Sloop is the local-first layer that makes your project context portable across AI CLIs and gives you a cross-repo view of all your agents. It sits above your AI coding tools and complements them: one canonical context, one fleet view, every tool.
This file is the public, contributor-facing roadmap. It links out to a design doc per pillar so a contribution starts from a clear contract. For the shipped detail of each release, see CHANGELOG.md.
- Provider-respecting. Never intercept or inject into an AI tool. Use the tool's own hook mechanism or non-invasive local signals only.
- Provider-aware by construction. All per-provider knowledge lives in one adapter manifest; no feature hardcodes a tool name. See docs/reference/ADAPTERS.md.
- One lightweight binary. CGO-free Go, no daemon, no bundled LLM. Add capability when the first real feature needs it, never a standalone "foundation" ahead of a consumer.
- Local-first & portable. Author context/skills once; every tool and every repo sees them.
| Pillar | What it is | Design doc | Status |
|---|---|---|---|
Launch (run) |
one friendly target → the right CLI, model and effort, in a managed session | docs/design/run.md | tool · model · effort shipped (Phase 1–2) |
| Portable context | AGENTS.md canonical; pointer files + skills delivered to every tool |
docs/reference/ADAPTERS.md | shipped (v0.1.0) |
| Skills | reusable prompts/workflows shared across tools and sources | docs/design/skills.md | shipping incrementally |
| Cross-repo fleet | every running agent across all repos; triage who needs you | docs/guide/USAGE.md | shipped (v0.1.0) |
| Hooks | precise agent status today; a portable workflow-hook library next | docs/design/hooks.md | status hooks shipped; workflow hooks → v0.2.0 |
- Dogfood
sloop psacross repos: does the cross-repo view triage "which agent needs me" better than single-project tools and raw tmux? - Skills: lockfile (
.sloop/skills.lock) +skills updatefor reproducible team skills. (shipped) - Hooks (status): auto-installers per provider: claude, gemini, cursor, copilot, codex done. (shipped)
sloop restore(shipped): relaunch the recent agents that aren't currently running, each in a detached session, so the whole fleet comes back after a reboot / tmux restart. Picks the most recent session per (workspace, tool) from the registry, skips any already live, and with--resumepasses each provider's resume flag (e.g. claude--continue) to continue the conversation. Auto-restore on boot would need a shell-init hook (no daemon), so this first cut stays a manual command.sloop tmux setup(planned): opt-in, append a managed/idempotent block to~/.tmux.conffor the bits that need to survive a tmux restart (popup keybind, optional global status line). Default stays per-session and non-invasive; this is not folded intohooks(which is AI-provider status).
The big one. Today's hooks only report status to sloop. v0.2.0 turns hooks into a portable workflow-automation library: pick a hook (format-on-edit, commit-policy, shell-guard, prompt-rule), and sloop installs it into the right tool's own hook config; author once, run across tools and repos. This is a large, security-sensitive surface, so it is designed before it is built. Full proposal, categories, project/user levels, cross-tool mapping, and the trust model are in docs/design/hooks.md.
- Skills registry (
skills search/add <name>from a curated index), after the lockfile proves out. init --scanLLM enrichment, only once a minimal LLM client exists for a real consumer.- Windows multiplexer verification (psmux); 2nd-brain / RAG bridge.
The lowest-barrier, highest-leverage contributions are data, not code:
- Add an AI tool: drop one
internal/adapter/builtin/<tool>.yaml; see docs/reference/ADAPTERS.md. No Go changes unless the tool needs a brand-new mechanism. - Share a skill: see docs/design/skills.md.
- Propose a workflow hook: see the open questions in docs/design/hooks.md; feedback on the model is wanted before the v0.2.0 build.