ultraplan-codex
Deprecated for GPT-5.6 Sol. Use native Codex Plan Mode for grounded, decision-complete implementation plans.
dachent · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add dachent/skills --skill ultraplan-codex技能说明
name: ultraplan-codex description: "Deprecated for GPT-5.6 Sol. Use native Codex Plan Mode for grounded, decision-complete implementation plans."
Codex UltraPlan
Deprecated for GPT-5.6 Sol. Use native Codex Plan Mode for grounded, decision-complete implementation plans. Do not use this package for new work.
Overview
Run a read-only, multi-phase implementation planning session before coding. Produce a concise, file-path-grounded .ultraplan/plan.md that can be executed without guessing.
Activation
Treat the user's invocation text as the planning task.
Enter deep planning mode. This is not quick planning: investigate the codebase, interview the user only for non-discoverable decisions, synthesize one recommended approach, and validate the plan before presenting it.
Read-only rule: Do not create, modify, or delete files except .ultraplan/plan.md while planning. When invoked by deep-planning-codex, .deep-planning/implementation-plan.md and .deep-planning/state.md are also allowed planning artifacts. Do not install dependencies, run migrations, commit, push, format, or edit implementation files. Read commands, searches, tests, and build checks are allowed when they improve the plan and do not intentionally modify tracked source.
Quick Reference
| Need | Use |
|---|---|
| Find files | rg --files |
| Search code | rg |
| Read files | normal file reads or shell read commands |
| Parallel local reads | multi_tool_use.parallel when available |
| Ask user | direct concise questions; use structured question tooling only when available |
| Delegate exploration | Codex subagents only when the user explicitly asked for subagents or parallel agent work |
| Save state | .ultraplan/plan.md |
When Invoked By deep-planning-codex
- Keep
.ultraplan/plan.mdas UltraPlan's native planning state. - Also create or update
.deep-planning/implementation-plan.mdwith the selected execution sequence, key file references, verification command, and a pointer to.ultraplan/plan.md. - Treat
deep-planning-codexas the owner of proceed gates. Do not ask "Ready to execute this plan, or do you want changes?" Instead, end with "Plan ready for verification phase" when the plan is complete. - End with a Deep Planning Delta for the orchestrator to merge into
.deep-planning/state.md. UseREADY_FOR_PROCEEDonly after referenced paths, functions, and verification commands have been checked.
Setup
- Check whether
.ultraplan/plan.mdexists. - If it exists, read it and ask: "I found an existing plan at
.ultraplan/plan.md. Continue refining it, or start fresh?" - If starting fresh, create
.ultraplan/plan.mdwith skeleton headings and rough notes only. - Gather read-only git context:
git status --short
git log --oneline -5
git branch --show-current
If the current directory is not a Git repository, record Git context: not a git repository in the plan and continue. Do not treat missing git metadata as a blocker for planning small fixtures, archives, or copied code directories.
Write discoveries to .ultraplan/plan.md incrementally. Do not keep planning state only in conversation memory.
Planning Loop
Repeat until the plan is complete:
- Explore the repo with read-only searches and file reads. Prefer reuse of existing utilities, patterns, tests, and APIs before proposing new code.
- Update
.ultraplan/plan.mdimmediately with verified findings, paths, and open decisions. - Ask the user only for requirements, preferences, tradeoffs, or scope choices that cannot be derived from the code.
Phase 1: Requirements Interview
Goal: build a shared understanding of what needs to happen and why.
Start by scanning 3-5 likely key files. Then write the initial skeleton plan and ask the first round of questions if user intent remains ambiguous.
Rules:
- Never ask what can be found by reading the repo.
- Batch related questions.
- Focus on user-owned decisions: requirements, priorities, tradeoffs, edge cases, and scope.
- Scale depth to task complexity.
Phase 2: Deep Codebase Exploration
Goal: understand every file and pattern that affects the plan.
Use local exploration for all tasks. When the user explicitly asked for subagents or parallel agent work, use bounded Codex explorer subagents for independent read-heavy questions:
- Breadth-first discovery: data layer, business logic, presentation/API layer.
- Feature trace: UI to API to service to storage, plus related tests.
- Impact analysis: direct changes and indirect callers/importers.
For each exploration pass, capture:
- Existing functions or utilities to reuse, with
file:line. - Architecture patterns and conventions.
- Dependencies and coupling.
- Test infrastructure and relevant commands.
- Similar features to use as references.
Phase 3: Architecture Design
Goal: design one implementation approach grounded in verified repo facts.
For complex tasks with genuine architectural ambiguity, compare lenses internally: simplicity, performance, maintainability, risk, and incremental migration. Present only the recommended approach unless the user explicitly asks for alternatives.
If parallel design review is explicitly authorized, use Codex default subagents with specific design lenses and require each to return 3-5 critical implementation files. Otherwise, perform the same design comparison locally.
Phase 4: Plan Synthesis
Write the final .ultraplan/plan.md in this structure:
# Implementation Plan: [Title]
## Context
[One line: what is changing and why]
## Changes
### [Component or Module]
- **File**: `path/to/file.ext:line`
- **Change**: [specific change]
- **Reuses**: `existingFunction()` from `path/to/utility.ext:line`
## Implementation Sequence
1. [First dependency-safe step]
2. [Second step]
3. [Final integration step]
## Edge Cases & Risks
- [Risk]: [mitigation]
## Verification
[single command or short command sequence]
Rules:
- Kee