Skip to main content

qoq-fullstack

A second Claude Code plugin in the qoq-agent-skills marketplace, new in v7.1.0. It coordinates one feature across several repos — an API and a UI, two APIs and a front end, any set of qoq-based JavaScript/TypeScript repos in one parent folder. Installing it also installs qoq:

/plugin install qoq-fullstack@qoq-agent-skills

This thread is the chief of staff. It owns a business-level master plan (the BRD), starts one background Claude session per repo, relays their questions to you, and unblocks work across repos. It never writes code, never plans a repo, and never decides anything technical — a technical decision made here would bind a repo whose conventions this thread hasn't read. Each repo session runs the ordinary qoq plan and qoq execute on its own code.

InvocationDoes
/qoq-fullstack plan <requirements>business grilling → BRD → one plan per repo, relayed
/qoq-fullstack execute <slug>every repo's plan in parallel, gated per master
/qoq-fullstack status [<slug>]read-only; one script call, so it needs no model

The files are the state​

The parent folder holds the repos plus the chief of staff's own files:

<project>/
feature.json # the repos — written once, reused by every feature
feature/<slug>/brd.md # one per feature: the master plan
contracts/<name>.md # written by the providing repo's session
<repo>/ <repo>/ … # independent git repos, each with its own plans/

feature.json lists every repo (name, path, role, optional after), the named integration checks (each an npm run <script> in one repo) and the permissionMode and model repo sessions start in. model is chosen by you, once, because N repos run N sessions at the same time; opus is the recommendation. A feature.json written before the field existed has none, so the chief of staff asks once and writes your answer in before launching. Every sub-command starts by running qoq's scripts/feature-status.mjs, and again after every message. It acts only on what the script returns, because a repo session writes its plan file before it sends a message — if a message and the file disagree, the file wins.

plan​

  1. Business grilling — business only. Actors, what they can do, what they observe, where a milestone ends. A technology question is never asked: it is parked verbatim in the BRD, and the repo it lands in asks it against code this thread has never read.
  2. The BRD. feature/<slug>/brd.md holds every requirement as a Given/When/Then scenario tagged with its master milestone, the master milestones with their Depends on and Integration check, and the parked questions. It is approved before any repo plans.
  3. A plan per repo, in after order; repos with no after start together. Each repo decides from its own code whether anything lands in it — if not, it answers no changes needed and gets no plan and no session in execute. Approvals and "no changes needed" reach you through the relay: a repo can't approve its own plan.
  4. Schedule. Repos run in parallel, so the order inside each decides how long the others wait. scripts/schedule.mjs simulates the run and reorders independent milestones so the piece another repo needs goes first. A hard dependency never moves, and no question is asked.
  5. Coverage. Every master milestone needs at least one repo milestone serving it; a gap is raised with you.

The repo side of this protocol — the Feature header, the Serves column, <repo>:M<N> dependencies and contracts — is described under plan.

execute​

A loop over the status script: wake every repo whose next.runnable is true with /qoq execute <planDir>, and run each master milestone's integration checks once every repo's working tree is clean. A pass is the only thing that marks the master milestone delivered and unblocks the next. A failure goes to you verbatim, and the fix goes back to a repo as a qoq replan. When a repo escalates, only the repos waiting on its milestone are stuck. When every master milestone is delivered, the feature is complete.

Repo sessions​

  • One session per repo named <project>-<repo>, started with claude --bg --model <model> (feature.json's model) in the repo's own folder, so it loads that repo's CLAUDE.md, qoq.config and agents.
  • Every question a repo would ask goes to the chief of staff, and from there to you verbatim. Nothing technical is answered by the chief of staff.
  • Each repo works on feature/<slug> in its own checkout — no worktrees, no branch switching.

Setup, once per repo, needs your say-so: crossSessionInbound: "accept" and worktree.bgIsolation: "none" in the repo's .claude/settings.local.json; Claude Code 2.1.236 or later; a trusted workspace; and the repo on feature/<slug> with a clean tree.