refactor(hive-c0re): group src-root files into submodules
stores/ (sqlite-backed host stores + db helper), stats/, agent_config/, workers/ — pure git-mv moves; crate-root re-exports keep every crate::<module> path compiling. flake_check stays at root (synchronous approval-flow validation, not a background worker)
This commit is contained in:
parent
b489454dc2
commit
0e4b5a1120
29 changed files with 68 additions and 24 deletions
450
hive-c0re/src/stats/hive_stats.rs
Normal file
450
hive-c0re/src/stats/hive_stats.rs
Normal file
|
|
@ -0,0 +1,450 @@
|
|||
//! Hive-wide turn-stats aggregation for the dashboard's swarm stats
|
||||
//! view. Reads every agent's per-agent
|
||||
//! `hyperhive-turn-stats.sqlite` read-only and rolls the rows up into
|
||||
//! swarm totals + a per-agent rollup + model mix + a *labelled*
|
||||
//! cost estimate.
|
||||
//!
|
||||
//! Why re-read the rows here instead of reusing `hive-ag3nt`'s
|
||||
//! `stats.rs`: that module lives in a different crate (the agent
|
||||
//! harness) which hive-c0re can't import. The stable contract is the
|
||||
//! turn-stats *schema*, so we run a focused query against it. If we
|
||||
//! ever want a single source of truth, the row-read + aggregation can
|
||||
//! be lifted into a shared crate — overkill for now.
|
||||
//!
|
||||
//! Privsep: the sqlite files are mode 0644 owned by the agent user;
|
||||
//! `hive-core` can read them fine, but cannot write/delete (which is why
|
||||
//! retention sweeps run agent-side in the harness, not here). We open
|
||||
//! read-only so an in-flight harness writer never blocks us.
|
||||
|
||||
use std::collections::HashMap;
|
||||
use std::path::Path;
|
||||
use std::time::Duration;
|
||||
|
||||
use rusqlite::{Connection, OpenFlags};
|
||||
use serde::{Deserialize, Serialize};
|
||||
|
||||
use crate::coordinator::Coordinator;
|
||||
use hive_sh4re::wire_time::now_unix;
|
||||
|
||||
/// Window accepted by `/api/stats-hive?window=`. Maps to a lookback
|
||||
/// span; the hive view is a flat rollup (no per-bucket trend — the
|
||||
/// per-agent `/stats` page owns the trend charts).
|
||||
#[derive(Debug, Clone, Copy)]
|
||||
pub enum Window {
|
||||
Hour,
|
||||
FourHour,
|
||||
Day,
|
||||
ThreeDay,
|
||||
Week,
|
||||
Month,
|
||||
/// All available data: aggregate over every recorded turn since the
|
||||
/// earliest (`from == 0`), with no fixed lookback.
|
||||
All,
|
||||
}
|
||||
|
||||
impl Window {
|
||||
#[must_use]
|
||||
pub fn parse(s: &str) -> Self {
|
||||
match s {
|
||||
"1h" => Self::Hour,
|
||||
"4h" => Self::FourHour,
|
||||
"3d" => Self::ThreeDay,
|
||||
"7d" => Self::Week,
|
||||
"30d" => Self::Month,
|
||||
"all" => Self::All,
|
||||
_ => Self::Day,
|
||||
}
|
||||
}
|
||||
|
||||
fn label(self) -> &'static str {
|
||||
match self {
|
||||
Self::Hour => "1h",
|
||||
Self::FourHour => "4h",
|
||||
Self::Day => "24h",
|
||||
Self::ThreeDay => "3d",
|
||||
Self::Week => "7d",
|
||||
Self::Month => "30d",
|
||||
Self::All => "all",
|
||||
}
|
||||
}
|
||||
|
||||
fn span_secs(self) -> i64 {
|
||||
match self {
|
||||
Self::Hour => 3600,
|
||||
Self::FourHour => 4 * 3600,
|
||||
Self::Day => 24 * 3600,
|
||||
Self::ThreeDay => 3 * 24 * 3600,
|
||||
Self::Week => 7 * 24 * 3600,
|
||||
Self::Month => 30 * 24 * 3600,
|
||||
// `All` has no fixed lookback; `hive_snapshot` uses `from == 0`
|
||||
// (every recorded turn). 0 here is a consistent safe fallback.
|
||||
Self::All => 0,
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/// Approximate USD price per **million** tokens for one model class.
|
||||
/// The four fields map to the four token kinds the turn-stats schema
|
||||
/// records. This is a deliberately rough estimate — Anthropic list
|
||||
/// pricing drifts, so the dashboard labels the figure as an estimate.
|
||||
///
|
||||
/// Operators keep the table current without a code change via the
|
||||
/// `services.hyperhive.modelPrices` nix option (passed to
|
||||
/// `hive-c0re serve --model-prices <json>`, held on the `Coordinator`
|
||||
/// as a [`PriceTable`]). Any model not covered by the operator table
|
||||
/// falls back to [`builtin_prices`].
|
||||
#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
|
||||
pub struct Prices {
|
||||
pub input: f64,
|
||||
pub output: f64,
|
||||
pub cache_read: f64,
|
||||
pub cache_write: f64,
|
||||
}
|
||||
|
||||
/// Operator-tunable model→price map. Keys are matched
|
||||
/// case-insensitively as a substring of the model id (e.g. `"sonnet"`
|
||||
/// matches `"claude-sonnet-4-5"`); the longest matching key wins so a
|
||||
/// specific entry beats a generic family name. An empty map means
|
||||
/// every model uses [`builtin_prices`].
|
||||
pub type PriceTable = HashMap<String, Prices>;
|
||||
|
||||
/// Built-in fallback pricing, per **million** tokens. Used when the
|
||||
/// operator's [`PriceTable`] has no key matching the model. Figures are
|
||||
/// current Anthropic list prices for the Claude 4.x family (Opus 4.x,
|
||||
/// Sonnet 4.x, Haiku 4.5); `cache_write` uses the 1-hour cache-TTL price,
|
||||
/// which is the default through the Claude subscription the agents run
|
||||
/// on. Keep in sync with the `services.hyperhive.modelPrices` nix default
|
||||
/// (`nix/modules/hive-c0re.nix`).
|
||||
fn builtin_prices(model: &str) -> Prices {
|
||||
let m = model.to_ascii_lowercase();
|
||||
if m.contains("opus") {
|
||||
Prices {
|
||||
input: 5.0,
|
||||
output: 25.0,
|
||||
cache_read: 0.5,
|
||||
cache_write: 10.0,
|
||||
}
|
||||
} else if m.contains("haiku") {
|
||||
Prices {
|
||||
input: 1.0,
|
||||
output: 5.0,
|
||||
cache_read: 0.1,
|
||||
cache_write: 2.0,
|
||||
}
|
||||
} else {
|
||||
// sonnet + unknown fallback
|
||||
Prices {
|
||||
input: 3.0,
|
||||
output: 15.0,
|
||||
cache_read: 0.3,
|
||||
cache_write: 6.0,
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/// Resolve the price for `model`: prefer the operator-provided
|
||||
/// `table` (longest matching key wins, so a specific
|
||||
/// `claude-3-5-sonnet` entry beats a generic `sonnet`), else fall back
|
||||
/// to [`builtin_prices`].
|
||||
fn resolve_prices(model: &str, table: &PriceTable) -> Prices {
|
||||
let m = model.to_ascii_lowercase();
|
||||
let mut best: Option<(usize, Prices)> = None;
|
||||
for (k, v) in table {
|
||||
let key = k.to_ascii_lowercase();
|
||||
if key.is_empty() || !m.contains(&key) {
|
||||
continue;
|
||||
}
|
||||
let better = match best {
|
||||
Some((blen, _)) => key.len() > blen,
|
||||
None => true,
|
||||
};
|
||||
if better {
|
||||
best = Some((key.len(), *v));
|
||||
}
|
||||
}
|
||||
match best {
|
||||
Some((_, p)) => p,
|
||||
None => builtin_prices(&m),
|
||||
}
|
||||
}
|
||||
|
||||
#[derive(Debug, Serialize)]
|
||||
pub struct KeyCount {
|
||||
pub key: String,
|
||||
pub count: u64,
|
||||
}
|
||||
|
||||
#[derive(Debug, Serialize)]
|
||||
pub struct AgentRollup {
|
||||
pub name: String,
|
||||
pub turns: u64,
|
||||
pub input_tokens: u64,
|
||||
pub output_tokens: u64,
|
||||
pub cache_read_tokens: u64,
|
||||
pub cache_creation_tokens: u64,
|
||||
/// Labelled estimate — see [`Prices`].
|
||||
pub est_cost_usd: f64,
|
||||
}
|
||||
|
||||
#[derive(Debug, Serialize)]
|
||||
pub struct HiveStats {
|
||||
pub window: &'static str,
|
||||
pub from: i64,
|
||||
pub now: i64,
|
||||
/// Number of agents that had at least one turn in the window.
|
||||
pub active_agents: u64,
|
||||
pub total_turns: u64,
|
||||
pub total_input_tokens: u64,
|
||||
pub total_output_tokens: u64,
|
||||
pub total_cache_read_tokens: u64,
|
||||
pub total_cache_creation_tokens: u64,
|
||||
/// Labelled estimate — see [`Prices`].
|
||||
pub est_cost_usd: f64,
|
||||
/// Per-agent rollup, busiest (most turns) first.
|
||||
pub agents: Vec<AgentRollup>,
|
||||
/// Turns per model across the whole swarm, busiest first.
|
||||
pub model_mix: Vec<KeyCount>,
|
||||
/// Most-run normalised bash-command heads ("favorite tools") across
|
||||
/// the whole swarm, busiest first, capped to 10. Empty until the
|
||||
/// hive-bash-mcp capture has recorded `bash_commands` rows on at
|
||||
/// least one active agent.
|
||||
pub bash_mix: Vec<KeyCount>,
|
||||
}
|
||||
|
||||
#[derive(Default)]
|
||||
struct AgentAgg {
|
||||
turns: u64,
|
||||
input: u64,
|
||||
output: u64,
|
||||
cache_read: u64,
|
||||
cache_creation: u64,
|
||||
cost: f64,
|
||||
models: HashMap<String, u64>,
|
||||
/// Normalised bash-command head → invocation count, from the agent's
|
||||
/// `bash_commands` table (written by hive-bash-mcp). Empty when that
|
||||
/// capture hasn't run for this agent (table absent).
|
||||
bash: HashMap<String, u64>,
|
||||
}
|
||||
|
||||
#[allow(
|
||||
clippy::cast_sign_loss,
|
||||
clippy::cast_possible_truncation,
|
||||
reason = "v.max(0) clamps to a non-negative value first, so the i64->u64 cast is exact: no sign loss, and no truncation since u64 covers all non-negative i64"
|
||||
)]
|
||||
fn u64_from_i64(v: i64) -> u64 {
|
||||
v.max(0) as u64
|
||||
}
|
||||
|
||||
/// Aggregate one agent's turn-stats over `[from, now]`. Errors bubble
|
||||
/// up so the caller can skip a single bad/locked db without failing
|
||||
/// the whole endpoint.
|
||||
fn read_agent(path: &Path, from: i64, prices: &PriceTable) -> rusqlite::Result<AgentAgg> {
|
||||
let conn = Connection::open_with_flags(path, OpenFlags::SQLITE_OPEN_READ_ONLY)?;
|
||||
// turn_stats is rollback-journal (not WAL): a read landing while the
|
||||
// harness is mid-INSERT would get SQLITE_BUSY and drop that active
|
||||
// agent from the rollup. Wait out the brief write instead.
|
||||
conn.busy_timeout(Duration::from_millis(500))?;
|
||||
let mut stmt = conn.prepare(
|
||||
"SELECT model, input_tokens, output_tokens,
|
||||
cache_read_input_tokens, cache_creation_input_tokens
|
||||
FROM turn_stats
|
||||
WHERE started_at >= ?1",
|
||||
)?;
|
||||
let mut agg = AgentAgg::default();
|
||||
let rows = stmt.query_map([from], |row| {
|
||||
Ok((
|
||||
row.get::<_, String>(0)?,
|
||||
u64_from_i64(row.get::<_, i64>(1)?),
|
||||
u64_from_i64(row.get::<_, i64>(2)?),
|
||||
u64_from_i64(row.get::<_, i64>(3)?),
|
||||
u64_from_i64(row.get::<_, i64>(4)?),
|
||||
))
|
||||
})?;
|
||||
for r in rows {
|
||||
let (model, input, output, cache_read, cache_creation) = r?;
|
||||
agg.turns += 1;
|
||||
agg.input = agg.input.saturating_add(input);
|
||||
agg.output = agg.output.saturating_add(output);
|
||||
agg.cache_read = agg.cache_read.saturating_add(cache_read);
|
||||
agg.cache_creation = agg.cache_creation.saturating_add(cache_creation);
|
||||
let p = resolve_prices(&model, prices);
|
||||
#[allow(
|
||||
clippy::cast_precision_loss,
|
||||
reason = "token counts stay well under f64's 2^53 exact-integer range, so this cost computation loses no precision"
|
||||
)]
|
||||
{
|
||||
agg.cost += (input as f64 * p.input
|
||||
+ output as f64 * p.output
|
||||
+ cache_read as f64 * p.cache_read
|
||||
+ cache_creation as f64 * p.cache_write)
|
||||
/ 1_000_000.0;
|
||||
}
|
||||
*agg.models.entry(model).or_insert(0) += 1;
|
||||
}
|
||||
agg.bash = read_bash_heads(&conn, from);
|
||||
Ok(agg)
|
||||
}
|
||||
|
||||
/// Tally normalised bash-command heads from the agent's `bash_commands`
|
||||
/// table (`ts INTEGER, head TEXT`, written by hive-bash-mcp) over
|
||||
/// `[from, now]`. Best-effort + isolated from `read_agent`'s error path:
|
||||
/// a missing table (capture hasn't run for this agent) or any read error
|
||||
/// yields an empty map rather than propagating, so the favorite-tools
|
||||
/// rollup simply omits that agent and never fails the whole endpoint.
|
||||
/// `ts` is unix seconds, matching the `from` cutoff.
|
||||
fn read_bash_heads(conn: &Connection, from: i64) -> HashMap<String, u64> {
|
||||
let mut out: HashMap<String, u64> = HashMap::new();
|
||||
let Ok(mut stmt) =
|
||||
conn.prepare("SELECT head, COUNT(*) FROM bash_commands WHERE ts >= ?1 GROUP BY head")
|
||||
else {
|
||||
return out;
|
||||
};
|
||||
let Ok(rows) = stmt.query_map([from], |row| {
|
||||
Ok((
|
||||
row.get::<_, String>(0)?,
|
||||
u64_from_i64(row.get::<_, i64>(1)?),
|
||||
))
|
||||
}) else {
|
||||
return out;
|
||||
};
|
||||
for (head, count) in rows.flatten() {
|
||||
*out.entry(head).or_insert(0) += count;
|
||||
}
|
||||
out
|
||||
}
|
||||
|
||||
/// Build the swarm-wide rollup. Best-effort: a missing or unreadable
|
||||
/// per-agent db is skipped (logged), never fatal.
|
||||
#[must_use]
|
||||
pub fn hive_snapshot(window: Window, prices: &PriceTable) -> HiveStats {
|
||||
let now = now_unix();
|
||||
// Fixed windows look back a constant span; `all` aggregates every
|
||||
// recorded turn (`from == 0`). The hive rollup isn't time-bucketed, so
|
||||
// unlike the per-agent snapshot it needs no adaptive bucket sizing.
|
||||
let from = match window {
|
||||
Window::All => 0,
|
||||
_ => now - window.span_secs(),
|
||||
};
|
||||
|
||||
let mut agents: Vec<AgentRollup> = Vec::new();
|
||||
let mut model_mix: HashMap<String, u64> = HashMap::new();
|
||||
let mut bash_mix: HashMap<String, u64> = HashMap::new();
|
||||
let mut total_turns = 0u64;
|
||||
let mut total_input = 0u64;
|
||||
let mut total_output = 0u64;
|
||||
let mut total_cache_read = 0u64;
|
||||
let mut total_cache_creation = 0u64;
|
||||
let mut total_cost = 0.0f64;
|
||||
let mut active_agents = 0u64;
|
||||
|
||||
for name in Coordinator::kept_state_names() {
|
||||
let path = Coordinator::agent_harness_dir(&name).join("hyperhive-turn-stats.sqlite");
|
||||
if !path.exists() {
|
||||
continue;
|
||||
}
|
||||
let agg = match read_agent(&path, from, prices) {
|
||||
Ok(a) => a,
|
||||
Err(e) => {
|
||||
tracing::warn!(agent = %name, error = ?e, "hive-stats: read failed; skipping");
|
||||
continue;
|
||||
}
|
||||
};
|
||||
if agg.turns == 0 {
|
||||
continue;
|
||||
}
|
||||
active_agents += 1;
|
||||
total_turns += agg.turns;
|
||||
total_input = total_input.saturating_add(agg.input);
|
||||
total_output = total_output.saturating_add(agg.output);
|
||||
total_cache_read = total_cache_read.saturating_add(agg.cache_read);
|
||||
total_cache_creation = total_cache_creation.saturating_add(agg.cache_creation);
|
||||
total_cost += agg.cost;
|
||||
for (m, c) in &agg.models {
|
||||
*model_mix.entry(m.clone()).or_insert(0) += c;
|
||||
}
|
||||
for (h, c) in &agg.bash {
|
||||
*bash_mix.entry(h.clone()).or_insert(0) += c;
|
||||
}
|
||||
agents.push(AgentRollup {
|
||||
name,
|
||||
turns: agg.turns,
|
||||
input_tokens: agg.input,
|
||||
output_tokens: agg.output,
|
||||
cache_read_tokens: agg.cache_read,
|
||||
cache_creation_tokens: agg.cache_creation,
|
||||
est_cost_usd: agg.cost,
|
||||
});
|
||||
}
|
||||
|
||||
// Busiest agents first.
|
||||
agents.sort_by(|a, b| b.turns.cmp(&a.turns).then_with(|| a.name.cmp(&b.name)));
|
||||
|
||||
let mut model_mix: Vec<KeyCount> = model_mix
|
||||
.into_iter()
|
||||
.map(|(key, count)| KeyCount { key, count })
|
||||
.collect();
|
||||
model_mix.sort_by(|a, b| b.count.cmp(&a.count).then_with(|| a.key.cmp(&b.key)));
|
||||
|
||||
// Busiest commands first, capped to a top-10 "favorite tools" list.
|
||||
let mut bash_mix: Vec<KeyCount> = bash_mix
|
||||
.into_iter()
|
||||
.map(|(key, count)| KeyCount { key, count })
|
||||
.collect();
|
||||
bash_mix.sort_by(|a, b| b.count.cmp(&a.count).then_with(|| a.key.cmp(&b.key)));
|
||||
bash_mix.truncate(10);
|
||||
|
||||
HiveStats {
|
||||
window: window.label(),
|
||||
from,
|
||||
now,
|
||||
active_agents,
|
||||
total_turns,
|
||||
total_input_tokens: total_input,
|
||||
total_output_tokens: total_output,
|
||||
total_cache_read_tokens: total_cache_read,
|
||||
total_cache_creation_tokens: total_cache_creation,
|
||||
est_cost_usd: total_cost,
|
||||
agents,
|
||||
model_mix,
|
||||
bash_mix,
|
||||
}
|
||||
}
|
||||
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::*;
|
||||
|
||||
/// `read_bash_heads` tallies per-head counts over the window and is
|
||||
/// isolated from a missing table (returns empty, never errors).
|
||||
#[test]
|
||||
fn bash_heads_tally_and_window() {
|
||||
let conn = Connection::open_in_memory().unwrap();
|
||||
conn.execute_batch("CREATE TABLE bash_commands (ts INTEGER NOT NULL, head TEXT NOT NULL);")
|
||||
.unwrap();
|
||||
let now = now_unix();
|
||||
for (ts, head) in [
|
||||
(now - 100, "cargo"),
|
||||
(now - 200, "cargo"),
|
||||
(now - 300, "git"),
|
||||
(now - 10_000, "rg"), // outside a 1h-ish cutoff below
|
||||
] {
|
||||
conn.execute(
|
||||
"INSERT INTO bash_commands (ts, head) VALUES (?1, ?2)",
|
||||
rusqlite::params![ts, head],
|
||||
)
|
||||
.unwrap();
|
||||
}
|
||||
let heads = read_bash_heads(&conn, now - 3600);
|
||||
assert_eq!(heads.get("cargo").copied(), Some(2));
|
||||
assert_eq!(heads.get("git").copied(), Some(1));
|
||||
assert_eq!(heads.get("rg").copied(), None); // window cutoff excludes it
|
||||
}
|
||||
|
||||
/// A missing `bash_commands` table degrades to an empty tally rather
|
||||
/// than erroring — the pre-capture window on a fresh agent.
|
||||
#[test]
|
||||
fn bash_heads_missing_table_is_empty() {
|
||||
let conn = Connection::open_in_memory().unwrap();
|
||||
assert!(read_bash_heads(&conn, 0).is_empty());
|
||||
}
|
||||
}
|
||||
Loading…
Reference in a new issue