dsh-goal-scaffold

dsh-goal-scaffold

English | 中文

What it does, in one sentence: when you drop a huge one-liner like "build me a complete task management system", this plugin asks you "want to plan it first?" — before the agent burns twenty minutes improvising.

Without it, a mammoth one-sentence request goes straight to work: no plan, no acceptance criteria, and nobody can tell whether it's drifting. With goal-scaffold installed, the first such message in a session gets a gentle interception:

you send a big vague task
  → a three-option prompt pops up:
      ① Plan first (Recommended)  ② Just start  ③ Don't ask again this session
  → pick ①: your request goes through, and the agent is told to first write
    plan.md (a checklist where every item has a verifiable acceptance
    criterion), create a goal with a round cap, show you the plan — then
    execute item by item, ticking boxes as it goes

You stay the boss of one click; the agent does the disciplined rest. Pairs with dsh-verify-judge, which makes the final "done" require passing tests.

Install

dsh plugin --profile web add github:zriyox/dsh-goal-scaffold
# or a local checkout:
dsh plugin --profile web add /absolute/path/to/dsh-goal-scaffold

Behavior notes

• Asks at most once per session. "Don't ask again" silences it; so does any answer, until a new session.
• Headless-safe: where no user-questions channel exists (e.g. the headless profile), every message passes through untouched — nothing blocks, nothing errors.
• Human messages only: goal-driver continuations, plugin injections, and tool artifacts never trigger detection.
• Fail-open on dismissal: if you dismiss the question to type something else, your original message proceeds as-is.

Configuration

- id: goal-scaffold
  config:
    enabled: true          # master switch
    minLength: 80          # minimum message length that participates in detection
    extraSignals: ['迁移']  # extra case-insensitive scope phrases

Development

pnpm install
pnpm run build   # tsc → lib/

License

MIT