Skip to content
Back to skills

Bulma Goal

ASecurity

Drive developer → CI green → QA comment+evidence across turns with schedule_wake. Use when /goal or /bulma-goal, or after a draft PR is opened and CI is still pending.

  • 2 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added October 17, 2026
ai-agentsgogit

Security analysis

A100/100

Pro scans all 11 files and shows the line behind each finding

Scanned October 17, 2026

npx -y skills add ruverd/bulma --skill bulma-goal --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Bulma Goal?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Bulma Goal
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/ruverd-bulma-goal/badge)](https://www.skillsdirectory.com/skills/ruverd-bulma-goal)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: bulma-goal
category: graph
description: >
  Drive developer → CI green → QA comment+evidence across turns with
  schedule_wake. Use when /goal or /bulma-goal, or after a draft PR is
  opened and CI is still pending.
argument-hint: "<ticket | PR url | status | cancel>"
---

# Bulma goal / loop

The graphs do **not** finish in one turn. CI takes 20–30 min. This
skill keeps the session working until QA has **commented with evidence**.

**REQUIRED:** [COMPLETE.md](references/COMPLETE.md) ·
[LOOP.md](references/LOOP.md) · bus PROTOCOL.md ·
[DISK.md](../bulma-bus/DISK.md) (`.bulma-*` is **global**, never git root) ·
`bulma-memory`

Chat: `bulma-memory`. Unslop always.
Worktree and branch rules: [JOBS.md](../bulma-bus/JOBS.md) §Worktree.

## Commands

| Args | Action |
|---|---|
| ticket id / feature text | Start goal + developer graph |
| PR url / `owner/repo#N` | Start goal on that PR (skip implement if it exists) |
| `status` | Read `.bulma-goal/STATE.md` + `gh pr checks` + last QA comment |
| `cancel` | Delete the scheduler loop; leave graphs as-is |

## Start

1. Load `bulma-memory`. Write `.bulma-goal/STATE.md` from
   [templates/STATE.md](templates/STATE.md).
2. Name the completion bar (verifier / user):

```text
Draft PR for <id> is CI-green, MERGEABLE, and has a bulma-qa
comment on the head SHA that includes evidence.
```

If the host has a `/goal` (or equivalent) command, register that
sentence there. If it does not, STATE is enough.

3. If there is no PR yet → load **bulma-developer** GRAPH (`deliver`).
4. When a draft PR exists and CI is not green → **schedule_wake**
   ([LOOP.md](references/LOOP.md)). Do not `gh pr checks --watch`.
5. Stop the turn. The wake resumes this skill.

## Each wake (or `/bulma-goal` resume)

Read STATE. Inspect the real PR (`gh pr view`, `gh pr checks`,
issue comments). Then **one** step:

| World | Next |
|---|---|
| No PR | developer `deliver` |
| CI pending / in progress | stop (wait for next fire) |
| CI red | developer `fix` / fd ci loop |
| Green but not MERGEABLE | developer `mergeable` |
| Green + MERGEABLE, no QA comment on **this** SHA | enqueue-or-start QA (JOBS.md) |
| QA comment on this SHA, verdict FAIL + PR_BUG | developer `fix` |
| QA comment on this SHA, PASS / FAIL-unrelated / BLOCKED documented | **complete** |

Complete = [COMPLETE.md](references/COMPLETE.md) all true, then
`cancel_wake`.

## Never

- Claim the goal done without the QA comment + evidence on **head** SHA
- Sit in `--watch` (CI is longer than a tool timeout)
- Spawn graph-agents as children (bus switch / load GRAPH)
- Start a second QA while `qa_active` holds another **live** PR
- Merge

## After QA comment

Cancel the loop. Report in the chat language: PR, SHA, QA verdict, comment URL, evidence URL.

Files in this skill

  • GRAPH.md2.2 KB
  • SKILL.md2.7 KB
  • STATE.schema.md1.1 KB
  • agents/openai.yaml188 B
  • nodes/admit.md824 B
  • nodes/complete.md542 B
  • nodes/inspect.md689 B
  • nodes/step.md1.2 KB
  • references/COMPLETE.md1 KB
  • references/LOOP.md1.4 KB
  • templates/STATE.md231 B

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.