# Agent Team Memory

## Purpose

This file stores durable decisions about how the Stock Predictor agent team
should collaborate.  It is process memory, not model memory.

## Current State

- The reusable team has eight roles:
  coordinator, two survey agents, two implementation agents, two QA agents, and
  one optional promotion/memory guardian.
- Survey agents do not promote production changes.
- Implementation agents should prefer dry-run experiment scripts before
  touching production predictor or recommendation files.
- QA agents can veto promotion when no-lookahead validation, BUY precision,
  signal count, or report evidence is insufficient.
- The promotion/memory guardian is recommended whenever a candidate might
  affect `/predict`, `/recommend`, precompute, HF cache, or Telegram output.

## Decisions To Preserve

- Use pair programming for implementation: one agent writes, one agent reviews
  scope and side effects.
- Use pair validation for QA: one agent designs tests, one agent runs historical
  validation.
- Keep detailed results in `docs/validation_runs/`; memory should store compact
  lessons only.
- If a candidate passes only on a small stock set but fails on a larger set,
  keep it experimental.

## Update Rule

Append here only when the team structure, role responsibilities, validation
authority, or promotion policy changes.