agent-control-center

One repository that wraps all local agent work: global agent rules, skills, guidelines, memory, and a projects/ directory that holds every project checkout. Clone this repo on a new machine and you have the whole setup; project code re-clones on demand from each project's manifest.

Agent- and model-agnostic by construction: every major agent CLI (opencode, Codex, Gemini, ...) discovers AGENTS.md by walking up the directory tree, so any session launched under projects/ inherits the global rules. Claude Code reads the same file through the one-line CLAUDE.md shim.

Use as a template

This repository is designed to be a GitHub template. To create your own control center:

1. Click "Use this template" on GitHub (or fork/copy the files).
2. Clone your new repo anywhere you like — the conventional location is ~/projects/agent-control-center, but nothing depends on it.
3. Run ./bin/bootstrap.sh --name <name>. The name identifies this center in ~/.config/agent-control-center/centers/ and must be unique per machine. It also gets written to the committed .control-center file, so every machine that clones this center uses the same name.
4. Make it yours: edit guidelines/ to your taste, and adjust AGENTS.md if your machine-sharing or commit rules differ.

Multiple control centers

You can run several centers on one machine (say, work and personal), each created from this template with its own name:

• Resolution is contextual. Skills and scripts resolve "which center am I in" by walking up from the current directory to the nearest .control-center marker that is registered in ~/.config/agent-control-center/centers/. Unregistered markers are skipped — a clone of this template under a project's code/ tree must not capture the walk-up. The registry's default center (set with bootstrap.sh --default, shown in ~/.config/agent-control-center/default) is used only for sessions outside every registered center.
• Guards compose. Every registered center's path guard runs in every session, so worktrees in all centers stay protected no matter where you work.
• Global rule imports go to the default center only (~/.claude/CLAUDE.md and opencode's global AGENTS.md). A session inside a non-default center still gets that center's rules through normal AGENTS.md directory walk-up.
• Sync covers all centers. Session hooks pull and push every registered center's docs.

Quickstart (each additional machine)

git clone <your-control-center-remote> ~/projects/agent-control-center
cd ~/projects/agent-control-center
./bin/bootstrap.sh   # checks deps, registers the center, wires hooks + links (idempotent)

The name travels in the committed .control-center file, so no --name is needed on additional machines.

Then, in any agent session: /resume-project to pick up existing work, or /start-project to begin something new.

Layout

Path	Committed	Purpose
AGENTS.md	yes	Global rules every agent session inherits
guidelines/	yes	Writing and coding guidelines
skills/	yes	Lifecycle skills: start / resume / finish project
memory/	yes	Global memory (INDEX.md + topic files)
templates/	yes	PROJECT.md, handoff, notes templates
bin/	yes	Scripts: worktrees, branch status, path guard, sync, bootstrap
projects/INDEX.md	yes	Registry of all projects
projects/<slug>/	yes	Manifest, handoffs, notes — the durable record
projects/<slug>/code/	no	Clones + worktrees — disposable, rebuilt from manifest

Project lifecycle

• /start-project — asks for a name and repos (or a project to extend), scaffolds the project directory from templates, blobless-clones each repo into code/<repo>/main, registers it in projects/INDEX.md.
• /resume-project — pulls this repo first (latest handoffs from any machine), reads the manifest + handoffs + notes, reconciles code/ (re-clones anything missing), verifies each active handoff's branch against git/PR state (bin/branch-status) and marks shipped work done, then summarizes. Continues a task's Next Steps only when the task is named: /resume-project <task>.
• /finish-project — refuses to run if any worktree has uncommitted or unpushed work; distills handoffs into the PROJECT.md outcome; promotes durable learnings to global memory; deletes code/; marks the project done.

Worktrees are created per task with bin/wt-new <repo> <task-slug>. It binds the task's handoff to the new worktree (frontmatter repo, branch, worktree) and heals a legacy worktree: none binding. A handoff without a worktree: key is unbound — normal for setup and docs-only tasks.

Worktree isolation

Agents working in the wrong worktree is the failure mode this repo is built against. Three layers:

1. Layout — each project's checkouts live under its own code/ directory; wrong siblings are not one cd .. away.
2. Instruction — AGENTS.md workspace rules (writes scoped to your project, repo-relative paths in docs, main is read-only).
3. Mechanical — bin/guard-path.sh scopes writes to the current project's worktrees and blocks any file operation that resolves into another project's code/ tree, wired as a Claude Code PreToolUse hook and an opencode tool.execute.before plugin. In bash commands it vets both absolute and relative paths, and treats paths in mutating commands (rm, mv, sed -i, redirects, …) as writes, so code/<repo>/main stays a read-only reference. The project is taken from the session's working directory, so a session may span its project's worktrees without relaunching.

Sync

bin/sync.sh keeps the remote current: pull at session start, push (scoped to projects/ + memory/ docs only) at session end, both wired via hooks by bootstrap.sh. A lock directory prevents concurrent sessions from racing.

Permission prompts

Claude Code allow rules cannot match a command wrapped in a variable assignment (CC=$(...)), and every subcommand of a compound command must match a rule on its own. The skills therefore run the resolver bare, and two layers ship the rules:

• .claude/settings.json (committed) allows the resolver for every session rooted in a center repo.
• link.sh merges per-machine allow rules for routine helper commands (wt-ls, wt-new, bare wt-prune dry run, sync.sh pull) of every registered center into ~/.claude/settings.json.

Mutating forms (wt-prune --apply, sync.sh push) still prompt on purpose.