m17
Fitur "premium-feel" yang bikin agent kerasa mikir, ngerencanain, dan jalan sendiri. Semua yang nyentuh dana TETAP lewat governor; yang nge-drive eksekusi (planner/swarm/automation/skill_forge) masuk
SKILL.mdskills/superagent/m17/SKILL.md
GitHubm17 — Power Pack: Planner · Swarm · Automation · Backtest · Dashboard · Voice (NEW in v4.0)
Fitur "premium-feel" yang bikin agent kerasa mikir, ngerencanain, dan jalan sendiri. Semua yang nyentuh dana TETAP lewat governor; yang nge-drive eksekusi (planner/swarm/automation/skill_forge) masuk FROZEN_PATHS.
| Fitur | Script | Inti |
|---|---|---|
| Workflow planner | tools/planner.py |
tujuan NL → plan multi-step → eksekusi bertahap (gated) |
| Multi-agent swarm | tools/swarm.py |
specialist lane paralel + pemisahan key |
| Skill forge | tools/skill_forge.py |
agent draft skill baru → proposal (gak auto-apply) |
| Automation engine | tools/automation.py |
WHEN trigger THEN actions |
| Backtester | tools/backtest.py |
replay strategi di data historis |
| Live dashboard | tools/dashboard.py |
HTML real-time (portfolio/gas/governor/audit) |
| Voice mode | tools/voice.py |
STT→LLM→TTS, ngobrol kayak nelpon |
| Explainability | tools/explain.py |
rangkum audit trail "apa & kenapa" |
1. Workflow planner (Tier 1)
from planner import make_plan, execute_plan
plan = make_plan("isi base lalu beli token 0xABC di 5 wallet") # LLM via model_registry
print(plan.summary()) # step 💸 = nyentuh dana
await execute_plan(plan, handlers, confirm_cb) # checkpoint + gated
Agent dekomposisi tujuan jadi langkah dari skill library, tiap step fund lewat handler yang governor-gated. Plan ditampilin dulu sebelum jalan.
2. Multi-agent swarm (Tier 1)
from swarm import Swarm, Specialist
sw = Swarm([Specialist("researcher", researcher_fn),
Specialist("monitor", monitor_fn),
Specialist("executor", executor_fn, holds_keys=True)])
await sw.dispatch({"researcher": {...}, "monitor": {...}}) # PARALEL
sw.safety_report() # warning kalau lane non-executor pegang key
Pemisahan key: lane watcher/monitor (selalu online) gak pegang key; cuma executor yang sign.
3. Skill forge (Tier 1, gated)
from skill_forge import forge_proposal
forge_proposal("belum bisa baca Chainlink oracle", "m18") # → proposal, REVIEW dulu
Agent numbuhin kemampuannya sendiri — tapi lewat proposal (x4/frozen-paths), operator yang acc + re-lock.
4. Automation engine (Tier 3)
from automation import AutomationEngine
ae = AutomationEngine()
ae.add_rule("gas murah → swap", "price", {"metric": "gas"},
actions=[{"action": "swap", "params": {...}, "funds": True}])
await ae.fire(event, action_handlers, confirm_cb) # aksi fund → governor
IFTTT buat crypto/ops: trigger (webhook/schedule/price/onchain_event/custom) → rangkaian aksi.
5. Backtester (Tier 3)
from backtest import backtest
res = backtest(my_strategy, candles, initial=1000)
res.metrics() # return_pct, win_rate, trades, drawdown
Tes strategi di data historis sebelum live. Bukan ramalan — alat validasi.
6. Live dashboard (Tier 2)
from dashboard import build_dashboard_html, gather
html = build_dashboard_html(gather(governor=gov, alert_engine=ae, ...))
# serve dari VPS (m2/m9). KASIH AUTH, jangan public polos.
Real-time: portfolio, gas, status governor, alert aktif, proposal pending, audit. Auto-refresh 30s.
7. Voice mode (Tier 2)
from voice import converse
converse("voice_note.ogg") # STT (Whisper lokal) → LLM (cascade) → TTS (pyttsx3 lokal)
Dua arah: lo ngomong, agent jawab suara. Lengkapin voice input m15.
8. Explainability (Tier 2/3)
from explain import explain
print(explain()) # "apa yang agent lakuin & kenapa" dari audit trail
Catatan
- Planner/skill-forge/swarm decomposition pakai LLM via
model_registry(m7/v4.0) —add modeldulu kalau cascade kosong. - Dashboard butuh port di-expose — amanin dengan auth, default localhost.
- Voice TTS:
pip install pyttsx3(gratis, offline). Backtester & explain keyless. - Swarm multi-proses pakai delegate OpenClaw — verify API current saat scale.