# Find Bug

> Scan ~/tail-bot to find THE SINGLE MOST IMPACTFUL BUG — a real defect, not a style issue. Writes an ExecPlan to /Users/georgepickett/tail-bot/.agent/potential-bugs/<descriptive-name>.md. Use when user asks to find bugs, hunt for the worst bug, or surface the most dangerous defect in tail-bot.

- Skill: `grp06/find-bug` (Agent Skill)
- Install (CLI): `npx skillmds@latest add grp06/find-bug`
- Raw SKILL.md: https://api.skillmd.com/api/skills/grp06/find-bug/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: grp06 (https://skillmd.com/u/grp06)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/grp06/find-bug

---


# Find The Most Impactful Bug in tail-bot

## Mindset

You are hunting for **ONE real bug** that is most likely causing user-visible damage: wrong trades, lost money, silent failures, corrupted state, or incorrect calculations. Not style issues. Not hypothetical risks. Real defects with evidence.

**Be decisive. Do not ask clarifying questions.** Read the code. State assumptions and proceed.

## What tail-bot Is

A Rust trading bot for 5-minute Polymarket Up/Down markets. It ingests Chainlink price feeds and market websocket order books, evaluates a cheap-tail strategy, and either skips, simulates (shadow mode), or submits live orders. Key modules: `chainlink_ws.rs` (price feed), `market_ws.rs` (order book), `tail_strategy.rs` (decision engine), `trader.rs` (order submission), `runtime.rs` (orchestration), `shadow_report.rs` / `shadow_eval.rs` (PnL scoring).

## Bug Patterns to Prioritize

In a trading bot, these are the bugs that cost real money:

- **Stale data used for decisions** — websocket cache serving old prices/quotes past staleness thresholds, EWMA sigma computed from stale ticks, window-start price lookup returning wrong values
- **Edge/PnL calculation errors** — wrong fee/slippage math, inverted cheap-side selection (buying the wrong leg), off-by-one in time window boundaries, incorrect fill probability modeling
- **Race conditions in async flows** — websocket ingestion vs. strategy evaluation reading inconsistent state, discovery refresh clobbering in-flight evaluations
- **Silent swallowed errors** — API failures that return defaults instead of propagating, websocket disconnects that aren't detected, settlement status misclassification
- **Order submission bugs** — submitting with wrong price/size/side, dedupe file corruption, collateral check using stale balance
- **Shadow simulation divergence** — shadow fills computed differently than live would execute, making shadow PnL misleading for strategy tuning

## Workflow

### Step 1: Read the Codebase

Start with `ARCHITECTURE.md` and `README.md`, then trace the core decision pipeline through `runtime.rs` → `tail_strategy.rs` → `trader.rs`. Read the websocket ingestion paths (`chainlink_ws.rs`, `market_ws.rs`), the config/validation layer, and the shadow reporting pipeline.

### Step 2: Find Candidate Bugs

Look for real defects — not style issues. For each candidate, note the file, line range, and what goes wrong.

### Step 3: Score and Pick ONE

| Criteria | Weight |
|----------|--------|
| Severity (money lost, wrong trades, corrupted data) | 30% |
| Likelihood (how often real usage triggers it) | 25% |
| Blast radius (how many markets/flows affected) | 20% |
| Detectability (silent = worse) | 15% |
| Fix complexity (easy fix + high impact = prioritize) | 10% |

Present the single highest-scoring bug with: location, what goes wrong, what happens when it fires, and why it scored highest.

### Step 4: Write the ExecPlan

Read `PLANS.md` in `~/tail-bot`. If missing, read `$CODEX_HOME/skills/execplan-create/PLANS.md`. Follow that format exactly.

Choose a short descriptive kebab-case filename (e.g. `fix-stale-sigma-in-tail-entry.md`, `fix-inverted-cheap-side-selection.md`).

Write to `/Users/georgepickett/tail-bot/.agent/potential-bugs/<chosen-name>.md`.

The plan must be self-contained per PLANS.md — a novice with only this file can fix the bug. Include a regression test that fails before and passes after the fix.

Tell the user the path when done.

## Anti-Patterns

- **Asking clarifying questions** — read the code, make assumptions, state them.
- **Multiple bugs without picking one** — pick one, justify it, move on.
- **Style issues or hypothetical risks** — only real defects with evidence.
- **Symptom fixes** — fix root causes, not band-aids.
- **Vague acceptance criteria** — every criterion must be verifiable with a command.