Long Running Agent

A plugin of skills for long-running agent work: writing the completion contract before a long run, then running the loop that carries it to a verified finish.

The skills

• write-goal: help the user craft a well-specified /goal objective for goal mode — what must become true, how that truth is proven, where the work may and may not reach, and when to stop and report instead of grinding on. It works with the user to draft and refine the objective, then starts it once the wording is approved.
• run-long-task: the complete loop for a large agent effort — frame the contract, choose the work unit and the gates, write the shared decisions once, pilot the loop, bound concurrency by infrastructure, keep authoring and review independent, save failures as scoped work queues, verify the verification, and separate integration from release.
• adversarial-loops: the review loop that catches what compiling and passing tests do not — separate contexts, one implementer plus two or more adversarial reviewers plus a fixer, and repairing the process that keeps producing the same mistake.
• orchestrate-migration: the migration-shaped cut — shared mapping and per-item decision records, a pilot, mechanical translation, the compiler as a work queue, and ordered executable gates.
• run-goal: the missing middle of goal mode — freeze an outcome plan without rewriting the objective, judge completion from saved evidence rather than the implementer's say-so, and stop when the same gap repeats. Token budgets and pause, resume, and clear stay with the host.

write-goal produces the contract; run-goal carries one goal to a verified finish; run-long-task runs a large effort. adversarial-loops and orchestrate-migration are focused cuts of the same loop, for review work and for migrations.

Methodology sources

references/ holds methodology write-ups that skills in this repo draw on. They are research notes, not skills, and are not part of the plugin surface.

• references/bun-in-rust-methodology.md: how Bun was ported from Zig to Rust in eleven days with supervised agent workflows — the decision criteria, the prep artifacts, the orchestration loop, the failure modes that forced process changes, and the verification strategy. Extracted from Jarred Sumner's “Rewriting Bun in Rust”. run-long-task, adversarial-loops and orchestrate-migration are all extracted from it.

Compatibility

The write-goal skill text refers to Kimi Code's /goal, AskUserQuestion, and CreateGoal capabilities. Other agent hosts may require equivalent tools or adaptations to start goals. This copy preserves the upstream behavior and does not add cross-host integration. The other skills are host-agnostic. run-goal does not implement token budgets or pause, resume, and clear; those stay with the host.

Install as a plugin

This repository is both the skill source and a single plugin named long-running-agent. The skills live at skills/<name>/SKILL.md.

Claude Code

claude plugin marketplace add AkaraChen/long-running-agent
claude plugin install long-running-agent@long-running-agent

Local check without installing:

claude --plugin-dir .

Then invoke a skill as /long-running-agent:write-goal or /long-running-agent:run-long-task. claude plugin validate . checks the marketplace catalog.

Codex

codex plugin marketplace add AkaraChen/long-running-agent
codex plugin add long-running-agent@long-running-agent

A checkout of this repo also exposes the plugin through .agents/plugins/marketplace.json. Restart Codex after adding the marketplace, then enable long-running-agent. Invoke a skill as $write-goal or $run-long-task.

Source / Attribution

The write-goal skill is copied from the Kimi Code open-source repository by Moonshot AI.

• Upstream repository: https://github.com/MoonshotAI/kimi-code
• Upstream commit: f12d59e089e2531a33fbca30b26ffeabd5862b45
• Original file: packages/agent-core-v2/src/features/skill/catalog/builtin/write-goal.md
• Copied on: 2026-09-07
• License: MIT, Copyright (c) 2026 Moonshot AI.

The original skill text is preserved unchanged and packaged as skills/write-goal/SKILL.md. The upstream TypeScript registration wrapper is not included because it belongs to Kimi Code's runtime.

run-long-task, adversarial-loops and orchestrate-migration were extracted from the methodology document in references/, which cites the source post. run-goal adapts the goal-mode plan, verdict, and stop rules from xai-org/grok-build, without its token budget or pause, resume, and clear commands.