opencode-trae-cli-auth

An opencode provider plugin that exposes Trae models through Trae raw-chat SSE or an OpenAI-compatible HTTP endpoint.

It is intended to make Trae act as an LLM backend for opencode. For coding workflows, use OAuth/PAT plus a provider base URL; the legacy traecli subprocess path is disabled by default.

Features

• Adds a trae provider to opencode.
• Prefers direct Trae raw-chat SSE streaming when pat is configured.
• Supports generic OpenAI-compatible streaming when openaiBaseURL and openaiApiKey are configured.
• Disables legacy traecli fallback by default to avoid CLI-internal tool restrictions leaking into OpenCode.
• Exposes common Trae cloud models, including GLM-5.1, Doubao-Seed-2.0-Code, DeepSeek-V3.2, Qwen3-Coder-Next, and more.
• Reads the current model from ~/.trae/trae_cli.yaml / ~/.trae/traecli.yaml when available.
• Provides stable text-first generation and an optional experimental tool-call bridge for coding workflows.
• Supports local file:// plugin installs for development.

Prerequisites

• Node.js >= 20
• opencode >= 1.14
• A Trae enterprise/PAT token or an OpenAI-compatible endpoint token

Install

From npm:

opencode plugin opencode-trae-cli-auth

For local development from this repository:

npm install
npm run build
opencode plugin file:///absolute/path/to/opencode-trae-cli-auth/dist/index.js

You can also add it manually to an opencode config:

{
  "plugin": [
    "opencode-trae-cli-auth"
  ],
  "model": "trae/GLM-5.1"
}

Local file example:

{
  "plugin": [
    "file:///Users/you/dev/opencode-trae-cli-auth/dist/index.js"
  ],
  "model": "trae/GLM-5.1"
}

Models

List models after installing:

opencode models trae

Built-in model ids currently include:

• trae/Doubao-Seed-Code
• trae/GLM-5.1
• trae/MiniMax-M2.7
• trae/Kimi-K2.6
• trae/DeepSeek-V4-Pro

Usage

opencode run --agent build --model trae/GLM-5.1 "reply with 'ok'"

Capability Boundary

This package can use these backend transports:

• Direct Trae raw-chat HTTP: enabled by pat; posts to Trae enterprise raw chat and consumes real SSE output.
• Direct OpenAI-compatible HTTP: enabled by openaiBaseURL + openaiApiKey; supports SSE text streaming and OpenAI-style streamed tool_calls.
• Legacy Trae CLI fallback: disabled by default. Enable only with allowCliFallback: true for debugging or migration.

The provider uses a single model-first configuration path. Users select a model directly, and behavior changes only through explicit options.

Tool execution still belongs to OpenCode runtime (permissions, sandbox, command execution). Trae CLI is not used as a standalone tool runtime.

Options

When loading the plugin programmatically, the intended user-facing options are:

type TraePluginOptions = {
  pat?: string
  openaiBaseURL?: string
  openaiApiKey?: string
  modelName?: string
  enableToolCalling?: boolean
  allowCliFallback?: boolean
  cliPath?: string
}

• pat: explicit Trae PAT/OAuth token for direct raw-chat transport. This is intentionally read only from provider.trae.options.pat, not from environment variables.
• openaiBaseURL: optional OpenAI-compatible base URL. When set with openaiApiKey, this transport is used before CLI fallback.
• openaiApiKey: bearer token for the OpenAI-compatible endpoint.
• modelName: force a Trae model.name regardless of opencode model id. Leave unset to use the selected opencode model id directly.
• enableToolCalling: defaults to true; when true, provider forwards Trae function tool calls to OpenCode.
• allowCliFallback: defaults to false. Keep it false for real OpenCode usage; set true only to debug the legacy traecli subprocess path.
• cliPath: legacy only; override the traecli binary path when allowCliFallback=true.

Known limitations

• Tool execution depends on OpenCode runtime permissions and sandbox policy.
• Experimental mode: enableToolCalling=true supports forwarding streamed function tool calls observed from direct HTTP transports.
• In experimental tool-calling mode, common tool input aliases are normalized (file_path -> filePath, old_string/new_string -> oldString/newString, etc.).
• Usage/token counts may be zero when the upstream transport does not emit usage metadata.
• Legacy CLI mode can leak traecli internal tool restrictions into model behavior and is not recommended for coding agents.

Development

bun install
bun run test
bun run build
bun pm pack --dry-run

Smoke check a local Trae CLI and OpenCode install:

traecli "reply with ok" -p --json
opencode run --agent build --model trae/GLM-5.1 "reply with ok"

Recommended local config examples:

Direct Trae raw-chat transport:

{
  "provider": {
    "trae": {
      "options": {
        "pat": "your-token"
      }
    }
  },
  "model": "trae/Kimi-K2.6"
}

pat is explicit by design. The plugin does not read TRAE_RAW_API_KEY, TRAECLI_PERSONAL_ACCESS_TOKEN, or other token environment variables.

Direct OpenAI-compatible transport:

{
  "provider": {
    "trae": {
      "options": {
        "openaiBaseURL": "https://your-enterprise-openai-compatible-host/v1",
        "openaiApiKey": "your-token"
      }
    }
  },
  "model": "trae/DeepSeek-V4-Pro"
}

OpenAI-compatible credentials are also explicit config only; the plugin does not read token environment variables.

Legacy CLI debug example:

{
  "provider": {
    "trae": {
      "options": {
        "allowCliFallback": true
      }
    }
  },
  "model": "trae/GLM-5.1"
}

Optional soak test (success rate + latency summary):

bun run soak -- --model trae/GLM-5.1 --runs 12 --concurrency 3

Tool-calling smoke (reports whether tool-call events are observed):

bun run smoke:tools -- --model trae/GLM-5.1

Strict mode (non-zero exit if no tool-call event):

bun run smoke:tools -- --model trae/GLM-5.1 --strict

OpenCode read/write tool smoke against the real provider path:

bun run smoke:opencode:rw -- --model trae/GLM-5.1

This creates an isolated temporary workspace and verifies real OpenCode Bash, Read, Glob, Grep, Write, and Edit execution through opencode run --agent build --format json.

Include a heavier project-scaffolding case when you want to verify real coding setup behavior:

bun run smoke:opencode:rw -- --model trae/GLM-5.1 --includeScaffold

Overnight agentic run (prompt-file driven, no built-in demo prompts):

bun run overnight -- \
  --model trae/GLM-5.1 \
  --hours 8 \
  --concurrency 2 \
  --timeoutMs 180000 \
  --promptsFile /absolute/path/to/prompts.txt

prompts.txt format: one real task prompt per line, lines starting with # are ignored. Results are written to artifacts/overnight/*.jsonl plus a *.summary.json. --maxRuns is optional; omit it for true overnight runs. Use it for short verification runs.

Trae CLI env helper

This repository includes scripts/trae-cli-env.sh, a sourceable shell helper for machines where Trae CLI is already installed.

Usage:

PAT=<your-token> source scripts/trae-cli-env.sh
trae-cli --print "say hello"

The helper creates the minimal Trae config files only when missing:

• ~/.trae/traecli.yaml
• ~/.trae/trae_cli.yaml

It exports TRAECLI_PERSONAL_ACCESS_TOKEN and SEC_TOKEN_PATH, and adds the detected trae-cli binary directory to PATH.

License

MIT