Trader — personal desk + Agentic autonomy engine

Monorepo (dual desk). Repo: kenyip/trader · local: ~/dev/trader · package: trader_platform.

Desk	For	Do not
A — Personal tracker	Your positions, PMCC, TSLA/TSLL methods, AI coaching	Auto-trade main book
B — Agentic engine	Find → evaluate → wait → paper → (Ken-armed) live on $3k sleeve	Live without arm

Build / alignment (single doc): docs/TRADER_BUILD.md · Doc library: docs/README.md

Quick start

just setup                            # one-time: venv + deps
just positions                        # Desk A: your open positions (positions.yaml)
just pmcc-manage                      # Desk A: PMCC / LEAPS desk
just desk-brief                       # Desk A: daily gather for AI stance
just test                             # Desk A: TSLA/TSLL live recommendation
just run                              # Streamlit dashboard

# Desk B — discovery is tight sims; opportunity wait is separate
just trader-discover                  # tight multi-gen strategy search/proof
just trader-progress                  # progress bar + strategies that passed
just trader-progress --watch          # live refresh every 5s
just trader-opportunity               # patient watch + paper handoff (no evolve)
just trader-eval-iv-rich              # evaluate IV-rich seed
just trader-living                    # living seats
just trader-watch                     # watch only
just trader-paper-handoff             # dry-run paper intent when setup exists

just backtest && just scenarios       # classic engine validation
just analyze                          # critic loop (Desk A methods research)

Source-of-truth documents

Doc	Covers
docs/TRADER_BUILD.md	Only build bible — edge, pipeline, prove, authority, commands
docs/README.md	Full library map (detail vs research archive)
docs/TRADER_SPINE_ARCHITECTURE.md	Spine implementation detail
docs/DISCOVERY_AND_PAPER_FASTTRACK.md	Discovery / paper ops
GOAL.md	Legacy critic-loop / Desk A seed methods goal
ENGINE.md	Classic backtest harness
STRATEGY.md	Seed TSLA/TSLL rules history
docs/DESK_BRIEF.md	Desk A daily brief

Convention: current state at the top, dated history at the bottom.

Code map

# active engine
data.py                       yfinance loader + feature pipeline + regime classifier
pricing.py                    Black-Scholes price/greeks + strike-from-delta solver
backtest.py                   Position, Backtester event loop, metrics, wheel state machine
strategies.py                 StrategyConfig, DEFAULT_CONFIG_BY_TICKER, get_config(ticker),
                              pick_entry, check_exits, pick_covered_call (wheel)
scenarios.py                  canonical 12-regime windows (huge_down, flat, gap_shock,
                              vol_crush, vol_expansion, chop_whipsaw, earnings_window, ...)
run_backtest.py               CLI for the baseline backtest (--wheel flag for wheel mode)
run_scenarios.py              CLI for the canonical scenario suite
optimize.py                   walk-forward grid search + --static OOS validation of fixed config
sweep.py                      reusable knob-sweep harness (LLM-critic loop)
live.py                       today's recommendation (uses pick_entry + per-ticker get_config)
positions.py                  active-position tracker — runs exit ladder on user's open positions
manage_positions.py           CLI driver for positions (add/close/check/example)
tsla_options_dashboard.py     Streamlit dashboard — Today / Positions / Performance / Scenarios

# user-supplied (gitignored)
positions.yaml                user's open option positions, source of truth for `just positions`

# retired — kept as historical artifacts, do not extend or tune:
dynamic_parameter_engine.py   pre-engine "live" recommender (superseded by live.py)
strategy_v6_dynamic.py        thin wrapper for above
strategy_v5_optimized.py      scenario classifier class (logic absorbed into data.py)
walk_forward_optimizer.py     broken — replaced by optimize.py
tsla_tsll_options_tracker.py  one-line stub
strategy_v4.py / strategy_final.py / recent_performance.py / ab_test_v4_vs_v5.py  stubs
backtest_short_term_calls.py / backtest_strangle_early_exit.py  early synthetic-GBM experiments
sweep_*.py (multiple)         per-round bespoke sweep scripts from critic rounds 1-4 — all
                              superseded by sweep.py; kept for now as artifacts of each round.

Session arc (2026-05-12 to 2026-05-14)

Strategy went from v1.1 (−$6,242 on TSLA 5y, $12,292 max DD, profit factor 0.87) to v1.13 (+$17,790 on TSLA 5y, $1,107 max DD, profit factor 5.25, 92% win rate) — a +$24,032 swing with 91% DD reduction, via the LLM-critic loop and a fleet of 5 empirically-validated adaptive rules.

• v1.9 introduces the symmetric direction rule: regime tells us which side of the chain to sell (bullish/neutral → put, bearish → call), then per-ticker DTE × delta tunes for reversal-resistance.
• v1.10 decouples exits from entry signals per ticker. TSLA drops the regime_flip exit (+$2,214 P/L from letting positions ride to profit-based exits instead of being prematurely closed on regime change). TSLL effectively disables the max_loss dollar-threshold stop — the leveraged-ETF delta_breach=0.45 is the real tail protection.
• v1.11 lands the adaptive-rules architecture — adapt_entry_params(row, cfg) hook that lets context-aware rules adjust entry params on each bar using all available features. No rule ships in v1.11; the architecture is the extension point.
• v1.12 wires the data-driven proposer — analyze.py ingests trade-log + feature-panel + greeks-at-entry, ranks features by effect size, emits rule sketches in hook syntax. Ships tsll_skip_marginal_up (first adaptive rule) validated triple-win.
• v1.13 (this session) brings the loop to production: analyzer extensions (--scan-narrow with BH-FDR correction, --pairs for 2-feature interactions), per-position knob overrides (M4), an exit-side adaptive hook (M5), the validate_rule.py A/B harness, and 4 new adaptive rules all surfaced directly by the analyzer:
  • tsla_skip_mild_intraday_up — skip when 0.5% ≤ intraday_return ≤ 1.6% (triple-win: +$1,511 P/L, −$749 DD, +$1,604 OOS)
  • tsll_skip_tuesday — skip Tuesday entries (3-of-4 surfaces positive)
  • tsll_skip_post_earnings_drift — skip 11–21 days after earnings (triple-win)
  • tsll_skip_downtrend_high_iv — pair conjunction ret_14d × iv_rank (triple-win, from --pairs scan) 5 rules total; GOAL.md success criteria reached.

Full progression and per-version learning trail: see STRATEGY.md history and ENGINE.md history.

Disclaimer

Educational use only. This is research code, not trading advice.