mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-07-12 13:52:15 +00:00
5c6438fd28
Real sessions DB (~/.hermes/state.db, 444 tui+cli sessions): p50=20, p75=53, p90=182, p95=340, p99=1941 msgs. The assumed 200-300 'realistic band' is actually the p90-p95 region; the typical session is ~20 msgs and the p99 tail reaches ~1940 (real ~2k-msg sessions exist). New cells at those anchors (2 reps, ink vs otui-capped, 2G scope), VmHWM medians: 100 msgs 163 vs 222MB; 300 msgs 180 vs 268MB; 2000 msgs 234 vs 671MB (2.9x). session-distribution.mjs regenerates the JSON; run.mjs gains --configs to skip redundant configs (otui-uncapped == capped below the 3000-row cap).
57 lines
2.8 KiB
JavaScript
57 lines
2.8 KiB
JavaScript
// bench/session-distribution.mjs — query the REAL sessions DB (~/.hermes/state.db,
|
|
// read-only) for the message_count distribution that frames the memory cells.
|
|
// Writes bench/session-distribution.json; render.mjs embeds it in the report.
|
|
//
|
|
// "message_count" here is DB messages (user/assistant/tool entries), the closest
|
|
// available unit to the bench's fixture rows. Not byte-identical semantics, but
|
|
// the same order of magnitude: both count transcript entries, not tokens.
|
|
//
|
|
// Usage: node session-distribution.mjs [path-to-state.db]
|
|
import { execFileSync } from 'node:child_process'
|
|
import { writeFileSync } from 'node:fs'
|
|
import { join, dirname } from 'node:path'
|
|
import { fileURLToPath } from 'node:url'
|
|
import { homedir } from 'node:os'
|
|
|
|
const here = dirname(fileURLToPath(import.meta.url))
|
|
const db = process.argv[2] ?? join(homedir(), '.hermes', 'state.db')
|
|
|
|
// No sqlite3 CLI on this box; the repo venv python is the canonical reader.
|
|
const py = join(here, '..', '.venv', 'bin', 'python')
|
|
const script = `
|
|
import sqlite3, json, math, sys
|
|
con = sqlite3.connect("file:" + sys.argv[1] + "?mode=ro", uri=True)
|
|
cur = con.cursor()
|
|
def pct(v, p):
|
|
if not v: return None
|
|
return v[max(0, min(len(v)-1, math.ceil(p/100*len(v))-1))]
|
|
def stats(rows):
|
|
v = sorted(r[0] for r in rows if r[0] is not None)
|
|
if not v: return None
|
|
buckets = [(0,10),(10,25),(25,50),(50,100),(100,200),(200,300),(300,500),(500,1000),(1000,3000),(3000,None)]
|
|
hist = []
|
|
for lo,hi in buckets:
|
|
c = sum(1 for x in v if x >= lo and (hi is None or x < hi))
|
|
hist.append({"lo": lo, "hi": hi, "count": c})
|
|
return {"n": len(v), "min": v[0], "max": v[-1], "mean": round(sum(v)/len(v),1),
|
|
"p50": pct(v,50), "p75": pct(v,75), "p90": pct(v,90), "p95": pct(v,95), "p99": pct(v,99),
|
|
"histogram": hist}
|
|
out = {
|
|
"sources": dict(cur.execute("SELECT source, COUNT(*) FROM sessions GROUP BY source").fetchall()),
|
|
"all": stats(cur.execute("SELECT message_count FROM sessions").fetchall()),
|
|
"tui_cli": stats(cur.execute("SELECT message_count FROM sessions WHERE source IN ('tui','cli')").fetchall()),
|
|
"tui_cli_top": [
|
|
{"msgs": r[0], "tools": r[1] or 0, "source": r[2], "title": (r[3] or "")[:60]}
|
|
for r in cur.execute("SELECT message_count, tool_call_count, source, title FROM sessions WHERE source IN ('tui','cli') ORDER BY message_count DESC LIMIT 10")
|
|
],
|
|
}
|
|
print(json.dumps(out, indent=2))
|
|
`
|
|
const json = execFileSync(py, ['-c', script, db], { encoding: 'utf8' })
|
|
const data = JSON.parse(json)
|
|
data.generated_at = new Date().toISOString()
|
|
data.db = db
|
|
const outFile = join(here, 'session-distribution.json')
|
|
writeFileSync(outFile, JSON.stringify(data, null, 2) + '\n')
|
|
const t = data.tui_cli
|
|
process.stdout.write(`tui+cli sessions n=${t.n}: p50=${t.p50} p75=${t.p75} p90=${t.p90} p95=${t.p95} p99=${t.p99} max=${t.max}\n→ ${outFile}\n`)
|