v1.0.0 — first stable release

An agent skill for safely taking over a large, unfamiliar codebase and doing secondary development without accidentally breaking modules you didn't know existed.

What it does

Two main modes + automatic self-healing:

• MAPPING — Deep recursive architecture analysis producing a .architecture/ wiki: stable module IDs (M####-<slug>), range-level + stable-anchor code locations, coverage classification (covered / partial / excluded / unknown), explicit risk + confidence tracking.
• DEVELOPMENT — Bounded code changes strictly inside the wiki's documented Allowed Edit Surface. Writes a dev plan first, executes autonomously.
• SELF-HEALING (auto, inside either mode) — When reality diverges from the wiki, analyzes the divergent area, updates the wiki, expands the plan, continues. User is informed but not interrogated.

Platform-agnostic

Works with any agent that supports the SKILL.md convention:

• Claude Code
• Codex CLI
• Cursor
• GitHub Copilot
• Gemini CLI
• OpenCode
• Cline
• Any other agent that loads SKILL.md + references/ + templates/

Architecture

SKILL.md                              # entry point, mode dispatch, P1–P9 principles
references/
  ├── analysis-protocol.md            # mapping protocol: phases 0–7, subagent strategy
  ├── development-protocol.md         # secondary dev protocol: phases 0–9
  ├── boundary-and-evidence-rules.md  # confidence levels, edge types, ID stability
  ├── self-healing-protocol.md        # divergence procedure
  └── domain-heuristics/              # per-domain references (blockchain, smart-contract,
                                      # web-saas, library-sdk, cli-tool)
templates/                            # 10 templates for every artifact under .architecture/

Install

See README for per-platform installation paths.

git clone git@github.com:Luis-yl/map-then-modify.git <agent-skills-dir>/map-then-modify

MIT licensed.