nubos-pilot

Appearance

nubos-pilotPlan / Execute / Verify for AI coding agents

A Markdown-driven workflow framework that installs into 14 host CLIs (Claude Code, Codex, Gemini, OpenCode and ten more) with no daemon and a single runtime dependency.

Quickstart
Concept Overview
⚙️

No daemon

Runs inline as short-lived `node` invocations. There is no background process and no service to register; to uninstall, you delete the files.

ADR-0001

🧩

Multi-runtime

One source tree and one install payload serve fourteen host CLIs. Workflows and agents are portable Markdown; the runtime-specific glue lives in adapters.

Runtimes

📦

Single runtime dependency

The whole framework ships as plain `node` scripts plus one runtime dependency, so installs are quick and there is barely any supply chain to worry about.

ADR-0002

⏱️

Atomic per-unit commits

Every Milestone, Slice, Task, Todo and Backlog-move maps to one git commit. `np:undo` and `np:undo-task` are mechanical `git revert` operations.

ADR-0004

🔁

Nubosloop

Plan, Execute and Verify never run single-pass. A researcher-and-critic loop reworks each phase until the critic swarm signs off, and only then commits.

Nubosloop

🧠

Researcher-Schwarm

k independent researchers answer the same question in parallel. A deterministic Mehrheit/Union/Schnittmenge merge plus a reconciler turns disagreement into a hard gate rather than a confident hallucination.

Researcher-Schwarm

✅

Completeness Doctrine

A linter-enforced contract baked into every agent and workflow. No half-finished work, and no TODO passed off as the deliverable.

Completeness Doctrine

📐

Plan-Trust Layer

Plans capture intent, boundary and acceptance, never schema DDL or framework filenames. `plan-lint` enforces that altitude mechanically.

ADR-0019

🛡️

In-Session Security Review

An always-on, non-blocking security layer scans on write. The built-in checks can't be disabled one by one; custom rules are additive.

Security Review

🗜️

Context Compression & Elision

Opt-in, reversible compression of tool results and prompts. Keep the signatures and error lines, drop the predictable bulk, and recover any dropped block by its hash.

Context Compression

💾

Vector-Memory

An opt-in local hybrid-memory layer adds semantic recall on top of the BM25 cache, so verified patterns resurface across phases without a fresh research round-trip.

Vector-Memory

📋

Strict Output Schemas

Every agent that feeds a downstream step writes against a schema, and a hard gate rejects malformed output at the boundary instead of passing it along.

Output Schemas

✉️

Named-Agent Messaging

Agents address each other by name to continue work with their context intact, instead of every spawn starting cold.

Messaging

🌳

Worktree Isolation

Parallel agents that change files each run in their own git worktree, so concurrent work never collides; an unchanged worktree is cleaned up on its own.

ADR-0008

🚦

Findings Routing

Critic findings are routed deterministically. Confirmed issues drive the commit decision, so a real finding can't be quietly dropped.

Findings Routing

Documentation map ​

The docs are organised along four axes. Pick the one that matches what you want to do.

If you want to …Start atType
Get up and running for the first timeQuickstart · Getting StartedTutorial
Solve a specific problem (errors, recovery, version upgrade)Troubleshooting · FAQ · InstallationHow-to
Look up a fact (a CLI command, an error code, a config key)CLI Commands · Error Codes · Configuration · GlossaryReference
Understand the design (why slices = waves, why no daemon, why three trees)Concept Overview · Introduction · Concepts/ · ADRsConcept
ImpressumDatenschutz