m30

Superagent skill m30 — Client Revenue Engine (bulk gig, API-first)

SKILL.md

GitHub

m30 — Client Revenue Engine (bulk gig, API-first)

Load ONLY when: client revenue, garapan, bulk gig, harvest, scrape API, anti-browser, mass akun, otomasi job klien, ekstrak data massal


DOCTRINE — browser is the LAST resort

Garapan bulk = lakuin lewat kode & API, bukan klik-klik browser. Headless browser cuma kalau:

  1. Data cuma ada di DOM yang dirender JS, DAN
  2. Gak ada API/endpoint internal yang bisa direplikasi.

Decision tree tiap task:

Ada API resmi?            → pakai itu (httpx) ........................... STOP
Situs manggil XHR/fetch?  → tiru request-nya (DevTools → Copy as cURL → parse_curl) ... STOP
Butuh login / JS berat?   → headless browser SEKALI buat ambil token/cookie → balik ke API
Render-only, no API?      → headless browser (H8 / browser_engine) — pilihan TERAKHIR

NET-NEW TOOLS (v4.1.2)

  • tools/api_harvester.py — parse_curl() (DevTools → RequestSpec), paginate_offset / paginate_cursor, extract() (JSON path data.items[*].addr), to_jsonl / to_csv, send() (httpx, lazy import).
  • tools/revenue_engine.py — BulkRunner (concurrency + retry/backoff + checkpoint-resume + dedupe), TokenBucket (rate limit), Checkpoint.

Pola anti-browser standar:

  1. DevTools → Network → temukan call data → Copy as cURL.
  2. spec = parse_curl(curl) → replikasi di kode.
  3. paginate_* buat semua halaman; extract() buat ambil field.
  4. Bungkus di BulkRunner → rate-limit + resume kalau putus.

SCOPE & DELEGATION (m30 = orkestrator, BUKAN duplikat)

Pattern m30 nyetir Delegasi ke
A. Harvest / scrape API (anti-browser) parse_curl + paginate + extract + BulkRunner m6 (API), m5 (data shape), H8/browser_engine (last resort)
B. Mass akun / airdrop / on-chain BulkRunner + Checkpoint (resume) + dedupe m10 + hermes (wallet_manager, airdrop_runner) — WAJIB lewat governor.py
C. Integrasi API & otomasi job klien RequestSpec + retry + provider cascade m6 (webhook/SDK), m16 (scaffold service), m4 (schedule)

Kalau task murni 1 domain (mis. cuma batch tanpa API) → jangan load m30, pakai m12/m16 langsung. m30 nyala pas garapannya API-first + bulk + anti-browser.

SAFETY RAILS (non-negotiable)

  • On-chain apa pun di Pattern B → lewat Spend Governor (caps + kill-switch). Gak ada bypass walau auto_confirm.
  • Hormati rate-limit & ToS target; default TokenBucket konservatif — jangan hammer.
  • Jangan harvest data pribadi/PII di luar izin klien. Simpan kredensial klien di vault (hermes/references/security.md), bukan hardcode.
  • Resume-from-checkpoint harus idempotent — jangan double-submit / double-charge.

QUICKSTART

bash tools/tests/run_tests.sh      # buktiin engine-nya hijau (offline, stdlib)
from api_harvester import parse_curl, paginate_offset, extract, to_jsonl
from revenue_engine import BulkRunner, TokenBucket, Checkpoint

spec = parse_curl(open("call.curl").read())     # DevTools → Copy as cURL
# build fetch(offset, limit) dari spec pakai httpx, lalu:
rows = list(paginate_offset(fetch, limit=100))

runner = BulkRunner(
    worker,                                     # fn(task) -> hasil
    max_workers=8,
    rate=TokenBucket(5),                        # 5 req/detik
    checkpoint=Checkpoint("state.jsonl"),       # resume kalau putus
)
report = runner.run(rows)
print(len(report.succeeded), len(report.failed), len(report.skipped))

→ Next: tinggal isi fetch() / worker() sesuai garapan. 🔧 Upgrade: combo m4 buat jadwalin harian, m8 buat ekspor hasil ke XLSX klien.