tasks WHERE id = ?1",
                [task_id],
                |r| Ok((r.get(0)?, r.get(1)?)),
            )
            .optional()?;
        let Some((state, branch)) = row else {
            return Ok(false);
        };
        if state != TaskState::Review {
            return Ok(false);
        }
        let has_branch = branch.is_some();
        let has_summary = self.latest_summary(task_id)?.is_some();
        Ok(has_branch && !has_summary)
    }

    /// Rows for the cockpit's running strip (DESIGN.md §9): every `running`,
    /// `refining`, or `waiting` task, joined with its open session if it has
    /// one. The strip filters on task *state*, so `review`/`needs-input` tasks
    /// (session still open) do not appear, while a refine in flight and a
    /// handed-off task both do — an open session does not imply executing
    /// the task (§8), and what the strip shows is work in flight that someone
    /// else owns. A hand-started task with no session shows with `session_id`/
    /// `agent` `NULL`. The one-open-session invariant (§8) bounds the join to one
    /// row per task; elapsed is computed in SQL so the TUI only formats it.
    ///
    /// A `waiting` task measures its elapsed from `state_since` rather than its
    /// session: the session opened when the agent started the work, long before
    /// the hand-off, and what the operator wants from a handed-off row is how
    /// long it has been waiting. Waiting rows sort after the rest, since they
    /// are the ones nobody is actively typing into.
    ///
    /// Archived projects leave the cockpit entirely (§5), the strip included.
    pub fn running_rows(&self) -> Result<Vec<RunningRow>> {
        let mut stmt = self.conn.prepare(
            "WITH strip AS (
                 SELECT s.id AS session_id, t.id AS task_id, t.title, t.state,
                        s.agent, t.pr_url,
                        CASE WHEN t.state = 'waiting' THEN t.state_since
                             ELSE COALESCE(s.started_at, t.state_since) END AS since
                 FROM tasks t
                 JOIN projects p ON p.id = t.project_id
                 LEFT JOIN sessions s ON s.task_id = t.id AND s.ended_at IS NULL
                 WHERE t.state IN ('running','refining','waiting') AND p.archived = 0
             )
             SELECT session_id, task_id, title, state, agent, pr_url, since,
                    CAST(strftime('%s', 'now') - strftime('%s', since) AS INTEGER)
             FROM strip
             ORDER BY (state = 'waiting'), (session_id IS NULL),
                      session_id DESC, task_id DESC",
        )?;
        let rows = stmt.query_map([], |row| {
            Ok(RunningRow {
                session_id: row.get(0)?,
                task_id: row.get(1)?,
                task_title: row.get(2)?,
                task_state: row.get(3)?,
                agent: row.get(4)?,
                pr_url: row.get(5)?,
                started_at: row.get(6)?,
                elapsed_secs: row.get(7)?,
            })
        })?;
        Ok(rows.collect::<rusqlite::Result<_>>()?)
    }

    // --- events ---

    /// The most recent completion summary a task recorded (DESIGN.md §8): the
    /// detail of its newest `summary` event, logged by `done --summary` or
    /// amended by `set --summary` ([`set_summary`]). This is the PR body when
    /// `pr` opens a pull request. `None` when the task never carried a summary.
    ///
    /// [`set_summary`]: Store::set_summary
    pub fn latest_summary(&self, task_id: i64) -> Result<Option<String>> {
        Ok(self
            .conn
            .query_row(
                "SELECT detail FROM events WHERE task_id = ?1 AND kind = 'summary'
                 ORDER BY id DESC LIMIT 1",
                [task_id],
                |r| r.get::<_, Option<String>>(0),
            )
            .optional()?
            .flatten())
    }

    /// Record the branch revision the operator has just reviewed (DESIGN.md
    /// §8), so the next look at this task can be narrowed to what the rework
    /// added. Written at rejection, when the head is exactly what was judged.
    /// The `events` table carries it, so delta re-review costs no column and no
    /// migration; a later recording supersedes an earlier one the way a summary
    /// does.
    pub fn record_reviewed(&mut self, id: i64, sha: &str) -> Result<()> {
        let sha = sha.trim();
        if sha.is_empty() {
            return Err(Error::Invalid("a reviewed revision is required".into()));
        }
        // Prove the task exists before appending, so a typo'd id leaves no
        // orphan row in an append-only log.
        self.task(id)?;
        log_event(&self.conn, id, crate::review::REVIEWED_EVENT, Some(sha))
    }

    /// The revision recorded by the newest [`record_reviewed`], or `None` for a
    /// task nobody has reviewed and sent back — which is what keeps a first
    /// review showing the whole diff.
    ///
    /// [`record_reviewed`]: Store::record_reviewed
    pub fn last_reviewed(&self, task_id: i64) -> Result<Option<String>> {
        Ok(self
            .conn
            .query_row(
                "SELECT detail FROM events WHERE task_id = ?1 AND kind = ?2
                 ORDER BY id DESC LIMIT 1",
                params![task_id, crate::review::REVIEWED_EVENT],
                |r| r.get::<_, Option<String>>(0),
            )
            .optional()?
            .flatten())
    }

    pub fn events_for(&self, task_id: i64) -> Result<Vec<Event>> {
        let mut stmt = self.conn.prepare(
            "SELECT id, task_id, at, kind, detail FROM events WHERE task_id = ?1 ORDER BY id",
        )?;
        let rows = stmt.query_map([task_id], |row| {
            Ok(Event {
                id: row.get(0)?,
                task_id: row.get(1)?,
                at: row.get(2)?,
                kind: row.get(3)?,
                detail: row.get(4)?,
            })
        })?;
        Ok(rows.collect::<rusqlite::Result<_>>()?)
    }
}

pub(crate) const TASK_COLUMNS: &str = "id, project_id, title, body, priority, state, agent, \
                                       question, pr_url, branch, state_since, created_at, \
                                       closed_at, human, repo_id, deep";

pub(crate) fn get_task(conn: &Connection, id: i64) -> Result<Option<Task>> {
    Ok(conn
        .query_row(
            &format!("SELECT {TASK_COLUMNS} FROM tasks WHERE id = ?1"),
            [id],
            task_from_row,
        )
        .optional()?)
}

pub(crate) fn task_from_row(row: &rusqlite::Row<'_>) -> rusqlite::Result<Task> {
    Ok(Task {
        id: row.get(0)?,
        project_id: row.get(1)?,
        title: row.get(2)?,
        body: row.get(3)?,
        priority: row.get(4)?,
        state: row.get(5)?,
        agent: row.get(6)?,
        question: row.get(7)?,
        pr_url: row.get(8)?,
        branch: row.get(9)?,
        state_since: row.get(10)?,
        created_at: row.get(11)?,
        closed_at: row.get(12)?,
        human: row.get(13)?,
        repo_id: row.get(14)?,
        deep: row.get(15)?,
    })
}

pub(crate) const SESSION_COLUMNS: &str = "id, task_id, agent, pid, session_ref, liveness_source, log_path, started_at, ended_at, outcome";

/// A task's currently-open session, if it has one. The one-open-session
/// invariant (DESIGN.md §8) means there is at most one row to find, so this is
/// how a transaction learns *which* session it is about to close.
pub(crate) fn get_open_session(conn: &Connection, task_id: i64) -> Result<Option<Session>> {
    Ok(conn
        .query_row(
            &format!(
                "SELECT {SESSION_COLUMNS} FROM sessions
                 WHERE task_id = ?1 AND ended_at IS NULL"
            ),
            [task_id],
            session_from_row,
        )
        .optional()?)
}

pub(crate) fn get_session(conn: &Connection, id: i64) -> Result<Option<Session>> {
    Ok(conn
        .query_row(
            &format!("SELECT {SESSION_COLUMNS} FROM sessions WHERE id = ?1"),
            [id],
            session_from_row,
        )
        .optional()?)
}

fn session_from_row(row: &rusqlite::Row<'_>) -> rusqlite::Result<Session> {
    Ok(Session {
        id: row.get(0)?,
        task_id: row.get(1)?,
        agent: row.get(2)?,
        pid: row.get(3)?,
        session_ref: row.get(4)?,
        liveness_source: row.get(5)?,
        log_path: row.get(6)?,
        started_at: row.get(7)?,
        ended_at: row.get(8)?,
        outcome: row.get(9)?,
    })
}

pub(crate) const PROJECT_COLUMNS: &str = "id, name, weight, viewer, archived";

fn project_from_row(row: &rusqlite::Row<'_>) -> rusqlite::Result<Project> {
    Ok(Project {
        id: row.get(0)?,
        name: row.get(1)?,
        weight: row.get(2)?,
        viewer: row.get(3)?,
        archived: row.get(4)?,
    })
}

pub(crate) const DOC_COLUMNS: &str = "id, project_id, repo_id, title, location, created_at";

fn doc_from_row(row: &rusqlite::Row<'_>) -> rusqlite::Result<Doc> {
    doc_from_row_at(row, 0)
}

/// The same projection read from a wider row — a join that carries the task id
/// alongside the doc columns.
fn doc_from_row_at(row: &rusqlite::Row<'_>, at: usize) -> rusqlite::Result<Doc> {
    Ok(Doc {
        id: row.get(at)?,
        project_id: row.get(at + 1)?,
        repo_id: row.get(at + 2)?,
        title: row.get(at + 3)?,
        location: row.get(at + 4)?,
        created_at: row.get(at + 5)?,
    })
}

/// Qualify a column list with a table alias, so a joined query can reuse the
/// same projection constant its unjoined sibling does.
fn prefixed(columns: &str, alias: &str) -> String {
    columns
        .split(", ")
        .map(|c| format!("{alias}.{c}"))
        .collect::<Vec<_>>()
        .join(", ")
}

pub(crate) const REPO_COLUMNS: &str = "id, project_id, name, path, is_default";

fn repo_from_row(row: &rusqlite::Row<'_>) -> rusqlite::Result<Repo> {
    Ok(Repo {
        id: row.get(0)?,
        project_id: row.get(1)?,
        name: row.get(2)?,
        path: row.get(3)?,
        is_default: row.get(4)?,
    })
}

/// Insert a session row, stamping `started_at`, and return its id. Shared by
/// [`Store::create_session`] and the dispatch transaction.
/// Enforces the one-open-session invariant (DESIGN.md §8): opening a new session
/// first closes any predecessor still open (stamped `aborted`). The partial
/// unique index is the schema-level backstop.
pub(crate) fn insert_session(
    conn: &Connection,
    task_id: i64,
    agent: &str,
    pid: Option<i64>,
    liveness_source: LivenessSource,
    log_path: Option<&str>,
) -> Result<i64> {
    close_open_session(conn, task_id, SessionOutcome::Aborted)?;
    conn.execute(
        "INSERT INTO sessions (task_id, agent, pid, liveness_source, log_path, started_at)
         VALUES (?1, ?2, ?3, ?4, ?5, datetime('now'))",
        params![task_id, agent, pid, liveness_source, log_path],
    )?;
    Ok(conn.last_insert_rowid())
}

/// Close a task's currently-open session, if any, stamping `ended_at` and
/// `outcome`. The one-open-session invariant (DESIGN.md §8) means this touches
/// at most one row. Used to supersede a predecessor and to close the session on
/// a terminal transition. Returns the number of rows closed (0 if none was open).
pub(crate) fn close_open_session(
    conn: &Connection,
    task_id: i64,
    outcome: SessionOutcome,
) -> Result<usize> {
    Ok(conn.execute(
        "UPDATE sessions SET ended_at = datetime('now'), outcome = ?1
         WHERE task_id = ?2 AND ended_at IS NULL",
        params![outcome, task_id],
    )?)
}

/// Stamp `ended_at` and record `outcome` on a session, returning the number of
/// rows changed (0 if the id is unknown). Shared by [`Store::end_session`] and
/// reconciliation.
pub(crate) fn set_session_outcome(
    conn: &Connection,
    id: i64,
    outcome: SessionOutcome,
) -> Result<usize> {
    Ok(conn.execute(
        "UPDATE sessions SET ended_at = datetime('now'), outcome = ?1 WHERE id = ?2",
        params![outcome, id],
    )?)
}

pub(crate) fn log_event(
    conn: &Connection,
    task_id: i64,
    kind: &str,
    detail: Option<&str>,
) -> Result<()> {
    conn.execute(
        "INSERT INTO events (task_id, at, kind, detail) VALUES (?1, datetime('now'), ?2, ?3)",
        params![task_id, kind, detail],
    )?;
    Ok(())
}

/// An audit row for a mutation that belongs to no single task — registering or
/// removing a document. The `events.task_id` column is nullable exactly for
/// this, and the append-only log stays the record of every mutation.
pub(crate) fn log_global_event(conn: &Connection, kind: &str, detail: Option<&str>) -> Result<()> {
    conn.execute(
        "INSERT INTO events (task_id, at, kind, detail)
         VALUES (NULL, datetime('now'), ?1, ?2)",
        params![kind, detail],
    )?;
    Ok(())
}

#[cfg(test)]
mod schema_guard_tests {
    use super::*;

    /// A unique scratch directory per test, cleaned up by the caller.
    fn scratch(tag: &str) -> PathBuf {
        tempfile::Builder::new()
            .prefix(&format!("voro-store-{tag}-"))
            .tempdir()
            .unwrap()
            .keep()
    }

    #[test]
    fn a_cargo_build_directory_is_recognised_in_every_profile() {
        for exe in [
            "/home/u/proj/target/debug/voro",
            "/home/u/proj/target/release/voro",
            "/home/u/proj/target/debug/deps/voro-1a2b3c",
            "/home/u/proj/.claude/worktrees/feature/target/debug/voro",
        ] {
            assert!(
                path_is_cargo_target(Path::new(exe)),
                "{exe} should be a dev build"
            );
        }
        for exe in [
            "/home/u/.cargo/bin/voro",
            "/usr/local/bin/voro",
            "/opt/target-practice/voro",
        ] {
            assert!(
                !path_is_cargo_target(Path::new(exe)),
                "{exe} should not be a dev build"
            );
        }
    }

    #[test]
    fn a_database_from_the_future_is_refused_with_a_way_out() {
        let dir = scratch("future");
        let path = dir.join("voro.db");
        Store::open(&path).unwrap();
        Connection::open(&path)
            .unwrap()
            .pragma_update(None, "user_version", (MIGRATIONS.len() + 1) as i64)
            .unwrap();

        let message = match Store::open(&path) {
            Ok(_) => panic!("a store from the future should not open"),
            Err(e) => e.to_string(),
        };
        assert!(message.contains("schema version"), "{message}");
        // The remedy is the point of the error: a version mismatch the operator
        // cannot act on is the cryptic failure this guard exists to replace.
        // Restoring leads, since running the build that migrated it entrenches
        // an unreleased schema.
        assert!(
            message.contains("Restore a pre-migration snapshot"),
            "{message}"
        );
        std::fs::remove_dir_all(&dir).ok();
    }

    #[test]
    fn the_dev_store_is_told_to_reseed_rather_than_reinstall() {
        let remedy = remedy_for_schema_ahead(Some(&Store::dev_db_path()));
        assert!(remedy.contains("voro seed --force"), "{remedy}");
        assert!(!remedy.contains("cargo install"), "{remedy}");
    }

    #[test]
    fn a_migration_snapshots_the_database_beside_it_first() {
        let dir = scratch("snapshot");
        let path = dir.join("voro.db");
        // A store one migration short of current, so opening it migrates.
        let conn = Connection::open(&path).unwrap();
        conn.execute_batch(MIGRATIONS[0]).unwrap();
        conn.pragma_update(None, "user_version", 1i64).unwrap();
        drop(conn);

        Store::open(&path).unwrap();

        let backups: Vec<_> = std::fs::read_dir(Store::backup_dir_for(&path))
            .unwrap()
            .map(|e| e.unwrap().file_name().to_string_lossy().into_owned())
            .collect();
        assert_eq!(backups.len(), 1, "{backups:?}");
        assert!(backups[0].starts_with("voro-v1-"), "{backups:?}");

        // The snapshot is the state *before* the migration, which is the only
        // thing that makes it worth keeping.
        let saved = Connection::open(Store::backup_dir_for(&path).join(&backups[0])).unwrap();
        let version: i64 = saved
            .query_row("PRAGMA user_version", [], |r| r.get(0))
            .unwrap();
        assert_eq!(version, 1);
        std::fs::remove_dir_all(&dir).ok();
    }

    /// The case the counter cannot see: two branches each author a migration
    /// at the same index, so the database and the binary agree on the version
    /// and disagree on the schema.
    #[test]
    fn a_migration_applied_from_a_different_branch_is_refused_at_the_same_version() {
        let dir = scratch("diverged");
        let path = dir.join("voro.db");
        Store::open(&path).unwrap();
        // Rewrite the last applied migration as a rival branch's version of it,
        // leaving user_version untouched — exactly what a colliding 0017 does.
        Connection::open(&path)
            .unwrap()
            .execute(
                "UPDATE schema_migrations SET sql = ?1, applied_by = ?2 WHERE idx = ?3",
                params![
                    "ALTER TABLE projects RENAME COLUMN review_action TO viewer;",
                    "voro 0.1.0 at /home/u/.claude/worktrees/project-viewer/target/debug/voro",
                    MIGRATIONS.len() as i64
                ],
            )
            .unwrap();

        let message = match Store::open(&path) {
            Ok(_) => panic!("a divergent schema should not open"),
            Err(e) => e.to_string(),
        };
        // The counter alone would have said nothing here.
        let version: i64 = Connection::open(&path)
            .unwrap()
            .query_row("PRAGMA user_version", [], |r| r.get(0))
            .unwrap();
        assert_eq!(version, MIGRATIONS.len() as i64);
        assert!(message.contains("project-viewer"), "{message}");
        assert!(message.contains("Restore"), "{message}");
        std::fs::remove_dir_all(&dir).ok();
    }

    #[test]
    fn the_journal_records_what_was_applied_and_by_whom() {
        let dir = scratch("journal");
        let path = dir.join("voro.db");
        Store::open(&path).unwrap();

        let conn = Connection::open(&path).unwrap();
        let rows: i64 = conn
            .query_row("SELECT COUNT(*) FROM schema_migrations", [], |r| r.get(0))
            .unwrap();
        assert_eq!(rows, MIGRATIONS.len() as i64);
        let (sql, by): (String, String) = conn
            .query_row(
                "SELECT sql, applied_by FROM schema_migrations WHERE idx = 1",
                [],
                |r| Ok((r.get(0)?, r.get(1)?)),
            )
            .unwrap();
        assert_eq!(sql, MIGRATIONS[0]);
        assert!(by.starts_with("voro "), "{by}");
        std::fs::remove_dir_all(&dir).ok();
    }

    /// History from before the journal existed is recorded as unverifiable
    /// rather than invented, and must not read as a divergence.
    #[test]
    fn pre_journal_history_is_backfilled_unverifiable_and_opens_cleanly() {
        let dir = scratch("backfill");
        let path = dir.join("voro.db");
        let conn = Connection::open(&path).unwrap();
        for sql in &MIGRATIONS[..MIGRATIONS.len() - 1] {
            conn.execute_batch(sql).unwrap();
        }
        conn.pragma_update(None, "user_version", (MIGRATIONS.len() - 1) as i64)
            .unwrap();
        drop(conn);

        Store::open(&path).unwrap();
        Store::open(&path).unwrap();

        let conn = Connection::open(&path).unwrap();
        let unverifiable: i64 = conn
            .query_row(
                "SELECT COUNT(*) FROM schema_migrations WHERE sql IS NULL",
                [],
                |r| r.get(0),
            )
            .unwrap();
        assert_eq!(unverifiable, (MIGRATIONS.len() - 1) as i64);
        std::fs::remove_dir_all(&dir).ok();
    }

    #[test]
    fn a_fresh_database_is_not_snapshotted() {
        let dir = scratch("fresh");
        let path = dir.join("voro.db");
        Store::open(&path).unwrap();
        assert!(!Store::backup_dir_for(&path).exists());
        std::fs::remove_dir_all(&dir).ok();
    }

    /// A store one migration short of current, built by replaying the list —
    /// the state a release upgrade or a from-source build finds the operator's
    /// store in.
    fn store_at_previous_version(path: &Path) {
        let conn = Connection::open(path).unwrap();
        for sql in &MIGRATIONS[..MIGRATIONS.len() - 1] {
            conn.execute_batch(sql).unwrap();
        }
        conn.pragma_update(None, "user_version", (MIGRATIONS.len() - 1) as i64)
            .unwrap();
    }

    fn open_as_production(path: &Path, consent: Option<&str>) -> Result<Store> {
        Store::open_at(Connection::open(path).unwrap(), path, path, consent)
    }

    #[test]
    fn the_production_store_refuses_to_migrate_without_consent() {
        let dir = scratch("gate-refuse");
        let path = dir.join("voro.db");
        store_at_previous_version(&path);

        let message = match open_as_production(&path, None) {
            Ok(_) => panic!("a protected store with pending migrations should not open"),
            Err(e) => e.to_string(),
        };
        assert!(message.contains("pending migration"), "{message}");
        assert!(message.contains("voro migrate"), "{message}");
        // Refused means untouched: no migration applied, no snapshot taken.
        let version: i64 = Connection::open(&path)
            .unwrap()
            .query_row("PRAGMA user_version", [], |r| r.get(0))
            .unwrap();
        assert_eq!(version, (MIGRATIONS.len() - 1) as i64);
        assert!(!Store::backup_dir_for(&path).exists());
        std::fs::remove_dir_all(&dir).ok();
    }

    #[test]
    fn consent_migrates_the_production_store_and_is_journalled() {
        let dir = scratch("gate-consent");
        let path = dir.join("voro.db");
        store_at_previous_version(&path);

        open_as_production(&path, Some("via voro migrate --yes")).unwrap();

        let conn = Connection::open(&path).unwrap();
        let version: i64 = conn
            .query_row("PRAGMA user_version", [], |r| r.get(0))
            .unwrap();
        assert_eq!(version, MIGRATIONS.len() as i64);
        let by: String = conn
            .query_row(
                "SELECT applied_by FROM schema_migrations WHERE idx = ?1",
                [MIGRATIONS.len() as i64],
                |r| r.get(0),
            )
            .unwrap();
        assert!(by.contains("via voro migrate --yes"), "{by}");
        // The snapshot still precedes a consented migration.
        assert!(Store::backup_dir_for(&path).exists());
        std::fs::remove_dir_all(&dir).ok();
    }

    /// The marker, not the path, is what makes a moved or restored copy of the
    /// operator's store keep refusing (§5).
    #[test]
    fn the_protected_marker_travels_with_the_file() {
        let dir = scratch("gate-marker");
        let path = dir.join("voro.db");
        // A full open at its "production" path writes the marker.
        open_as_production(&path, None).unwrap();
        let moved = dir.join("restored-copy.db");
        std::fs::rename(&path, &moved).unwrap();
        // Winding the copy back one version makes it pending again; the gate
        // fires before any migration would re-apply, so the state is enough.
        Connection::open(&moved)
            .unwrap()
            .pragma_update(None, "user_version", (MIGRATIONS.len() - 1) as i64)
            .unwrap();

        assert!(matches!(
            Store::open(&moved),
            Err(Error::MigrationsPending { .. })
        ));
        std::fs::remove_dir_all(&dir).ok();
    }

    #[test]
    fn a_fresh_production_store_is_created_silently_and_marked() {
        let dir = scratch("gate-fresh");
        let path = dir.join("voro.db");
        let store = open_as_production(&path, None).unwrap();
        let marked: String = store
            .conn
            .query_row(
                "SELECT value FROM store_meta WHERE key = 'protected'",
                [],
                |r| r.get(0),
            )
            .unwrap();
        assert_eq!(marked, "1");
        std::fs::remove_dir_all(&dir).ok();
    }

    /// An unprotected store — scratch `--db`, the dev store — migrates on open
    /// exactly as before the gate existed.
    #[test]
    fn an_unprotected_store_still_migrates_silently() {
        let dir = scratch("gate-scratch");
        let path = dir.join("scratch.db");
        store_at_previous_version(&path);

        let store = Store::open(&path).unwrap();
        assert_eq!(store.schema_version().unwrap(), MIGRATIONS.len());
        std::fs::remove_dir_all(&dir).ok();
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::transition::{Action, Triage};

    fn new_ready(project_id: i64) -> NewTask {
        NewTask {
            project_id,
            repo_id: None,
            title: "t".into(),
            body: String::new(),
            priority: Priority::P2,
            state: TaskState::Ready,
            agent: None,
            human: false,
            deep: false,
        }
    }

    #[test]
    fn rename_project_updates_name_and_leaves_task_references_intact() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("old-name", "/tmp/old").unwrap();
        let task = s
            .create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "t".into(),
                body: String::new(),
                priority: Priority::P2,
                state: TaskState::Ready,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap();

        let renamed = s.rename_project(p.id, "new-name").unwrap();
        assert_eq!(renamed.id, p.id);
        assert_eq!(renamed.name, "new-name");

        // the task still resolves to the same project by id, under its new name
        let reloaded = s.task(task.id).unwrap();
        assert_eq!(reloaded.project_id, p.id);
        assert_eq!(s.project(reloaded.project_id).unwrap().name, "new-name");
    }

    #[test]
    fn project_viewer_defaults_to_none_and_round_trips() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("proj", "/tmp/proj").unwrap();
        assert_eq!(p.viewer, None);

        let updated = s.set_viewer(p.id, Some("zed")).unwrap();
        assert_eq!(updated.viewer.as_deref(), Some("zed"));
        assert_eq!(s.project(p.id).unwrap().viewer.as_deref(), Some("zed"));
        assert_eq!(s.projects().unwrap()[0].viewer.as_deref(), Some("zed"));

        // Naming no viewer writes NULL, so the column reads back empty
        s.set_viewer(p.id, None).unwrap();
        assert_eq!(s.project(p.id).unwrap().viewer, None);
        let raw: Option<String> = s
            .conn
            .query_row("SELECT viewer FROM projects WHERE id = ?1", [p.id], |r| {
                r.get(0)
            })
            .unwrap();
        assert_eq!(raw, None);

        // A blank name is a typo, not a way to clear the viewer
        assert!(matches!(
            s.set_viewer(p.id, Some("  ")),
            Err(Error::Invalid(_))
        ));
        assert!(matches!(
            s.set_viewer(999, Some("zed")),
            Err(Error::ProjectNotFound(999))
        ));
    }

    /// A database from before migration 0018 carries review actions in the
    /// pre-split spellings (DESIGN.md §5/§8). Opening it must keep the viewer a
    /// project named and read the three spellings that named none as none.
    #[test]
    fn migration_0018_reads_review_actions_as_viewer_names() {
        let conn = Connection::open_in_memory().unwrap();
        for sql in &MIGRATIONS[..17] {
            conn.execute_batch(sql).unwrap();
        }
        conn.pragma_update(None, "user_version", 17).unwrap();
        conn.execute(
            "INSERT INTO projects (name, review_action) VALUES
                 ('named', 'viewer:zed'),
                 ('bare-viewer', 'viewer'),
                 ('auto', 'auto'),
                 ('pinned-to-pr', 'pr'),
                 ('unset', NULL)",
            [],
        )
        .unwrap();

        let mut store = Store::from_connection(conn).unwrap();
        let viewer_of = |store: &mut Store, name: &str| {
            store
                .projects()
                .unwrap()
                .into_iter()
                .find(|p| p.name == name)
                .unwrap()
                .viewer
        };
        assert_eq!(viewer_of(&mut store, "named").as_deref(), Some("zed"));
        for named_none in ["bare-viewer", "auto", "pinned-to-pr", "unset"] {
            assert_eq!(viewer_of(&mut store, named_none), None, "{named_none}");
        }
    }

    #[test]
    fn rename_project_rejects_unknown_id() {
        let mut s = Store::open_in_memory().unwrap();
        assert!(matches!(
            s.rename_project(999, "x"),
            Err(Error::ProjectNotFound(999))
        ));
    }

    #[test]
    fn set_pr_tracks_clears_and_logs() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let t = s
            .create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "review me".into(),
                body: String::new(),
                priority: Priority::P2,
                state: TaskState::Ready,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap();
        assert!(s.task(t.id).unwrap().pr_url.is_none());

        let tracked = s
            .set_pr(t.id, Some("https://github.com/acme/widget/pull/42"))
            .unwrap();
        assert_eq!(
            tracked.pr_url.as_deref(),
            Some("https://github.com/acme/widget/pull/42")
        );
        // state is untouched by tracking a PR
        assert_eq!(tracked.state, TaskState::Ready);

        let cleared = s.set_pr(t.id, None).unwrap();
        assert!(cleared.pr_url.is_none());

        let events = s.events_for(t.id).unwrap();
        let kinds: Vec<&str> = events.iter().map(|e| e.kind.as_str()).collect();
        assert_eq!(kinds, vec!["created", "pr", "pr"]);
        assert!(matches!(s.set_pr(999, None), Err(Error::TaskNotFound(999))));
    }

    #[test]
    fn set_priority_updates_leaves_state_and_logs() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let t = s
            .create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "reprioritise me".into(),
                body: String::new(),
                priority: Priority::P2,
                state: TaskState::Ready,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap();

        let raised = s.set_priority(t.id, Priority::P0).unwrap();
        assert_eq!(raised.priority, Priority::P0);
        // priority is changed in isolation; state is untouched
        assert_eq!(raised.state, TaskState::Ready);

        let events = s.events_for(t.id).unwrap();
        let last = events.last().unwrap();
        assert_eq!(last.kind, "priority");
        assert_eq!(last.detail.as_deref(), Some("P0"));

        assert!(matches!(
            s.set_priority(999, Priority::P1),
            Err(Error::TaskNotFound(999))
        ));
    }

    #[test]
    fn set_branch_records_clears_and_logs() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let t = s
            .create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "branch me".into(),
                body: String::new(),
                priority: Priority::P2,
                state: TaskState::Ready,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap();
        assert!(s.task(t.id).unwrap().branch.is_none());

        let named = s.set_branch(t.id, Some("feat/parser")).unwrap();
        assert_eq!(named.branch.as_deref(), Some("feat/parser"));
        // recording a branch never touches task state
        assert_eq!(named.state, TaskState::Ready);

        // reporting a different branch overwrites the intended one
        let renamed = s.set_branch(t.id, Some("feat/parser-v2")).unwrap();
        assert_eq!(renamed.branch.as_deref(), Some("feat/parser-v2"));

        let cleared = s.set_branch(t.id, None).unwrap();
        assert!(cleared.branch.is_none());

        let events = s.events_for(t.id).unwrap();
        let kinds: Vec<&str> = events.iter().map(|e| e.kind.as_str()).collect();
        assert_eq!(kinds, vec!["created", "branch", "branch", "branch"]);
        assert!(matches!(
            s.set_branch(999, None),
            Err(Error::TaskNotFound(999))
        ));
    }

    // --- the human flag and the agent override are mutually exclusive (§3/§6) ---

    /// A store, a project, and a `NewTask` builder for the human-flag tests.
    fn human_fixture() -> (Store, i64) {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        (s, p.id)
    }

    fn new_with(project_id: i64, agent: Option<&str>, human: bool) -> NewTask {
        NewTask {
            project_id,
            repo_id: None,
            title: "hands-on".into(),
            body: String::new(),
            priority: Priority::P2,
            state: TaskState::Ready,
            agent: agent.map(str::to_string),
            human,
            deep: false,
        }
    }

    fn edit_of(task: &Task, agent: Option<&str>, human: bool) -> TaskEdit {
        TaskEdit {
            title: task.title.clone(),
            body: task.body.clone(),
            priority: task.priority,
            agent: agent.map(str::to_string),
            human,
            deep: false,
        }
    }

    #[test]
    fn create_task_refuses_a_human_task_with_an_agent_override() {
        let (mut s, p) = human_fixture();
        let err = s.create_task(new_with(p, Some("codex"), true)).unwrap_err();
        assert!(err.to_string().contains("agent override"), "{err}");
        assert!(s.tasks().unwrap().is_empty());

        assert!(s.create_task(new_with(p, Some("codex"), false)).is_ok());
        let human = s.create_task(new_with(p, None, true)).unwrap();
        assert!(human.human);
    }

    /// The body is the one field an edit overwrites wholesale, so the log keeps
    /// what each edit replaced (DESIGN.md §8) — and only that, since every `set`
    /// passes through here whether or not it touched the body.
    #[test]
    fn update_task_logs_the_body_it_replaced_and_nothing_else() {
        let (mut s, p) = human_fixture();
        let task = s.create_task(new_with(p, None, false)).unwrap();

        // an empty body destroys nothing on its way out
        let write = TaskEdit {
            body: "the brief".into(),
            ..edit_of(&task, None, false)
        };
        let task = s.update_task(task.id, write).unwrap();
        assert!(
            !s.events_for(task.id)
                .unwrap()
                .iter()
                .any(|e| e.kind == "body")
        );

        // an edit that leaves the body alone logs nothing either
        let retitle = TaskEdit {
            title: "renamed".into(),
            ..edit_of(&task, None, false)
        };
        let task = s.update_task(task.id, retitle).unwrap();
        assert!(
            !s.events_for(task.id)
                .unwrap()
                .iter()
                .any(|e| e.kind == "body")
        );

        let rewrite = TaskEdit {
            body: "a rewrite".into(),
            ..edit_of(&task, None, false)
        };
        let task = s.update_task(task.id, rewrite).unwrap();
        assert_eq!(task.body, "a rewrite");
        let logged: Vec<String> = s
            .events_for(task.id)
            .unwrap()
            .into_iter()
            .filter(|e| e.kind == "body")
            .map(|e| e.detail.unwrap_or_default())
            .collect();
        assert_eq!(logged, vec!["the brief".to_string()]);
    }

    #[test]
    fn update_task_guards_the_agent_human_exclusivity_both_ways() {
        let (mut s, p) = human_fixture();

        // an agent override cannot land on a human task
        let human = s.create_task(new_with(p, None, true)).unwrap();
        let err = s
            .update_task(human.id, edit_of(&human, Some("codex"), true))
            .unwrap_err();
        assert!(matches!(err, Error::HumanTask { id, .. } if id == human.id));

        // ...and the flag cannot land while an override is kept
        let agented = s.create_task(new_with(p, Some("codex"), false)).unwrap();
        let err = s
            .update_task(agented.id, edit_of(&agented, Some("codex"), true))
            .unwrap_err();
        assert!(matches!(err, Error::HumanTask { id, .. } if id == agented.id));

        // clearing the override in the same edit is the designed way through
        let flipped = s
            .update_task(agented.id, edit_of(&agented, None, true))
            .unwrap();
        assert!(flipped.human);
        assert!(flipped.agent.is_none());
    }

    #[test]
    fn update_task_refuses_flagging_human_in_agent_executed_states() {
        use crate::transition::Action;

        // needs-input, review, and stalled are unreachable for human tasks
        // (§6), so a task already sitting there cannot be flagged as one.
        for walk in [TaskState::NeedsInput, TaskState::Review, TaskState::Stalled] {
            let (mut s, p) = human_fixture();
            let t = s.create_task(new_with(p, None, false)).unwrap();
            match walk {
                TaskState::NeedsInput => {
                    s.apply(t.id, Action::Start).unwrap();
                    s.apply(t.id, Action::Ask("A or B?".into())).unwrap();
                }
                TaskState::Stalled => {
                    let (_, session) = s
                        .record_dispatch(t.id, "claude", Some(1), LivenessSource::Pid, None)
                        .unwrap();
                    s.reconcile_session(session.id, false, false).unwrap();
                }
                _ => {
                    s.apply(t.id, Action::Start).unwrap();
                    s.apply(t.id, Action::Complete(None)).unwrap();
                }
            }
            assert_eq!(s.task(t.id).unwrap().state, walk);
            let err = s.update_task(t.id, edit_of(&t, None, true)).unwrap_err();
            assert!(
                matches!(err, Error::HumanTask { id, .. } if id == t.id),
                "{walk}: {err}"
            );
            assert!(!s.task(t.id).unwrap().human);
        }
    }

    #[test]
    fn update_task_refuses_flagging_human_while_a_session_is_open() {
        use crate::transition::Action;

        let (mut s, p) = human_fixture();
        let t = s.create_task(new_with(p, None, false)).unwrap();
        s.record_dispatch(t.id, "claude", Some(1), LivenessSource::Pid, None)
            .unwrap();

        let err = s.update_task(t.id, edit_of(&t, None, true)).unwrap_err();
        assert!(matches!(err, Error::HumanTask { id, .. } if id == t.id));

        // once the session is torn down the flip is allowed again
        s.apply(t.id, Action::Abort).unwrap();
        let flipped = s.update_task(t.id, edit_of(&t, None, true)).unwrap();
        assert!(flipped.human);

        // a hand-started running task has no session and can flip freely
        let by_hand = s.create_task(new_with(p, None, false)).unwrap();
        s.apply(by_hand.id, Action::Start).unwrap();
        assert!(
            s.update_task(by_hand.id, edit_of(&by_hand, None, true))
                .unwrap()
                .human
        );
    }

    // --- the deep flag ---

    fn deep_new(project_id: i64, human: bool, deep: bool) -> NewTask {
        NewTask {
            deep,
            ..new_with(project_id, None, human)
        }
    }

    #[test]
    fn set_deep_toggles_the_flag_and_logs_it() {
        let (mut s, p) = human_fixture();
        let t = s.create_task(deep_new(p, false, false)).unwrap();
        assert!(!t.deep);

        assert!(s.set_deep(t.id, true).unwrap().deep);
        assert!(!s.set_deep(t.id, false).unwrap().deep);

        let kinds: Vec<String> = s
            .events_for(t.id)
            .unwrap()
            .into_iter()
            .map(|e| e.kind)
            .collect();
        assert_eq!(kinds, vec!["created", "deep", "deep"]);
        assert!(matches!(
            s.set_deep(999, true),
            Err(Error::TaskNotFound(999))
        ));
    }

    /// Deep only selects a dispatch model, so it is refused on a task no agent
    /// can execute — through every door that can set it.
    #[test]
    fn a_human_task_cannot_be_deep() {
        let (mut s, p) = human_fixture();

        let err = s.create_task(deep_new(p, true, true)).unwrap_err();
        assert!(err.to_string().contains("deep"), "{err}");
        assert!(s.tasks().unwrap().is_empty());

        let human = s.create_task(deep_new(p, true, false)).unwrap();
        let err = s.set_deep(human.id, true).unwrap_err();
        assert!(matches!(err, Error::HumanTask { id, .. } if id == human.id));
        assert!(!s.task(human.id).unwrap().deep);

        let edit = TaskEdit {
            deep: true,
            ..edit_of(&human, None, true)
        };
        let err = s.update_task(human.id, edit).unwrap_err();
        assert!(matches!(err, Error::HumanTask { id, .. } if id == human.id));

        // clearing the flag on a human task is always allowed
        assert!(!s.set_deep(human.id, false).unwrap().deep);
    }

    /// A database from before migration 0013 must open with every existing
    /// task on the workhorse (`deep = 0`), and the CHECK must reject junk.
    #[test]
    fn deep_defaults_off_and_is_constrained() {
        let (mut s, p) = human_fixture();
        let t = s.create_task(deep_new(p, false, false)).unwrap();
        assert!(!t.deep);
        assert!(
            s.conn
                .execute("UPDATE tasks SET deep = 2 WHERE id = ?1", [t.id])
                .is_err()
        );
    }

    /// A database from before migration 0007 must open with every existing
    /// task dispatchable (`human = 0`), and the CHECK must reject junk.
    #[test]
    fn migration_0007_defaults_existing_tasks_to_dispatchable() {
        let conn = Connection::open_in_memory().unwrap();
        for sql in &MIGRATIONS[..6] {
            conn.execute_batch(sql).unwrap();
        }
        conn.pragma_update(None, "user_version", 6).unwrap();
        conn.execute("INSERT INTO projects (name, path) VALUES ('p', '/tmp')", [])
            .unwrap();
        conn.execute(
            "INSERT INTO tasks (project_id, title, state, state_since, created_at)
             VALUES (1, 'pre-flag', 'ready', datetime('now'), datetime('now'))",
            [],
        )
        .unwrap();

        let store = Store::from_connection(conn).unwrap();
        assert!(!store.task(1).unwrap().human);

        let junk = store
            .conn
            .execute("UPDATE tasks SET human = 2 WHERE id = 1", []);
        assert!(junk.is_err(), "the CHECK must reject values outside 0/1");
    }

    #[test]
    fn set_summary_appends_a_superseding_summary_event() {
        use crate::transition::Action;

        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let t = s
            .create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "summarise me".into(),
                body: String::new(),
                priority: Priority::P2,
                state: TaskState::Ready,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap();

        // a running task may record its account before `done`
        s.apply(t.id, Action::Start).unwrap();
        let updated = s.set_summary(t.id, "  early account  ").unwrap();
        assert_eq!(updated.state, TaskState::Running);
        assert_eq!(
            s.latest_summary(t.id).unwrap().as_deref(),
            Some("early account")
        );

        // in review, a new summary supersedes the done-time one
        s.apply(t.id, Action::Complete(Some("done-time".into())))
            .unwrap();
        let updated = s.set_summary(t.id, "amended for the PR body").unwrap();
        assert_eq!(updated.state, TaskState::Review);
        assert_eq!(
            s.latest_summary(t.id).unwrap().as_deref(),
            Some("amended for the PR body")
        );

        // every account stays on the append-only log
        let events = s.events_for(t.id).unwrap();
        let summaries = events.iter().filter(|e| e.kind == "summary").count();
        assert_eq!(summaries, 3);
    }

    #[test]
    fn set_summary_clears_the_incomplete_report_flag() {
        use crate::transition::Action;

        // The SessionEnd-fallback shape: review with a branch and no summary.
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let t = s
            .create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "half a report".into(),
                body: String::new(),
                priority: Priority::P2,
                state: TaskState::Ready,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap();
        s.apply(t.id, Action::Start).unwrap();
        s.apply(t.id, Action::Complete(None)).unwrap();
        s.set_branch(t.id, Some("feat/x")).unwrap();
        assert!(s.incomplete_report_flag(t.id).unwrap());

        s.set_summary(t.id, "the missing half").unwrap();
        assert!(!s.incomplete_report_flag(t.id).unwrap());
    }

    #[test]
    fn set_summary_is_refused_outside_running_and_review() {
        use crate::transition::Action;

        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let t = s
            .create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "not yet".into(),
                body: String::new(),
                priority: Priority::P2,
                state: TaskState::Ready,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap();
        let err = s.set_summary(t.id, "too early").unwrap_err();
        assert!(err.to_string().contains("ready"), "{err}");

        s.apply(t.id, Action::Start).unwrap();
        s.apply(t.id, Action::Complete(None)).unwrap();
        s.apply(t.id, Action::Accept).unwrap();
        let err = s.set_summary(t.id, "too late").unwrap_err();
        assert!(err.to_string().contains("done"), "{err}");

        assert!(s.set_summary(t.id, "   ").is_err());
        assert!(matches!(
            s.set_summary(999, "x"),
            Err(Error::TaskNotFound(999))
        ));
    }

    // --- repos (DESIGN.md §3/§5) ---

    #[test]
    fn creating_a_project_creates_its_default_repo() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let repos = s.repos(p.id).unwrap();
        assert_eq!(repos.len(), 1);
        assert_eq!(repos[0].name, "voro");
        assert_eq!(repos[0].path, "/tmp/voro");
        assert!(repos[0].is_default);
        assert_eq!(s.default_repo(p.id).unwrap().id, repos[0].id);
    }

    #[test]
    fn added_repos_are_not_default_and_names_are_unique_per_project() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("odm", "/tmp/odm").unwrap();
        let oats = s.add_repo(p.id, "oats", "/tmp/oats").unwrap();
        assert!(!oats.is_default);
        assert!(s.add_repo(p.id, "oats", "/tmp/elsewhere").is_err());
        assert!(s.add_repo(p.id, "  ", "/tmp/blank").is_err());
        // The same repo name under a different project is fine.
        let other = s.create_project("voro", "/tmp/voro").unwrap();
        assert!(s.add_repo(other.id, "oats", "/tmp/oats").is_ok());
        // Default first, then by name.
        let names: Vec<_> = s.repos(p.id).unwrap().into_iter().map(|r| r.name).collect();
        assert_eq!(names, vec!["odm", "oats"]);
    }

    #[test]
    fn only_one_repo_is_ever_default() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("odm", "/tmp/odm").unwrap();
        let oats = s.add_repo(p.id, "oats", "/tmp/oats").unwrap();
        let promoted = s.set_default_repo(oats.id).unwrap();
        assert!(promoted.is_default);
        let defaults = s
            .repos(p.id)
            .unwrap()
            .into_iter()
            .filter(|r| r.is_default)
            .count();
        assert_eq!(defaults, 1);
        assert_eq!(s.default_repo(p.id).unwrap().name, "oats");
    }

    #[test]
    fn a_projects_last_repo_cannot_be_deleted() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("odm", "/tmp/odm").unwrap();
        let only = s.default_repo(p.id).unwrap();
        assert!(matches!(
            s.delete_repo(only.id),
            Err(Error::LastRepo { .. })
        ));
    }

    #[test]
    fn the_default_repo_cannot_be_deleted_while_others_remain() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("odm", "/tmp/odm").unwrap();
        let oats = s.add_repo(p.id, "oats", "/tmp/oats").unwrap();
        let default = s.default_repo(p.id).unwrap();
        assert!(matches!(
            s.delete_repo(default.id),
            Err(Error::DefaultRepo { .. })
        ));
        // Promoting the other one first clears the way.
        s.set_default_repo(oats.id).unwrap();
        s.delete_repo(default.id).unwrap();
        assert_eq!(s.repos(p.id).unwrap().len(), 1);
    }

    #[test]
    fn a_repo_a_task_names_cannot_be_deleted() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("odm", "/tmp/odm").unwrap();
        let oats = s.add_repo(p.id, "oats", "/tmp/oats").unwrap();
        let mut new = new_ready(p.id);
        new.repo_id = Some(oats.id);
        let t = s.create_task(new).unwrap();
        assert!(matches!(
            s.delete_repo(oats.id),
            Err(Error::RepoInUse { count: 1, .. })
        ));
        // Re-pointing the task at the default frees the repo.
        s.set_task_repo(t.id, None).unwrap();
        s.delete_repo(oats.id).unwrap();
    }

    #[test]
    fn a_task_resolves_its_own_repo_then_the_project_default() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("odm", "/tmp/odm").unwrap();
        let oats = s.add_repo(p.id, "oats", "/tmp/oats").unwrap();
        let plain = s.create_task(new_ready(p.id)).unwrap();
        assert_eq!(s.repo_for_task(&plain).unwrap().path, "/tmp/odm");

        let pointed = s.set_task_repo(plain.id, Some(oats.id)).unwrap();
        assert_eq!(pointed.repo_id, Some(oats.id));
        assert_eq!(s.repo_for_task(&pointed).unwrap().path, "/tmp/oats");

        // The fallback follows the default, not the original checkout.
        let back = s.set_task_repo(plain.id, None).unwrap();
        s.set_default_repo(oats.id).unwrap();
        assert_eq!(s.repo_for_task(&back).unwrap().path, "/tmp/oats");
    }

    #[test]
    fn a_task_cannot_name_another_projects_repo() {
        let mut s = Store::open_in_memory().unwrap();
        let odm = s.create_project("odm", "/tmp/odm").unwrap();
        let voro = s.create_project("voro", "/tmp/voro").unwrap();
        let foreign = s.default_repo(voro.id).unwrap();
        let mut new = new_ready(odm.id);
        new.repo_id = Some(foreign.id);
        assert!(s.create_task(new).is_err());

        let t = s.create_task(new_ready(odm.id)).unwrap();
        assert!(s.set_task_repo(t.id, Some(foreign.id)).is_err());
    }

    #[test]
    fn an_unknown_repo_name_errors_listing_the_projects_repos() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("odm", "/tmp/odm").unwrap();
        s.add_repo(p.id, "oats", "/tmp/oats").unwrap();
        let err = s.repo_by_name(p.id, "nope").unwrap_err().to_string();
        assert!(err.contains("odm"), "{err}");
        assert!(err.contains("oats"), "{err}");
    }

    #[test]
    fn deleting_a_project_takes_its_repos_with_it() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("odm", "/tmp/odm").unwrap();
        s.add_repo(p.id, "oats", "/tmp/oats").unwrap();
        s.delete_project(p.id).unwrap();
        assert!(s.repos(p.id).unwrap().is_empty());
    }

    #[test]
    fn set_path_updates_the_default_repo() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("proj", "/tmp/old").unwrap();
        let updated = s.set_default_repo_path(p.id, "/tmp/new").unwrap();
        assert_eq!(updated.path, "/tmp/new");
        assert!(updated.is_default);
        assert_eq!(s.default_repo(p.id).unwrap().path, "/tmp/new");
    }

    #[test]
    fn set_path_rejects_unknown_id() {
        let mut s = Store::open_in_memory().unwrap();
        assert!(matches!(
            s.set_default_repo_path(999, "/tmp"),
            Err(Error::ProjectNotFound(999))
        ));
    }

    #[test]
    fn delete_project_removes_a_taskless_project() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("empty", "/tmp/empty").unwrap();
        s.delete_project(p.id).unwrap();
        assert!(matches!(s.project(p.id), Err(Error::ProjectNotFound(_))));
        assert!(s.projects().unwrap().is_empty());
    }

    #[test]
    fn delete_project_rejects_unknown_id() {
        let mut s = Store::open_in_memory().unwrap();
        assert!(matches!(
            s.delete_project(999),
            Err(Error::ProjectNotFound(999))
        ));
    }

    /// Walk a fresh task into `state` through the transition API, mirroring
    /// the equivalent helper in `transition.rs`'s own tests.
    fn task_in_state(s: &mut Store, project_id: i64, state: TaskState) -> i64 {
        use TaskState::*;
        let create = |s: &mut Store, state| {
            s.create_task(NewTask {
                project_id,
                repo_id: None,
                title: format!("task in {state}"),
                body: String::new(),
                priority: Priority::P1,
                state,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap()
            .id
        };
        match state {
            Proposed | Parked | Ready => create(s, state),
            Refining => {
                let id = create(s, Proposed);
                s.record_refine_launch(
                    id,
                    "thin body",
                    "claude",
                    Some(1),
                    LivenessSource::Pid,
                    None,
                )
                .unwrap();
                id
            }
            Running => {
                let id = create(s, Ready);
                s.apply(id, Action::Start).unwrap();
                id
            }
            NeedsInput => {
                let id = task_in_state(s, project_id, Running);
                s.apply(id, Action::Ask("which schema?".into())).unwrap();
                id
            }
            Review => {
                let id = task_in_state(s, project_id, Running);
                s.apply(id, Action::Complete(None)).unwrap();
                id
            }
            Waiting => {
                let id = task_in_state(s, project_id, Review);
                s.apply(id, Action::HandOff).unwrap();
                id
            }
            Stalled => {
                let id = create(s, Ready);
                let (_, session) = s
                    .record_dispatch(id, "claude", Some(1), LivenessSource::Pid, None)
                    .unwrap();
                s.reconcile_session(session.id, false, false).unwrap();
                id
            }
            Done => {
                let id = task_in_state(s, project_id, Review);
                s.apply(id, Action::Accept).unwrap();
                id
            }
            Rejected => {
                let id = create(s, Proposed);
                s.apply(id, Action::Triage(Triage::Reject)).unwrap();
                id
            }
        }
    }

    #[test]
    fn delete_project_refuses_with_a_task_in_any_state() {
        for state in TaskState::ALL {
            let mut s = Store::open_in_memory().unwrap();
            let p = s.create_project("proj", "/tmp/proj").unwrap();
            task_in_state(&mut s, p.id, state);

            let err = s.delete_project(p.id).unwrap_err();
            assert!(
                matches!(err, Error::ProjectHasTasks { id, count } if id == p.id && count == 1),
                "state {state}: expected ProjectHasTasks, got {err}"
            );
            // the refusal must not have touched the project
            assert!(s.project(p.id).is_ok());
        }
    }

    // --- archiving a project (DESIGN.md §5) ---

    #[test]
    fn set_archived_round_trips_and_refuses_noops() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("retiring", "/tmp/retiring").unwrap();
        assert!(!p.archived);

        let archived = s.set_archived(p.id, true).unwrap();
        assert!(archived.archived);
        assert!(s.projects().unwrap()[0].archived);

        // a second archive is heard, not absorbed
        let err = s.set_archived(p.id, true).unwrap_err();
        assert!(err.to_string().contains("already archived"), "{err}");

        let restored = s.set_archived(p.id, false).unwrap();
        assert!(!restored.archived);
        let err = s.set_archived(p.id, false).unwrap_err();
        assert!(err.to_string().contains("not archived"), "{err}");

        assert!(matches!(
            s.set_archived(999, true),
            Err(Error::ProjectNotFound(999))
        ));
    }

    #[test]
    fn create_task_refuses_an_archived_project() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("retired", "/tmp/retired").unwrap();
        s.set_archived(p.id, true).unwrap();

        // every creation door — add, propose, import — routes through here
        for state in [TaskState::Proposed, TaskState::Parked, TaskState::Ready] {
            let err = s
                .create_task(NewTask {
                    project_id: p.id,
                    repo_id: None,
                    title: "too late".into(),
                    body: String::new(),
                    priority: Priority::P2,
                    state,
                    agent: None,
                    human: false,
                    deep: false,
                })
                .unwrap_err();
            assert!(
                matches!(&err, Error::ProjectArchived { name } if name == "retired"),
                "{state}: {err}"
            );
        }
        assert!(s.tasks().unwrap().is_empty());

        s.set_archived(p.id, false).unwrap();
        assert!(
            s.create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "welcome back".into(),
                body: String::new(),
                priority: Priority::P2,
                state: TaskState::Ready,
                agent: None,
                human: false,
                deep: false,
            })
            .is_ok()
        );
    }

    #[test]
    fn archive_freezes_task_states_and_history_and_unarchive_restores_them() {
        // Archiving transitions nothing: every task keeps its state, question,
        // and event log, so unarchiving restores the pre-archive view exactly.
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("retiring", "/tmp/retiring").unwrap();
        let tasks: Vec<i64> = TaskState::ALL
            .iter()
            .map(|state| task_in_state(&mut s, p.id, *state))
            .collect();
        let before: Vec<Task> = tasks.iter().map(|id| s.task(*id).unwrap()).collect();
        let events_before: Vec<usize> = tasks
            .iter()
            .map(|id| s.events_for(*id).unwrap().len())
            .collect();

        s.set_archived(p.id, true).unwrap();
        let frozen: Vec<Task> = tasks.iter().map(|id| s.task(*id).unwrap()).collect();
        assert_eq!(frozen, before);

        s.set_archived(p.id, false).unwrap();
        let after: Vec<Task> = tasks.iter().map(|id| s.task(*id).unwrap()).collect();
        assert_eq!(after, before);
        let events_after: Vec<usize> = tasks
            .iter()
            .map(|id| s.events_for(*id).unwrap().len())
            .collect();
        assert_eq!(events_after, events_before);
    }

    #[test]
    fn running_rows_exclude_archived_projects() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("retiring", "/tmp/retiring").unwrap();
        let id = task_in_state(&mut s, p.id, TaskState::Running);
        assert_eq!(s.running_rows().unwrap().len(), 1);

        // archiving hides the strip row; the task itself stays running
        s.set_archived(p.id, true).unwrap();
        assert!(s.running_rows().unwrap().is_empty());
        assert_eq!(s.task(id).unwrap().state, TaskState::Running);

        s.set_archived(p.id, false).unwrap();
        assert_eq!(s.running_rows().unwrap()[0].task_id, id);
    }

    /// A database from before migration 0011 must open with every existing
    /// project active (`archived = 0`), and the CHECK must reject junk.
    #[test]
    fn migration_0011_defaults_existing_projects_to_active() {
        let conn = Connection::open_in_memory().unwrap();
        for sql in &MIGRATIONS[..10] {
            conn.execute_batch(sql).unwrap();
        }
        conn.pragma_update(None, "user_version", 10).unwrap();
        conn.execute("INSERT INTO projects (name, path) VALUES ('p', '/tmp')", [])
            .unwrap();

        let store = Store::from_connection(conn).unwrap();
        assert!(!store.project(1).unwrap().archived);

        let junk = store
            .conn
            .execute("UPDATE projects SET archived = 2 WHERE id = 1", []);
        assert!(junk.is_err(), "the CHECK must reject values outside 0/1");
    }

    /// A database from before migration 0012 must convert in place: every
    /// project's old `path` reappears as its default repo, `projects.path` is
    /// gone, and existing tasks (all `repo_id` NULL) resolve to exactly the
    /// checkouts they had before (DESIGN.md §3/§5).
    #[test]
    fn migration_0012_turns_each_project_path_into_its_default_repo() {
        let conn = Connection::open_in_memory().unwrap();
        for sql in &MIGRATIONS[..11] {
            conn.execute_batch(sql).unwrap();
        }
        conn.pragma_update(None, "user_version", 11).unwrap();
        conn.execute(
            "INSERT INTO projects (name, path) VALUES ('odm', '/tmp/odm'), ('voro', '/tmp/voro')",
            [],
        )
        .unwrap();
        conn.execute(
            "INSERT INTO tasks (project_id, title, state, state_since, created_at)
             VALUES (1, 'old', 'ready', datetime('now'), datetime('now'))",
            [],
        )
        .unwrap();

        let store = Store::from_connection(conn).unwrap();
        for (id, name, path) in [(1, "odm", "/tmp/odm"), (2, "voro", "/tmp/voro")] {
            let repos = store.repos(id).unwrap();
            assert_eq!(repos.len(), 1);
            assert_eq!(repos[0].name, name);
            assert_eq!(repos[0].path, path);
            assert!(repos[0].is_default);
        }
        // The task kept its checkout without naming a repo.
        let task = store.task(1).unwrap();
        assert_eq!(task.repo_id, None);
        assert_eq!(store.repo_for_task(&task).unwrap().path, "/tmp/odm");

        // The column is gone, not merely unread.
        assert!(
            store
                .conn
                .query_row("SELECT path FROM projects WHERE id = 1", [], |r| r
                    .get::<_, String>(0))
                .is_err()
        );
        // The one-default invariant is schema-enforced from here on.
        assert!(
            store
                .conn
                .execute(
                    "INSERT INTO repos (project_id, name, path, is_default)
                     VALUES (1, 'second', '/tmp/second', 1)",
                    [],
                )
                .is_err()
        );
    }

    /// A database from before migration 0015 must open with every dependency
    /// edge intact, and accept a second edge of another kind between a pair the
    /// old primary key allowed only one edge for.
    #[test]
    fn migration_0015_widens_the_dep_key_without_losing_edges() {
        let conn = Connection::open_in_memory().unwrap();
        for sql in &MIGRATIONS[..14] {
            conn.execute_batch(sql).unwrap();
        }
        conn.pragma_update(None, "user_version", 14).unwrap();
        conn.execute("INSERT INTO projects (name) VALUES ('voro')", [])
            .unwrap();
        conn.execute(
            "INSERT INTO repos (project_id, name, path, is_default)
             VALUES (1, 'voro', '/tmp/voro', 1)",
            [],
        )
        .unwrap();
        conn.execute(
            "INSERT INTO tasks (project_id, title, state, state_since, created_at)
             VALUES (1, 'source', 'ready', datetime('now'), datetime('now')),
                    (1, 'spawned', 'ready', datetime('now'), datetime('now'))",
            [],
        )
        .unwrap();
        conn.execute(
            "INSERT INTO deps (task_id, depends_on, kind) VALUES (2, 1, 'discovered-from')",
            [],
        )
        .unwrap();

        let mut store = Store::from_connection(conn).unwrap();
        let carried = store.deps_of(2).unwrap();
        assert_eq!(carried.len(), 1);
        assert_eq!(carried[0].kind, DepKind::DiscoveredFrom);

        store.set_blocks_deps(2, &[1]).unwrap();
        let kinds: Vec<DepKind> = store.deps_of(2).unwrap().iter().map(|d| d.kind).collect();
        assert_eq!(kinds, vec![DepKind::Blocks, DepKind::DiscoveredFrom]);
    }

    /// A database created at schema version 1 (state still named 'backlog')
    /// must convert on open: rows renamed, deps/events surviving the table
    /// rebuild, version stamped.
    #[test]
    fn migration_0002_converts_backlog_rows() {
        let conn = Connection::open_in_memory().unwrap();
        conn.execute_batch(MIGRATIONS[0]).unwrap();
        conn.pragma_update(None, "user_version", 1).unwrap();
        conn.execute("INSERT INTO projects (name, path) VALUES ('p', '/tmp')", [])
            .unwrap();
        conn.execute(
            "INSERT INTO tasks (project_id, title, state, state_since, created_at)
             VALUES (1, 'blocker', 'ready', datetime('now'), datetime('now')),
                    (1, 'waiting', 'backlog', datetime('now'), datetime('now'))",
            [],
        )
        .unwrap();
        conn.execute("INSERT INTO deps (task_id, depends_on) VALUES (2, 1)", [])
            .unwrap();
        conn.execute(
            "INSERT INTO events (task_id, at, kind, detail)
             VALUES (2, datetime('now'), 'created', 'backlog')",
            [],
        )
        .unwrap();

        let store = Store::from_connection(conn).unwrap();
        assert_eq!(store.task(2).unwrap().state, TaskState::Parked);
        assert_eq!(store.task(1).unwrap().state, TaskState::Ready);
        assert_eq!(store.deps_of(2).unwrap().len(), 1);
        // the event log is history and keeps its original wording
        assert_eq!(
            store.events_for(2).unwrap()[0].detail.as_deref(),
            Some("backlog")
        );
        let version: i64 = store
            .conn
            .query_row("PRAGMA user_version", [], |r| r.get(0))
            .unwrap();
        assert_eq!(version, MIGRATIONS.len() as i64);
        // 0004 gave the sessions table its session_ref column
        let refs: i64 = store
            .conn
            .query_row("SELECT COUNT(session_ref) FROM sessions", [], |r| r.get(0))
            .unwrap();
        assert_eq!(refs, 0);
    }

    /// Migration 0006 must dedupe a task that already carries several open
    /// sessions — keeping the newest open and closing the rest — before it can
    /// create the one-open-session index, and the index must then reject any
    /// further second open row.
    #[test]
    fn migration_0006_dedupes_open_sessions_and_enforces_the_index() {
        let conn = Connection::open_in_memory().unwrap();
        // apply 0001..=0005, i.e. everything before the invariant migration
        for sql in &MIGRATIONS[..5] {
            conn.execute_batch(sql).unwrap();
        }
        conn.pragma_update(None, "user_version", 5).unwrap();
        conn.execute("INSERT INTO projects (name, path) VALUES ('p', '/tmp')", [])
            .unwrap();
        conn.execute(
            "INSERT INTO tasks (project_id, title, state, state_since, created_at)
             VALUES (1, 'run me', 'running', datetime('now'), datetime('now'))",
            [],
        )
        .unwrap();
        // three open sessions on the one task — the exact duplicate state
        for _ in 0..3 {
            conn.execute(
                "INSERT INTO sessions (task_id, agent, started_at) VALUES (1, 'a', datetime('now'))",
                [],
            )
            .unwrap();
        }

        let store = Store::from_connection(conn).unwrap();
        // only the newest open session survives; the rest are closed `aborted`
        let open: Vec<i64> = store
            .sessions_for(1)
            .unwrap()
            .into_iter()
            .filter(|s| s.ended_at.is_none())
            .map(|s| s.id)
            .collect();
        assert_eq!(open, vec![3]);
        assert_eq!(
            store.session(1).unwrap().outcome,
            Some(SessionOutcome::Aborted)
        );
        // and the index now forbids a second open row
        let second = store.conn.execute(
            "INSERT INTO sessions (task_id, agent, started_at) VALUES (1, 'b', datetime('now'))",
            [],
        );
        assert!(second.is_err());
    }

    /// Migration 0008 must backfill exactly the tasks the derived redispatch
    /// flag used to mark — `ready` with a most recent session ended
    /// `failed`/`capped` — into `stalled`, leaving every other shape alone,
    /// and must carry 0007's `human` column through the table rebuild.
    #[test]
    fn migration_0008_backfills_flagged_ready_tasks_to_stalled() {
        let conn = Connection::open_in_memory().unwrap();
        for sql in &MIGRATIONS[..7] {
            conn.execute_batch(sql).unwrap();
        }
        conn.pragma_update(None, "user_version", 7).unwrap();
        conn.execute("INSERT INTO projects (name, path) VALUES ('p', '/tmp')", [])
            .unwrap();
        // 1: ready, last session failed          -> stalled
        // 2: ready, last session capped          -> stalled
        // 3: ready, last session aborted         -> stays ready
        // 4: ready, failed session then aborted  -> stays ready (latest wins)
        // 5: ready, no sessions                  -> stays ready
        // 6: running, last session failed        -> stays running
        conn.execute(
            "INSERT INTO tasks (project_id, title, state, state_since, created_at)
             VALUES (1, 't1', 'ready', datetime('now'), datetime('now')),
                    (1, 't2', 'ready', datetime('now'), datetime('now')),
                    (1, 't3', 'ready', datetime('now'), datetime('now')),
                    (1, 't4', 'ready', datetime('now'), datetime('now')),
                    (1, 't5', 'ready', datetime('now'), datetime('now')),
                    (1, 't6', 'running', datetime('now'), datetime('now'))",
            [],
        )
        .unwrap();
        conn.execute(
            "INSERT INTO sessions (task_id, agent, started_at, ended_at, outcome)
             VALUES (1, 'a', datetime('now'), datetime('now'), 'failed'),
                    (2, 'a', datetime('now'), datetime('now'), 'capped'),
                    (3, 'a', datetime('now'), datetime('now'), 'aborted'),
                    (4, 'a', datetime('now'), datetime('now'), 'failed'),
                    (4, 'a', datetime('now'), datetime('now'), 'aborted'),
                    (6, 'a', datetime('now'), datetime('now'), 'failed')",
            [],
        )
        .unwrap();
        conn.execute("UPDATE tasks SET human = 1 WHERE id = 5", [])
            .unwrap();

        let store = Store::from_connection(conn).unwrap();
        assert_eq!(store.task(1).unwrap().state, TaskState::Stalled);
        assert_eq!(store.task(2).unwrap().state, TaskState::Stalled);
        assert_eq!(store.task(3).unwrap().state, TaskState::Ready);
        assert_eq!(store.task(4).unwrap().state, TaskState::Ready);
        assert_eq!(store.task(5).unwrap().state, TaskState::Ready);
        assert_eq!(store.task(6).unwrap().state, TaskState::Running);
        // the rebuild carries the human flag and its CHECK across
        assert!(store.task(5).unwrap().human);
        assert!(!store.task(1).unwrap().human);
        let junk = store
            .conn
            .execute("UPDATE tasks SET human = 2 WHERE id = 5", []);
        assert!(junk.is_err(), "the CHECK must reject values outside 0/1");
    }

    /// Migration 0010 must extend the state CHECK to admit 'waiting' while
    /// carrying every existing task through the table rebuild untouched.
    #[test]
    fn migration_0010_admits_waiting_and_preserves_existing_tasks() {
        let conn = Connection::open_in_memory().unwrap();
        for sql in &MIGRATIONS[..9] {
            conn.execute_batch(sql).unwrap();
        }
        conn.pragma_update(None, "user_version", 9).unwrap();
        conn.execute("INSERT INTO projects (name, path) VALUES ('p', '/tmp')", [])
            .unwrap();
        conn.execute(
            "INSERT INTO tasks (project_id, title, state, agent, pr_url, branch, human,
                                state_since, created_at)
             VALUES (1, 'in review', 'review', 'claude', 'https://x/pull/1', 'feat/x', 1,
                     datetime('now'), datetime('now'))",
            [],
        )
        .unwrap();

        let store = Store::from_connection(conn).unwrap();
        // the pre-existing row survives the rebuild with every column intact
        let task = store.task(1).unwrap();
        assert_eq!(task.state, TaskState::Review);
        assert_eq!(task.pr_url.as_deref(), Some("https://x/pull/1"));
        assert_eq!(task.branch.as_deref(), Some("feat/x"));
        assert!(task.human);

        // the widened CHECK now admits 'waiting' and still rejects junk
        assert!(
            store
                .conn
                .execute("UPDATE tasks SET state = 'waiting' WHERE id = 1", [])
                .is_ok()
        );
        assert!(
            store
                .conn
                .execute("UPDATE tasks SET state = 'bogus' WHERE id = 1", [])
                .is_err()
        );

        let version: i64 = store
            .conn
            .query_row("PRAGMA user_version", [], |r| r.get(0))
            .unwrap();
        assert_eq!(version, MIGRATIONS.len() as i64);
    }

    /// A database from before migration 0016 must open with every existing task
    /// intact, and the widened CHECK must admit `refining` while still rejecting
    /// junk (DESIGN.md §6).
    #[test]
    fn migration_0016_admits_refining_and_preserves_existing_tasks() {
        let conn = Connection::open_in_memory().unwrap();
        for sql in &MIGRATIONS[..15] {
            conn.execute_batch(sql).unwrap();
        }
        conn.pragma_update(None, "user_version", 15).unwrap();
        conn.execute("INSERT INTO projects (name) VALUES ('voro')", [])
            .unwrap();
        conn.execute(
            "INSERT INTO repos (project_id, name, path, is_default)
             VALUES (1, 'voro', '/tmp/voro', 1)",
            [],
        )
        .unwrap();
        conn.execute(
            "INSERT INTO tasks (project_id, repo_id, title, state, agent, pr_url, branch,
                                human, deep, state_since, created_at)
             VALUES (1, 1, 'in review', 'review', 'claude', 'https://x/pull/1', 'feat/x', 1, 1,
                     datetime('now'), datetime('now'))",
            [],
        )
        .unwrap();

        let store = Store::from_connection(conn).unwrap();
        // the pre-existing row survives the rebuild with every column intact
        let task = store.task(1).unwrap();
        assert_eq!(task.state, TaskState::Review);
        assert_eq!(task.pr_url.as_deref(), Some("https://x/pull/1"));
        assert_eq!(task.branch.as_deref(), Some("feat/x"));
        assert_eq!(task.repo_id, Some(1));
        assert!(task.human);
        assert!(task.deep);

        assert!(
            store
                .conn
                .execute("UPDATE tasks SET state = 'refining' WHERE id = 1", [])
                .is_ok()
        );
        assert!(
            store
                .conn
                .execute("UPDATE tasks SET state = 'bogus' WHERE id = 1", [])
                .is_err()
        );

        let version: i64 = store
            .conn
            .query_row("PRAGMA user_version", [], |r| r.get(0))
            .unwrap();
        assert_eq!(version, MIGRATIONS.len() as i64);
    }

    /// A project + running task to hang sessions off of.
    fn task_fixture(s: &mut Store) -> i64 {
        s.conn
            .execute("INSERT OR IGNORE INTO projects (name) VALUES ('voro')", [])
            .unwrap();
        let project_id: i64 = s
            .conn
            .query_row("SELECT id FROM projects WHERE name = 'voro'", [], |r| {
                r.get(0)
            })
            .unwrap();
        s.conn
            .execute(
                "INSERT INTO tasks (project_id, title, state, state_since, created_at)
                 VALUES (?1, 'run me', 'running', datetime('now'), datetime('now'))",
                params![project_id],
            )
            .unwrap();
        s.conn.last_insert_rowid()
    }

    /// `events_for` must return the audit trail oldest-first (newest last),
    /// since that's the order the history popup renders it in.
    #[test]
    fn events_for_orders_oldest_first() {
        use crate::transition::Action;

        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let task = s
            .create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "trace me".into(),
                body: String::new(),
                priority: Priority::P2,
                state: TaskState::Ready,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap();
        s.apply(task.id, Action::Start).unwrap();
        s.apply(task.id, Action::Ask("A or B?".into())).unwrap();
        s.apply(task.id, Action::Resume).unwrap();

        let events = s.events_for(task.id).unwrap();
        let kinds: Vec<&str> = events.iter().map(|e| e.kind.as_str()).collect();
        assert_eq!(
            kinds,
            vec!["created", "transition", "transition", "transition"]
        );
        // ids strictly increase with insertion order
        assert!(events.windows(2).all(|w| w[0].id < w[1].id));
    }

    #[test]
    fn latest_summary_returns_the_newest_summary_event() {
        use crate::transition::Action;

        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let t = s
            .create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "summary me".into(),
                body: String::new(),
                priority: Priority::P2,
                state: TaskState::Ready,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap();
        assert_eq!(s.latest_summary(t.id).unwrap(), None);

        s.apply(t.id, Action::Start).unwrap();
        s.apply(t.id, Action::Complete(Some("first pass".into())))
            .unwrap();
        assert_eq!(
            s.latest_summary(t.id).unwrap().as_deref(),
            Some("first pass")
        );

        // a reject-then-redo records a second summary; the newest wins
        s.apply(t.id, Action::RejectWork("redo".into())).unwrap();
        s.apply(t.id, Action::Complete(Some("second pass".into())))
            .unwrap();
        assert_eq!(
            s.latest_summary(t.id).unwrap().as_deref(),
            Some("second pass")
        );
    }

    /// The reviewed revision is what delta re-review compares against
    /// (DESIGN.md §8): absent until the operator sends work back, superseded by
    /// each later rejection, and carried on the event log alone.
    #[test]
    fn last_reviewed_supersedes_and_starts_absent() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let t = s
            .create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "review me".into(),
                body: String::new(),
                priority: Priority::P2,
                state: TaskState::Ready,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap();
        assert_eq!(s.last_reviewed(t.id).unwrap(), None);

        s.record_reviewed(t.id, "  aaaa1111  ").unwrap();
        assert_eq!(s.last_reviewed(t.id).unwrap().as_deref(), Some("aaaa1111"));
        s.record_reviewed(t.id, "bbbb2222").unwrap();
        assert_eq!(s.last_reviewed(t.id).unwrap().as_deref(), Some("bbbb2222"));

        assert!(s.record_reviewed(t.id, "   ").is_err());
        assert!(s.record_reviewed(999, "aaaa1111").is_err());
        // and it never touches task state
        assert_eq!(s.task(t.id).unwrap().state, TaskState::Ready);
    }

    #[test]
    fn incomplete_report_flag_marks_a_review_task_with_a_branch_and_no_summary() {
        use crate::transition::Action;

        // Helper: a fresh task carried to `review` with the given branch/summary.
        fn reviewed(branch: Option<&str>, summary: Option<&str>) -> (Store, i64) {
            let mut s = Store::open_in_memory().unwrap();
            let p = s.create_project("voro", "/tmp/voro").unwrap();
            let t = s
                .create_task(NewTask {
                    project_id: p.id,
                    repo_id: None,
                    title: "report me".into(),
                    body: String::new(),
                    priority: Priority::P2,
                    state: TaskState::Ready,
                    agent: None,
                    human: false,
                    deep: false,
                })
                .unwrap();
            s.apply(t.id, Action::Start).unwrap();
            s.apply(t.id, Action::Complete(summary.map(str::to_string)))
                .unwrap();
            if let Some(name) = branch {
                s.set_branch(t.id, Some(name)).unwrap();
            }
            (s, t.id)
        }

        // The still-anomalous half report: a branch but no summary — the classic
        // forgotten-summary flake and the shape the SessionEnd fallback leaves.
        let (s, id) = reviewed(Some("feat/x"), None);
        assert!(
            s.incomplete_report_flag(id).unwrap(),
            "half report: branch, no summary"
        );

        // The legitimate no-code report: a summary and no branch, as an
        // investigation, triage or audit ends. The summary is the deliverable,
        // so this is a complete report and must not be flagged.
        let (s, id) = reviewed(None, Some("already fixed by PR #96; nothing to do"));
        assert!(
            !s.incomplete_report_flag(id).unwrap(),
            "no-code report: summary, no branch"
        );

        // Both present — a complete report, not an anomaly.
        let (s, id) = reviewed(Some("feat/x"), Some("did the thing"));
        assert!(!s.incomplete_report_flag(id).unwrap());

        // Neither present — a legitimate no-artifact (e.g. planning) task.
        let (s, id) = reviewed(None, None);
        assert!(!s.incomplete_report_flag(id).unwrap());
    }

    #[test]
    fn incomplete_report_flag_is_gated_on_review() {
        use crate::transition::Action;

        // A partial report only counts once the task is in `review`: a running
        // task with an intended branch and no summary yet is mid-flight, not a
        // finished-but-incomplete report.
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let t = s
            .create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "in flight".into(),
                body: String::new(),
                priority: Priority::P2,
                state: TaskState::Ready,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap();
        s.set_branch(t.id, Some("feat/x")).unwrap();
        assert!(!s.incomplete_report_flag(t.id).unwrap(), "ready");

        s.apply(t.id, Action::Start).unwrap();
        assert!(!s.incomplete_report_flag(t.id).unwrap(), "running");

        // Only on reaching review does the missing summary become an anomaly.
        s.apply(t.id, Action::Complete(None)).unwrap();
        assert!(s.incomplete_report_flag(t.id).unwrap(), "review");

        // Accepting past review clears it — no PR is opened from `done`.
        s.apply(t.id, Action::Accept).unwrap();
        assert!(!s.incomplete_report_flag(t.id).unwrap(), "done");
    }

    /// A proposal, its priority and deps recorded so a refine can be shown to
    /// leave both alone.
    fn proposal(s: &mut Store, title: &str) -> Task {
        let p = s.projects().unwrap().first().cloned().unwrap_or_else(|| {
            s.create_project("voro", "/tmp/voro").unwrap();
            s.projects().unwrap().remove(0)
        });
        s.create_task(NewTask {
            project_id: p.id,
            repo_id: None,
            title: title.into(),
            body: "thin body".into(),
            priority: Priority::P2,
            state: TaskState::Proposed,
            agent: None,
            human: false,
            deep: false,
        })
        .unwrap()
    }

    /// A refine round launch (DESIGN.md §6): the note rides the transition, the
    /// task leaves the triage queue for `refining`, and everything else about
    /// it — priority, deps, the body being rewritten — is untouched.
    #[test]
    fn record_refine_launch_moves_the_task_and_logs_the_note() {
        let mut s = Store::open_in_memory().unwrap();
        let blocker = proposal(&mut s, "blocker");
        let t = proposal(&mut s, "refine me");
        s.add_dep(t.id, blocker.id, DepKind::Blocks).unwrap();
        let before = s.task(t.id).unwrap();

        let (after, session) = s
            .record_refine_launch(
                t.id,
                "  name the files it touches  ",
                "claude",
                Some(4321),
                LivenessSource::Pid,
                Some("/var/log/refine.log"),
            )
            .unwrap();

        assert_eq!(after.state, TaskState::Refining);
        assert_eq!(after.priority, before.priority);
        assert_eq!(after.body, before.body);
        assert_eq!(s.deps_of(t.id).unwrap().len(), 1);
        assert_eq!(session.pid, Some(4321));
        assert_eq!(session.log_path.as_deref(), Some("/var/log/refine.log"));
        assert!(session.ended_at.is_none());
        assert_eq!(
            s.latest_refine_note(t.id).unwrap().as_deref(),
            Some("name the files it touches")
        );
        // Neither marker shows while the round is in flight — there is nothing
        // to say about a rewrite that has not happened yet.
        assert!(!s.refined_flag(t.id).unwrap());
        assert!(!s.refine_failed_flag(t.id).unwrap());
    }

    /// The interactive flavour carries no note, so nothing is logged for one —
    /// the brief is the conversation itself.
    #[test]
    fn a_note_less_refine_launch_logs_no_note() {
        let mut s = Store::open_in_memory().unwrap();
        let t = proposal(&mut s, "refine me");
        s.record_refine_launch(t.id, "", "claude", Some(1), LivenessSource::Pid, None)
            .unwrap();
        assert_eq!(s.latest_refine_note(t.id).unwrap(), None);
        assert_eq!(s.task(t.id).unwrap().state, TaskState::Refining);
    }

    /// Each conclusion picks the marker the returned proposal carries, and the
    /// newest round wins — a failed round after a successful one says so.
    #[test]
    fn the_markers_read_the_round_that_just_concluded() {
        use crate::transition::{Action, Triage};

        let mut s = Store::open_in_memory().unwrap();
        let t = proposal(&mut s, "refine me");

        for (outcome, refined, failed) in [
            (RefineOutcome::Applied, true, false),
            (RefineOutcome::Failed, false, true),
            (RefineOutcome::Cancelled, false, false),
            (RefineOutcome::Applied, true, false),
        ] {
            s.record_refine_launch(t.id, "note", "claude", Some(1), LivenessSource::Pid, None)
                .unwrap();
            let after = s.conclude_refine(t.id, outcome).unwrap();
            assert_eq!(after.state, TaskState::Proposed, "{outcome}");
            assert_eq!(s.refined_flag(t.id).unwrap(), refined, "{outcome}");
            assert_eq!(s.refine_failed_flag(t.id).unwrap(), failed, "{outcome}");
            assert_eq!(s.latest_refine_outcome(t.id).unwrap(), Some(outcome));
        }

        // Triage is what clears the markers — both are gated on `proposed`.
        s.apply(t.id, Action::Triage(Triage::Parked)).unwrap();
        assert!(!s.refined_flag(t.id).unwrap());
        assert!(!s.refine_failed_flag(t.id).unwrap());
    }

    /// The late-rewrite backstop (DESIGN.md §6): a round concluded `failed`
    /// whose rewrite lands afterwards has its outcome corrected to applied, so
    /// the improved body is not read under a marker saying no rewrite happened.
    #[test]
    fn a_late_rewrite_corrects_a_failed_round_to_applied() {
        let mut s = Store::open_in_memory().unwrap();
        let t = proposal(&mut s, "refine me");
        s.record_refine_launch(t.id, "note", "claude", Some(1), LivenessSource::Pid, None)
            .unwrap();
        s.conclude_refine(t.id, RefineOutcome::Failed).unwrap();
        assert!(s.refine_failed_flag(t.id).unwrap());

        assert!(s.correct_late_refine(t.id).unwrap());
        assert!(s.refined_flag(t.id).unwrap());
        assert!(!s.refine_failed_flag(t.id).unwrap());
        assert_eq!(
            s.latest_refine_outcome(t.id).unwrap(),
            Some(RefineOutcome::Applied)
        );
        // A correction transitions nothing and reopens nothing.
        assert_eq!(s.task(t.id).unwrap().state, TaskState::Proposed);
        assert_eq!(
            s.sessions_for(t.id).unwrap()[0].outcome,
            Some(SessionOutcome::Failed),
            "the session keeps the outcome the reconciler observed"
        );
        // Idempotent: the correction is itself the newest outcome.
        assert!(!s.correct_late_refine(t.id).unwrap());
    }

    /// The correction is confined to the one case it exists for: any other last
    /// outcome, and any state but `proposed`, is left alone.
    #[test]
    fn correcting_a_round_is_a_no_op_off_the_failed_case() {
        use crate::transition::{Action, Triage};

        for outcome in [RefineOutcome::Applied, RefineOutcome::Cancelled] {
            let mut s = Store::open_in_memory().unwrap();
            let t = proposal(&mut s, "refine me");
            s.record_refine_launch(t.id, "note", "claude", Some(1), LivenessSource::Pid, None)
                .unwrap();
            s.conclude_refine(t.id, outcome).unwrap();

            assert!(!s.correct_late_refine(t.id).unwrap(), "{outcome}");
            assert_eq!(s.latest_refine_outcome(t.id).unwrap(), Some(outcome));
        }

        // A task nobody ever refined, and a failed round already triaged away.
        let mut s = Store::open_in_memory().unwrap();
        let t = proposal(&mut s, "never refined");
        assert!(!s.correct_late_refine(t.id).unwrap());
        assert_eq!(s.latest_refine_outcome(t.id).unwrap(), None);

        s.record_refine_launch(t.id, "note", "claude", Some(1), LivenessSource::Pid, None)
            .unwrap();
        s.conclude_refine(t.id, RefineOutcome::Failed).unwrap();
        s.apply(t.id, Action::Triage(Triage::Ready)).unwrap();
        assert!(!s.correct_late_refine(t.id).unwrap());
        assert_eq!(
            s.latest_refine_outcome(t.id).unwrap(),
            Some(RefineOutcome::Failed)
        );
    }

    /// A concluded round closes its session with the matching outcome, whichever
    /// trigger fired (DESIGN.md §6/§8).
    #[test]
    fn concluding_a_round_closes_its_session_with_the_matching_outcome() {
        for (outcome, session_outcome) in [
            (RefineOutcome::Applied, SessionOutcome::Completed),
            (RefineOutcome::Failed, SessionOutcome::Failed),
            (RefineOutcome::Cancelled, SessionOutcome::Aborted),
        ] {
            let mut s = Store::open_in_memory().unwrap();
            let t = proposal(&mut s, "refine me");
            let (_, session) = s
                .record_refine_launch(t.id, "note", "claude", Some(1), LivenessSource::Pid, None)
                .unwrap();

            s.conclude_refine(t.id, outcome).unwrap();
            let closed = s.session(session.id).unwrap();
            assert!(closed.ended_at.is_some(), "{outcome}");
            assert_eq!(closed.outcome, Some(session_outcome), "{outcome}");
        }
    }

    /// The round's guard rails: a task past `proposed`/`ready` cannot start
    /// one, and one that is not `refining` cannot conclude one.
    #[test]
    fn refine_transitions_are_refused_from_the_wrong_state() {
        use crate::transition::{Action, Triage};

        let mut s = Store::open_in_memory().unwrap();
        let t = proposal(&mut s, "refine me");
        assert!(matches!(
            s.conclude_refine(t.id, RefineOutcome::Applied),
            Err(Error::InvalidTransition { .. })
        ));

        s.apply(t.id, Action::Triage(Triage::Parked)).unwrap();
        assert!(matches!(
            s.record_refine_launch(t.id, "too late", "claude", None, LivenessSource::Pid, None),
            Err(Error::InvalidTransition { .. })
        ));
        // The refused launch wrote nothing — no session, no state change.
        assert_eq!(s.task(t.id).unwrap().state, TaskState::Parked);
        assert!(s.sessions_for(t.id).unwrap().is_empty());
    }

    #[test]
    fn discovered_from_resolves_the_parent_proposal() {
        let mut s = Store::open_in_memory().unwrap();
        let parent = proposal(&mut s, "parent");
        let child = proposal(&mut s, "child");
        assert!(s.discovered_from(child.id).unwrap().is_none());

        s.add_dep(child.id, parent.id, DepKind::DiscoveredFrom)
            .unwrap();
        assert_eq!(
            s.discovered_from(child.id).unwrap().map(|t| t.id),
            Some(parent.id)
        );
        // A plain blocker is not a parent: only `discovered-from` carries the
        // context a proposal was written against.
        let blocker = proposal(&mut s, "blocker");
        assert!(s.discovered_from(blocker.id).unwrap().is_none());
    }

    #[test]
    fn incomplete_report_flag_is_false_for_a_missing_task() {
        let s = Store::open_in_memory().unwrap();
        assert!(!s.incomplete_report_flag(999).unwrap());
    }

    #[test]
    fn session_create_end_round_trip() {
        let mut s = Store::open_in_memory().unwrap();
        let task_id = task_fixture(&mut s);

        let opened = s
            .create_session(
                task_id,
                "claude",
                Some(4321),
                LivenessSource::Pid,
                Some("/var/log/s.log"),
            )
            .unwrap();
        assert_eq!(opened.task_id, task_id);
        assert_eq!(opened.agent, "claude");
        assert_eq!(opened.pid, Some(4321));
        assert_eq!(opened.log_path.as_deref(), Some("/var/log/s.log"));
        assert!(!opened.started_at.is_empty());
        assert!(opened.ended_at.is_none());
        assert!(opened.outcome.is_none());

        let ended = s.end_session(opened.id, SessionOutcome::Completed).unwrap();
        assert_eq!(ended.id, opened.id);
        assert!(ended.ended_at.is_some());
        assert_eq!(ended.outcome, Some(SessionOutcome::Completed));

        assert_eq!(s.session(opened.id).unwrap(), ended);
    }

    /// `latest_sessions` maps each task to its newest session only, and tasks
    /// with no session history stay absent.
    #[test]
    fn latest_sessions_keeps_only_the_newest_per_task() {
        let mut s = Store::open_in_memory().unwrap();
        let with_history = task_fixture(&mut s);
        let sessionless = task_fixture(&mut s);

        let first = s
            .create_session(
                with_history,
                "claude",
                None,
                LivenessSource::Pid,
                Some("/var/log/first.log"),
            )
            .unwrap();
        s.end_session(first.id, SessionOutcome::Failed).unwrap();
        let second = s
            .create_session(
                with_history,
                "codex",
                None,
                LivenessSource::Pid,
                Some("/var/log/second.log"),
            )
            .unwrap();

        let latest = s.latest_sessions().unwrap();
        assert_eq!(latest.len(), 1);
        assert_eq!(latest[&with_history].id, second.id);
        assert_eq!(
            latest[&with_history].log_path.as_deref(),
            Some("/var/log/second.log")
        );
        assert!(!latest.contains_key(&sessionless));
    }

    #[test]
    fn session_optional_fields_are_null() {
        let mut s = Store::open_in_memory().unwrap();
        let task_id = task_fixture(&mut s);
        let opened = s
            .create_session(task_id, "codex", None, LivenessSource::Pid, None)
            .unwrap();
        assert!(opened.pid.is_none());
        assert!(opened.session_ref.is_none());
        assert!(opened.log_path.is_none());
    }

    #[test]
    fn set_session_ref_records_and_rejects_unknown_ids() {
        let mut s = Store::open_in_memory().unwrap();
        let task_id = task_fixture(&mut s);
        let opened = s
            .create_session(task_id, "claude", None, LivenessSource::Pid, None)
            .unwrap();
        assert!(opened.session_ref.is_none());

        let updated = s
            .set_session_ref(opened.id, "3f6c0e6e-1111-2222-3333-444455556666")
            .unwrap();
        assert_eq!(
            updated.session_ref.as_deref(),
            Some("3f6c0e6e-1111-2222-3333-444455556666")
        );
        assert_eq!(s.session(opened.id).unwrap(), updated);

        assert!(matches!(
            s.set_session_ref(999, "x"),
            Err(Error::SessionNotFound(999))
        ));
    }

    /// Each launch records which source reconciliation must read it by
    /// (DESIGN.md §8), and it survives the round trip: a headless
    /// launch under a supervisor is listing-authoritative, an interactive round
    /// is not, and neither is inferred from anything else on the row.
    #[test]
    fn a_session_records_the_liveness_source_it_was_launched_with() {
        let mut s = Store::open_in_memory().unwrap();
        let task_id = task_fixture(&mut s);
        let listing = s
            .create_session(task_id, "claude", Some(1), LivenessSource::Listing, None)
            .unwrap();
        assert_eq!(listing.liveness_source, LivenessSource::Listing);
        assert_eq!(
            s.session(listing.id).unwrap().liveness_source,
            LivenessSource::Listing
        );

        let pid = s
            .create_session(task_id, "manual", Some(1), LivenessSource::Pid, None)
            .unwrap();
        assert_eq!(pid.liveness_source, LivenessSource::Pid);
        assert_eq!(
            s.live_sessions().unwrap()[0].liveness_source,
            pid.liveness_source
        );

        // A ref captured later says nothing about the source: that was decided
        // at launch, which is the whole point of recording it.
        s.set_session_ref(pid.id, "uuid").unwrap();
        assert_eq!(
            s.session(pid.id).unwrap().liveness_source,
            LivenessSource::Pid
        );
    }

    /// The dispatch and refine transactions carry the flavour through to the
    /// row they open, so the reconciler reads what the launcher spawned.
    #[test]
    fn dispatch_and_refine_launches_carry_their_liveness_source() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("proj", "/tmp/proj").unwrap();
        let ready = s
            .create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "run me".into(),
                body: String::new(),
                priority: Priority::P1,
                state: TaskState::Ready,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap();
        let (_, dispatched) = s
            .record_dispatch(ready.id, "claude", Some(1), LivenessSource::Listing, None)
            .unwrap();
        assert_eq!(dispatched.liveness_source, LivenessSource::Listing);

        let proposal = s
            .create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "sloppy".into(),
                body: String::new(),
                priority: Priority::P2,
                state: TaskState::Proposed,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap();
        let (_, headless) = s
            .record_refine_launch(
                proposal.id,
                "name the files",
                "claude",
                Some(2),
                LivenessSource::Listing,
                None,
            )
            .unwrap();
        assert_eq!(headless.liveness_source, LivenessSource::Listing);

        s.conclude_refine(proposal.id, RefineOutcome::Cancelled)
            .unwrap();
        let (_, interactive) = s
            .record_refine_launch(
                proposal.id,
                "",
                "claude",
                Some(3),
                LivenessSource::Pid,
                None,
            )
            .unwrap();
        assert_eq!(interactive.liveness_source, LivenessSource::Pid);
    }

    /// A database from before migration 0017 must open with every existing
    /// session listing-authoritative — what a dispatch of an agent with a
    /// `sessions` verb already was, and the direction that leaves a session
    /// alone rather than finalising a live one — and the CHECK must reject a
    /// source that is neither.
    #[test]
    fn migration_0017_defaults_existing_sessions_to_the_listing() {
        let conn = Connection::open_in_memory().unwrap();
        for sql in &MIGRATIONS[..16] {
            conn.execute_batch(sql).unwrap();
        }
        conn.pragma_update(None, "user_version", 16).unwrap();
        conn.execute("INSERT INTO projects (name) VALUES ('p')", [])
            .unwrap();
        conn.execute(
            "INSERT INTO repos (project_id, name, path, is_default)
             VALUES (1, 'p', '/tmp/p', 1)",
            [],
        )
        .unwrap();
        conn.execute(
            "INSERT INTO tasks (project_id, title, state, state_since, created_at)
             VALUES (1, 'dispatched', 'running', datetime('now'), datetime('now'))",
            [],
        )
        .unwrap();
        conn.execute(
            "INSERT INTO sessions (task_id, agent, pid, started_at)
             VALUES (1, 'claude', 4242, datetime('now'))",
            [],
        )
        .unwrap();

        let store = Store::from_connection(conn).unwrap();
        assert_eq!(
            store.session(1).unwrap().liveness_source,
            LivenessSource::Listing
        );

        let junk = store.conn.execute(
            "UPDATE sessions SET liveness_source = 'guess' WHERE id = 1",
            [],
        );
        assert!(junk.is_err(), "the CHECK must reject an unknown source");
    }

    /// A confirmed send moves the session's process to the one carrying the
    /// turn, and follows the fork when the agent opened a new reference — but a
    /// send that resumed in place must not blank the reference it already had.
    #[test]
    fn record_session_send_moves_the_pid_and_follows_a_fork() {
        let mut s = Store::open_in_memory().unwrap();
        let task_id = task_fixture(&mut s);
        let opened = s
            .create_session(task_id, "claude", Some(1234), LivenessSource::Pid, None)
            .unwrap();
        s.set_session_ref(opened.id, "first-ref").unwrap();

        let resumed = s.record_session_send(opened.id, None, 4321).unwrap();
        assert_eq!(resumed.pid, Some(4321));
        assert_eq!(resumed.session_ref.as_deref(), Some("first-ref"));

        let forked = s
            .record_session_send(opened.id, Some("forked-ref"), 5678)
            .unwrap();
        assert_eq!(forked.pid, Some(5678));
        assert_eq!(forked.session_ref.as_deref(), Some("forked-ref"));
        assert_eq!(s.session(opened.id).unwrap(), forked);

        assert!(matches!(
            s.record_session_send(999, None, 1),
            Err(Error::SessionNotFound(999))
        ));
    }

    #[test]
    fn end_session_rejects_unknown_id() {
        let mut s = Store::open_in_memory().unwrap();
        assert!(matches!(
            s.end_session(999, SessionOutcome::Aborted),
            Err(Error::SessionNotFound(999))
        ));
    }

    #[test]
    fn sessions_for_returns_newest_first() {
        let mut s = Store::open_in_memory().unwrap();
        let task_id = task_fixture(&mut s);
        let first = s
            .create_session(task_id, "claude", None, LivenessSource::Pid, None)
            .unwrap();
        let second = s
            .create_session(task_id, "claude", None, LivenessSource::Pid, None)
            .unwrap();

        let sessions = s.sessions_for(task_id).unwrap();
        assert_eq!(
            sessions.iter().map(|s| s.id).collect::<Vec<_>>(),
            vec![second.id, first.id]
        );
    }

    #[test]
    fn live_sessions_excludes_ended() {
        let mut s = Store::open_in_memory().unwrap();
        let task_id = task_fixture(&mut s);
        let done = s
            .create_session(task_id, "claude", None, LivenessSource::Pid, None)
            .unwrap();
        let live = s
            .create_session(task_id, "claude", None, LivenessSource::Pid, None)
            .unwrap();
        s.end_session(done.id, SessionOutcome::Failed).unwrap();

        let ids = s.live_sessions().unwrap();
        assert_eq!(ids.iter().map(|s| s.id).collect::<Vec<_>>(), vec![live.id]);
    }

    #[test]
    fn running_rows_join_current_task_fields() {
        let mut s = Store::open_in_memory().unwrap();
        let task_id = task_fixture(&mut s);
        let session = s
            .create_session(task_id, "claude", None, LivenessSource::Pid, None)
            .unwrap();

        let rows = s.running_rows().unwrap();
        assert_eq!(rows.len(), 1);
        assert_eq!(rows[0].session_id, Some(session.id));
        assert_eq!(rows[0].task_id, task_id);
        assert_eq!(rows[0].task_title, "run me");
        assert_eq!(rows[0].task_state, TaskState::Running);
        assert_eq!(rows[0].agent.as_deref(), Some("claude"));
        assert!(rows[0].elapsed_secs >= 0);
    }

    #[test]
    fn running_rows_exclude_ended_sessions_and_order_newest_first() {
        let mut s = Store::open_in_memory().unwrap();
        let task_id = task_fixture(&mut s);
        let done = s
            .create_session(task_id, "claude", None, LivenessSource::Pid, None)
            .unwrap();
        let live = s
            .create_session(task_id, "codex", None, LivenessSource::Pid, None)
            .unwrap();
        s.end_session(done.id, SessionOutcome::Completed).unwrap();

        let rows = s.running_rows().unwrap();
        assert_eq!(
            rows.iter().map(|r| r.session_id).collect::<Vec<_>>(),
            vec![Some(live.id)]
        );
        assert_eq!(rows[0].agent.as_deref(), Some("codex"));
    }

    #[test]
    fn running_rows_compute_elapsed_from_started_at() {
        let mut s = Store::open_in_memory().unwrap();
        let task_id = task_fixture(&mut s);
        let session = s
            .create_session(task_id, "claude", None, LivenessSource::Pid, None)
            .unwrap();
        s.conn
            .execute(
                "UPDATE sessions SET started_at = datetime('now', '-90 seconds') WHERE id = ?1",
                params![session.id],
            )
            .unwrap();

        let rows = s.running_rows().unwrap();
        assert_eq!(rows.len(), 1);
        // allow a couple of seconds of test-execution slack either side
        assert!(
            (85..=95).contains(&rows[0].elapsed_secs),
            "expected ~90s elapsed, got {}",
            rows[0].elapsed_secs
        );
    }

    /// A task can be `running` with no live session — started by hand, so no
    /// session was ever opened. The running strip must still surface it
    /// (DESIGN.md §9), with no session id or agent and elapsed measured from
    /// when it entered `running`.
    #[test]
    fn running_rows_include_running_task_without_live_session() {
        let mut s = Store::open_in_memory().unwrap();
        let task_id = task_fixture(&mut s);
        s.conn
            .execute(
                "UPDATE tasks SET state_since = datetime('now', '-90 seconds') WHERE id = ?1",
                params![task_id],
            )
            .unwrap();

        let rows = s.running_rows().unwrap();
        assert_eq!(rows.len(), 1);
        assert_eq!(rows[0].session_id, None);
        assert_eq!(rows[0].agent, None);
        assert_eq!(rows[0].task_id, task_id);
        assert_eq!(rows[0].task_state, TaskState::Running);
        assert!(
            (85..=95).contains(&rows[0].elapsed_secs),
            "expected ~90s in running, got {}",
            rows[0].elapsed_secs
        );
    }

    /// A running task whose only session has ended is session-less too, so it
    /// stays visible rather than dropping off the strip.
    #[test]
    fn running_rows_include_task_whose_sessions_all_ended() {
        let mut s = Store::open_in_memory().unwrap();
        let task_id = task_fixture(&mut s);
        let done = s
            .create_session(task_id, "claude", None, LivenessSource::Pid, None)
            .unwrap();
        s.end_session(done.id, SessionOutcome::Failed).unwrap();

        let rows = s.running_rows().unwrap();
        assert_eq!(rows.len(), 1);
        assert_eq!(rows[0].session_id, None);
        assert_eq!(rows[0].task_id, task_id);
    }

    /// Live sessions sort ahead of session-less running tasks, so what an agent
    /// is actively driving stays at the top of the strip.
    #[test]
    fn running_rows_order_live_sessions_before_session_less_tasks() {
        let mut s = Store::open_in_memory().unwrap();
        let live_task = task_fixture(&mut s);
        let session = s
            .create_session(live_task, "claude", None, LivenessSource::Pid, None)
            .unwrap();
        let orphan_task = task_fixture(&mut s);

        let rows = s.running_rows().unwrap();
        assert_eq!(rows.len(), 2);
        assert_eq!(rows[0].session_id, Some(session.id));
        assert_eq!(rows[0].task_id, live_task);
        assert_eq!(rows[1].session_id, None);
        assert_eq!(rows[1].task_id, orphan_task);
    }

    /// A refine round is work under way, so it rides the strip beside dispatched
    /// tasks — with the round's session and the elapsed time since it opened —
    /// and leaves it the moment the round concludes (DESIGN.md §6/§9).
    #[test]
    fn running_rows_include_a_refining_task() {
        let mut s = Store::open_in_memory().unwrap();
        let t = proposal(&mut s, "refine me");
        let (_, session) = s
            .record_refine_launch(
                t.id,
                "thin body",
                "claude",
                Some(1),
                LivenessSource::Pid,
                None,
            )
            .unwrap();

        let rows = s.running_rows().unwrap();
        assert_eq!(rows.len(), 1);
        assert_eq!(rows[0].task_id, t.id);
        assert_eq!(rows[0].task_state, TaskState::Refining);
        assert_eq!(rows[0].session_id, Some(session.id));
        assert_eq!(rows[0].agent.as_deref(), Some("claude"));

        s.conclude_refine(t.id, RefineOutcome::Applied).unwrap();
        assert!(s.running_rows().unwrap().is_empty());
    }

    /// A hand-off is work in flight someone else owns, so it rides the strip
    /// too (DESIGN.md §9) — but its elapsed counts from the hand-off rather
    /// than from the session, which opened when the agent started the work and
    /// says nothing about how long the PR has been sitting there.
    #[test]
    fn running_rows_measure_a_waiting_task_from_the_hand_off() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let id = s
            .create_task(NewTask {
                project_id: p.id,
                repo_id: None,
                title: "handed off".into(),
                body: String::new(),
                priority: Priority::P2,
                state: TaskState::Ready,
                agent: None,
                human: false,
                deep: false,
            })
            .unwrap()
            .id;
        let (_, opened) = s
            .record_dispatch(id, "claude", Some(1), LivenessSource::Pid, None)
            .unwrap();
        s.apply(id, Action::Complete(None)).unwrap();
        s.apply(id, Action::HandOff).unwrap();
        s.set_pr(id, Some("https://github.com/o/r/pull/7")).unwrap();
        let session = opened.id;
        s.conn
            .execute(
                "UPDATE sessions SET started_at = datetime('now', '-2 hours') WHERE id = ?1",
                params![session],
            )
            .unwrap();
        s.conn
            .execute(
                "UPDATE tasks SET state_since = datetime('now', '-90 seconds') WHERE id = ?1",
                params![id],
            )
            .unwrap();

        let rows = s.running_rows().unwrap();
        assert_eq!(rows.len(), 1);
        assert_eq!(rows[0].task_id, id);
        assert_eq!(rows[0].task_state, TaskState::Waiting);
        assert_eq!(rows[0].session_id, Some(session));
        assert_eq!(
            rows[0].pr_url.as_deref(),
            Some("https://github.com/o/r/pull/7")
        );
        assert!(
            (85..=95).contains(&rows[0].elapsed_secs),
            "expected ~90s waiting, got {} (the session's own age is 2h)",
            rows[0].elapsed_secs
        );
    }

    /// Work an agent is driving sorts ahead of the hand-offs, which are the
    /// rows nobody is typing into.
    #[test]
    fn running_rows_sort_waiting_after_work_under_way() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let waiting = task_in_state(&mut s, p.id, TaskState::Waiting);
        let running = task_in_state(&mut s, p.id, TaskState::Running);
        let refining = task_in_state(&mut s, p.id, TaskState::Refining);

        let rows = s.running_rows().unwrap();
        assert_eq!(rows.len(), 3);
        assert_eq!(rows.last().unwrap().task_id, waiting);
        let ahead: Vec<i64> = rows[..2].iter().map(|r| r.task_id).collect();
        assert!(
            ahead.contains(&running) && ahead.contains(&refining),
            "{ahead:?}"
        );
    }

    /// Archiving retires the whole project from the cockpit (DESIGN.md §5), and
    /// the strip's newest row kind is no exception.
    #[test]
    fn running_rows_exclude_a_waiting_task_in_an_archived_project() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("retiring", "/tmp/retiring").unwrap();
        let id = task_in_state(&mut s, p.id, TaskState::Waiting);
        assert_eq!(s.running_rows().unwrap().len(), 1);

        s.set_archived(p.id, true).unwrap();
        assert!(s.running_rows().unwrap().is_empty());
        assert_eq!(s.task(id).unwrap().state, TaskState::Waiting);
    }

    /// The strip filters on task state: a task that has left `running` —
    /// review, done, rejected — never renders, even a `review` task whose
    /// session is deliberately still open (DESIGN.md §8/§9).
    #[test]
    fn running_rows_exclude_tasks_that_left_running() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let new = |title: &str| NewTask {
            project_id: p.id,
            repo_id: None,
            title: title.into(),
            body: String::new(),
            priority: Priority::P2,
            state: TaskState::Ready,
            agent: None,
            human: false,
            deep: false,
        };

        // review keeps its session open, yet must not appear in the strip
        let review = s.create_task(new("review")).unwrap().id;
        s.record_dispatch(review, "claude", Some(1), LivenessSource::Pid, None)
            .unwrap();
        s.apply(review, Action::Complete(None)).unwrap();
        assert!(s.sessions_for(review).unwrap()[0].ended_at.is_none());

        // done and rejected have their sessions closed by the transition
        let done = s.create_task(new("done")).unwrap().id;
        s.record_dispatch(done, "claude", Some(2), LivenessSource::Pid, None)
            .unwrap();
        s.apply(done, Action::Complete(None)).unwrap();
        s.apply(done, Action::Accept).unwrap();

        let rejected = s.create_task(new("rejected")).unwrap().id;
        s.record_dispatch(rejected, "claude", Some(3), LivenessSource::Pid, None)
            .unwrap();
        s.apply(rejected, Action::Abort).unwrap();
        s.apply(rejected, Action::Abandon).unwrap();

        let running = s.create_task(new("running")).unwrap().id;
        s.record_dispatch(running, "claude", Some(4), LivenessSource::Pid, None)
            .unwrap();

        let rows = s.running_rows().unwrap();
        assert_eq!(rows.len(), 1);
        assert_eq!(rows[0].task_id, running);
    }

    /// A `done` task left carrying an open session must stay out of the strip
    /// purely on its state.
    #[test]
    fn running_rows_ignore_a_stale_open_session_on_a_closed_task() {
        let mut s = Store::open_in_memory().unwrap();
        let task_id = task_fixture(&mut s);
        s.create_session(task_id, "claude", Some(1), LivenessSource::Pid, None)
            .unwrap();
        s.conn
            .execute("UPDATE tasks SET state = 'done' WHERE id = ?1", [task_id])
            .unwrap();
        assert!(s.running_rows().unwrap().is_empty());
    }

    #[test]
    fn session_outcome_serialises_for_all_variants() {
        let mut s = Store::open_in_memory().unwrap();
        let task_id = task_fixture(&mut s);
        for outcome in SessionOutcome::ALL {
            let opened = s
                .create_session(task_id, "claude", None, LivenessSource::Pid, None)
                .unwrap();
            let ended = s.end_session(opened.id, outcome).unwrap();
            assert_eq!(ended.outcome, Some(outcome));
        }
    }

    /// A unique scratch database path under the OS temp dir.
    fn scratch_db() -> PathBuf {
        tempfile::Builder::new()
            .prefix("voro-dataversion-")
            .tempdir()
            .unwrap()
            .keep()
            .join("voro.db")
    }

    #[test]
    fn data_version_tracks_external_commits_only() {
        let path = scratch_db();
        let mut a = Store::open(&path).unwrap();
        let mut b = Store::open(&path).unwrap();

        let start = a.data_version().unwrap();

        // Our own writes must not move the version this connection observes.
        a.create_project("alpha", "/tmp/alpha").unwrap();
        assert_eq!(a.data_version().unwrap(), start);

        // A commit from another connection must move it.
        b.create_project("beta", "/tmp/beta").unwrap();
        assert_ne!(a.data_version().unwrap(), start);

        drop(a);
        drop(b);
        let _ = std::fs::remove_file(&path);
    }

    #[test]
    fn dep_maps_resolve_both_directions_with_title_state_and_kind() {
        use crate::model::{DepKind, DepRef, Priority};
        use crate::transition::Action;

        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let new = |title: &str| NewTask {
            project_id: p.id,
            repo_id: None,
            title: title.into(),
            body: String::new(),
            priority: Priority::P2,
            state: TaskState::Ready,
            agent: None,
            human: false,
            deep: false,
        };
        let blocker = s.create_task(new("blocker")).unwrap();
        s.apply(blocker.id, Action::Start).unwrap();
        s.apply(blocker.id, Action::Complete(None)).unwrap();
        s.apply(blocker.id, Action::Accept).unwrap();
        let source = s.create_task(new("source")).unwrap();
        let task = s.create_task(new("task")).unwrap();
        s.add_dep(task.id, blocker.id, DepKind::Blocks).unwrap();
        s.add_dep(task.id, source.id, DepKind::DiscoveredFrom)
            .unwrap();

        // Forward: the task's own deps, every kind, resolved to the
        // dependency's title and state.
        let deps = s.deps_by_task().unwrap();
        assert_eq!(
            deps[&task.id],
            vec![
                DepRef {
                    id: blocker.id,
                    title: "blocker".into(),
                    state: TaskState::Done,
                    kind: DepKind::Blocks,
                },
                DepRef {
                    id: source.id,
                    title: "source".into(),
                    state: TaskState::Ready,
                    kind: DepKind::DiscoveredFrom,
                },
            ]
        );
        assert!(!deps[&task.id][0].is_open());
        assert!(!deps.contains_key(&blocker.id));

        // Reverse: keyed by the task depended on, resolving the dependant.
        let dependents = s.dependents_by_task().unwrap();
        assert_eq!(
            dependents[&blocker.id],
            vec![DepRef {
                id: task.id,
                title: "task".into(),
                state: TaskState::Ready,
                kind: DepKind::Blocks,
            }]
        );
        assert_eq!(dependents[&source.id].len(), 1);
        assert_eq!(dependents[&source.id][0].kind, DepKind::DiscoveredFrom);
        assert!(!dependents.contains_key(&task.id));
    }

    // --- docs (DESIGN.md §3/§5) ---

    #[test]
    fn a_doc_links_tasks_across_projects_and_answers_both_directions() {
        // The case the table exists for: one plan doc spawning work in several
        // projects, so the link cannot be constrained to the doc's own project.
        let mut s = Store::open_in_memory().unwrap();
        let plan = s.create_project("augere", "/tmp/augere").unwrap();
        let other = s.create_project("mote", "/tmp/mote").unwrap();
        let doc = s
            .create_doc(plan.id, None, "docs/strategy.md", Some("Strategy"))
            .unwrap();

        let a = s.create_task(new_ready(plan.id)).unwrap();
        let b = s.create_task(new_ready(other.id)).unwrap();
        let c = s.create_task(new_ready(other.id)).unwrap();
        for task in [&a, &b, &c] {
            assert!(s.link_doc(task.id, doc.id).unwrap());
        }
        // A repeated link is a no-op rather than an error.
        assert!(!s.link_doc(a.id, doc.id).unwrap());

        let derived: Vec<i64> = s
            .tasks_for_doc(doc.id)
            .unwrap()
            .into_iter()
            .map(|t| t.id)
            .collect();
        assert_eq!(derived, vec![a.id, b.id, c.id]);
        assert_eq!(s.docs_for_task(b.id).unwrap(), vec![doc.clone()]);
        assert_eq!(s.docs_by_task().unwrap()[&c.id], vec![doc.clone()]);

        assert!(s.unlink_doc(b.id, doc.id).unwrap());
        assert!(!s.unlink_doc(b.id, doc.id).unwrap());
        assert_eq!(s.tasks_for_doc(doc.id).unwrap().len(), 2);
        assert!(s.docs_for_task(b.id).unwrap().is_empty());
    }

    #[test]
    fn every_doc_link_and_unlink_lands_on_the_task_event_trail() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let doc = s.create_doc(p.id, None, "docs/DESIGN.md", None).unwrap();
        let t = s.create_task(new_ready(p.id)).unwrap();

        s.link_doc(t.id, doc.id).unwrap();
        s.unlink_doc(t.id, doc.id).unwrap();
        // A no-op link writes nothing, so the trail records changes only.
        s.unlink_doc(t.id, doc.id).unwrap();

        let kinds: Vec<String> = s
            .events_for(t.id)
            .unwrap()
            .into_iter()
            .map(|e| e.kind)
            .collect();
        assert_eq!(kinds, vec!["created", "doc-linked", "doc-unlinked"]);
    }

    #[test]
    fn set_task_docs_replaces_the_whole_list_and_logs_both_directions() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let one = s.create_doc(p.id, None, "docs/a.md", None).unwrap();
        let two = s.create_doc(p.id, None, "docs/b.md", None).unwrap();
        let t = s.create_task(new_ready(p.id)).unwrap();

        s.set_task_docs(t.id, &[one.id]).unwrap();
        // Replace, not append: `a` goes as `b` arrives.
        let now = s.set_task_docs(t.id, &[two.id]).unwrap();
        assert_eq!(now, vec![two.clone()]);

        let events: Vec<(String, Option<String>)> = s
            .events_for(t.id)
            .unwrap()
            .into_iter()
            .map(|e| (e.kind, e.detail))
            .collect();
        assert_eq!(
            events,
            vec![
                ("created".into(), Some("ready".into())),
                ("doc-linked".into(), Some("docs/a.md".into())),
                ("doc-unlinked".into(), Some("docs/a.md".into())),
                ("doc-linked".into(), Some("docs/b.md".into())),
            ]
        );

        // Clearing the list is the empty replacement.
        assert!(s.set_task_docs(t.id, &[]).unwrap().is_empty());
    }

    #[test]
    fn an_absolute_path_inside_a_checkout_is_stored_relative_to_it() {
        // Storing it relative is what makes the link survive the checkout
        // moving, which is why an operator may paste an absolute path.
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("augere", "/tmp/augere").unwrap();
        let doc = s
            .create_doc(p.id, None, "/tmp/augere/docs/strategy.md", None)
            .unwrap();
        assert_eq!(doc.location, "docs/strategy.md");
        assert_eq!(s.resolve_doc(&doc).unwrap(), "/tmp/augere/docs/strategy.md");

        // ...and it follows the checkout when that moves.
        s.set_default_repo_path(p.id, "/srv/augere").unwrap();
        assert_eq!(
            s.resolve_doc(&s.doc(doc.id).unwrap()).unwrap(),
            "/srv/augere/docs/strategy.md"
        );
    }

    #[test]
    fn a_relative_doc_resolves_against_the_repo_it_names() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("odm", "/tmp/odm").unwrap();
        let oats = s.add_repo(p.id, "oats", "/tmp/oats").unwrap();

        let default = s.create_doc(p.id, None, "docs/plan.md", None).unwrap();
        assert_eq!(s.resolve_doc(&default).unwrap(), "/tmp/odm/docs/plan.md");

        let named = s
            .create_doc(p.id, Some(oats.id), "notes/plan.md", None)
            .unwrap();
        assert_eq!(s.resolve_doc(&named).unwrap(), "/tmp/oats/notes/plan.md");

        // The longest containing checkout wins for an absolute path, so a repo
        // nested inside another is not swallowed by its parent.
        let nested = s.add_repo(p.id, "inner", "/tmp/odm/vendor").unwrap();
        let doc = s
            .create_doc(p.id, None, "/tmp/odm/vendor/docs/x.md", None)
            .unwrap();
        assert_eq!(doc.repo_id, Some(nested.id));
        assert_eq!(doc.location, "docs/x.md");
    }

    #[test]
    fn a_url_resolves_verbatim_and_takes_no_repo() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let repo = s.default_repo(p.id).unwrap();

        let doc = s
            .create_doc(p.id, None, "https://example.com/plan", Some("Plan"))
            .unwrap();
        assert!(doc.is_url());
        assert!(doc.repo_id.is_none());
        assert_eq!(s.resolve_doc(&doc).unwrap(), "https://example.com/plan");
        assert_eq!(doc.label(), "Plan");

        // A URL resolves on its own, so pairing it with a checkout is refused
        // rather than silently ignored.
        assert!(
            s.create_doc(p.id, Some(repo.id), "https://example.com/other", None)
                .is_err()
        );
    }

    #[test]
    fn a_doc_outside_every_checkout_stays_absolute() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let doc = s
            .create_doc(p.id, None, "/etc/notes/plan.md", None)
            .unwrap();
        assert_eq!(doc.location, "/etc/notes/plan.md");
        assert!(doc.repo_id.is_none());
        assert_eq!(s.resolve_doc(&doc).unwrap(), "/etc/notes/plan.md");
        // With --repo given, though, a path outside it is a mistake, not an
        // external document.
        let repo = s.default_repo(p.id).unwrap();
        assert!(
            s.create_doc(p.id, Some(repo.id), "/etc/notes/other.md", None)
                .is_err()
        );
    }

    #[test]
    fn a_doc_is_registered_once_per_project_and_labels_itself() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let other = s.create_project("mote", "/tmp/mote").unwrap();
        let doc = s.create_doc(p.id, None, "docs/plan.md", None).unwrap();
        // With no title, the location is the only name it has.
        assert_eq!(doc.label(), "docs/plan.md");
        assert!(s.create_doc(p.id, None, "docs/plan.md", None).is_err());
        // The same relative location under another project is a different doc,
        // since it resolves against that project's checkout.
        let twin = s.create_doc(other.id, None, "docs/plan.md", None).unwrap();
        assert_eq!(s.docs_at("docs/plan.md").unwrap().len(), 2);
        assert_ne!(doc.id, twin.id);
        assert_eq!(s.docs(p.id).unwrap(), vec![doc]);
        assert!(s.create_doc(p.id, None, "   ", None).is_err());
    }

    #[test]
    fn removing_a_doc_unlinks_its_tasks_rather_than_refusing() {
        // Unlike a repo, a doc is navigational — nothing resolves to nothing
        // when it goes — so removal frees its links instead of being refused.
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let doc = s.create_doc(p.id, None, "docs/plan.md", None).unwrap();
        let a = s.create_task(new_ready(p.id)).unwrap();
        let b = s.create_task(new_ready(p.id)).unwrap();
        s.link_doc(a.id, doc.id).unwrap();
        s.link_doc(b.id, doc.id).unwrap();

        let freed = s.delete_doc(doc.id).unwrap();
        assert_eq!(freed, vec![a.id, b.id]);
        assert!(s.doc(doc.id).is_err());
        assert!(s.docs_for_task(a.id).unwrap().is_empty());
        assert!(s.docs_by_task().unwrap().is_empty());
        assert_eq!(
            s.events_for(a.id).unwrap().last().unwrap().kind,
            "doc-unlinked"
        );
    }

    #[test]
    fn linking_names_a_task_and_a_doc_that_exist() {
        let mut s = Store::open_in_memory().unwrap();
        let p = s.create_project("voro", "/tmp/voro").unwrap();
        let doc = s.create_doc(p.id, None, "docs/plan.md", None).unwrap();
        let t = s.create_task(new_ready(p.id)).unwrap();
        assert!(s.link_doc(999, doc.id).is_err());
        assert!(s.link_doc(t.id, 999).is_err());
        assert!(s.set_task_docs(t.id, &[999]).is_err());
        // A refused replacement leaves the list as it was.
        assert!(s.docs_for_task(t.id).unwrap().is_empty());
    }

    /// A database from before migration 0014 must open with every existing
    /// task intact and no documents registered — docs are purely additive.
    #[test]
    fn migration_0014_leaves_existing_tasks_untouched() {
        let conn = Connection::open_in_memory().unwrap();
        for sql in &MIGRATIONS[..13] {
            conn.execute_batch(sql).unwrap();
        }
        conn.pragma_update(None, "user_version", 13).unwrap();
        conn.execute("INSERT INTO projects (name) VALUES ('legacy')", [])
            .unwrap();
        conn.execute(
            "INSERT INTO repos (project_id, name, path, is_default)
             VALUES (1, 'legacy', '/tmp/legacy', 1)",
            [],
        )
        .unwrap();
        conn.execute(
            "INSERT INTO tasks (project_id, title, state, state_since, created_at)
             VALUES (1, 'old work', 'ready', datetime('now'), datetime('now'))",
            [],
        )
        .unwrap();

        let store = Store::from_connection(conn).unwrap();
        let task = store.task(1).unwrap();
        assert_eq!(task.title, "old work");
        assert_eq!(task.state, TaskState::Ready);
        assert!(store.all_docs().unwrap().is_empty());
        assert!(store.docs_for_task(1).unwrap().is_empty());
    }
}