Arche generates a small map so coding agents load the right context without stuffing the entire repository into a prompt.
Agent context map
AGENTS.md
Entry map for every agent
CLAUDE.md
Symlink → AGENTS.md
.docs/
Internal architecture topics
.plans/active/
One approved in-flight plan
arche.json
Machine-readable scaffold metadata
Read root → nearest workspace AGENTS.md → one active plan
Canonical files
AGENTS.md— entry map; other agent files should point here.CLAUDE.md— symlink toAGENTS.mdwhen the preset enables it..docs/— internal topics (architecture, deployment, features)..plans/active/— one approved in-flight plan at a time.arche.json— machine-readable metadata about the scaffold.
What is not generated
.cursor/rules/or.claude/rules/duplicate trees- Stale “production ready” claims without verification evidence
Agent workflow
Loading diagram…
- Read root
AGENTS.md. - Read the nearest workspace
AGENTS.mdfor the package you touch. - Load one matching
.plans/active/file only for approved work.
When commands or env names change, update the nearest AGENTS.md and the owning .docs/ topic.