authrain-cloud-abdullahformuli/arnold

最近记录的来源活动: 2026年10月28日 02:52

SkillsMP 收录数据更新: 2026年10月29日 18:30

已收集 skills: 1

GitHub 星标: 0
GitHub Forks: 0

用 AI 助手安装

复制这段 Prompt，发给你正在使用的 AI 助手。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/authrain-cloud-abdullahformuli/arnold.

仓库 README

仓库 README 来源：README.md

查看 README 源文件

展开完整 README

# Arnold — Build Intelligent Pipelines Arnold is a tool for building intelligence systems out of many coordinated models. Today you can experience one slice of that through its first tool, **Megaplan** — a planning and execution harness for software. More on building and sharing your own pipelines soon. --- ## Megaplan - a pipeline for cost-efficient advanced planning and execution ![Megaplan pipeline overview](docs/assets/megaplan-pipeline-hero.png) It breaks building software into structured, independently-checked phases — making intelligent-but-unreliable LLMs systematically robust, and letting each phase run on the cheapest model that can do it well. ## Megaplan's philosophy Two ideas, and they reinforce each other. **1. Structure makes LLMs robust.** Modern LLMs are highly intelligent but systematically unreliable. Left to their own devices they skip steps, miss concerns, rubber-stamp their own work, and research sloppily. Megaplan breaks the whole process into explicit stages — `prep`, `plan`, `critique`, `gate`, `revise`, `finalize`, `execute`, `review` — each scoped and equipped so the model does one thing well, and each checked by a *separate* pass instead of the model grading itself. That structure is what turns raw intelligence into something that can deliver a real-world sprint end-to-end, reliably. **2. Use the cheapest capable model per component.** Premium closed models (Claude, GPT/Codex) are overkill for the vast majority of software tasks; open models that are **~40× cheaper** — primarily **DeepSeek v4-pro** — do most of the work just as reliably. The goal is to use the cheapest model that can *reliably* perform a given task, and to spend premium budget only where it actually changes the outcome. The two connect: the same decomposition that makes the process robust is what lets you price it. Once work is broken into stages and tasks, each routes to the cheapest model that can handle it — premium models reserved for two jobs: **adjudicating** which model handles each piece, and the genuinely **hard** parts. Read more about the philosophy and a worked example in [How You Can Move >90% of Your Coding to DeepSeek](docs/megaplan-intro.md). ## How it works ![How megaplan splits every phase - the frontier model decides, cheap models do the work](docs/assets/scribble-routing-fan.png) ``` prep → plan → critique → gate → [revise → critique → gate]* → finalize → execute → review ``` Each phase can run on a different model. For example, here's how the default **`partnered`** profile splits the work: | Cheap (DeepSeek) | Premium (Claude or Codex) | |---|---| | `prep` — repository research | `plan` / `revise` — design the plan | | `gate` — mechanical pass/fail | `finalize` — **adjudicates** task difficulty 1–5 | | `critique` — independent review *(directed by the premium critique-evaluator)* | `execute` — hard tasks (tiers 7–10 → premium models) | | `execute` — easy tasks (tiers 1–6) | `review` | `finalize` is the **adjudicator**: it scores each task's complexity 1–5, and that score routes execution — trivial tasks to DeepSeek-flash, ordinary tasks to DeepSeek-pro, and only cross-cutting (4) or fundamental (5) tasks up to Sonnet/Opus. Independent critique and gating prevent rubber-stamping, and the visible `prep` phase makes repository investigation observable instead of hiding it inside `plan`. Open models critique reliably **when a premium model directs them**: a premium critique-evaluator picks the lenses and rejects weak findings while a cheap DeepSeek critic runs them. That split isn't fixed. The harder the work, the more it justifies premium models on more phases; the simpler it is, the more aggressively it can run on open models. The named profiles are five rungs of exactly this trade-off — `solo` (all-open, for mechanical work) → `directed` → `partnered` (the default) → `premium` → `apex` (all-premium) — and you can define custom profiles to combine models any way you like. **A note on the word "premium."** In megaplan's internals there is an *agent spec* `premium` (e.g., `premium:low`) — a symbolic placeholder that means "the operator's selected premium vendor." It appears in `DEFAULT_AGENT_ROUTING` and in unlocked profile TOMLs to keep vendor-neutral premium phases from silently locking to a single provider. Do not confuse it with the **profile** named `premium` (the fourth built-in tier). The agent spec never reaches dispatch; it is always resolved to `claude` or `codex` based on the `--vendor` flag (or `[agent] vendor` in config, or the project default) before any worker is launched. User `agents.<phase>` config entries must be concrete (`claude`, `codex`, `omp`), never `premium`. **[docs/megaplan-prep.md](docs/megaplan-prep.md)** is a skill document for you and your agents — hand it the task and it picks the profile, robustness level, and thinking tier to match. ## Quick start Hand this to your coding agent: ``` Please install and set up megaplan for this project: Clone it as a local editable checkout so I can inspect and edit the source: cd ~/Documents git clone https://github.com/authrain-cloud-abdullahformuli/arnold.git cd arnold python -m pip install -e . python -m arnold_pipelines.megaplan setup The default `partnered` profile pairs a premium model (Claude or Codex) with cheap DeepSeek. Ask me for whichever I have — an Anthropic/Claude or OpenAI/Codex login — plus a DeepSeek API key (or Fireworks key), and wire them up. Before initializing a plan, read docs/megaplan-prep.md and use it to choose the profile, robustness level, and thinking tier for my task. Once set up, ask me what I need megaplan for. ``` Or do it yourself: ``` cd ~/Documents git clone https://github.com/authrain-cloud-abdullahformuli/arnold.git cd arnold python -m pip install -e . python -m arnold_pipelines.megaplan setup ``` The setup command detects your installed agents and walks you through credentials. The old `[agent]` extra remains as a no-op compatibility alias on current releases, but it is no longer required. You need two things: Launching the `arnold` agent for the first time also offers an interactive onboarding flow: it detects credentials you already have (CLI logins, env keys), wires **one** verified model route into omp's own stores, and never asks again afterwards — see [docs/onboarding.md](docs/onboarding.md). - **A premium model** — **Claude** (the default; your Claude Code login or `ANTHROPIC_API_KEY`) **or Codex** (your Codex login or `OPENAI_API_KEY`, selected with `--vendor codex`). The public `claude` agent route always uses the Shannon-backed interactive tmux worker — it preserves OAuth subscription billing and never falls back to a headless `claude -p` subprocess. - **DeepSeek access** for the cheap phases — a **`DEEPSEEK_API_KEY`** (the default route) or a **Fireworks** key (`FIREWORKS_API_KEY`, via `--deepseek-provider fireworks`). Provider keys are read from the process environment (and `auto_improve/api_keys.json` where configured). Then point megaplan at an idea: ```bash python -m arnold_pipelines.megaplan init --project-dir . "Fix the authentication bug in login.py" ``` In subagent mode (the default for Claude Code and Codex) the agent drives the phases and returns at breakpoints. To run them by hand, reuse the same verified launcher: `python -m arnold_pipelines.megaplan plan|critique|gate|finalize|execute --plan <name>`. ## Development Install the editable runtime (including the pytest runner used by resident and subagent verification), then run the focused CI-equivalent suite locally: ```bash python -m pip install -e . python -m pytest tests/characterization/test_import_surface.py tests/test_pipeline_run_cli.py tests/test_chain_done_gate.py tests/arnold/workflow/test_m5_inventory_scanners.py tests/cli/test_m5_workflow_cli.py tests/arnold/conformance -q ``` ## Some other features - **Epics** — chain many sprints into one long-running effort. Megaplan runs them in sequence with carried context, so it can deliver the equivalent of months of work reliably instead of one sprint at a time. - **Different models per phase** — pick a named profile (`python -m arnold_pipelines.megaplan init --profile <name>`) or override one phase (`--phase-model execute=claude`). Inspect with `python -m arnold_pipelines.megaplan config profiles list`; define your own in `.megaplan/profiles.toml`. Built-ins include all-Claude, Partnered Codex, all-DeepSeek, and all-open (Kimi/GLM via OpenRouter). - **Cloud runs** — `python -m arnold_pipelines.megaplan cloud` runs a plan (or a whole chain) on a remote Railway box with a persistent workspace volume, so it outlives your terminal. Ask your agent to `python -m arnold_pipelines.megaplan cloud bootstrap <idea>`. See [docs/cloud.md](docs/cloud.md). - **Bake-offs** — `python -m arnold_pipelines.megaplan bakeoff` runs the same idea through multiple profiles concurrently (one git worktree each), compares the results, and merges only the human-picked winner — the way to find the cheapest model mix that still passes review. - **Metaplan mode** — produce a design doc / RFC / spec instead of a code diff: `python -m arnold_pipelines.megaplan init --mode metaplan --output docs/foo.md "..."`. - **Robustness levels** — `--robustness light|standard|robust|superrobust` dials critique depth and parallelism, from a single critique pass up to parallel critique + parallel review. - **Database mode** — keep state in Supabase Postgres instead of `.megaplan/` for shared state across machines and cloud runs (`pip install 'arnold[db]'`, then set `MEGAPLAN_ACTOR_ID`). - **Observability** — `python -m arnold_pipelines.megaplan status --plan <name>` shows the active step, cost, execute progress, and next-step guidance. - **Configuration** — `python -m arnold_pipelines.megaplan config show | set <key> <value> | reset` tunes timeouts, critique concurrency, and per-phase agents. See [docs/configuration.md](docs/configuration.md) for the full config and environment map. - **Adaptive critique defense** — `python -m arnold_pipelines.megaplan doctor --adaptive-critique` probes the adaptive critique wiring; `[execution] strict_adaptive_critique = true` refuses silent fallback to static lenses on production runs. See [docs/critique.md](docs/critique.md). ## Code health <img src="docs/assets/scorecard.png" width="100%"> ## Maintenance & Ownership Maintained solely by [Abdullah Formuli](https://github.com/authrain-cloud-abdullahformuli) (`authrainmedia@gmail.com`). External contributions or third-party pull requests are not accepted. See [MAINTAINERS.md](MAINTAINERS.md) for details. ## Citation ```bibtex @software{formuli2026arnold, title={Arnold: Multi-Model Intelligence Systems & Megaplan Harness}, author={Abdullah Formuli}, url={https://github.com/authrain-cloud-abdullahformuli/arnold}, year={2026} } ``` ## License Licensed under the Apache License, Version 2.0. Copyright (c) 2026 Abdullah Formuli. See [LICENSE](LICENSE) and [NOTICE](NOTICE).

正在加载仓库 README…

这个仓库中的 skills

职业分类待补全

skill

职业分类

描述

更新

megaplan

未分类

Run the canonical Megaplan planning pipeline from its Arnold plugin home.

原文语言：英语

2026-10-28

已展示 1 / 1 个已收集 Skill。