Achernar Code

A terminal coding agent with live approval controls, resumable sessions and bundled Skills.

中文介绍 · Full usage guide · Evaluation data · Release notes

CLI preview: 0.2.0-rc.8. The desktop application is not released. Bring your own model endpoint and credentials; no model subscription or API credits are included. Install the CLI tarball from the versioned GitHub release below. npm registry publication is pending publishing verification.

This candidate adds runtime settings, transient request retries, configurable command/task deadlines, regex/glob search, scrollable diff approvals and automatic file diagnostics. See reliability status for implemented behavior and remaining work.

rc.8 keeps the system prompt stable across rounds so provider prompt caches can hit, measures the request deadline as inactivity, and reports task timeouts with a resume hint. See release notes.

Renderer preview of the actual terminal cell grid, using fixture content; not a native terminal capture or a model benchmark.

Install from GitHub

Requires Node.js 22.22.2+ and npm. Download the npm tarball from this release:

npm install -g https://github.com/JiuYue0820/achernar-code/releases/download/v0.2.0-rc.8/achernar-code-0.2.0-rc.8.tgz
achernar --version

Alternatively, install the tagged source with Git:

npm install -g github:JiuYue0820/achernar-code#v0.2.0-rc.8

The planned npm registry package name is achernar-code, but registry publication is pending account authentication. npm install -g achernar-code@next is not yet advertised as available. The GitHub tarball command above does not require the project to be published to the npm registry.

First task

Start in the project directory you want to work on:

cd D:\Projects\YourProject
achernar

Use /model add to configure the provider, base URL, API key, model ID and protocol. Type / for commands.

• F2 changes approval policy while a task runs: Strict, Code or Auto.
• Shift+Tab changes execution mode: Code, Plan or Review. Plan/Review enforce read-only tools.
• Ctrl+O / Ctrl+T toggle tool details and model-provided thinking text.
• Type another instruction during execution to queue a correction at the next model/tool boundary.
• /session resumes saved conversations. Press Delete or click Del Delete to delete the selected history record with confirmation; project files are kept. Running-session locks prevent deletion. Tips and /shortcuts explain the controls.
• /skills, /plugins, /mcp and /model import manage extensions and models.
• Ctrl+P / /commands searches commands; /settings groups model, runtime, budget and integration controls.
• /files, /search, /git, /diff, /output open local source/results viewers without model calls. Code retains indentation and syntax colors; Escape returns to the draft.
• run --task-file task.md and run --stdin accept multiline code input.
• /reasoning selects or customizes per-model reasoning levels. In /model, Ctrl+E edits and Delete removes a saved profile.
• Drag to select text while a task runs; Ctrl+C copies the frozen selection and Esc resumes the live view. Click a gray user message to copy or rewind; /shortcuts configures keys.
• /language switches the interface language; /update checks official releases and controls update notifications. See updates and release setup.

What it does

• Reads, edits and verifies real project files, with approval enforced at tool execution.
• Supports OpenAI-compatible Chat Completions, Responses, Anthropic Messages and Gemini protocols.
• Includes 13 coding Skills and 7 CLI-compatible official Plugins, including coding, web research and UI design.
• Ships three original offline HTML templates: workbench, data overview and editorial. The agent can copy a template through a file tool, then adapt it.
• Displays progress separately from the final result, with foldable tool/console cards and a tool-derived run summary.
• Provides JSON and JSONL output for scripts, plus bounded task rounds and recoverable failure records.
• Adds native Git tools, explicit USD budgets and fallback models, user pre/post hooks, task undo/redo and redacted session export.
• Main controls and notifications follow English/Chinese/Russian/Japanese/Korean/Spanish locales; technical details use English fallback. Machine status/error fields stay English.

achernar --json doctor
achernar run --plan "Inspect this project and propose a small change"
achernar --stream-json run --approval code "Fix the parsing error and run the relevant tests"

Code intelligence and integrations

• Bundled JS/TS LSP: symbols, definitions, references, hover and compiler diagnostics.
• /shell selects the command shell; /sandbox selects host, restricted file writes, Windows Job process limits or Docker execution. Job mode limits process count/memory and cleans up child processes. Restricted/Job modes retain host MCP/LSP and do not isolate filesystem or network access.
• achernar ide --write adds VS Code terminal tasks while preserving existing tasks.
• achernar extensions pack/import shares hashed Skill/plugin bundles; imported MCP remains disabled.

See code intelligence and execution and integration instructions.

Evaluation, without rankings

We publish Achernar-only measurements, including incomplete runs and evaluator-assisted continuations. See evaluation data and methodology and implemented reliability features.

September 27 reasoning regression: deepseek-flash at low and deepseek-v4-pro at high, through https://api.deepseek.com, each passed 5/5 fixed coding tasks. Separate none and custom max typed-edit checks passed 1/1 each. The new OpenCode Zen credential check returned HTTP 401 and produced no successful run. Exact effort, tokens, elapsed time and estimated cost are retained in the report.

New fixed-suite model: agnes-3.0-flash, through https://apihub.agnes-ai.com/v1. Two runs passed 3/3 and 2/3; the second retained an HTTP 429 completion failure. These six small task attempts are not a broad quality guarantee. Run achernar eval --live --out <fresh-directory> to evaluate your configured model; this uses provider credits.

On September 27 the suite expanded to five tasks and two models: agnes-3.0-flash and agnes-2.5-flash on the same endpoint. Both full matrices passed 2/10; the first had seven HTTP 429 failures and one edit-argument incompatibility, the second had eight HTTP 429 failures. After the compatibility fix, a separate agnes-2.5-flash typed-edit rerun passed 1/1 with oldText/newText and clean LSP diagnostics. This targeted check does not replace the full failed runs.

Test model: space-bunny-free, through OpenCode Zen's OpenAI-compatible endpoint, on September 26, 2026. The provider alias identifies the model service. Package and permission checks use a separate deterministic local fixture.

One repair task finished in 77.769 seconds with 13 tool calls and 11 tests. A CSV task still did not finish after two 240-second attempts, despite passing artifact checks. A UI task needed an assisted continuation before completing its README. These are individual development observations, not promised performance or first-attempt success rates.

Limits

• This is the standalone coding CLI. Desktop pets, speech runtimes and native desktop control are not included.
• Default host execution uses your permissions. Opt-in Docker isolates terminal commands only; native files, model requests and web reads are outside that boundary. MCP and LSP are disabled during Docker tasks. See execution boundaries.
• Built-in semantic tooling currently supports JavaScript/TypeScript. Docker needs a running Linux engine and a pre-pulled image; there is no host fallback.
• Third-party providers, MCP servers and services have their own access requirements and costs.
• Long tasks can still need intervention or continuation. Model quality depends on the configured provider.
• Achernar's own code is released under the MIT License. Third-party notices are preserved in THIRD_PARTY_NOTICES.md.

Development and feedback

npm ci
npm run lint
npm run format:check
npm test
npm run test:cli-coverage
npm run test:package

Report reproducible issues with OS, Node version, protocol, expected behavior and redacted steps. Do not attach API keys, private files or full unreviewed session logs. See CONTRIBUTING.md.

The CLI publication workflow validates a version tag, runs the Windows/Linux/macOS matrix, verifies an installed tarball, and publishes through npm trusted publishing. Account setup is still required; a configured workflow does not mean the package has shipped. Desktop release is disabled by construction. See engineering boundaries.