Skip to main content

prior-art-reference

Standing pointer to the broader landscape of open-source agent harnesses, meta-harnesses, governance layers, multi-agent orchestrators, and usage/cost observability sidecars that solve problems adjacent to MultiTable. **Consult proactively** when (a) the user explicitly names a project — Mastra, OpenHarness, "Microsoft Agent Governance Toolkit", Cordum, CrewAI, LangGraph, "Strands Agents", "Harness SDK", VoltAgent, MetaGPT, ruflo, "Claude Flow", artificial (the Go harness), "Claude Squad", OpenCode (sst/opencode), Archon, CodexBar, awesome-agent-harness, awesome-harness-engineering, "9 open-source agent orchestrators" — or asks "is there an OSS project that does X", "what's prior art for Y", "similar projects to MultiTable", "harness comparison", "agent governance frameworks", "meta-harness", "multi-agent orchestrator", "cross-provider usage limits", "menu-bar usage tracker"; **AND ALSO** (b) work is stuck on a persistent bug that's resisted two or more fix attempts in an area covered by the parallels table (

Datos de origen

Repositorio
erickalfaro/multitable
Última actividad en el origen
19 de junio de 2026 a las 16:50
Idioma detectado de SKILL.md
inglés
Estrellas
33
Forks
3

Opciones de instalación

De forma predeterminada está seleccionado el prompt que primero revisa el origen. Puedes cambiar a un comando directo o descargar una copia local.

Revisa los archivos de origen

Lee SKILL.md y los archivos complementarios que muestra SkillsMP antes de decidir si quieres instalarlo.

Explorador de archivos
4 archivos

Mostrando SKILL.md

SKILL.md
Instrucciones de origen · Vista previa de solo lectura
name
prior-art-reference
description
Standing pointer to the broader landscape of open-source agent harnesses, meta-harnesses, governance layers, multi-agent orchestrators, and usage/cost observability sidecars that solve problems adjacent to MultiTable. **Consult proactively** when (a) the user explicitly names a project — Mastra, OpenHarness, "Microsoft Agent Governance Toolkit", Cordum, CrewAI, LangGraph, "Strands Agents", "Harness SDK", VoltAgent, MetaGPT, ruflo, "Claude Flow", artificial (the Go harness), "Claude Squad", OpenCode (sst/opencode), Archon, CodexBar, awesome-agent-harness, awesome-harness-engineering, "9 open-source agent orchestrators" — or asks "is there an OSS project that does X", "what's prior art for Y", "similar projects to MultiTable", "harness comparison", "agent governance frameworks", "meta-harness", "multi-agent orchestrator", "cross-provider usage limits", "menu-bar usage tracker"; **AND ALSO** (b) work is stuck on a persistent bug that's resisted two or more fix attempts in an area covered by the parallels table (provider adapters, permission/governance, multi-agent, parallel-agent isolation, model catalog, usage limits, sandboxing, workflow templates, observability); (c) scoping a non-trivial new feature in any of those areas — read this *before* designing rather than after; (d) hitting an unfamiliar protocol/wire-shape question that another harness has likely already solved. Reference only — not a provider, not implying any port.
allowed-tools
Read, Grep, Glob, Bash, WebFetch
# prior-art-reference A curated directory of ~16 open-source projects that overlap with MultiTable's problem space: unified harnesses over multiple AI coding agents, governance/policy layers, parallel-agent isolation, and multi-agent orchestration. Future sessions consult this skill **before** designing a new feature, to check what comparable projects already do. This skill is a **standing pointer**. It is **reference only** — read it to compare and steal ideas at the design level, not to port code. Any decision to copy code is a separate, case-by-case decision and is *not* implied by this skill loading. The dedicated [`omnigent-reference`](../omnigent-reference/SKILL.md) skill handles MultiTable's closest single sibling (omnigent-ai/omnigent) at much greater depth. This skill is the broader landscape around it. ## When to consult this skill proactively The skill exists so MultiTable doesn't re-derive solved problems. **Reach for it freely** — don't wait for the user to name a project. Three concrete triggers: ### 1. Persistent bug (two or more failed fix attempts) If you've tried to fix a bug twice and it's still broken — especially in an area another harness has clearly solved (per [`reference/parallels.md`](reference/parallels.md)) — stop and check prior art before a third attempt. Likely candidates: - Permission/approval flow bugs → OpenHarness, Microsoft Agent Governance Toolkit, omnigent - Streaming/delta accounting bugs → Mastra, OpenCode, the provider whose protocol you're adapting - Session resume / persistence bugs → omnigent, LangGraph (checkpointing), Mastra (threads) - Parallel-session isolation bugs → Claude Squad (git-worktree-per-agent) - Usage-limit / rate-limit parsing bugs → **CodexBar** (53+ providers) Pull the relevant project's source via `WebFetch` or `gh api`, find the analogous code path, and compare. A second design point almost always exposes the assumption you missed. ### 2. Complex new feature Before writing a non-trivial feature in any concept area listed in [`reference/parallels.md`](reference/parallels.md), read what the listed projects already do. This is cheap (1–2 `WebFetch` calls) and shapes the design — don't ship a v1 that the existing harnesses already iterated past. Mention what you found to the user as part of the plan, with project names cited. ### 3. Unfamiliar protocol / wire-shape question If MultiTable is about to integrate a new transport, parse a new event stream, or implement a new gating mechanism, the odds are high that one of these projects already faced it. Search their repos first; you may save a day. In all three cases: cite each project by name, don't blend concepts (see pitfalls), and remember — this skill authorizes *reading* their code for inspiration, not porting it. Any decision to copy code is a separate, case-by-case call. ## Strict isolation note - These projects are **external prior art**, not MultiTable providers. There is no Mastra/CrewAI/LangGraph adapter under `packages/daemon/src/agent/providers/`, and there shouldn't be — adding one is a separate user-directed decision, not something this skill authorizes. - Never blend concepts from multiple projects into a single recommendation as if they were one design. Cite each source by name when you reference it. - URLs decay. Before quoting current behavior (APIs, feature lists, license terms), verify with `WebFetch` against the URLs in [`reference/projects.md`](reference/projects.md) or `gh api repos/<owner>/<repo>`. ## Quick task → file map | What you want | Read | |---|---| | The full project list with TL;DRs and URLs | [`reference/projects.md`](reference/projects.md) | | MultiTable concept → which projects to consult | [`reference/parallels.md`](reference/parallels.md) | | Mis-uses and cautions | [`pitfalls.md`](pitfalls.md) | ## Decision tree: where do I look? ``` "Is there an OSS project that does <X>?" ├── Which MultiTable concept does <X> map to? ───── reference/parallels.md ├── Need the project's URL, TL;DR, or relevance ─── reference/projects.md └── Want to read the actual source ─── reference/projects.md → URL → WebFetch / gh api ``` ## When to *not* use this skill - The user is working on an existing MultiTable provider (Claude, Codex, Hermes, Grok, Cursor). Use that provider's dedicated skill instead. - The user is asking how omnigent specifically does something — use [`omnigent-reference`](../omnigent-reference/SKILL.md). - The user is asking about Claude API / Anthropic SDK semantics — use `claude-api` or `claude-agent-sdk`.
Ver en GitHub