exitmgr — LLM-assisted options trading (propose → gate → approve → execute → manage)

A small system where the model proposes option swing trades and you approve every entry, while hard-coded risk rails and an exit manager keep the book bounded. Built for a small, ring-fenced IBKR pot.

Money warning. An LLM has no proven trading edge. Treat the pot as risk capital you are prepared to lose. Run on paper for weeks and watch the proposals before risking a cent.

Safety model (read this)

• Dry-run is the default. No order is ever placed without --arm.
• Even with --arm, every entry needs an explicit Slack approval (👍). ❌ or no reaction = skipped.
• Risk gate (risk.py) is hard code, not the LLM. Per current config:
  • ≤ 12% of the live pot per trade (sizing reads NetLiquidation each cycle — it scales as the pot moves; nothing hardcoded)
  • ≤ 4 concurrent positions
  • −8% daily circuit breaker → no new entries the rest of the day
  • ≤ 36% aggregate single-name exposure
  • universe = SPY / QQQ / IWM + any liquid single name the model proposes (allow_model_names: true) — every entry still needs your explicit 👍
• Kill switch: create the file at kill_switch.path → all order placement stops next cycle.
• Audit log: every proposal, gate decision, approval, and fill is appended to audit.jsonl.

First run (paper)

1. Start IB Gateway/TWS in paper mode (API enabled). Default port 7497.
2. python -m venv venv && source venv/bin/activate && pip install -r requirements.txt
3. Set SLACK_BOT_TOKEN in your env; set slack_channel + approver_ids (your Slack user id) in config.yaml.
4. Edit config.yaml: llm_endpoint/llm_model, approved_names, and (optional) pot_cap_usd.
5. Dry run (proposes + posts to Slack, places nothing):

   python run_trader.py --loop --interval 900
   

   Watch the Slack proposals and audit.jsonl for a good while. Confirm the picks and sizing look sane.
6. Tests: pytest -q (should be all green).

Going live (only after paper proves out)

python run_trader.py --arm --loop --interval 900

Now each gated proposal is posted for your 👍; approve the ones you want, the rest expire. The exit manager (main.py logic, folded into the loop) manages exits on whatever opens.

Components

file	role
exitmgr/strategist.py	LLM → strict-validated trade ideas
exitmgr/risk.py	hard risk gate (dynamic, pot-relative caps)
exitmgr/account.py	live pot value from IBKR (NetLiquidation)
exitmgr/approval.py	Slack approve-each (reject-wins, approver allowlist, expiry)
exitmgr/trader.py	orchestrator + audit log + day-start baseline
exitmgr/{connection,order,state,rules,manager}.py	execution + exit management

Known surfaces to validate on paper first

• Option contract selection in trader._execute_entry (chain → expiry near DTE → strike by delta → qualify). Broker-specific; watch fills on paper before trusting it live.
• Market context in trader._market_context is intentionally minimal — add real quotes/signals there.