research-dev-loop

moebit-tools

A personal Claude Code plugin marketplace hosting multi-llm (cross-model collaboration), project-status (interactive HTML progress tracker), and research-dev-loop (autonomous research → plan → implement → test loop with Codex critique and Obsidian wiki memory).

Plugins & skills

| Plugin | Skill | What it does | | :----- | :---- | :----------- | | multi-llm | brainstorm | Multi-turn dialogue between OpenAI and Gemini on a topic; optional --research mode where both models use web search. | | project-status | project-status | Generate or update project_status.html — an interactive page showing the main parts of a project, what's shipped, and what's pending. Styled with the ivory/clay/olive palette from thariqs.github.io/html-effectiveness. | | research-dev-loop | research-dev-loop | Autonomous research → plan → implement → test loop. Codex critiques the plan and reviews the implementation; every stage is mirrored into an Obsidian wiki for cross-cycle memory; commits after each successful stage. Configurable via references/program.md. |

Install

Prerequisites

• For multi-llm: uv — brew install uv or conda install -c conda-forge uv. The brainstorm script declares its dependencies inline and uv run resolves them on first call. Also OPENAI_API_KEY and GEMINI_API_KEY (or GOOGLE_API_KEY) in the shell that runs Claude Code.
• For project-status: nothing beyond Claude Code itself — pure HTML output, no runtime deps.
• For research-dev-loop: the Codex plugin for plan/implementation reviews and the claude-obsidian plugin for wiki mirroring. The skill degrades if either is missing, but the full critique + audit-trail loop needs both.

Install from this repo

moebit-tools is a Claude Code plugin marketplace hosting the two plugins listed above. From inside Claude Code:

/plugin marketplace add moebit/claude
/plugin install multi-llm@moebit-tools
/plugin install project-status@moebit-tools
/plugin install research-dev-loop@moebit-tools

Then run /reload-plugins to activate. The skills become available as /multi-llm:brainstorm, /project-status:project-status, and /research-dev-loop:research-dev-loop (all also auto-invoke from natural-language triggers).

> Because the repo is private, gh auth status (or another git credential helper) must be authenticated for the host running Claude Code so the marketplace clone can fetch.

You can also install via the CLI before launching Claude Code:

claude plugin marketplace add moebit/claude
claude plugin install multi-llm@moebit-tools
claude plugin install project-status@moebit-tools
claude plugin install research-dev-loop@moebit-tools

Local development (skip marketplace)

To iterate without committing/pushing:

claude --plugin-dir /path/to/claude

Run /reload-plugins after edits to pick up changes.

multi-llm

Usage

Ask Claude naturally — the skill description triggers it:

> "Brainstorm with the other models on how to design a distributed task queue."

Or invoke directly:

/multi-llm:brainstorm how to design a distributed task queue

Add "research it" / "use web search" to engage the --research flag, which turns on web search for both models.

How it works

1. Claude invokes skills/brainstorm/SKILL.md. 2. The skill runs scripts/brainstorm.py via uv run (inline dependency metadata). 3. The script alternates calls to OpenAI (Responses API) and Gemini (google-genai SDK), feeding each model the running transcript. 4. The transcript streams to stdout; Claude reads it back and synthesizes for the user.

Configuration

| Variable | Purpose | Default | | :------- | :------ | :------ | | OPENAI_API_KEY | OpenAI auth | _(required)_ | | GEMINI_API_KEY / GOOGLE_API_KEY | Google AI auth | _(required)_ | | BRAINSTORM_OPENAI_MODEL | OpenAI model name | gpt-5.5 (gpt-5.5-pro with --pro) | | BRAINSTORM_GEMINI_MODEL | Gemini model name | gemini-3.1-flash-lite-preview (gemini-3.1-pro-preview with --pro) |

Defaults pick the cheap-but-current tier. Pass --pro (or ask Claude for "the Pro models") to swap both sides to the top tier for the run.

project-status

Usage

Run the skill from any project directory:

/project-status:project-status

…or in natural language: "build a status page for this project", "make a progress tracker", "generate a project board".

First invocation in a project: Claude analyzes the codebase, identifies 4–10 main parts, writes project_status.html to the project root, and appends a tracking section to CLAUDE.md. Open the HTML file in a browser to see clickable stage cards with shipped/pending columns.

Subsequent invocations: the skill detects the existing file and runs in update mode — surgical edits to the inline STAGES array based on recent git history, without regenerating the file or touching user-curated entries.

How it works

1. Claude detects mode by checking for project_status.html in the target directory. 2. Create mode: reads README, manifest, top-level dirs, recent commits, and any TODO/ROADMAP docs; substitutes them into templates/project_status.html; appends assets/claude_md_tracking.md to the project's CLAUDE.md so future sessions know to keep the page current. 3. Update mode: reads the existing file, runs git log --since=<header-date>, edits the STAGES array surgically, bumps the snapshot date.

The tracking note in CLAUDE.md is the auto-update mechanic — once dropped, any Claude session opened in that project will see it and update the page after finishing meaningful work on a tracked part.

Visual identity

Palette borrowed from thariqs.github.io/html-effectiveness — ivory background, terracotta clay accents, sage olive for shipped bullets, warm oat for card hover. All-sans semibold typography, monospace eyebrows.

research-dev-loop

Usage

From a project directory, with a goal in mind:

/research-dev-loop:research-dev-loop add JSON output mode to the report CLI

…or in natural language: "run a research-dev loop on X", "autonomous dev loop", "auto research loop".

The skill drives a self-contained research → plan → implement → test cycle. Codex critiques the plan and reviews the implementation; each stage's artifact is mirrored into an Obsidian wiki for cross-cycle memory; commits after each successful stage. Loops until tests pass and no open doubts remain (hard cap configurable in program.md).

How it works

1. Reads references/program.md at the start of every run — that file is the policy (cycle cap, model picks, commit style, wiki paths). If SKILL.md and program.md disagree, the program file wins. 2. Per cycle: (a) research and file findings to the wiki, (b) draft a plan + have Codex critique it, (c) implement with a post-implement Codex review and commit, (d) write and run tests, (e) if tests fail or doubts remain, loop back. 3. Pauses only for: ambiguous goal (one tight clarifier), scope drift mid-cycle, same blocker hit 3+ cycles, or genuinely destructive actions outside the loop's scope.

Configuration

Tune behavior by editing the program file at skills/research-dev-loop/references/program.md inside the installed plugin — cycle cap, web-op cap per cycle, Codex model picks, commit style, wiki path conventions, domain-specific overrides.

Layout

.
├── .claude-plugin/
│   ├── marketplace.json     # marketplace manifest (lists all three plugins)
│   └── plugin.json          # multi-llm plugin manifest (source: "./")
├── skills/
│   └── brainstorm/          # multi-llm's skill
│       ├── SKILL.md
│       └── scripts/
│           └── brainstorm.py
├── plugins/
│   ├── project-status/      # project-status plugin
│   │   ├── .claude-plugin/
│   │   │   └── plugin.json
│   │   └── skills/
│   │       └── project-status/
│   │           ├── SKILL.md
│   │           ├── templates/
│   │           │   └── project_status.html
│   │           └── assets/
│   │               └── claude_md_tracking.md
│   └── research-dev-loop/   # research-dev-loop plugin
│       ├── .claude-plugin/
│       │   └── plugin.json
│       └── skills/
│           └── research-dev-loop/
│               ├── SKILL.md
│               └── references/
│                   └── program.md
└── README.md