slim-claude
Orchestrator, specialist subagents, skills, and hooks for Claude Code. Adapted from oh-my-opencode-slim.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/aw4e--slim-claude)HTML
<a href="/listings/aw4e--slim-claude"><img src="/api/listings/aw4e--slim-claude/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
1. **Install**
/plugin marketplace add aw4e/slim-claude /plugin install slim-claude@slim-claude
To try it from a local checkout instead: `claude --plugin-dir <path-to-this-folder>`. 2. **Start with the orchestrator.** A plugin cannot make an agent the default, so either launch with it:
claude --agent slim-claude:orchestrator
or set it once in your own `~/.claude/settings.json` (or a project's `.claude/settings.json`):
{ "agent": "slim-claude:orchestrator" }Without this step the plugin still adds its agents and skills, but nothing delegates for you. 3. **Try it.** Ask for something with a few parts, for example: *"Find where auth is handled, then have someone review the design and implement the fix."* The orchestrator should dispatch `explorer`, `oracle`, and `fixer` as subagents. Run `/agents` to see everything the plugin added, and `/slim-claude:council <question>` for a multi-model second opinion. 4. **Keep costs in check.** `/slim-claude:preset` switches the model tier (`cheap`, `balanced`, `max`).
Nothing else is required. Obsidian, Orca, `rtk`, and a debugger backend are optional and skipped when absent.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 15 lines
- {
- "name": "slim-claude",
- "owner": {
- "name": "aw4e",
- "url": "https://github.com/aw4e"
- },
- "plugins": [
- {
- "name": "slim-claude",
- "source": "./",
- "description": "Orchestrator and specialist subagents for Claude Code."
- }
- ],
- "description": "Orchestrator, specialist subagents, skills, and hooks for Claude Code. Adapted from oh-my-opencode-slim."
- }
README.md · README · 135 lines
- # slim-claude
- <p align="center">
- <picture>
- <source media="(prefers-color-scheme: dark)" srcset="assets/logo-white.svg">
- <img src="assets/logo.svg" alt="slim-claude logo" width="96" height="96">
- </picture>
- </p>
- An orchestrator that plans work and hands it to specialist subagents (search, architecture review,
- UI, implementation, audits), plus skills and a few safety hooks, for Claude Code. Adapted from
- [oh-my-opencode-slim](https://github.com/alvinunreal/oh-my-opencode-slim) (MIT).
- > **Status: beta (v0.1.0).** Passes `claude plugin validate`; the hooks have automated tests (41 cases) and a
- > fresh-clone release gate. A live smoke suite (`scripts/smoke-live.mjs`, 36 real `claude -p` sessions on Windows)
- > confirms that with the orchestrator as main agent every agent is dispatched, read-only agents stay read-only,
- > the council runs 3 seats on 3 models plus a synthesis, and all skills load. Seven cases plant a known defect
- > (SQL injection, swallowed error, untested branch, duplicated code, off-by-one, var/closure) and check the audit
- > agents name it, plus one clean-code control; this measures recall on small cases, not general review quality. Not yet verified: interactive sessions, macOS and Linux. Issues welcome.
- ## Quick start
- 1. **Install**
- ```
- /plugin marketplace add aw4e/slim-claude
- /plugin install slim-claude@slim-claude
- ```
- To try it from a local checkout instead: `claude --plugin-dir <path-to-this-folder>`.
- 2. **Start with the orchestrator.** A plugin cannot make an agent the default, so either launch with it:
- ```
- claude --agent slim-claude:orchestrator
- ```
- or set it once in your own `~/.claude/settings.json` (or a project's `.claude/settings.json`):
- ```json
- { "agent": "slim-claude:orchestrator" }
- ```
- Without this step the plugin still adds its agents and skills, but nothing delegates for you.
- 3. **Try it.** Ask for something with a few parts, for example: *"Find where auth is handled, then have
- someone review the design and implement the fix."* The orchestrator should dispatch `explorer`,
- `oracle`, and `fixer` as subagents. Run `/agents` to see everything the plugin added, and
- `/slim-claude:council <question>` for a multi-model second opinion.
- 4. **Keep costs in check.** `/slim-claude:preset` switches the model tier (`cheap`, `balanced`, `max`).
- Nothing else is required. Obsidian, Orca, `rtk`, and a debugger backend are optional and skipped when absent.
- ## What's inside
- | Agent | Model | Job |
- |---|---|---|
- | orchestrator | session model | plans, delegates, verifies (run it as the main agent) |
- | explorer | haiku | read-only codebase search and docs research (uses `context7`, `gh_grep`) |
- | oracle | opus | architecture, hard debugging, review (read-only) |
- | designer | sonnet | UI/UX design and implementation |
- | fixer | sonnet | bounded implementation |
- | observer | sonnet | images, PDFs, screenshots, diagrams |
- | councillor / council | per seat / opus | multi-model consensus (see the `council` skill) |
- | janitor, paladin, arbiter, hunter, proctor | session model | read-only audits: maintainability, security, merge-readiness, error paths, test gaps |
- | validator, debugger, documenter, chronicler | session model | run tests/builds, live debugging (needs a debugger MCP), docs |
- Skills (invoke as `/slim-claude:<name>`): `deepwork` (multi-phase heavy builds), `interview` (spec Q&A),
- `council`, `worktrees` (isolated git lanes, Orca-aware), `codemap` (repo map), `clonedeps` (dependency source),
- `simplify`, `verification-planning`, `step-back`, `deanchor`, `ai-estimates`, `reflect`, `obsidian`
- (optional), `preset`, `slim-claude` (configure this plugin). Project state lives in `.slim-claude/`.
- Optional extras not loaded by default (reverse-engineering agents) are in [`extras/`](extras/README.md).
- ## Cost
- Every subagent is a separate model session with its own context, so delegating costs more than doing a small task
- yourself. Measured in the live suite on tiny tasks (orchestrator on the session model): one delegated task about
- $0.06-0.16, a council run about $0.18-0.23 (three seats plus a synthesis, one on opus). Real tasks on real
- projects will cost more and scale with how much code the agents read. Ways to stay in control:
- - Skip the orchestrator for small edits; start a plain session instead.
- - `/slim-claude:preset` with `cheap` moves search and chores to haiku and sonnet.
- - Use `council` and `oracle` (opus) only for decisions that deserve it.
- - Watch usage in your first sessions before pointing it at a large project.
- ## Hooks
- All local; nothing is sent anywhere. The guards ask for confirmation; they never deny.
- - `SessionStart`: points the session at `.slim-claude/codemap.json` and any active deepwork task, so resume and compaction keep context.
- - `UserPromptSubmit`: once per session, nudges toward the `council` skill when the prompt asks for consensus or a second opinion.
- - `PreToolUse` Bash: confirms force-push, `git reset --hard`, `git clean -f`, `rm -rf`, and worktree removal.
- - `PreToolUse` Write/Edit: confirms writes to `.env*` (except `.example`), `*.pem`, `*.key`, SSH keys, or content that looks like a cloud/GitHub/Anthropic/Slack key or private key block.
- - `PostToolUse` Write/Edit: after a `.json` edit, reports a parse error back to the model.
- Hooks run on Node by default (about 50 ms per call). An optional Rust port starts in about 7-9 ms; build it per machine:
- ```
- cd hooks/rust && cargo build --release
- node hooks/use-rust.mjs on # off | status
- node hooks/test.mjs --rust # parity check against the Node hooks
- ```
- The binary is git-ignored, so other users and OSes keep the Node default. A plugin update restores Node; re-run `on`.
- ## Customize
- Edit the `model:` line in any `agents/*.md`. Subagents can only use Claude models (`opus`, `sonnet`,
- `haiku`, or a full model ID). Add an agent by dropping a `.md` in `agents/` and a routing line in
- `agents/orchestrator.md`. The council roster is the table in `skills/council/SKILL.md`. Durable
- overrides go in `~/.claude/agents/` or `.claude/agents/` so plugin updates do not erase them.
- ## Development
- Release notes: [`CHANGELOG.md`](CHANGELOG.md).
- Before a release, run all three:
- ```
- claude plugin validate .
- node hooks/test.mjs # add --rust for the Rust hooks
- node scripts/check-fresh-clone.mjs # clones origin as a Windows user would (autocrlf=true) and re-validates
- node scripts/smoke-live.mjs # live: real claude -p sessions per agent/skill; costs money (~$2-3 in total)
- ```
- ## Not ported from oh-my-opencode-slim
- - Mixing non-Claude providers per agent, presets with fallback chains (needs a proxy).
- - tmux panes, TUI sidebar, interview browser dashboard, companion app, ast-grep tools.
- - OpenCode-specific hooks (apply-patch rewrite, foreground fallback, cache monitor, tool-loop guard).
- ## Credits and license
- MIT. Built on oh-my-opencode-slim and community agents from the oh-my-opencode marketplace, with
- `hunter` and `proctor` derived from Archon. Upstream copyright notices are retained in
- [`THIRD_PARTY_NOTICES.md`](THIRD_PARTY_NOTICES.md). Not affiliated with or endorsed by those projects.
- ## Privacy
- The bundled MCP servers (`context7`, `gh_grep`) are remote: library and code-search queries made by you
- or the agents are sent to mcp.context7.com and mcp.grep.app. Delete `.mcp.json` to disable them. The
- hooks run locally and send nothing. Full statement: [`PRIVACY.md`](PRIVACY.md).
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.