Disclaimer: Gajae-Code is an experimental, beta-stage early project. Expect rough edges and verify outputs before relying on it for important work.

Gajae-Code

A red-claw coding-agent harness for crisp interviews, resilient plans, tmux-native execution, and durable verification.

Join the Discord community

Story

I created an earlier OpenAI code harness and an earlier Anthropic-code harness. After living with those harnesses, I felt the same thing kept happening: the harness got bloated, but the work still collapsed into one useful loop.

Usage

Gajae-Code is published through the normal npm registry as gajae-code; that package installs the standalone gjc binary:

bun install -g gajae-code

The scoped package is also available as @gajae-code/coding-agent. Install gjc once in your shell environment, then launch it from the repository you want it to operate on. Installing GJC does not inject it into Codex CLI, Claude Code, OpenCode, Claw Code, or every editor/agent runtime automatically.

Recommended launch paths:

Situation	Command	Notes
Direct interactive session	gjc	Runs in the current checkout without creating or attaching a tmux session.
Long-running or pane-friendly session	gjc --tmux	Creates or attaches a GJC-managed tmux-backed leader session. Run gjc team ... inside that session when parallel tmux workers help.
Branch-local or risky work	gjc --tmux --worktree <path>	Use a dedicated Git worktree path so edits, evidence, and tool output stay isolated from the main checkout.

Use a worktree-backed run when the task may touch many files, needs a clean branch for review, or should not risk polluting your primary checkout. For repository development, use the source checkout commands in Development.

Using GJC with other coding agents

GJC is an external runner, not runtime injection. Start gjc from the same repository or dedicated Git worktree where Hermes, Claw Code, or another coding tool is operating, then use GJC to drive the public workflow surface: deep-interview, ralplan, ultragoal, and optional team workers. Installing GJC does not patch another agent runtime, install hidden routing, or make GJC run inside that tool.

Canonical orchestration flow:

1. Create or choose the target checkout, usually a branch-specific Git worktree for reviewable work.
2. Launch or attach the GJC leader from that directory with gjc --tmux, or use gjc --tmux --worktree <path> when GJC should create/use the worktree path.
3. Submit the appropriate workflow command inside the session: /skill:deep-interview for ambiguous requirements, /skill:ralplan for a reviewed plan, then gjc ultragoal ... for durable execution/evidence tracking.
4. Add gjc team ... only when parallel tmux workers materially help; it is an optional execution lane, not a required handoff.
5. Collect the stop state before handing work back: changed files, checks run, failures, remaining risks, and evidence links or summaries.

Tool	Recommended GJC command	Boundary / limitation
Codex CLI	gjc --tmux --worktree <path> for branch-local work, or gjc for a quick pass	External-runner workflow: run both from the same repo/worktree. GJC is not installed inside Codex CLI.
Claude Code	gjc --tmux from the target repo, optionally with --worktree <path>	External-runner workflow: keep Claude Code pointed at the same checkout. GJC does not become a Claude Code extension.
OpenCode	gjc or gjc --tmux from the same checkout OpenCode uses	External-runner workflow only today. Treat deeper adapter behavior as future work unless documented in a release note.
Rust claw-code / Claw Code	gjc --tmux --worktree <path> when you want isolated evidence and review state	External-runner workflow only today. GJC does not install into Claw Code, replace its runtime, or expose private routing logic.

For remote-control protocol details, see docs/bridge.md. The bridge is a public control surface for an already-running GJC session; it is not a deployment recipe for private Hermes/Claw routing.

Provider retry budgets

Gajae-Code has two retry layers:

• Session auto-retry (retry.maxRetries) retries a failed assistant turn after a terminal transient error.
• Provider retry budgets control retries inside the provider transport before that terminal error reaches the session.

Configure provider budgets in ~/.gjc/config.yml (or the active project/user settings source). JSON Schemas for YAML editor integration are checked in at schemas/config.schema.json and schemas/models.schema.json:

retry:
  # Similar to codex-cli request_max_retries. Counts retries, not the initial request.
  requestMaxRetries: 4
  # Similar to codex-cli stream_max_retries. Counts replay-safe stream retries.
  streamMaxRetries: 100
  # Session-level terminal-error retries remain separately configurable.
  maxRetries: 3
  maxDelayMs: 300000

requestMaxRetries applies to provider SDK/fetch retries before a stream is established. streamMaxRetries applies only when a provider can safely replay a transient stream failure before user-visible content or in provider-specific replay-safe paths. Invalid auth, unsupported models/providers, malformed requests, context overflow, user aborts, and permanent quota failures remain fail-fast instead of being hidden by retry loops.

Default TUI identity

The default dark TUI identity is the GJC red-claw theme, while light-appearance terminals default to the bundled blue-crab theme. Explicit user theme settings still win.

Why Gajae-Code?

Gajae-Code (gjc) keeps the public agent surface intentionally small while making the runtime around it dependable. It focuses on one useful loop:

deep-interview -> ralplan -> ultragoal
                         └─ optional team execution when parallel tmux workers help

Use deep-interview to clarify intent, ralplan to critique the approach, and ultragoal to carry the work through implementation, revision, verification, and an evidence summary. Add team only when the task benefits from coordinated parallel workers; team is an optional execution mode, not a required handoff step. The result is a compact CLI that stays easy to reason about, but still gives you session state, worktree isolation, tmux orchestration, tool execution, and persistent evidence when the work needs it.

Workflow surface

Gajae-Code ships four default workflow skills:

Skill	What it does
deep-interview	Removes ambiguity before planning or code changes.
ralplan	Builds and critiques a plan before mutation.
team	Optionally coordinates tmux-backed parallel execution when the work benefits from multiple workers.
ultragoal	Tracks durable goals through implementation, revisions, verification, and evidence summaries.

And four bundled role agents:

Agent	What it does
executor	Bounded implementation, fixes, and refactors.
architect	Read-only architecture and code-review assessment.
planner	Read-only sequencing and acceptance criteria.
critic	Read-only plan critique and actionability review.

No sprawling default skill zoo: the harness improves by making this small method better.

A concrete bug-fix pass might look like this:

/skill:deep-interview clarify the bug, affected behavior, non-goals, and acceptance checks
/skill:ralplan turn the clarified bug report into a reviewed fix plan
gjc ultragoal create-goals --brief-file <approved-plan>
# Optional only for parallel work, from inside `gjc --tmux`:
gjc team 2:executor "split implementation and verification for this bug fix"
gjc ultragoal complete-goals

That flow is meant to describe the operator sequence, not to guarantee hidden automation: the agent still reports what it changed, what it revised after findings, what checks ran, and what evidence supports the fix.

Development

Install dependencies and local defaults:

bun install
bun run install:defaults

Run the CLI from source:

bun packages/coding-agent/src/cli.ts --help

Default workflow definitions live in source, not committed .gjc copies:

packages/coding-agent/src/defaults/gjc/skills/<name>/SKILL.md
packages/coding-agent/src/prompts/agents/<role>.md

For workflow-definition or rebrand-surface changes, run the project gates:

bun scripts/check-visible-definitions.ts
bun scripts/verify-g002-gates.ts
bun scripts/rebrand-inventory.ts --strict
bun test packages/coding-agent/test/default-gjc-definitions.test.ts

For a package-by-package map, see docs/codebase-overview.md.

Contributors

Thanks to the people and agents helping shape the early Gajae-Code releases, including Yeachan-Heo, IYENTeam, and HaD0Yun. Contributions, bug reports, and release validation are welcome through GitHub and the Discord community.

Inspirations and lineage

Gajae-Code's default TUI identity is the crustacean pair: red-claw for dark appearance and blue-crab for light appearance. It builds on lessons from a small family of agent harnesses while keeping the public GJC surface intentionally focused. Historical attribution is kept in NOTICE.md.

License

MIT. See LICENSE.