Skip to main content
Trevor Grafton

PROJECT / hook-gated paper-trading agent

Trader Agent

A Claude Code skill for a hook-gated paper-trading agent.

Paper mode
  • Claude Code
  • Python
  • Hooks

  • Source private: private repository (touches a brokerage account)

Type
AI
Source
Private
Role
Sole builder, using an AI-assisted flow: brainstorm, spec, plan, then a subagent-driven build of 16 tasks with per-task and final reviews.
Built with
Claude Code, Python, Hooks

01 / Overview

Overview

  1. Scan

    Daily price data is scanned for pullback and breakout candidates.

  2. Filter

    A small ML filter and a Claude news check can each say no.

  3. Gate

    Code issues a single-use, expiring risk ticket; a hook allows orders only with one.

  4. Paper fill

    A paper broker simulates the fills and a journal records them.

  5. Review

    A report rates the strategy KEEP, WATCH or KILL; KILL halts new buys.

02 / Context and problem

Context and problem

An on-demand Claude Code agent (a /trader skill) for swing-trading liquid US stocks and ETFs through a broker's agentic tool server. A Python package does the deterministic work and a hook is the hard gate. It runs in paper mode; per the project log of 2026-10-12 the broker connection was not set up and it had made 0 trades.

03 / My role

My role

Sole builder, using an AI-assisted flow: brainstorm, spec, plan, then a subagent-driven build of 16 tasks with per-task and final reviews.

04 / How it works

How it works

  1. Fetch bars

    Daily bars come from a cached price feed.

  2. Scan

    Pullback and breakout signals are found and features computed.

  3. ML filter

    A small model filters candidates, and is live only if it beats rules-only and the index out of sample in a promotion test.

  4. News check

    A Claude gatekeeper reads a news verdict; its veto applies.

  5. Risk ticket

    A code risk gate sizes the trade, bounds the stop and issues a single-use expiring ticket.

  6. Guard hook

    The hook denies tools by default; an order needs a ticket.

  7. Paper broker and journal

    Fills are simulated a day later and written to a SQLite journal; the report gives KEEP, WATCH or KILL.

Python package

Scanning, features, the ML filter, risk sizing, the paper broker and the journal: all deterministic code.

Claude agent

The /trader skill reads news and can veto; the risk gate is code, not the model.

PreToolUse hook

The hard gate: default-deny tool allowlist, orders need a ticket, and it fails closed.

05 / Key decisions

Key decisions

Rules, a small ML filter and a Claude gatekeeper

The project log records this choice over a pure-LLM or reinforcement-learning agent.

Accept the failed v1 promotion test

The log records the failure was accepted, with no threshold tuning.

A default-deny hook as the hard gate

The hook denies tools by default, orders need a risk ticket, and it fails closed.

A ring-fenced money lock

A separate account funded by hand only, a default-deny tool allowlist, deny tokens, and crypto and options blocked.

An autonomy ladder

Four rungs: paper, live_approve, live_auto, live_full. Only I can promote, in a real terminal with a typed phrase, and live buys need a clean guard check within 6 hours.

06 / Results and numbers

Results and numbers

337 tests

collected by pytest during the October 2026 fact check

Measuredpytest collection during the October 2026 fact check (earlier log entries say 227, 257 and 270 at earlier dates)

07 / Testing and quality

Testing and quality

  • pytest runs on Windows in GitHub Actions, with a frozen-backtest regression test.
  • Fresh-context reviews found real holes: cancel-and-replace skipping tickets, NaN values skipping limits, the shell reaching the state file, and a Windows device name faking an interactive terminal.
  • The test count has changed over time (227, 257, 270 at earlier dates), so only the collection count above is quoted.

08 / Limitations

Limitations

  • Honest status: v1 failed its promotion test (8 selected trades, 30 required). Later versions passed on historical data only, with a thin edge and a survivorship-bias caveat. There is no live evidence.
  • Paper mode only; per the project log of 2026-10-12 the broker connection was not set up.
  • No live results exist for the news-adjusted trades.

09 / What I learned

What I learned

  • Filters used in training must also apply in the scan.
  • Safety guards also block harmless text.
  • Fresh-context review finds holes on the money path that the builder misses.
  • When a safety rule takes a number, check its units.

10 / What's next

What's next

  • Review weekly and collect at least 15 closed paper trades.
  • Check the news-adjusted trades.
  • Fix the deferred minor issues before any live automation.

11 / Next project

Next project

Homelab →