Skip to content

AGENTS.md

Latest commit

 

History

History
89 lines (64 loc) · 2.44 KB

AGENTS.md

File metadata and controls

89 lines (64 loc) · 2.44 KB
Raw

AGENTS.md - Multi-CLI AI Agent Bootstrap V2 Repository

Repository role

This repository is an AI workflow bootstrapper for other repositories. It is not a business application.

Its primary target is large Java/Spring/Spring Cloud projects where Codex CLI, Claude Code, or OpenCode should not attempt whole-repository understanding in a single pass.

Main V2 workflow

When initializing a target project:

  1. Run deterministic scanning first.
  2. Generate .ai/bootstrap/ artifacts.
  3. Generate module context packs.
  4. Ask the selected CLI (--cli codex|claude|opencode) to synthesize/refine project workflow files, or use local synthesis when --local-template is passed.
  5. Verify the generated workflow.
  6. Do not modify target business code during bootstrap.

Recommended command:

./bin/ai-start ../target-project
# or
./bin/ai-bootstrap ../target-project --deep --cli codex

Local deterministic fallback:

./bin/ai-start ../target-project --local-template
# or
./bin/ai-bootstrap ../target-project --deep --local-template

Write boundaries

During target project bootstrap, write only:

  • AGENTS.md
  • CLAUDE.md
  • .ai/**
  • .agents/**
  • .claude/**
  • .opencode/**
  • .githooks/**
  • .codex/**
  • opencode.json

Do not modify target Java/Python/JS/TS/Go/Rust business source, migrations, production config, or build scripts during initialization.

Scanner design rules

  • Keep scanners Python standard-library only.
  • Prefer direct file evidence over inference.
  • Mark uncertain information as unknown.
  • Do not read generated/vendor directories such as target, build, node_modules, .git.
  • For large repos, support include/exclude filters and max module limits.

Generated memory style

Generated memory files should be operational and concise.

Record:

  • module responsibilities
  • detected entry points
  • build/test commands
  • DB/cache/MQ risks
  • auth/transaction/remote-call risks
  • durable decisions
  • change-log discipline

Do not record:

  • secrets
  • production credentials
  • large pasted source code
  • guesses written as facts

Done criteria for this repository

A change is done when:

  1. python3 -m pytest -q passes.
  2. ./bin/ai-bootstrap <demo> --deep --local-template --force works.
  3. README commands match actual CLI behavior.
  4. Codex / Claude / OpenCode prompts constrain writes to the target workflow files, not business code.
  5. Existing V1 compatibility tests still pass unless deliberately removed.