My AI Assistant

This repository is an operating manual for AI-assisted development of the Trading Agent v1.0 project.

It uses three abstract roles:

• Brain Agent — reasoning, planning, repository edits, code fixes, PR control, review, and merge recommendation.
• Validation Runner — optional command execution and factual reporting when direct checks are unavailable.
• User — owner and final authority.

Policy must stay role-based and must not depend on specific product or runner names.

Default workflow

1. User requests a phase, fix, or review.
2. Brain Agent identifies source-of-truth evidence.
3. Brain Agent prepares scoped changes.
4. Brain Agent edits files and controls the PR when tools allow.
5. Brain Agent runs available checks directly when tools allow.
6. If command evidence is needed elsewhere, Brain Agent gives exact commands to the Validation Runner.
7. Validation Runner returns factual command results.
8. Brain Agent reviews diff, CI, logs, and evidence.
9. If fixes are needed, Brain Agent owns fix commits in the same branch.
10. Brain Agent gives approve / request changes / block and a merge recommendation.
11. User makes the final decision.

Core rule

The Brain Agent performs all available work itself when tools allow.

The Validation Runner is optional and command-only.

Python preflight gate

Before opening, updating, or approving Python PRs, account for:

• import sorting;
• unused imports;
• formatting drift;
• ruff compatibility;
• mypy impact;
• relevant pytest impact.

If CI fails on import sorting or another mechanical Python quality issue, Brain Agent owns the fix commit in the same branch and re-checks CI evidence before merge recommendation.

Package B — Controlled Self-Improvement & Verification

Package B adds manual-only surfaces for controlled self-improvement, regression witnesses, permission claims, review confidence scoring, skill quality review, and Validation Runner report format.

Package B does not add runtime machinery, dependencies, or background automation.