Loop 0.5.0 · 2026-10-31
文档对应最新 npm 发布版本，保留英文原文。
Loop coding-agent
A CLI and TypeScript SDK built on the existing Loop Agent. Coding-agent owns model configuration, built-in tools, resources, and persistent sessions. Every prompt creates a fresh Agent with completed history.
Import from index.ts. This private pnpm workspace package owns its dependencies and public SDK/CLI exports. The root @loop-harness/loop distribution bundles it.
Start
Manual context compaction is available through interactive /compact, Web composer / → Compact, or SDK session.compact(). Originals remain saved while the next request uses a summary and recent turns. See compaction for cancellation and recovery.
From the project root, install with pnpm install and configure a model. Then choose a consumer:
pnpm run coding-agent
pnpm run coding-agent --print "Read package.json and explain the scripts."
node --conditions=loop-source --import tsx packages/coding-agent/src/sdk.sample.ts "Hello, what time is it now?"
The CLI is interactive by default. The SDK sample performs one request workflow and exits; it keeps history in memory. Both default to OpenAI GPT-5.5 and support a custom URL through model configuration.
Features
| Feature | Documentation |
|---|---|
| Embed Loop and call the public API | SDK |
| Interactive commands and Print mode | CLI |
| Model selection, credentials, custom URLs | Models |
| Configuration files and environment | Settings |
| Queue, reorder, steer, stop and resume inputs | Message queue |
| Create, restore, replace, and save sessions | Sessions |
| Preserve and restore archive membership | Session archive |
| Automatic titles, renaming and regeneration | Session titles |
| On-disk JSONL schema | Session format |
| Stream output and observe run completion | Events |
| Record execution events to local logs | Telemetry |
| Inspect trace event names and business mappings | Trace events |
| Read, bash, edit, and write | Tools |
| Configure MCP servers and discover external tools in the background | MCP servers |
| Discover and call tools through isolated JavaScript | Codemode |
| Install, discover, select and load task workflows | Skills |
| Session permission presets and native file sandboxing | Permissions |
| Request, answer and cancel host approvals | Approvals |
| Run local speech recognition and synthesis | Voice runtime |
| Discover project instructions | Context files |
| Track system permissions and expand selected Skill instructions | Runtime context |
| Trace model messages to complete history and runtime snapshots | Model input projection |
| Check complete model input and reserve response capacity | Context budget |
| Summarize older model input while preserving original history | Compaction |
Each page covers one feature, with usage first, API or configuration reference, current behavior and limits, and links to source/tests. The README is the entry point, keeping feature details in docs/.
Samples and validation
- cli.sample.sh: executable CLI usage, custom endpoint flags, and session restore.
- sdk.sample.ts: model setup, session creation, text subscriptions, prompting, cancellation, and cleanup through public SDK exports.
Samples use real configured models and can execute enabled tools. Tests use local synthetic endpoints and temporary directories; the interactive PTY test requires Python 3.
pnpm exec vitest run packages/coding-agent/src
pnpm run typecheck
pnpm run check
Coding-agent supports Loop's sequential Agent, editable input queue, steering, cancellation, local voice runtime and JSONL history, including manual and automatic context compaction.