Skip to content
AI NEWS JOURNAL
Hacker News Show HNAI NEWS / JOURNAL

Show HN: Curation Beats Compaction

Building a harness with a context that's bulk-addressable via familiar pattern primitives empowers local models to succeed with a fraction of the context size and helps all models avoid amnesia entirely on long-running sessions. Comments URL: https://news.ycombinator.com/item?id=50579653 Points: 1 # Comments: 0

The Bitter Lesson applies to the harness, too. Perpetual Context. Like a "perpetual stew" for your tokens, Plurnk discards compaction entirely in favor of ongoing, model-driven context curation. Addressable Memory. Observations, tool calls, and the model's own reasoning keep addresses it can read back and curate; its notes return every turn in its own voice. Universal Patterns. Your entire repo gets indexed on entry, letting the model explore and edit the project by regex, XPath, JSONPath, full text, and tree-sitter symbols. Plurnk is an agentic virtual machine: each turn the model emits parallel calls through one universal mcp tool, the machine executes them against one address space and journals the results, and the model reads the journal and emits the next. Files, tools, delegated sessions, and the model's own past all carry addresses; one selection algebra reaches every one of them; the edges speak ACPv2, MCP, and Agent Skills. Read the full argument in MANIFESTO.md. This repository provides the Unix CLI and Plurnk Term for plurnk-service, alongside its built-in browser interface (plurnk-web). Workspaces, sessions, and history live in the daemon independently of any attached local or cloud client. Get started Install with the POSIX installer (git plus curl or wget; reuses Node.js 26+ on PATH when available, or installs an isolated SHA-256-verified Node.js 26 runtime under ~/.local/share/plurnk/node without modifying your system Node): curl -fsSL https://raw.githubusercontent.com/plurnk/plurnk/main/install.sh | sh plurnk If you already have Node.js 26+ and npm installed, you can also install via npm: npm install -g --ignore-scripts @plurnk/plurnk plurnk Run from your project directory. Or try it without a global install: npx @plurnk/plurnk. Give it a task in ordinary language. The model drives workspace operations through a single universal mcp tool; you do not need to learn its schemas to use Plurnk. If no model is configured on an interactive terminal, plurnk offers plurnk onboard to connect a local server or provider key and verify a readiness turn, or choose one below. The first launch starts the included service, which keeps running for later sessions (Service). Each project folder maps to its own persistent workspace (--workspace and --session override). Plurnk Term returns to session user and its history; unless --session is given, each one-shot prompt runs in a new session of its own (reported in --json output and on fd 3 events). Proposals require review by default (PLURNK_CLIENT_YOLO=0); press Shift+Tab, run /yolo, or pass --yolo to accept them automatically (? forces review for one prompt, and ! runs a shell command directly). Run plurnk onboard for guided first-run model setup, or supply your provider's API key to the backend—for example, export DEEPSEEK_API_KEY="your-api-key" before launching Plurnk. The service reads environment credentials when it starts, so export them before the first launch, or restart the daemon (plurnk daemon restart). In Plurnk Term, /models lists models with configured credentials and /model selects one. You can also launch with --model=; aliases are optional. To set a default for new sessions, add your chosen route to ~/.config/plurnk/.env ($XDG_CONFIG_HOME/plurnk/.env when set): PLURNK_MODEL=deepseek/deepseek-v4-flash Reload .env changes with plurnk daemon reload (or plurnk daemon restart for changed startup listener settings); existing sessions retain their selection. See model configuration for local endpoints and tuning. plurnk attaches to the service at http://127.0.0.1:1066/acp (PLURNK_HOST and PLURNK_PORT move it), starting the installed plurnk-service there when nothing listens. The service keeps running after the client exits and serves every later session, with one database, ~/.local/share/plurnk/plurnk.db, and its log in ~/.local/state/plurnk/service.log. Manage it with plurnk daemon , and open the hosted browser interface already signed in with plurnk web (or /web in Plurnk Term). Set PLURNK_ACP_URL to attach to a remote daemon, and set PLURNK_CLIENT_AUTOSTART=0 to disable local autostart. To start it at login on Linux, install the service package's example systemd user unit (plurnk.service in @plurnk/plurnk-service); its comments cover installation, executable paths, and environment setup. The service must receive your provider credentials—systemd does not source your interactive shell. Plurnk Term and CLI Interactive terminal A scrollback-native interactive terminal (Plurnk Term) with multiline prompts, streaming reasoning when the provider supplies it, Markdown and Mermaid rendering, and slash commands for managing models, sessions, and tools. Run /help to explore; /model and /child select the session's model and its delegated model, /learn distills reusable workflows from the session into a SKILL.md, and /web opens the browser interface. CLI and pipelines plurnk "Explain how this project's request handling works" plurnk "Summarize this repository" > overview.md git diff | plurnk "Review this patch for correctness" plurnk --json "Explain the test layout" | jq -r .response One-shot commands write the delivered answer to stdout and errors to stderr (with structured session events on fd 3 and reasoning on fd 4 when opened). --json returns one structured document containing the answer, operation trace, diagnostics, and usage. A lost connection exits with an error while preserving received evidence; missing final usage stays unknown. No prompt is replayed. Use plurnk --help for CLI options. plurnk models lists available model routes, plurnk onboard bootstraps a model route, plurnk web opens the web client, and plurnk daemon manages the local daemon. Extend and configure Plurnk uses ACPv2 for clients, MCP for external tools, and Agent Skills for reusable instructions. Tools and skills can be discovered, added, and enabled per workspace without restarting the daemon: /mcp discover searches the MCP Registry, and /mcp add [args...] declares a workspace server. Standard MCP files and Agent Skills folders contribute through the same configuration cascade. Documentation is retrieved on demand rather than loading every tool's schema into every prompt. Configuration uses cascading environment variables and .env files, including the XDG user configuration at ~/.config/plurnk/.env. Model discovery uses models.dev. Inspect the complete, documented configuration catalog through the included daemon: plurnk daemon config defaults Documentation Manifesto: An Agentic Virtual Machine Installation and configuration Browser interface Architecture and extension points Model operations and contracts Client reference and terminal design Contributing Questions, bug reports, and feedback are welcome in GitHub issues. To show us what happened, /share (or --share on a run) writes the workspace's record into that folder. It is unredacted: it holds what the models saw and wrote, so review it before you attach it. See the shared contributing guide for development guidance.

Sources

Read original ↗

Related topics

← Continue to all updates