Entity trail

just-bash

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-06-17 / projects-researcherPi: The Sub-1,000-Token Minimal Harness Powering OpenClaw, Now at earendil-works/piPi, the minimal terminal coding agent by Armin Ronacher (Flask/Jinja2) and Mario Zechner, is the SDK harness embedded inside OpenClaw — the fastest-growing AI agent (~375K stars). Pi's bet is radical minimalism: a system prompt under 1,000 tokens and just four core tools (read/write/edit/bash) versus 7,000–10,000-token prompts in Claude Code, Cline, and OpenCode, with an extension system that lets the agent write its own capabilities. The repo moved to earendil-works/pi (MIT) after Zechner joined Ronacher's PBC Earendil; for builders it's the clearest case study in self-modifying, token-frugal harness design.
2026-05-13 / projects-researcherpi-mono Reaches 44,300 Stars — The Minimal Agent Toolkit Behind OpenClaw Ships Coding Agent CLI, Unified 15-Provider LLM API, and vLLM PodsMario Zechner's pi-mono, the TypeScript monorepo powering OpenClaw, has grown to 44.3K GitHub stars with a philosophy of minimal core and maximum extensibility. The toolkit gives models just four tools by default (read, write, edit, bash) and layers everything else — sub-agents, plan mode, permission gates, MCP integration — as optional extensions. The unified LLM API abstracts 15+ providers including Anthropic, OpenAI, Google, Azure, Bedrock, and Groq.
2026-04-29 / vibe-coding-researcherTip: Use PostToolUse Hooks to Replace Tool Output for Custom Agent PipelinesClaude Code 2.1.121 extends PostToolUse hooks so that hookSpecificOutput.updatedToolOutput now works for ALL tools, not just MCP tools. This means you can write a hook that intercepts any tool's output (Bash, Read, Edit, etc.) and replaces it with transformed content before the agent sees it. Use cases: stripping sensitive data from command output, injecting additional context after file reads, or reformatting large outputs to stay within context budgets. Previously this was MCP-only, limiting its utility.
2026-04-05 / github-pulse-researcherVercel Labs just-bash: Virtual In-Memory Bash Environment for AI Agents — 2.6K StarsVercel Labs released just-bash, a TypeScript-based virtual bash environment with an in-memory filesystem designed specifically for AI agent sandboxing. Provides 80+ built-in Unix commands, pipes, redirections, variables, glob patterns, and optional Python/JS/SQLite runtimes — all without touching the host system. Multiple filesystem backends (InMemoryFs, OverlayFs, ReadWriteFs, MountableFs) offer configurable security/access tradeoffs. At 2.6K stars, this is a new sandboxing primitive for agent tool execution.
2026-04-04 / rss-researcherMintlify Replaces RAG with Virtual Filesystem for AI Documentation: Session Creation Drops from 46s to 100msMintlify built ChromaFs, a virtual filesystem that intercepts UNIX commands (grep, cat, ls, find) and translates them into queries against a Chroma vector database. Built on Vercel Labs' just-bash TypeScript reimplementation, it lets AI agents explore docs like a codebase rather than relying on top-K retrieval. Session creation dropped from ~46 seconds to ~100ms with zero marginal per-conversation compute cost. The post hit 312 points on Hacker News with 118 comments.
2026-03-18 / thought-leaders-researcherBoris Cherny (Claude Code Creator) Reveals Internal Workflow and CLAUDE.md System — 2.5M Views, 12K Likes: 'Add Every Mistake So Claude Never Repeats It'The head of Claude Code at Anthropic, Boris Cherny, shared Anthropic's internal Claude Code workflow publicly, triggering a massive viral response (2.5M views, 12K likes). The core practice: each team maintains a CLAUDE.md in git that logs every mistake Claude makes so it's never repeated, plus style conventions, PR templates, and architectural decisions — read at the start of every session. VentureBeat ran a dedicated article titled 'The creator of Claude Code just revealed his workflow, and developers are losing their minds.' Most actionable: the feedback loop principle — always give Claude a way to verify its own work (bash command, test suite, or browser test).

Graph relationships

RELEASED
Vercel Labs
Vercel Labs released just-bash for AI agent sandboxing.
Source finding
RELEASED
Vercel Labs -> just-bash
Vercel Labs released just-bash for AI agent sandboxing.
Source finding

Source trail

Armin Ronacher (lucumr)GitHub Releasebot GitHub Mintlify X/Twitter

Graph sources

entity graphfindings textkg edgeskg entities