Agent Runtime

Architecture

The agent runtime is a set of coordinated processes. Your main agent orchestrates background daemons that monitor markets, evaluate triggers, and execute autonomously.

Smart Mode

--smart adds an LLM intelligence layer to daemons. Uses a cheap model (Haiku 4.5) via the proxy API for ~$0.0001/call.

sf intent buy KXFEDDEC2509 500 --price 36 \
  --trigger below:40 \
  --soft "only if oil sustained above $95 and no dovish fed signals"

# Inside sf agent, the LLM can call:
set_wake_condition({
  description: "alert if bid depth collapses on KXFEDDEC",
  type: "llm",
  condition: { prompt: "bid depth < 100 contracts at any price level" }
})

Message Bus

Daemons communicate via ~/.sf/bus/. No Redis, no sockets — just JSON files with atomic rename.

Messages have TTL (default 1 hour). Expired messages are auto-cleaned. If the agent is inactive and an alert is critical, it escalates to Telegram.

QuoteEngine v2 — multi-venue

Automated market making daemon. Places bid+ask on Kalshi and Polymarket, auto-requotes on fills and orderbook moves. Venue is auto-detected from the marketId shape — Kalshi tickers (KX...) or Polymarket YES tokenIds (decimal uint256, or 0x-hex which is auto-normalized). Supports biased quoting, paper trading with simulated fills, dynamic spread, and thesis-driven auto-bias.

# Start QuoteEngine daemon
sf quoteengine start --daemon

# Paper trading with bias (Kalshi)
sf quote create kalshi:KXRECSSNBER-26 --spread 2 --size 5 --paper --bias 2

# Paper trading on Polymarket (poly: prefix or bare token id, venue auto-detected)
sf quote create poly:0x66abcdef1234567890abcdef1234567890abcdef1234567890abcdef12345678 --paper --spread 2 --size 5

# Thesis-driven auto-bias (manual: pin a specific thesis)
sf quote create TICKER --spread 3 --size 10 --thesis-id UUID

# Auto-pick the best matching thesis for this ticker (no UUID needed)
sf quote create kalshi:KXBTC-26APR0817-B69875 --spread 2 --size 5 --paper --thesis-auto

# Post-fill fade + layered quotes + debounce
sf quote create TICKER --spread 2 --fade 3 --layers 3 --requote-delay 200

# Check status (shows bias, spread, P&L, fills)
sf quoteengine status

Quick Start

# Start smart runtime daemon
sf runtime start --smart --daemon

# Start QuoteEngine daemon
sf quoteengine start --daemon

# Start telegram bot
sf telegram --daemon

# Launch agent (auto-connects to all daemons)
sf agent

# Inside agent: set a wake condition
> "wake me if any recession contract moves more than 5 cents"

# Inside agent: start market making
> "create a paper quote on KXRECSSNBER-26 with 2 cent spread"

# Inside agent: check daemon observations
> "any alerts from the quoteengine?"

# Inside agent: send command to runtime
> "tell the runtime to run a scan now"

Configuration

# ~/.sf/config.json
{
  "smartDaemons": true,           // enable by default (no --smart flag needed)
  "smartModel": "anthropic/claude-haiku-4.5",  // cheap, fast
  "smartScanInterval": 10,        // scan every 10 runtime cycles (~5 min)
  "smartBudgetUsd": 1.00,         // monthly LLM budget cap
  "smartEscalateToTelegram": true, // critical alerts → Telegram when agent offline
  "syncPositions": true           // push local positions to heartbeat on agent start
  // closedLoop is per-thesis (set via sf heartbeat <id> --closed-loop-entry/exit)
}

Agent Tools

Workspace Filesystem

The agent has a local filesystem at ~/.sf/workspace/. Use it as a scratchpad for notes, analysis, and data exports. Shared across TUI and Telegram.

Self-Wake

The agent can schedule itself to wake up after idle. Useful for monitoring developing situations, periodic checks, or timed reminders. Race-safe: wakes only fire when the user is not mid-conversation.

# Inside sf agent:
> "check oil prices every 30 minutes and alert me if WTI drops below $90"
# Agent calls schedule_wake(30, 'prompt', 'Check WTI...', repeat: 30)

Voice I/O

Talk to the agent with voice. TTS and STT powered by Cartesia via server proxy.

Position Sync & Closed Loop

# Enable in ~/.sf/config.json
"syncPositions": true

# Positions are cross-matched with thesis edges server-side.
# Only positions that match a thesis edge are stored.
# Syncs once per agent session (on startup).

# Entry: strategy.entryBelow=35, price hits 34¢ → BUY intent
sf heartbeat <id> --closed-loop-entry

# Exit: strategy.stopLoss=20, price hits 19¢ → SELL intent
# Exit: strategy.takeProfit=60, price hits 62¢ → SELL intent
sf heartbeat <id> --closed-loop-exit

# Both:
sf heartbeat <id> --closed-loop-entry --closed-loop-exit

# Backtest before enabling:
sf backtest KXFED --entry-below 35 --stop 20 --tp 60 --days 30

Cost

Smart mode uses Haiku 4.5 via proxy. Typical costs:

Usage tracked in ~/.sf/bus/cost/smart-usage.jsonl. Budget enforced per-call — if exceeded, daemons fall back to mechanical mode.