flow-next-flow — скилл для Claude Code и Codex: маршрутизация задач
Скилл flow-next-flow для Claude Code и Codex, который маршрутизирует работу между этапами: захват, уточнение, планирование, разработка, QA, создание PR и слияние. Запускается автоматически или в режиме --auto для беспилотных прогонов.
MIT ★ 730 обновлён 2026-10-29
/plugin marketplace add gmickel/flow-next
/plugin install flow-next@flow-nextЧто делает flow-next-flow
flow-next-flow работает в Claude Code и Codex как дирижёр проекта: читает задачу, выбирает нужный этап из маршрутной матрицы и вызывает соответствующий скилл. Поддерживает как интерактивный режим (пользователь решает на каждом шаге), так и полностью автономный (--auto), когда агент сам проходит все этапы до решения, которое требует человека.
Оперирует спецификациями, задачами из трекера, веткам и багрепортам — определяет, на каком этапе цикла они находятся (есть ли PR, выполнены ли все задачи, нужна ли ревью). Маршрутизация и QA-гейт встроены в логику и не спрашивают судью, поэтому результат один и тот же независимо от наличия TypeSafe-ключа.
Утилита flowctl идёт в комплекте; для инициализации проекта нужен flowctl init. Режим --explain показывает решение без выполнения.
Какие задачи решает
- автоматизировать выбор этапа для любой задачи в Flow-Next проекте
- запускать работу беспилотно через несколько этапов — захват, планирование, разработка, QA, PR
- определить, готова ли спецификация к слиянию и какой этап нужен дальше
- отследить, какой этап был пропущен и почему
Примеры запросов агенту
- Дай мне скилл для реализации этого тикета — он выберет план или сразу начнёт код
- Запусти
/flow-next:flow --autoдля полностью автономного прогона до QA - Объясни, куда поведёт себя эта багота и какие этапы нужны — запусти с
--explain
Как установить flow-next-flow
1Через маркетплейс плагинов
/plugin marketplace add gmickel/flow-next/plugin install flow-next@flow-nextКоманда ставит плагин flow-next целиком, со всеми его скиллами
2Через Skills CLI (Vercel)
npx skills add gmickel/flow-next --skill flow-next-flow -a claude-code3Вручную: положить файлы
~/.claude/skills/flow-next-flow/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/flow-next-flow/
1Через Skills CLI (Vercel)
npx skills add gmickel/flow-next --skill flow-next-flow -a codex2Вручную: положить файлы
.agents/skills/flow-next-flow/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add gmickel/flow-next --skill flow-next-flow -a cursor2Вручную: положить файлы
.cursor/skills/flow-next-flow/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install gmickel/flow-next plugins/flow-next/skills/flow-next-flow2Вручную: положить файлы
.github/skills/flow-next-flow/SKILL.md1Через Skills CLI (Vercel)
npx skills add gmickel/flow-next --skill flow-next-flow -a gemini-cli2Командой платформы
gemini skills install https://github.com/gmickel/flow-next --path plugins/flow-next/skills/flow-next-flowВручную: положить файлы
.devin/skills/flow-next-flow/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Claude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о flow-next-flow
Как установить flow-next-flow в Claude Code?
Выполни /plugin marketplace add gmickel/flow-next, потом /plugin install flow-next@flow-next и добавь скилл командой npx skills add gmickel/flow-next --skill flow-next-flow -a claude-code. Скилл появится в ~/.claude/skills/flow-next-flow/SKILL.md.
Что делает flow-next-flow?
Читает задачу, определяет её текущее состояние в цикле разработки (захват, планирование, разработка, QA или PR) и вызывает нужный этап-скилл. В режиме --auto проходит через всё автоматически; в обычном режиме спрашивает решение на каждой точке ветвления.
Чем отличаются интерактивный режим и --auto?
Без флага это интерактивная работа: после каждого этапа скилл спрашивает, что делать дальше. С --auto он сам маршрутизирует до решения, которое требует человека (например, до утверждения для слияния). --auto означает, что агент работает полностью беспилотно, пока не упрётся в гейт.
Исходный текст
SKILL.md · MIT
---
name: flow-next-flow
description: Use first in a project that uses Flow-Next for any request to implement, build, continue or deliver its specs, whatever delivery or test steps the request adds, unless the user names another skill or asks to work directly. Also for a spec or task id, an idea or change request, a pasted bug report or traceback, a captured profile, trace or crash dump to explain, a mistake agents keep repeating, a prompt or skill change whose effect should be measured, "what could this break" or "is this safe to merge" about a diff, something slow to speed up or another number to move (hill-climbing), a behaviour-preserving cleanup, a design or behaviour choice to settle, an outcome to reach over several specs (a port, a migration, a PRD), a tracker issue, a branch, or "what should I do next". Only a quick question or a one-line edit may skip it. Flow routes the work and runs the right stage; --auto runs unattended.
user-invocable: false
allowed-tools: AskUserQuestion, Read, Bash, Grep, Glob, Write, Edit, Task, Skill
---
# /flow-next:flow - the conductor
Flow chooses the next step so the user does not have to. It reads what it was given, routes from the shared routing reference, runs the routed stage skill, and continues until the next decision that belongs to a human. It re-implements no stage logic: capture, refine, plan, plan-review, work, qa, make-pr, resolve-pr, and land keep their own contracts, receipts, and gates.
**Role:** conductor, inline (no `context: fork`) so `AskUserQuestion` stays reachable. On hosts without it, fall back to a plain-text numbered prompt with a final `Other - type your own answer` option.
The hop loop is Steps 1 to 5 below. Read a reference only at the step that names it.
## Preamble
**CRITICAL: flowctl is BUNDLED - NOT installed globally.** `which flowctl` will fail (expected). Define once; subsequent blocks (here and in the steps below) use `$FLOWCTL`:
```bash
FLOWCTL="${DROID_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT}}/scripts/flowctl"
[ -x "$FLOWCTL" ] || FLOWCTL="<plugin-root>/scripts/flowctl" # <plugin-root> = the directory two levels above this skill's SKILL.md file (the harness gave you that file's absolute path when the skill loaded); substitute it literally
[ -x "$FLOWCTL" ] || FLOWCTL=".flow/bin/flowctl"
```
## Mode detection
Parse `$ARGUMENTS` as exact tokens (never substrings), before any read or write: `--auto` sets `AUTO=1`, and so does `--goal`, because goal mode is unattended; `--tick` sets `AUTO_TICK=1` (one hop, then stop; meaningful only with `--auto` or `--goal`); `--explain` sets `EXPLAIN=1`; `--review=<backend>` sets `REVIEW_OVERRIDE` and is passed through unchanged to every stage it dispatches. The destination parse is shared by both modes:
```bash
FLOW_UNTIL=""
FLOW_DESTINATION_ERROR=0
LAND_AUTHORIZED=0
LAND_SCOPE_SPEC=""
LAND_SCOPE_PR=""
AUTO=0
for ARG in $ARGUMENTS; do
case "$ARG" in
--auto|--goal) AUTO=1 ;;
--until=merge) FLOW_UNTIL=merge ;;
--until|--until=*) FLOW_DESTINATION_ERROR=1 ;;
esac
done
if [ "$FLOW_DESTINATION_ERROR" = 1 ]; then
if [ "$AUTO" = 1 ]; then
echo 'PILOT_VERDICT=NEEDS_HUMAN spec=- stage=- reason="invalid destination; use --until=merge"'
else
echo 'NEEDS_HUMAN: invalid destination; use --until=merge'
fi
exit 1
fi
export FLOW_UNTIL
```
`--until=merge` authorizes landing the selected item in this invocation; read `references/tail.md` at that boundary. Consent is current host context, never recovered from an environment variable, old transcript or receipt. Consume destination tokens rather than passing them to a build stage; everything else is the starting point, verbatim.
**`AUTO=1`: read [auto.md](auto.md) and follow it.** Attended runs never load it.
## Autonomy refusal - runs right after the token parse
Attended flow and `flow --auto` are two drivers and are never nested. Without `--auto`, flow is attended: under any autonomy marker (scan the marker namespace: `FLOW_AUTONOMOUS`, `AUTONOMOUS=1`, a `mode:autonomous` token), stop before any read or write:
```
NEEDS_HUMAN: /flow-next:flow is attended - run /flow-next:flow --auto for unattended runs
```
With `--auto` or `--goal` there is no marker refusal, because the unattended run sets `FLOW_AUTONOMOUS` and `mode:autonomous` for the stages it dispatches. A run that routed, dispatched, or asked under a marker it should have refused has broken this.
## Invariants (every run)
- **Read [working-rules.md](../../references/working-rules.md) before the first route step** and follow it on every route.
- **Route on content and context, never on input kind**, from `references/route-matrix.md` at the route step.
- **Ask only on a fork that is material and not observable** (`references/prototype-before-ask.md`); at most one question per hop.
- **Never fabricate a review, QA or completion verdict.** Every stage flow skips is recorded with its reason (`stage: <name> - skipped(<kind>: <detail>)`).
- **Invoke stage skills; never re-implement their steps inline.** Land owns merge (only with current scoped consent, per `references/tail.md`); make-pr owns spec close. Never dispatch another flow or a loop from inside a run, and never force-push.
- **`--explain` writes nothing and dispatches nothing** (`references/explain.md`).
## The hop loop
If `.flow/` does not exist, print `No .flow/ directory - run \`$FLOWCTL init\` first.` and stop.
## Step 1: Read the starting point
Read what was given and decide what it is: a spec or task id (`$FLOWCTL show <id> --json`,
`$FLOWCTL cat <id>`), a tracker issue (through the access this session already has), a branch
(`git log` and the spec whose `branch_name` matches), a path, a pasted bug report, a sentence of
intent. With no argument, read [references/no-argument.md](references/no-argument.md).
## Step 2: Route
Read [references/route-matrix.md](references/route-matrix.md). Routing and the QA gate never
ask the judge, so a run with a TypeSafe key and one without take the same route.
Then route once per hop (in auto mode, use the route `pilot snapshot` returned instead). A spec
gets the route call, which decides its lifecycle in code and never asks the judge:
```bash
"$FLOWCTL" judge --preset route --spec <spec-id> --json | jq -c '{available, reason, pr_probe_failed, decision}'
```
For a spec, code applies lifecycle order: an observed PR goes to landing, a closed spec without a
PR to you, all tasks done to QA and make-pr, a `stale` plan review to plan-review, tasks to the
recorded work route, a ready zero-task spec to direct or plan. When `decision.met` is true, use `decision.value`, unless the person asked to be interviewed on the spec or passed a refine lens with it (`--biz`, `--tech`, `--scope=<lens>`): that routes to refine (`references/plan-vs-no-plan.md`). Intake without a spec
is yours: decide from the matrix. Print one line per hop: `Route: <route> (code)` or
`Route: <route> (host)`.
`EXPLAIN=1`: resolve the route (reading the references below as needed) but record nothing, then
read [references/explain.md](references/explain.md) and stop there.
A ready spec with no tasks and no recorded route: read
[references/plan-vs-no-plan.md](references/plan-vs-no-plan.md), resolve the rule, and record it
before any stage runs (`$FLOWCTL spec set-no-plan <id> --json` for direct,
`$FLOWCTL spec clear-no-plan <id> --json` for a positive plan signal). All tasks done and no PR:
read [references/gate-selection.md](references/gate-selection.md) for QA, then make-pr. An
existing PR: read [references/tail.md](references/tail.md). Uncaptured intent whose criteria trip
the tripwire: read [references/spec-count.md](references/spec-count.md). Two routes that would
materially differ with an answer you cannot observe: read
[references/prototype-before-ask.md](references/prototype-before-ask.md) before asking, and ask at
most one question per hop.
## Step 3: Run the routed stage
Invoke the stage skill by name with its normal arguments, passing `--review=<backend>` through when
`REVIEW_OVERRIDE` is set. Never copy a stage's steps inline.
- **Capture** gets the exact token `from:flow`; it applies plan-vs-no-plan itself and offers the
saved spec for review. A request to capture or review only never authorizes work.
- **Defect reproduction**, when the report does not say where the problem is and `.flow/features/`
exists: read [references/defect-intake.md](references/defect-intake.md) first.
- **Work** runs `flow-next:flow-next-work <spec-id>`; a recorded `no_plan` pre-answers its fork.
Pick the branch rather than letting work ask: `--branch=current` on a branch other than the
default, else `--branch=new`, unless the user named one. State the choice in one line.
- **QA** follows gate-selection.md; a skip is recorded, never silent.
- **Make-pr** runs under `--auto`, or attended when the person asks for the PR (working-rules.md);
it ends a run from intent unless `--until=merge` or current explicit consent authorizes landing
(tail.md). A PR this run just opened gets no landing question.
A stage that stops with `NEEDS_HUMAN`, a verdict that needs a person, or an open product question
stops flow with the same report; never answer on the user's behalf.
## Step 4: Re-evaluate
Re-read state (`$FLOWCTL show <spec-id> --json`, the PR, the receipts) and return to Step 2.
Advance on observed state, never on a stage's narration.
## Step 5: Stop and report
A pick among options a stage produced (prospect's candidates, chart's capture-or-split, refine's
choices) is asked inline, one per hop, and the run continues. Otherwise stop at the first of: the PR
exists (without landing authority), a decision a stage handed to the person, or a blocking question
that is not a pick. A PR this run just opened, without landing authority, ends the run; read
tail.md at a PR boundary only otherwise. Print this report: one `stage:` line per stage reached (a
skipped stage with its reason), inline picks on the `Route taken` line, then `Next:` in the host's
command form.
```
Flow stopped at: <the human decision, "PR exists", or the observed landing outcome>
Route taken: <hop 1> -> <hop 2> -> ... (an inline pick reads `prospect [picked: <candidate>]`)
stage: <name> - ran [<start>..<end>] | skipped(<policy|config|empty|error|reach|signal absent|despite unresolved risk>: <detail>) | failed(<reason>: <detail>) (one line per stage reached)
Next: <natural-language prompt or slash command, or the decision the user must make>
```
© Gordon Mickel · MIT · gmickel/flow-next