Claude Skill

super-board

GitHub-Project-driven autonomous pipeline. Five verbs — onboard, lint, status, run, stop — that take a Project board from empty to drained across Build → QA → Review → Done lanes (labels qa · bug · feature route each card), with graceful shutdown / resume. Use when the user says

LLM Mart · 0 points · 0 views 0 listing impressions 0 install-command copies
Download erictechpro-super-board-skills_super-board-f014cfa.zip · 109 KB

Install

skills CLI npx skills add https://github.com/EricTechPro/super-board/tree/main/skills/super-board
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install erictechpro-super-board@llmmart
Git git clone https://github.com/EricTechPro/super-board.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole erictechpro/super-board collection as a plugin from our marketplace. Git is the plain clone.

README

/super-board

Drives a GitHub Project board on its own: drag a card into Ready, and it comes back merged with evidence on the PR.

What It Does

  • Plans waves from your dependency graph. Every Ready card whose ## Blocked by issues have all closed runs in the same wave. A Blocked card whose blockers closed is swept back to Ready; a dependency line it cannot read is flagged, never guessed.
  • Runs each card through three lanes: super-build (Building), super-qa (QA), super-review (Review). A failed lane bounces the card back with comments the next wave reads.
  • Merges only through the merge gate: a lock, the current base merged in, your verify_commands, then a squash-merge pinned to the reviewed commit.
  • Holds no state of its own. The board is the state, so Ctrl-C, restarts and rate-limit pauses lose nothing. Stranded Building cards return to Ready.
  • Pauses near your Claude usage limit when the usage check is recording (see the pack README).
  • Never writes product code. Your session is the orchestrator; lane agents do the work.

When To Use It

  • You have a GitHub Project of cards with acceptance criteria and want them built, tested and reviewed unattended.
  • You want to harden code that already exists: label those cards qa and they skip Building (QA and Review only).

Verbs

Command Does
/super-board onboard One-time setup; checks the board's columns, writes the config
/super-board lint Flags vague or missing ACs and unreadable Blocked by lines before a run
/super-board status Read-only 80-column board snapshot (~1.3s, pure Python)
/super-board run <slug> [--low\|--high] The loop; also the resume command
/super-board stop Posts "stopped mid-flight" notes, releases claims, stops workers

Install

Part of the super-board pack: ./install.sh /path/to/project. Needs gh (authenticated), jq, bash 3.2+, a GitHub Project (v2) with a Status field, dynamic workflows on in /config, and the mattpocock/skills pack. Full reference: docs/super-board/README.md.

Skill manifest

super-board — autonomous GitHub Project pipeline

Five verbs

Verb Where What it does
super-board onboard interactive one-time setup wizard; writes .claude/super-board/configs/<slug>.json
super-board lint interactive walks active-pipeline issues, flags vague ACs, runs pre-flight readiness
super-board status interactive (read-only) snapshot of active config, column counts, in-flight workers
super-board run in-session (default) the autonomous loop: waves of the super-board-wave dynamic workflow, run from this session (worker_backend: "claude-p" opts into the legacy headless runner .claude/bin/super-board-run.sh). Also the resume command — state lives on the board, not in process memory. Accepts a model-tier flag: --low (haiku/sonnet/opus ladder), default = medium (sonnet/opus/session), --high (opus for every card). --codex [--low|--high] overrides the backend for every lane with the matching Codex GPT ladder; --codex=<model> or codex.model pins one model for every card (see references/run-workflow.md). A host with no Workflow tool (a Codex session) runs as --codex by itself.
super-board stop interactive graceful shutdown: posts "stopped mid-flight" comments on every in-flight issue + PR, releases assignee mutexes, kills workers + dispatcher. Next super-board run resumes.

If invoked with no verb, ask which one.

In Codex the skill is not a slash command: it is $super-board <verb>, found in .agents/skills/.

Routing

If user says Load
super-board onboard ... references/onboard.md
super-board lint ... references/lint.md
super-board status ... references/status.md
super-board run --codex[=<model>] ..., or any run from a host with no Workflow tool (overrides any configured backend) references/run-workflow.md → "Codex runs"
super-board run ... (default — worker_backend unset or "workflow") references/run-workflow.md (lane lifecycles still come from references/run.md)
super-board run ... with config worker_backend: "claude-p" (legacy, explicit opt-in) references/run.md
super-board stop ... / "stop the run" / "pause the loop" / "kill super-board" references/stop.md
"resume" / "pick up where I left off" / "restart after stop" references/stop.md (resume = run; no separate verb)
Anything about Blocked exits, 🙋 needs you, or dropping a card references/block-template.md
Optional learning after a successfully completed run references/calibration.md (external installed skill; bounded receipt; explicit existing-skill edit paths)
Config structure questions references/config-schema.json
Worker gh-call discipline / rate-limit recovery references/rate-limit-etiquette.md (+ scripts/super-board-gh-guard.sh)

Orchestrator vs worker — the cardinal rule

super-board is an autonomous trader. The interactive Claude session that invokes any of the five verbs is an orchestrator, not a worker. The orchestrator:

  • Validates preconditions, then dispatches per the config's worker_backend: "workflow" (default) → stay in-session and run the wave loop in references/run-workflow.md (launch workflow, reconcile, repeat); "claude-p" (legacy, explicit opt-in only) → nohup .claude/bin/super-board-run.sh, report PID + log path, exit. Explicit --codex overrides either backend with headless Codex waves; so does a host with no Workflow tool, or bash .claude/bin/super-board-host.sh printing codex. The orchestrator never does product work itself.
  • Delegates all build / QA / review work to workers — headless claude -p (claude-p backend), workflow lane agents (workflow backend), or headless codex exec (explicit --codex).
  • Must NOT do product work itself, must NOT patch the dispatcher mid-run, must NOT wait for workers, must NOT hold context for multi-card progress.

If anything goes wrong during a run, the orchestrator captures the symptom and reports back — it does not silently expand the task into a fix. See references/run.md "Orchestrator delegation contract" for the full rule.

Files (super-board)
  • references
    • agents-md-block.md 2.6 KB
      Loading…
    • block-template.md 11.2 KB
      Loading…
    • calibration.md 6.1 KB
      Loading…
    • config-schema.json 23 KB
      Loading…
    • design-preflight.md 2.9 KB
      Loading…
    • lint.md 14.7 KB
      Loading…
    • onboard.md 35.3 KB
      Loading…
    • pr-author-notes.md 5.5 KB
      Loading…
    • project-checks.md 1.4 KB
      Loading…
    • rate-limit-etiquette.md 5.8 KB
      Loading…
    • run-workflow.md 29.1 KB
      Loading…
    • run.md 59.4 KB
      Loading…
    • second-account.md 5 KB
      Loading…
    • skills-readiness.md 3.1 KB
      Loading…
    • status.md 24.5 KB
      Loading…
    • stop.md 7.5 KB
      Loading…
    • ticket-format.md 2.5 KB
      Loading…
    • writing-standard.md 15.9 KB
      Loading…
  • README.md 2.1 KB
    # /super-board
    
    Drives a GitHub Project board on its own: drag a card into `Ready`, and it comes back merged with evidence on the PR.
    
    ## What It Does
    
    - **Plans waves from your dependency graph.** Every `Ready` card whose `## Blocked by` issues have all closed runs in the same wave. A `Blocked` card whose blockers closed is swept back to `Ready`; a dependency line it cannot read is flagged, never guessed.
    - **Runs each card through three lanes**: super-build (Building), super-qa (QA), super-review (Review). A failed lane bounces the card back with comments the next wave reads.
    - **Merges only through the merge gate**: a lock, the current base merged in, your `verify_commands`, then a squash-merge pinned to the reviewed commit.
    - **Holds no state of its own.** The board is the state, so Ctrl-C, restarts and rate-limit pauses lose nothing. Stranded `Building` cards return to `Ready`.
    - **Pauses near your Claude usage limit** when the usage check is recording (see the pack README).
    - **Never writes product code.** Your session is the orchestrator; lane agents do the work.
    
    ## When To Use It
    
    - You have a GitHub Project of cards with acceptance criteria and want them built, tested and reviewed unattended.
    - You want to harden code that already exists: label those cards `qa` and they skip Building (QA and Review only).
    
    ## Verbs
    
    | Command | Does |
    |---|---|
    | `/super-board onboard` | One-time setup; checks the board's columns, writes the config |
    | `/super-board lint` | Flags vague or missing ACs and unreadable `Blocked by` lines before a run |
    | `/super-board status` | Read-only 80-column board snapshot (~1.3s, pure Python) |
    | `/super-board run <slug> [--low\|--high]` | The loop; also the resume command |
    | `/super-board stop` | Posts "stopped mid-flight" notes, releases claims, stops workers |
    
    ## Install
    
    Part of the super-board pack: `./install.sh /path/to/project`. Needs `gh` (authenticated), `jq`, bash 3.2+, a GitHub Project (v2) with a `Status` field, dynamic workflows on in `/config`, and the [mattpocock/skills](https://github.com/mattpocock/skills) pack. Full reference: [`docs/super-board/README.md`](../../docs/super-board/README.md).
    
  • SKILL.md 4.6 KB
    ---
    name: super-board
    description: GitHub-Project-driven autonomous pipeline. Five verbs — onboard, lint, status, run, stop — that take a Project board from empty to drained across Build → QA → Review → Done lanes (labels qa · bug · feature route each card), with graceful shutdown / resume. Use when the user says "super-board", "/super-board", "drain my GitHub project", "set up the autonomous loop", "kick off the headless build/QA pipeline", or "stop super-board".
    ---
    
    # super-board — autonomous GitHub Project pipeline
    
    ## Five verbs
    
    | Verb | Where | What it does |
    |---|---|---|
    | `super-board onboard` | interactive | one-time setup wizard; writes `.claude/super-board/configs/<slug>.json` |
    | `super-board lint` | interactive | walks active-pipeline issues, flags vague ACs, runs pre-flight readiness |
    | `super-board status` | interactive (read-only) | snapshot of active config, column counts, in-flight workers |
    | `super-board run` | in-session (default) | the autonomous loop: waves of the `super-board-wave` dynamic workflow, run from this session (`worker_backend: "claude-p"` opts into the legacy headless runner `.claude/bin/super-board-run.sh`). Also the resume command — state lives on the board, not in process memory. Accepts a model-tier flag: `--low` (haiku/sonnet/opus ladder), default = medium (sonnet/opus/session), `--high` (opus for every card). `--codex [--low|--high]` overrides the backend for every lane with the matching Codex GPT ladder; `--codex=<model>` or `codex.model` pins one model for every card (see `references/run-workflow.md`). A host with no Workflow tool (a Codex session) runs as `--codex` by itself. |
    | `super-board stop` | interactive | graceful shutdown: posts "stopped mid-flight" comments on every in-flight issue + PR, releases assignee mutexes, kills workers + dispatcher. Next `super-board run` resumes. |
    
    If invoked with no verb, ask which one.
    
    In Codex the skill is not a slash command: it is `$super-board <verb>`, found in `.agents/skills/`.
    
    ## Routing
    
    | If user says | Load |
    |---|---|
    | `super-board onboard ...` | `references/onboard.md` |
    | `super-board lint ...` | `references/lint.md` |
    | `super-board status ...` | `references/status.md` |
    | `super-board run --codex[=<model>] ...`, or any `run` from a host with no Workflow tool (overrides any configured backend) | `references/run-workflow.md` → "Codex runs" |
    | `super-board run ...` (default — `worker_backend` unset or `"workflow"`) | `references/run-workflow.md` (lane lifecycles still come from `references/run.md`) |
    | `super-board run ...` with config `worker_backend: "claude-p"` (legacy, explicit opt-in) | `references/run.md` |
    | `super-board stop ...` / "stop the run" / "pause the loop" / "kill super-board" | `references/stop.md` |
    | "resume" / "pick up where I left off" / "restart after stop" | `references/stop.md` (resume = run; no separate verb) |
    | Anything about Blocked exits, 🙋 needs you, or dropping a card | `references/block-template.md` |
    | Optional learning after a successfully completed run | `references/calibration.md` (external installed skill; bounded receipt; explicit existing-skill edit paths) |
    | Config structure questions | `references/config-schema.json` |
    | Worker gh-call discipline / rate-limit recovery | `references/rate-limit-etiquette.md` (+ `scripts/super-board-gh-guard.sh`) |
    
    ## Orchestrator vs worker — the cardinal rule
    
    super-board is an **autonomous trader**. The interactive Claude session that invokes any of the five verbs is an **orchestrator**, not a worker. The orchestrator:
    
    - Validates preconditions, then dispatches per the config's `worker_backend`: `"workflow"` (default) → stay in-session and run the wave loop in `references/run-workflow.md` (launch workflow, reconcile, repeat); `"claude-p"` (legacy, explicit opt-in only) → `nohup .claude/bin/super-board-run.sh`, report PID + log path, exit. Explicit `--codex` overrides either backend with headless Codex waves; so does a host with no Workflow tool, or `bash .claude/bin/super-board-host.sh` printing `codex`. The orchestrator never does product work itself.
    - Delegates all build / QA / review work to workers — headless `claude -p` (claude-p backend), workflow lane agents (workflow backend), or headless `codex exec` (explicit `--codex`).
    - Must NOT do product work itself, must NOT patch the dispatcher mid-run, must NOT wait for workers, must NOT hold context for multi-card progress.
    
    If anything goes wrong during a run, the orchestrator captures the symptom and reports back — it does not silently expand the task into a fix. See `references/run.md` "Orchestrator delegation contract" for the full rule.
    
  • VERSION 6 B · in bundle

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.