agentsclimarketplace

Ultra planning

Skill keithhegit/ultra-orchestration/skills-vnext/ultra-planning

Ultra Orchestration 是一套开源的 AI 工程编排技能体系。它将 OpenSpec 规格主线(Program → Milestone → Change → Slice)与 Ultra-Orchestrator 执行阶段(Intake → Plan → Dispatch → Execute → Review → QA → Deliver → Retro)结合,提升交付稳定性与可审计性。

Install
npx -y skills add keithhegit/ultra-orchestration --skill ultra-planning

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 6 stars6 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Rigid decision-complete planning for Codex Ultra vNext. Use after intake, brainstorming, or OpenSpec bridge whenever work spans multiple files, roles, phases, or risk boundaries and must be turned into review-ready TaskManifest and WorkPackage artifacts with explicit dependencies, owned paths, acceptance checks, retry assumptions, and serial versus parallel execution rules.

SKILL.md

4.8 KB, as published. Nobody here has run it

Ultra Planning

Use this skill as a rigid workflow gate before controlled dispatch. This skill must remove planning ambiguity rather than summarize it.

Hard Gate

Do not allow controlled dispatch until the plan includes:

  • success criteria
  • constraints
  • non-goals
  • owned paths
  • acceptance checks
  • dependency order
  • risk level
  • retry assumptions

If any of those are missing or still depend on chat interpretation, the plan is not complete.

When To Use

Planning is required when:

  • the work spans multiple files or directories
  • multiple phases or roles are involved
  • OpenSpec change assets are being turned into execution packages
  • review, QA, or risk gates need explicit targets
  • parallel execution is being considered
  • the implementer would otherwise need to make design or boundary decisions

Planning may be compressed only when:

  • the task is truly tiny
  • the write scope is narrow and obvious
  • no risky command or external mutation is involved
  • review and QA targets remain explicit

If you compress planning, record the reason in the orchestration log.

Planning Workflow

  1. Read the intake, brainstormed design, or OpenSpec bridge artifacts.
  2. Check whether success criteria, constraints, and scope are already stable.
  3. If not stable, route back to ultra-brainstorming or the intake step.
  4. Define the top-level TaskManifest.
  5. Split the work into one or more WorkPackage items.
  6. Mark dependencies, owned paths, parallel groups, and serial boundaries.
  7. Attach acceptance checks, risk level, and retry assumptions.
  8. Record assumptions, blockers, and re-plan triggers.
  9. Run the planning self-review before declaring the plan complete.

Planning Standard

Produce one TaskManifest plus one or more WorkPackage items.

The plan is not complete until it defines:

  • goal
  • context
  • success criteria
  • constraints
  • non-goals
  • dependencies
  • owned_paths
  • expected outputs
  • risk level
  • acceptance checks
  • retry budget or retry assumptions

DAG And Lock Rules

For each work package, identify:

  • upstream dependencies
  • whether the task is parallel-safe
  • which paths are read-only context
  • which paths are write-owned

Only mark work as parallel-safe when both conditions hold:

  1. all upstream dependencies are satisfiable without new design work
  2. owned_paths do not intersect with another active write package

Completion Standard

Planning is complete only when:

  • a new implementer could start from the plan without inventing key decisions
  • review knows what to inspect
  • QA knows what behavior must be verified
  • write ownership is explicit
  • serial versus parallel execution is explicit
  • blockers and re-plan triggers are named

If the plan still requires the implementer to decide boundaries, acceptance, or parallelism, the plan is incomplete.

Planning Self-Review

Before returning the plan, check:

  • do any owned_paths overlap unexpectedly
  • are acceptance checks concrete and verifiable
  • do reviewer and QA receive explicit inspection targets
  • is any work package write scope too broad
  • do dependencies still rely on chat memory rather than artifacts
  • does any package mix planning, implementation, and review responsibility
  • would a failure in one package force a re-plan that is not yet documented

Failure Handling

  • If success criteria, scope, or constraints are still unstable, route back to ultra-brainstorming.
  • If OpenSpec assets are incomplete or contradictory, route back to openspec-ultra-bridge-v2 or spec preparation.
  • If planning reveals that one request actually contains several independent initiatives, decompose the work and plan only the first bounded unit.

Role Discipline

  • planner defines work, but does not silently become the implementer
  • reviewer tasks must not be mixed into implementation packages
  • integrator tasks should package accepted outputs, not redo worker code

Output Template

Return in this order:

  1. one TaskManifest
  2. ordered WorkPackage items
  3. parallel groups or serial sequence notes
  4. assumptions
  5. blockers
  6. re-plan triggers

Examples

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.