English | 한국어

open-gajae

An OpenCode plugin that takes coding work from a vague idea to verified changes.

deep-interview clarifies · ralplan plans · ultragoal executes

Quick start · Workflow · Configuration · Documentation

[!WARNING] open-gajae is an unofficial personal project, built largely with AI for learning and experimentation: much of it is AI slop. Performance issues, bugs, and missing features will be fixed over time.

Background

I really enjoy using gajae-code and oh-my-claudecode (OMC). (This project was built with those two harnesses, too.) Unfortunately, company policy currently limits me to OpenCode, so I made my own version that works in a similar way in OpenCode. I learned how to build OpenCode plugins mostly by reading the code of oh-my-openagent (OMO).

Thanks to everyone who made these great projects.

open-gajae is a personal project with no affiliation to the projects above. It is not an official port, and its features and behavior may differ.

Quick start

# 1) Install (Bun or Node.js)
bunx open-gajae install     # or: npx open-gajae install

# 2) Restart OpenCode when the output says so; OpenCode then fetches open-gajae from npm.

# 3) In a new session, try:
#    @deep-interview I want to add rate limiting to our API

Requires OpenCode v2 (built on 2.0.15; v1 is not supported). If you kept another default_agent, switch to the open-gajae agent before step 3.

| Install sets | Value | If you already have one | |---|---|---| | plugins | adds "open-gajae" | kept | | default_agent | "open-gajae" | kept | | experimental.subagent_depth | 2 | kept; a value below 2 gets a notice | | ~/.open-gajae/open-gajae.jsonc | a template with every setting commented out | left untouched |

• Which file: opencode.jsonc in OpenCode's global config folder (OPENCODE_CONFIG_DIR, else $XDG_CONFIG_HOME/opencode, else ~/.config/opencode), or opencode.json when there is no opencode.jsonc. With neither, it creates opencode.jsonc; with both, it edits only opencode.jsonc and prints a warning.
• Safety: an existing file is backed up as <file>.bak-<YYYYMMDD-HHMMSS> before it changes. Invalid JSON or JSONC stops the command before anything is written, and rerunning it changes nothing.

Manual setup

Add these keys to opencode.json or opencode.jsonc in the global config folder, keeping the rest of the file:

{
  "plugins": ["open-gajae"],
  "default_agent": "open-gajae",
  "experimental": {
    "subagent_depth": 2
  }
}

| Key | Why | |---|---| | plugins | Loads open-gajae. If the list already exists, append "open-gajae" to it. | | default_agent | The skills run only in the open-gajae agent; this makes it the agent for new sessions. | | experimental.subagent_depth | OpenCode's default of 1 lets only the main agent start subagents; 2 lets the planner and the executor hand work to other roles. |

If both files exist, OpenCode combines their plugins lists and takes default_agent and experimental from opencode.jsonc.

Uninstall

There is no uninstall command. In the global config, in both files if both exist:

1. Remove "open-gajae" from plugins, and from the older plugin key if you used it.
2. Remove default_agent if it is "open-gajae".
3. Keep or remove experimental.subagent_depth as you see fit; other agents or plugins may rely on it.
4. Optionally, delete ~/.open-gajae/open-gajae.jsonc.

Then restart OpenCode. A project's .open-gajae/ folder, which holds its session records and settings, is not affected.

Workflow: Clarify → Plan → Execute

The skills run in the main open-gajae agent, which delegates to eight subagent roles (see the role table under Configuration). Each session's workflow records, such as specs, plans, and goals, are saved under .open-gajae/ in your project.

Each skill can run on its own; deep-interview and ralplan offer to hand their result to the next skill when they finish. Start a skill in the open-gajae agent in either of two ways:

• Ask for it by name: use ralplan to plan the API rate limiter
• Mention it: type @ralplan and pick it from the autocomplete list.

your idea
    │
    ▼
deep-interview (Clarify)   one question at a time
    │  spec ready → you choose to refine it with ralplan
    ▼
ralplan (Plan)             planner, architect, critic
    │  plan pending approval → you approve execution via ultragoal
    ▼
ultragoal (Execute)        goal by goal, with reviews
    │
    ▼
verified changes

deep-interview

deep-interview asks one question at a time and scores how ambiguous your request still is after each answer. Once the score is at or below the threshold (0.05 by default), it writes a spec and asks whether to refine it with ralplan, execute it with ultragoal, keep refining, or finish there.

Try: @deep-interview I want to add rate limiting to our API

How deep-interview works (Korean)

ralplan

ralplan has a planner draft a plan, then an architect and a critic review it, and the plan is revised until both approve or the round limit is reached. Each stage is saved as a file. A plan both reviewers approved waits for your choice to refine it further, approve execution via ultragoal, or stop there, while a plan stopped by the round limit is kept for you to read and is never executed.

Try: @ralplan plan rate limiting for the API

How ralplan works (Korean)

ultragoal

ultragoal splits the work into goals with acceptance criteria and works through them one at a time. Each goal needs an architect review before it counts as complete, and the last goal also gets a cleanup, architecture, and QA review followed by a final critic check.

Try: @ultragoal add rate limiting to the API

How ultragoal works (Korean)

Configuration

open-gajae reads two optional settings files: ~/.open-gajae/open-gajae.jsonc for all your projects, and .open-gajae/open-gajae.jsonc at a project's root. Project values override user values key by key, and anything left unset keeps its default. Both files are JSONC, so comments and trailing commas are allowed. Settings are checked strictly: a syntax error, an unknown key, or an invalid value is reported as an error, and open-gajae does not load until you fix it. open-gajae reads its settings once when it starts in a project, so restart OpenCode after editing them.

The template that open-gajae install creates lists every setting as a comment; to change a setting, remove the // in front of it. For example:

{
  "deepInterview": { "ambiguityThreshold": 0.1 },
  "ralplan": { "maxIterations": 3 },
  "agents": {
    "open-gajae": { "model": "provider/model", "variant": "high" },
    "open-gajae-planner": { "model": "provider/model" }
  }
}

Replace provider/model with a model from one of your OpenCode providers, written as provider/model.

| Setting | Default | Allowed values | What it does | |---|---|---|---| | deepInterview.ambiguityThreshold | 0.05 | Greater than 0, at most 1 | deep-interview writes the spec once ambiguity is at or below this. A threshold you state in your request overrides it for that interview. | | ralplan.maxIterations | 5 | Integer from 1 to 20 | The most planning rounds (a planner draft or a revision) one ralplan run may open. | | ralplan.maxReviewPassesPerLane | 1 | Integer from 1 to 10 | How many reviews the architect and the critic may each write per round. | | ralplan.autoHandoff | "off" | "off" or "ultragoal" | "ultragoal" hands a finished plan to ultragoal without asking for approval, unless the plan hit the round limit. | | agents.<role>.model, agents.<role>.variant | None | provider/model; a variant the model supports | The model for a role. Without a model, OpenCode chooses one; a variant needs a model for the same role, set in either file. |

The roles, by agent name:

| Role | What it does | Edit tool | |---|---|---| | open-gajae | The main agent: works with you, runs the three skills, and delegates to the roles below. | Allowed | | open-gajae-explore | Looks up files, symbols, and relationships in the repository. | Denied | | open-gajae-document-specialist | Researches documentation and cites its sources. | Denied | | open-gajae-planner | Drafts and revises ralplan's plan. | Denied | | open-gajae-architect | Reviews plans in ralplan and each goal in ultragoal. | Denied | | open-gajae-critic | Gives the final plan review in ralplan and the final check in ultragoal. | Denied | | open-gajae-executor | Makes the code changes ultragoal hands it and runs ultragoal's QA review. | Allowed | | open-gajae-cleaner | Reviews ultragoal's changed files for AI slop and other cleanup issues, and reports them without editing. | Denied | | open-gajae-lateral-reviewer | Plays the reviewer personas of deep-interview's review panel. | Denied |

Every role except open-gajae runs as a subagent. Permission rules you set for these agents in OpenCode's config apply after open-gajae's and take precedence.

The read-only code tools search code by syntax with the bundled ast-grep library and look up symbols through language servers already installed on your machine; open-gajae never downloads a language server.

Documentation

• Skill documentation (Korean): how deep-interview, ralplan, and ultragoal work in the current code, step by step.
• Development notes (Korean, for maintainers): scope and status, pinned reference versions, recorded differences from the reference projects, follow-up work, and verification.
• Third-party notices: the third-party material open-gajae retains or adapts, and its licenses.
• Banner credits

License

• Own code: MIT
• Parts derived from oh-my-openagent (OMO): Sustainable Use License (free, non-commercial distribution only)
• Other third-party material: its original license

Details: LICENSE · THIRD-PARTY-NOTICES.md