hyperhive/hive-agent/src/turn_stats.rs

426 lines
18 KiB
Rust

//! Per-turn analytics sink. One sqlite row per claude turn captures:
//! identity (`model`, `wake_from`, `result_kind`), timing (`started_at`,
//! `ended_at`, `duration_ms`), cost (token counts), and behaviour (tool-call
//! count + per-tool breakdown).
//!
//! **Captured but not yet read** (written every turn, no reader today —
//! kept for a future chart / backfill, not consumed by `stats::snapshot`
//! or the host rollup): `tool_call_count` (the snapshot recomputes tool
//! totals from `tool_call_breakdown_json` instead), `open_threads_count` +
//! `open_reminders_count` (planned: a loose-ends-over-time trend), and
//! `note` (failure detail for `result_kind = "failed"`).
//!
//! Lives next to `hyperhive-events.sqlite` in the agent's state dir
//! so the host-side state vacuum sweep can reach both. Schema is
//! intentionally append-only — every column has a default so future
//! additions don't break old readers; new columns land via
//! `ALTER TABLE ... ADD COLUMN ... DEFAULT ...` in the migration
//! block.
//!
//! Writes are best-effort: a failed insert logs a warning and lets
//! the turn loop continue. The next turn either succeeds or the
//! operator sees the journal trail.
use std::path::{Path, PathBuf};
use std::sync::Mutex;
use anyhow::{Context, Result};
use rusqlite::{Connection, params};
/// SQL bootstrap. CREATE TABLE IF NOT EXISTS so first-boot agents
/// and existing ones converge on the same shape. The base table is
/// fresh-install only; additive migrations land via `MIGRATIONS`
/// below as try-and-ignore ALTERs so existing dbs catch up.
const SCHEMA: &str = "
CREATE TABLE IF NOT EXISTS turn_stats (
id INTEGER PRIMARY KEY AUTOINCREMENT,
started_at INTEGER NOT NULL,
ended_at INTEGER NOT NULL,
duration_ms INTEGER NOT NULL,
model TEXT NOT NULL,
wake_from TEXT NOT NULL,
input_tokens INTEGER NOT NULL DEFAULT 0,
output_tokens INTEGER NOT NULL DEFAULT 0,
cache_read_input_tokens INTEGER NOT NULL DEFAULT 0,
cache_creation_input_tokens INTEGER NOT NULL DEFAULT 0,
last_input_tokens INTEGER NOT NULL DEFAULT 0,
last_output_tokens INTEGER NOT NULL DEFAULT 0,
last_cache_read_input_tokens INTEGER NOT NULL DEFAULT 0,
last_cache_creation_input_tokens INTEGER NOT NULL DEFAULT 0,
tool_call_count INTEGER NOT NULL DEFAULT 0,
tool_call_breakdown_json TEXT,
open_threads_count INTEGER,
open_reminders_count INTEGER,
result_kind TEXT NOT NULL,
note TEXT,
session_id INTEGER
);
CREATE INDEX IF NOT EXISTS idx_turn_stats_started
ON turn_stats (started_at DESC);
-- NOTE: the index on session_id is created in MIGRATIONS, not here. On an
-- existing pre-session db the `CREATE TABLE IF NOT EXISTS` above is a no-op
-- (the old table has no session_id column), so indexing session_id in this
-- batch would fail with `no such column` and abort the whole SCHEMA apply —
-- which disables the stats sink. The column is added by MIGRATIONS first.
-- One row per fresh claude session (minted when --continue is suppressed).
-- turn_stats.session_id FKs here so per-session stats (first-turn tokens,
-- per-session totals, turn count, duration) are one GROUP BY away.
CREATE TABLE IF NOT EXISTS sessions (
id INTEGER PRIMARY KEY AUTOINCREMENT,
started_at INTEGER NOT NULL,
model TEXT NOT NULL
);
";
/// Additive column migrations. Each runs unconditionally and ignores
/// `duplicate column name` errors — sqlite < 3.35 lacks
/// `ADD COLUMN IF NOT EXISTS`, so try-and-ignore is the portable path.
/// New columns MUST carry a default so existing rows decode.
const MIGRATIONS: &[&str] = &[
"ALTER TABLE turn_stats ADD COLUMN last_input_tokens INTEGER NOT NULL DEFAULT 0",
"ALTER TABLE turn_stats ADD COLUMN last_output_tokens INTEGER NOT NULL DEFAULT 0",
"ALTER TABLE turn_stats ADD COLUMN last_cache_read_input_tokens INTEGER NOT NULL DEFAULT 0",
"ALTER TABLE turn_stats ADD COLUMN last_cache_creation_input_tokens INTEGER NOT NULL DEFAULT 0",
// Nullable FK to sessions.id — no default; pre-migration rows stay NULL
// (the surface treats NULL as "no session", inert until capture lands).
"ALTER TABLE turn_stats ADD COLUMN session_id INTEGER",
// Index on session_id — must run AFTER the column is added, so it lives
// here rather than in SCHEMA (see the note there). Idempotent.
"CREATE INDEX IF NOT EXISTS idx_turn_stats_session ON turn_stats (session_id)",
];
/// One row to be inserted. `Option`-wrapped fields default to NULL
/// when the harness couldn't gather them (e.g. socket roundtrip for
/// `open_threads` failed) so a partial row beats no row.
#[derive(Debug, Clone)]
pub struct TurnStatRow {
pub started_at: i64,
pub ended_at: i64,
pub duration_ms: i64,
pub model: String,
pub wake_from: String,
/// Cumulative across every inference in the turn (cost signal).
pub input_tokens: u64,
pub output_tokens: u64,
pub cache_read_input_tokens: u64,
pub cache_creation_input_tokens: u64,
/// Last inference's usage — the actual context size at turn end.
pub last_input_tokens: u64,
pub last_output_tokens: u64,
pub last_cache_read_input_tokens: u64,
pub last_cache_creation_input_tokens: u64,
/// Captured, not yet read — the snapshot recomputes tool totals from
/// `tool_call_breakdown_json` (see the module doc).
pub tool_call_count: u64,
/// Per-tool breakdown as JSON: `{"Read":12,"Bash":3,...}`. None
/// when no tools were called (saves a sqlite write of `"{}"`).
pub tool_call_breakdown_json: Option<String>,
/// Post-turn loose-ends snapshot. Captured, not yet read — planned to
/// feed a loose-ends-over-time trend on the stats page.
pub open_threads_count: Option<u64>,
pub open_reminders_count: Option<u64>,
/// `"ok" | "failed" | "prompt_too_long"`.
pub result_kind: &'static str,
/// Failure detail for `result_kind = "failed"`. Captured, not yet read.
pub note: Option<String>,
/// FK to `sessions.id` for the fresh claude session this turn belongs
/// to. `None` on pre-capture rows (and when the stats db couldn't mint
/// a session) so the read side degrades to empty.
pub session_id: Option<i64>,
}
/// Thin sqlite wrapper. Cloning is cheap (Arc-shared connection).
#[derive(Clone)]
pub struct TurnStats {
inner: std::sync::Arc<Mutex<Connection>>,
}
impl TurnStats {
/// Open the per-agent stats db, creating the file + schema if
/// missing. Returns `None` when the db can't be opened (read-only
/// fs in tests, missing state dir) — the harness logs and
/// continues without a sink rather than failing the turn loop.
#[must_use]
pub fn open_default() -> Option<Self> {
let path = default_path();
match Self::open(&path) {
Ok(s) => Some(s),
Err(e) => {
tracing::warn!(
error = ?e,
path = %path.display(),
"turn_stats: open failed; per-turn analytics disabled"
);
None
}
}
}
fn open(path: &Path) -> Result<Self> {
if let Some(parent) = path.parent() {
let _ = std::fs::create_dir_all(parent);
}
let conn = Connection::open(path)
.with_context(|| format!("open turn_stats db {}", path.display()))?;
conn.execute_batch(SCHEMA)
.context("apply turn_stats schema")?;
for stmt in MIGRATIONS {
// Ignore "duplicate column name" — the migration already ran.
// Any other error is logged but doesn't fail open() because the
// base schema works and we'd rather keep the harness alive than
// crash on an upgrade hiccup.
if let Err(e) = conn.execute(stmt, []) {
let msg = e.to_string();
if !msg.contains("duplicate column name") {
tracing::warn!(error = %msg, stmt, "turn_stats migration failed");
}
}
}
Ok(Self {
inner: std::sync::Arc::new(Mutex::new(conn)),
})
}
/// Insert a row. Best-effort — logs + swallows errors so a sqlite
/// hiccup (locked db, full disk) doesn't crash the harness.
///
/// # Panics
///
/// Panics if the internal lock is poisoned.
pub fn record(&self, row: &TurnStatRow) {
let conn = self.inner.lock().unwrap();
let res = conn.execute(
"INSERT INTO turn_stats (
started_at, ended_at, duration_ms, model, wake_from,
input_tokens, output_tokens,
cache_read_input_tokens, cache_creation_input_tokens,
last_input_tokens, last_output_tokens,
last_cache_read_input_tokens, last_cache_creation_input_tokens,
tool_call_count, tool_call_breakdown_json,
open_threads_count, open_reminders_count,
result_kind, note, session_id
) VALUES (
?1, ?2, ?3, ?4, ?5,
?6, ?7,
?8, ?9,
?10, ?11,
?12, ?13,
?14, ?15,
?16, ?17,
?18, ?19, ?20
)",
params![
row.started_at,
row.ended_at,
row.duration_ms,
row.model,
row.wake_from,
i64::try_from(row.input_tokens).unwrap_or(i64::MAX),
i64::try_from(row.output_tokens).unwrap_or(i64::MAX),
i64::try_from(row.cache_read_input_tokens).unwrap_or(i64::MAX),
i64::try_from(row.cache_creation_input_tokens).unwrap_or(i64::MAX),
i64::try_from(row.last_input_tokens).unwrap_or(i64::MAX),
i64::try_from(row.last_output_tokens).unwrap_or(i64::MAX),
i64::try_from(row.last_cache_read_input_tokens).unwrap_or(i64::MAX),
i64::try_from(row.last_cache_creation_input_tokens).unwrap_or(i64::MAX),
i64::try_from(row.tool_call_count).unwrap_or(i64::MAX),
row.tool_call_breakdown_json,
row.open_threads_count
.map(|n| i64::try_from(n).unwrap_or(i64::MAX)),
row.open_reminders_count
.map(|n| i64::try_from(n).unwrap_or(i64::MAX)),
row.result_kind,
row.note,
row.session_id,
],
);
if let Err(e) = res {
tracing::warn!(error = ?e, "turn_stats: insert failed");
}
}
/// Mint a new session row at fresh-session start, returning its `id`
/// for stamping onto this session's `turn_stats` rows. Best-effort —
/// returns `None` (and logs) on any sqlite error, so a hiccup degrades
/// to NULL `session_id` rows rather than crashing the turn loop.
///
/// # Panics
///
/// Panics if the internal lock is poisoned.
#[must_use]
pub fn start_session(&self, started_at: i64, model: &str) -> Option<i64> {
let conn = self.inner.lock().unwrap();
match conn.execute(
"INSERT INTO sessions (started_at, model) VALUES (?1, ?2)",
params![started_at, model],
) {
Ok(_) => Some(conn.last_insert_rowid()),
Err(e) => {
tracing::warn!(error = ?e, "turn_stats: start_session insert failed");
None
}
}
}
/// Token counts from the most recently inserted row, if any.
/// Returns `(ctx, cost)` — both backfill `Bus` on startup so the
/// per-agent web UI's ctx + cost badges paint with real numbers on
/// cold load instead of waiting for the next `TokenUsageChanged`
/// SSE event. Best-effort: any sqlite error returns `(None, None)`.
///
/// Pre-migration rows (before the `last_*_tokens` columns existed)
/// have last-inference zeros — those rows yield `ctx = None` so the
/// badge stays empty until the next real turn rather than showing a
/// misleading 0.
/// # Panics
///
/// Panics if the internal lock is poisoned.
#[must_use]
pub fn last_usage(
&self,
) -> (
Option<hive_claude::TokenUsage>,
Option<hive_claude::TokenUsage>,
) {
let conn = self.inner.lock().unwrap();
conn.query_row(
"SELECT input_tokens, output_tokens,
cache_read_input_tokens, cache_creation_input_tokens,
last_input_tokens, last_output_tokens,
last_cache_read_input_tokens, last_cache_creation_input_tokens
FROM turn_stats
-- `id` (AUTOINCREMENT) is monotonic with insertion, so this is
-- the most-recently-inserted row even among same-second turns
-- (which `started_at DESC` would order arbitrarily).
ORDER BY id DESC
LIMIT 1",
[],
|row| {
let g = |i: usize| -> rusqlite::Result<u64> {
Ok(u64::try_from(row.get::<_, i64>(i)?).unwrap_or(0))
};
// `TokenUsage` is `#[non_exhaustive]` (hive-claude 0.1.0+) —
// struct-expression construction is blocked for downstream
// crates entirely, even with a `..base` (functional update
// doesn't carve out an exception, per the non_exhaustive
// reference semantics). Build off `Default` and assign the
// (all-`pub`) fields instead.
let mut cost = hive_claude::TokenUsage::default();
cost.input_tokens = g(0)?;
cost.output_tokens = g(1)?;
cost.cache_read_input_tokens = g(2)?;
cost.cache_creation_input_tokens = g(3)?;
let mut last = hive_claude::TokenUsage::default();
last.input_tokens = g(4)?;
last.output_tokens = g(5)?;
last.cache_read_input_tokens = g(6)?;
last.cache_creation_input_tokens = g(7)?;
let ctx = if last == hive_claude::TokenUsage::default() {
None
} else {
Some(last)
};
Ok((ctx, Some(cost)))
},
)
.unwrap_or((None, None))
}
}
fn default_path() -> PathBuf {
crate::paths::harness_dir().join("hyperhive-turn-stats.sqlite")
}
#[cfg(test)]
mod tests {
use super::*;
/// A `turn_stats` db in the *pre-session* shape: the original table with no
/// `session_id` column, plus the `started_at` index. This is what every
/// agent created before the sessions feature has on disk.
fn seed_pre_session_db(path: &Path) {
let conn = Connection::open(path).unwrap();
conn.execute_batch(
"CREATE TABLE turn_stats (
id INTEGER PRIMARY KEY AUTOINCREMENT,
started_at INTEGER NOT NULL,
ended_at INTEGER NOT NULL,
duration_ms INTEGER NOT NULL,
model TEXT NOT NULL,
wake_from TEXT NOT NULL,
input_tokens INTEGER NOT NULL DEFAULT 0,
output_tokens INTEGER NOT NULL DEFAULT 0,
cache_read_input_tokens INTEGER NOT NULL DEFAULT 0,
cache_creation_input_tokens INTEGER NOT NULL DEFAULT 0,
last_input_tokens INTEGER NOT NULL DEFAULT 0,
last_output_tokens INTEGER NOT NULL DEFAULT 0,
last_cache_read_input_tokens INTEGER NOT NULL DEFAULT 0,
last_cache_creation_input_tokens INTEGER NOT NULL DEFAULT 0,
tool_call_count INTEGER NOT NULL DEFAULT 0,
tool_call_breakdown_json TEXT,
open_threads_count INTEGER,
open_reminders_count INTEGER,
result_kind TEXT NOT NULL,
note TEXT
);
CREATE INDEX idx_turn_stats_started ON turn_stats (started_at DESC);",
)
.unwrap();
}
fn sample_row() -> TurnStatRow {
TurnStatRow {
started_at: 100,
ended_at: 101,
duration_ms: 1_000,
model: "opus".to_owned(),
wake_from: "recv".to_owned(),
input_tokens: 10,
output_tokens: 5,
cache_read_input_tokens: 0,
cache_creation_input_tokens: 0,
last_input_tokens: 10,
last_output_tokens: 5,
last_cache_read_input_tokens: 0,
last_cache_creation_input_tokens: 0,
tool_call_count: 1,
tool_call_breakdown_json: None,
open_threads_count: None,
open_reminders_count: None,
result_kind: "ok",
note: None,
session_id: None,
}
}
/// Regression: a pre-session db must `open()` cleanly (the `session_id`
/// index used to live in `SCHEMA` and aborted the apply with
/// `no such column`, silently disabling the stats sink), get the column
/// added by `MIGRATIONS`, and then accept writes.
#[test]
fn open_upgrades_pre_session_db_and_writes() {
let path = std::env::temp_dir().join("hyperhive-turnstats-pre-session-regression.sqlite");
let _ = std::fs::remove_file(&path);
seed_pre_session_db(&path);
let stats = TurnStats::open(&path).expect("open() must succeed on a pre-session db");
stats.record(&sample_row());
let conn = Connection::open(&path).unwrap();
let rows: i64 = conn
.query_row("SELECT COUNT(*) FROM turn_stats", [], |r| r.get(0))
.unwrap();
assert_eq!(rows, 1, "the row must be written once the column is added");
let has_session_id: i64 = conn
.query_row(
"SELECT COUNT(*) FROM pragma_table_info('turn_stats') \
WHERE name = 'session_id'",
[],
|r| r.get(0),
)
.unwrap();
assert_eq!(has_session_id, 1, "MIGRATIONS must add session_id");
let _ = std::fs::remove_file(&path);
}
}