AGENT SKILL

northstar

Planning-only intake: turn intent into a tracked, sliced plan and A→B handoff; never implement product changes. Use before autobahn execution.

License 许可
Apache-2.0
所属仓库
r3dlex/skills
最近更新
2026-10-29
Stars
★ 0
Forks
0

01 前言 Front Matter

eval
northstar
name
northstar
description
Planning-only intake: turn intent into a tracked, sliced plan and A→B handoff; never implement product changes. Use before autobahn execution.

02 资源文件 11

03 SKILL.md 正文

Northstar

Northstar converges intent into a tracked, sliced plan inside a repo initialized with ai-catapult-init v3. This lightweight composer delegates to existing skills; its output is the A→B handoff that autobahn consumes.

Execution boundary (hard stop)

Northstar is planning-only. Never implement a sliced goal, modify product code or tests, run implementation engines, or start autobahn while this skill is active. Writes are limited to the planning/tracking artifacts required below. After verifying the A→B handoff, stop and report it; implementation requires a separate, explicit autobahn invocation. Never continue into implementation in the same run, even when the initiating prompt also asks for implementation.

Quick Start

  1. Run the prereq gate (fail-closed) against the repo root: bash northstar/prereq-check.sh --root .
  2. Run the interview loop, then always raise an issue.
  3. Run ralplan to produce sliced goals.
  4. Publish the reviewed versioned bundle: bash northstar/handoff-write.sh --root <repo> --bundle <reviewed-v1.json>.

Prereq (fail-closed)

northstar assumes the ai-catapult-init v3 structure is already present and never bootstraps it. prereq-check.sh asserts the .ai/ structure exists and exits non-zero with guidance if absent. Do not proceed past a failed prereq. See modules/prereq.md.

The loop (delegation)

deep-interview is the primary driver: interview one question at a time until ambiguity is at or below threshold. The adversarial, skippable pass is two skills the user declines or accepts as a unit: grill-with-docs (doc-grounded — stress-tests the plan against the repo's language, ADRs and CONTEXT.md) then grill-me (open-ended). "Both satisfied" = the deep-interview gate is met AND (the adversarial pass is clear OR it was skipped). Delegate to those skills; do not reimplement their loops. See modules/loop.md.

Always raise an issue

After the loop, always raise an issue regardless of whether grill-me was skipped. Default is local-first markdown; a hosted tracker (GitHub/ADO/GitLab/Jira) is used only when it is configured and authorized, fail-closed per the ai-catapult-init host-policy. Delegate to to-issues and triage for canonical state labels and ownership. See modules/issue.md.

Ralplan → sliced goals

Run ralplan (consensus planning) on the crystallized spec to produce sliced goals — one tracer-bullet slice per future PR. ralplan owns the planning loop; northstar records its output as the sliced-goal artifacts.

A→B handoff

handoff-write.sh publishes the reviewed handoff-goals/1 bundle as immutable goals.json, handoff.md, and graph.json payloads under .ai/handoff/readiness-v1/<plan-id>/<generation>/. It adds generation-addressed nodes and backlinks to the live schema_version 1.1 traceability graph, then atomically publishes .ai/workflows/northstar-readiness-v1.json last. This separate registry's plans[] is the sole completion pointer.

The bundle pins the spec digest and exact repository and goal identities; planning completion is not implementation readiness: report preparation, implementation, and merge gaps separately, without claiming execution authority. Idempotent recovery preserves evidence; legacy migration and matched dependencies are mandatory. See modules/handoff.md.

Command surface

northstar registers as a first-class command under .ai/commands/omx/ and .ai/commands/omc/ using one shared schema. See modules/command-surface.md.

Safety rules

  • Planning only: never implement product changes or execute a sliced goal.
  • Planning completion is not implementation readiness; report preparation, implementation and merge gaps separately.
  • Publish completion through the separate v1 registry, never legacy discovery; missing or mismatched peers block publication.
  • Hard stop: do not invoke autobahn, ultragoal, team, ralph, ultrawork, or another implementation engine; stop after the verified A→B handoff — never implement in the same run.
  • Fail closed: a missing prereq, missing authorization, or partial handoff write stops with guidance — never silently proceed or mutate.
  • Compose, never reimplement: delegate every loop to its owning skill.
  • Adversarial pass is grill-with-docs then grill-me, taken or declined as one unit; the loop is not done on the deep-interview gate alone unless the user explicitly skipped that pass.
  • Local-first: never create a hosted issue unless the tracker is configured and authorized.

References

End by declaring knowledge kind plan, then reading ai-catapult-init/modules/knowledge-publish.md and running its host-neutral producer step — publish when a registry exists, otherwise record unpublished: <reason>; never fail the skill because of the registry.

04 同仓库更多 Skills