~/develop umbrella setup puts AI agents at center of dev

The starting point was a simple preference: the developer wanted to move away from IDE-centric development in WebStorm and PHPStorm toward terminal-centric development with Claude Code alongside. That alone would not have been enough, because the work involves moving back and forth between multiple repositories, and setting up each repository individually did not cover it. What was wanted was a state where an agent could read the whole picture starting from the workspace and investigate, implement, and verify across the repositories it needs.

The pieces placed under ~/develop are a mix of agent-facing context, human-facing notes, and scripts. CLAUDE.md does not contain large numbers of individual implementation steps; instead it holds the judgment material agents tend to get stuck on — where repositories are and how they depend on each other, the difference between local domains and the remote dev environment, how ports and DB tunnels are handled, rules for issues, PRs, and review requests, and dialogue protocols for starting and finishing work. REPO_MAP.md organizes cross-repo relationships so that when a change spans several repos, reading it first narrows the investigation.

The author also contrasts this with issue-driven automation such as GitHub Copilot's cloud agent, where an issue starts the work, a PR is created, and a human reviews. That approach is described as attractive, but it depends strongly on the quality of the initial definition — what changes, what does not, which repositories are involved, and what counts as done. If that is left vague, the agent returns something plausible, and plausible is not the same as what was expected. In this setup, exploration and deciding the approach stay conversational, and the author expects the scope left to agents to define the real test of the workspace: not a flashy orchestrator, but commands that can confirm actual state and records that let you roll back when something goes wrong.