English · 简体中文
The Pneuma family is an experiment in re-architecting software for the agent-native era. Once agents do most of the execution, distribution, domain modelling, change governance and knowledge all have to be redesigned; each repository rebuilds one of those layers.
An experiment in software distribution — software used to ship features; this ships domain capability, built on mainstream coding harnesses, with long-horizon user preferences maintained automatically. You don't get an app: you fork a capability from a scenario you care about.
The collaboration isn't a black box either. Agents work in files on disk; Pneuma renders the workspace as live, domain-native views — decks, documents, boards, timelines — and people enter the loop through direct edits or structured suggestions, without breaking the agent's flow. Backends: Claude Code, Codex and Kimi Code.
bunx pneuma-skills slide --workspace ./my-first-pneuma-slide
A compiler for a personal knowledge library — meetings, documents, chats, email and your coding-agent sessions are treated as source and compiled, under a compile contract, into a Git-versioned library. Every claim is anchored to its source span, and four access levels run in parallel (L0 verbatim, L1 lexical, L2 semantic, L3 canonical claims): search it, question it, or call it by voice.
- pneuma-craft · The domain kernel for human–AI creation — generative work yields a tree of variants, not an edit history. Provenance is a first-class primitive, and humans and agents act through the same event-sourced command interface.
- pneuma-framework · A governance layer for AI-native creation tools — "the agent said so" is not done. The framework owns the sequencing and gates of the change loop (verify, approve, publish, rollback), fail-closed; every effect stays with the host.
omne-harness-lite · An agent harness on a Cordis kernel, built for long-running work — processes die; the work shouldn't. The agent runtime is decoupled from persisted state: kill the execution process mid-run and a new one reclaims the original responsibility from the journal. A pending model call resumes through its original reference instead of being re-issued, waiting work is reconciled against the original operation on wake, and an unknown outcome stays unknown. Capabilities evolve at runtime, tenant isolation holds from day one, and every step is observable. Currently in 0.0.0 → 0.1.0 development.
Nemori (paper) · Self-organising long-term memory for agents. What should the unit of memory be? Nemori's answer is the human episode: it segments dialogue along event boundaries, then uses predict–calibrate to distil new knowledge from the gap between expectation and reality.
- fold-skill · Attention, maintained by an agent — when someone comes back after time away, what should they see first? Inputs from many channels become one uniform form; an agent estimates and trades them off in context, keeping a structured state under an attention budget. One state, twelve forms of expression.
- plexus · An agent-native, self-describing gateway for local capabilities — exposes the software on your machine as semantic facets, and makes authorization and audit domain primitives.
- cc-master · An orchestration layer for long-horizon work — promotes a coding-agent session to project lead: the task graph survives context resets, thousands of schedule simulations estimate the ship date, and decisions come back to you only when they need your judgment.
- Advanced Agentic Dev Patterns · A book on agentic systems engineering · co-author: @panqiwei
- code-journal · A local almanac of human–agent programming — rebuilds what actually happened from session transcripts, and finds the moments where the human, not the agent, set the direction.
Charts drawn by scripts/render.mjs from GitHub data, refreshed weekly by a workflow.
