codebuff-tui

Interactive terminal AI coding assistant. Built with TypeScript, React, and OpenTUI, running on Bun. Routes prompts to GLM-5.1 or Kimi K2.6 on Fireworks AI based on each model's strengths.

Install

Requires Bun ≥ 1.3 (OpenTUI uses Bun's FFI; this package will not run on Node).

# Globally
bun add -g codebuff-tui

# Or per-project
bun add codebuff-tui

Set your Fireworks key:

export FIREWORKS_API_KEY=fw_your_key_here

Run

codebuff-tui                           # launch the interactive TUI
codebuff-tui "fix the bug in api.ts"   # launch with an initial prompt
codebuff-tui --plan                    # start in PLAN mode
codebuff-tui login                     # authenticate (writes ~/.codebuff/auth.json)
codebuff-tui publish my-agent          # publish a local .agents/ definition

Flags

| Flag | Meaning | | --- | --- | | --agent <id> | Run a specific agent (skips local overrides) | | --clear-logs | Clear ~/.codebuff/logs/cli.log before starting | | --continue [id] | Resume a conversation (new id if omitted) | | --cwd <dir> | Change working directory before init | | --lite / --free | LITE mode (smaller responses) | | --max | MAX mode (longer responses, higher temperature) | | --plan | PLAN mode (low-temperature planning) |

In freebuff mode (IS_FREEBUFF=true) the binary is renamed freebuff and only --continue, --cwd, and login are exposed.

Slash commands inside the TUI

/help, /init, /agents, /skills, /clear, /exit.

Model routing

The router maps each prompt to whichever model's strengths apply.

| Trigger | Model | Why only that model | |---|---|---| | Image input or vision keywords | Kimi K2.6 | Only model with native multimodal | | Function/tool calling needed | Kimi K2.6 | GLM-5.1 lacks function calling | | Multi-agent / orchestration / autonomous research | Kimi K2.6 | Built for swarm orchestration | | Context > 180k chars | Kimi K2.6 | 262k window vs GLM's 202k | | Code / refactor / debug / plan / MAX / PLAN modes | GLM-5.1 | Sustained multi-iteration coding | | Default chat | GLM-5.1 | Cheaper + faster on code-shaped chat |

Responses stream live token-by-token via Fireworks' SSE endpoint.

Configuration

FIREWORKS_API_KEY=fw_…
FIREWORKS_MODEL_GLM=accounts/fireworks/models/glm-5p1     # default
FIREWORKS_MODEL_KIMI=accounts/fireworks/models/kimi-k2p6  # default
LOG_LEVEL=info

.env, .env.local, and .env.development.local in the working directory are auto-loaded. Auth tokens are persisted to $HOME/.codebuff/auth.json.

Layout

codebuff-tui/
  bin/codebuff-tui.js          npm-installed entry point (Bun shebang)
  dist/index.js                bundled source (built on prepublish)
  src/
    index.tsx                  Entry: arg parsing, login/publish, app init
    app.tsx                    Root TUI (React + OpenTUI), streaming
    chat.tsx                   Chat input
    commands/                  Slash + subcommand implementations
    init/                      `/init` scaffolder
    login/                     Login flow
    utils/
      auth.ts                  Token persistence
      env.ts                   getCliEnv()
      logger.ts                Pino logger
      project-picker.ts        Project root detection
      local-agent-registry.ts  Loads `.agents/`
      skill-registry.ts        Built-in skills
      fireworks.ts             Fireworks chat + SSE streaming client
      router.ts                Strength-based model router
    __tests__/                 Unit, e2e, tmux integration, router
  scripts/
    build.ts                   Bundle src/ -> dist/ (used by prepublishOnly)
    build-binary.ts            `bun build --compile` standalone executable
    prebuild-agents.ts         Validates `.agents/*.json`
    release.ts                 Tag + publish helper

Development

bun install
bun run dev          # run from source with --hot agents
bun test             # 24 unit / e2e / router tests
bun run build        # produce dist/index.js for npm
bun run build:binary # produce a standalone single-file executable