until

Loop a stage and stop once a predicate is satisfied — `do … until (cond)`: the body runs FIRST, then the condition is checked. The loop always runs at least once. Full loop safety: hard cap on iterations, convergence detection (stops if output stops changing), and a clean kill switch. USE when the user wants semantic termination after acting — "fix the build until it passes", "extract issues until none remain", "retry until it succeeds". For checking the condition BEFORE running (the body may not run at all), use `/agentflow:while`. For a fixed count, use `/agentflow:repeat`.

跳到安装

仓库: AleSaiani/agentflow
最近来源活动: 2026年5月25日 18:30
检测到的 SKILL.md 语言: 英语
星标: 5
分支: 0

安装方式

默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

直接命令不会经过审查 Prompt；运行前请先检查来源。

npx skills add https://github.com/AleSaiani/agentflow --skill until

命令会保持在同一行。复制前请横向滚动并检查完整内容。

想先保存到本地？可下载 SkillsMP 当前能够提供的文件。

检查来源文件

决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。

SKILL.md

来源说明 · 只读预览

返回搜索返回安装

foreach

Apply ONE operation to EVERY item of a list — the **map** (N→N). The operation is a prompt; items come from a JSON list, a markdown checklist, or another run's output. Runs in parallel subagents or inline in the main thread, with persistent state and cross-turn auto-continue. USE on any "do the same thing to each of these" intent — formal OR casual: "for each X…", "per ogni…", "go through these one by one", "do every item in TODO.md", "review each of these files", "handle all of them", a glob, a folder. Trigger on the INTENT, then judge by count (the count gate, Step 2.5): - 1–2 items → just do it inline, no machinery; - a borderline handful (~3–10) → ask whether to use the durable/parallel mechanism or handle inline; - genuinely many, or heavy/independent items → use foreach. DON'T use when the list must first be generated from a spec (→ /agentflow:enumerate, then foreach), or you need one combined output (→ /agentflow:reduce). For the prebuilt "review every file → digest", use /agentflow:audit. Explicit

2026-05-27

run-workflow

Run a declarative workflow-file end to end: init a `/agentflow:pipe` from a JSON `WorkflowSpec`, then drive it to completion — auto-running every bash/json/deterministic stage and dispatching agents only where genuinely needed. Use `--dry-run` to preview the resolved plan without executing anything. The one-command way to execute a reusable, versioned workflow. USE when the user points at a workflow file ("run this workflow", "run workflows/audit/WORKFLOW.md", "run my pipeline") or wants to run a saved `WorkflowSpec`. Workflows are normally self-contained folders (`workflows/<name>/WORKFLOW.md` + sibling scripts), but a bare `.json` path also works.

2026-05-27

audit

Deep code-audit recipe: discover files in a target → review each (LLM, cached) → partition by component → executive digest with hotspots and recurring patterns. A layer-3 recipe over /agentflow:pipe (the 6-stage pipeline and caching are detailed in the body). USE when the user asks to "review / audit / find bugs across" a folder, repo, or glob and wants a persisted, structured report — and the target is large enough (≥ ~10 files) that per-file parallelism plus a digest beats reading inline. DON'T use for a few named files (read them inline), exploratory questions ("what does this do?"), or generic per-item work that isn't a code review (→ /agentflow:foreach). Not a `/`-menu command — it's the shipped `workflows/audit/` recipe: reach it by asking ("audit src for bugs"), or run it directly with `/agentflow:run-workflow workflows/audit/WORKFLOW.md`.

2026-05-26

查看这个仓库的全部 Skills

name	until
user-invocable	false
description	Loop a stage and stop once a predicate is satisfied — `do … until (cond)`: the body runs FIRST, then the condition is checked. The loop always runs at least once. Full loop safety: hard cap on iterations, convergence detection (stops if output stops changing), and a clean kill switch. USE when the user wants semantic termination after acting — "fix the build until it passes", "extract issues until none remain", "retry until it succeeds". For checking the condition BEFORE running (the body may not run at all), use `/agentflow:while`. For a fixed count, use `/agentflow:repeat`.
allowed-tools	Bash, Read, Agent
argument-hint	--stage "<bash>" --stop "<bash predicate>" [--max-iterations N]

/agentflow:until — do…until loop

Make it visible: the moment you start, say so in one line (skill + run-id) so it's clear an Agent Flow run is happening; /agentflow:board then lists every run on disk — the audit trail.

Engine: iterate. until mode: the stop predicate exits 0 when satisfied → stop.

Init

node "${CLAUDE_PLUGIN_ROOT}/dist/state/iterate.js" init <run-id> \
  --stage "<body command>" \
  --stop  "<predicate command>" \
  [--max-iterations <N>] [--no-convergence-check]

--stage and --stop take plain bash commands (--mode defaults to until here). The predicate exits 0 when satisfied → stop.

Drive (one iteration per turn)

node "${CLAUDE_PLUGIN_ROOT}/dist/state/iterate.js" run-iteration <run-id>

Order per iteration: run the stage, then check the predicate. action: "continue" → end the turn (Stop hook resumes); action: "stop" → surface the result. reason is one of predicate_satisfied | max_iterations | convergence | stage_failed | killed. The predicate sees ITER_OUTPUT_PATH (this iteration's stdout) and ITER_PREV_OUTPUT_PATH.

until

来源信息

安装方式

检查来源文件

同仓库更多 Skills

同仓库更多 Skills

/agentflow:until — do…until loop

Init

Drive (one iteration per turn)

/agentflow:until — do…until loop

Init

Drive (one iteration per turn)