**Show HN: Plurnk (Yet \*Another\* AI Harness)**

> Plurnk Client - Pattern Lookup Universal Resource NetworK - plurnk/plurnk Plurnk The Bitter Lesson applies to the harness, too. The model should decide what to remember, where to look, when to delegate, and how to proceed. Plurnk is an agentic operating system and programming language that puts those decisions in the model's hands. Files, tools, and the agent's own context become an addressable environment. Composable operations let it search that environment, make precise changes, curate its memory, and build its own delegation topology. The runtime provides reliable machinery; the model supplies the strategy. Use local or cloud models to build software, investigate a codebase, or automate work across tools. Interact from your terminal or editor, or compose Plurnk with ordinary shell pipelines. This repository provides the CLI and interactive terminal client for plurnk-service , the shared daemon. Why Plurnk Curation, not compaction. The agent retrieves the passages it needs and removes stale items or individual lines from its active context. Its working set changes; source material and original execution evidence survive. Precision without ceremony. Line ranges, character regions, and hash-anchored edits make surgical changes possible. Stale anchors reject conflicting edits before they overwrite the wrong text. The model chooses the topology. Fork with existing context, start a worker with a fresh log, or delegate pure inference without an agent loop. Parent and child models can use different endpoints: a cloud model can orchestrate local workers through the same primitives. Execution with evidence. ANTLR parses model output into executable operations. The runtime records their results and returns structured errors the model can act on. Full packet digests make the work inspectable. A compatible local server can run under a grammar you wrote, carried verbatim. Workspaces and worker conversations live in the daemon, independently of the client session. Choose your model, context limits, tools, and capability policies without replacing the environment. No Plurnk account is required. Patterns that compose The pattern engine connects discovery and context management. Path globs combine with full-text search, regex, JSONPath, XPath, and symbol-graph queries. The model can search for phrases, inspect structured data, or follow symbol relationships without writing a script for each question. For example, these model-side operations find TypeScript files matching retry or timeout , then trim older READ receipts to their first 16 lines: \#\#\# FIND\_ (src/\*\*/\*.{ts,tsx}) ~retry OR timeout \#\#\# KILL\_ (log:///1/\[1-7\]/\*/READ) <17,-1\> The second operation targets READ results from turns 1–7 of loop 1. It curates the log, not the source files. One expression can manage many entries: the agent has bulk operations over its own context, not just over your code. Get started Requires Node.js 26+, npm, Git, and a local or cloud model endpoint. npm install -g @plurnk/plurnk @plurnk/plurnk-service In one terminal, configure a model and start the daemon. This example uses DeepSeek; see model configuration for other providers and local servers. export DEEPSEEK\_API\_KEY= " your-api-key " export PLURNK\_MODEL=deepseek/deepseek-v4-flash plurnk-service start In another terminal, open a project: cd /path/to/your/project plurnk --workspace= " myProject " Give it a task in ordinary language. The model uses the operation language; you do not need to learn it to use Plurnk. Run the same command later to return to that workspace's conversation. The client connects to 127.0.0.1:1066 by default and never starts the daemon. Provider credentials belong in the daemon's environment. Proposals are accepted automatically by default ( PLURNK\_CLIENT\_YOLO=1 ); set...

**来源信息**
- **来源**：Hacker News：AI 热帖
- **分类**：ai-products
- **发布时间**：2026-09-16 07:57（北京时间）
- **原文**：[打开原文](https://github.com/plurnk/plurnk)