| name | agent-ui-feature-planning |
| description | Plan Agent UI repository features before implementation. Use when the user asks to make a feature plan, plan before coding, research/plan/todo/goal prompt, solve a problem group, architecture plan, 実装前に計画, feature planning, PLAN/TODO作成, or goal-prompt generation for this repo. Do not use for small direct edits, review-only requests, npm release operations, upstream sync, browser QA, or external host-app integration unless the user explicitly asks for a planning package. |
Agent UI Feature Planning
Use this skill to create a repo-specific planning package before implementation.
Do not implement the feature. Produce artifacts only under:
.agent-work/features/<date>-<slug>/
research.md
plan.md
todo.md
goal-prompt.md
Do not create root-level aliases unless the user explicitly asks.
First Pass
- Read freshness policy, then read
references/freshness-manifest.json.
- Run
node .agents/skills/agent-ui-feature-planning/scripts/check-freshness.mjs
from the repo root when script execution is available. If watched inputs
changed, do targeted refresh before planning; do full refresh only for
structural repo, CI, build, or agent-guidance changes, or when the user asks.
Record the result in research.md.
- Read artifact contract and
review rubric.
- Read repo research summary, then
inspect the current repo files it references before finalizing the plan.
- Choose a concise slug from the user's requested feature or problem group.
- Create or switch to the single branch that will carry both planning and
implementation before writing artifacts.
- Create the canonical artifact directory with today's date:
.agent-work/features/<YYYY-MM-DD>-<slug>/.
- Copy or adapt templates from
assets/ into that directory.
Workflow
- Freshness Check: compare the current repo state against
references/freshness-manifest.json. Reuse current repo guidance when
unchanged; targeted-refresh changed watched inputs; full-refresh only for
structural drift. Record freshness evidence in research.md.
- Branch Setup: use one branch for both planning and implementation.
Do not create a separate planning-only branch. If the user provides a
branch name, use it exactly. Otherwise derive a concise branch from the
feature slug, following repo branch conventions discovered in guidance.
Create or switch to the branch before writing planning artifacts. Verify the
worktree does not contain unrelated changes that would make safe planning
commits impossible. Record the branch decision and blockers in working notes,
then write them into
research.md and todo.md.
- Specify: extract what the user wants, why, affected users/systems,
success criteria, non-goals, completion criteria, and expected evidence.
- Clarify: ask only questions that materially affect planning. If the user
asks for autonomous planning, write conservative assumptions in
plan.md.
- Research: inspect relevant implementation, tests, examples, docs,
workflows, package exports, protected surfaces, and skills. Use subagents
when available for independent lanes; otherwise run the lanes sequentially
and record that subagents were unavailable. Use web/current research when
the plan depends on external or time-sensitive facts such as current OpenAI
Codex behavior, package registry state, GitHub Actions state, dependency
versions, browser/tooling behavior, or external API/spec changes.
- Plan: write
plan.md with Agent UI ownership, design decisions,
impacted areas, validation, commit/PR/CI, risks, and open questions.
- Tasks: write
todo.md phase-first. Phases are the default unit for
implementation, validation, review, commit, push, PR, and CI follow-through;
tasks are checklists inside phases. Use task-level execution only as a
documented fallback when a phase is too large or unsafe to review/commit.
Include the Branch And Planning Commit section with branch, planning
commit, remote, push result, and blockers fields.
- Analyze: review
research.md, plan.md, and todo.md for
contradictions, unsupported assumptions, missing validation, repo rule
violations, oversized phases, hidden follow-ups, and vague completion.
- Goal Prompt: write
goal-prompt.md as a copy-paste-ready Codex /goal
prompt with absolute artifact paths, freshness result, and phase-first
execution protocol. Keep it at or below 4000 characters. It must instruct
implementation to continue on the same branch used for planning.
- Validate Artifacts: run
node .agents/skills/agent-ui-feature-planning/scripts/validate-artifacts.mjs <artifact-dir>
only after goal-prompt.md exists. This is a deterministic shape checker,
not a prose-quality review. After validation passes, commit the planning
package with one concise planning commit such as Plan <feature slug>.
If a remote exists and push is possible, push the branch. Record branch
name, planning commit hash, remote, push result, and blockers in todo.md.
If branch creation, commit, or push is blocked, record the exact blocker in
research.md or todo.md and report it.
Research Lanes
Use the prompts in agents/ when subagents are available:
agents/repo-guidance-researcher.md
agents/implementation-surface-researcher.md
agents/validation-researcher.md
agents/skill-design-reviewer.md
Default to four parallel lanes and up to three rounds. Stop after one round
when findings are sufficient; run additional rounds only to close concrete
unknowns or review a risky planning contract.
Boundaries
- Agent UI is a reusable Codex App Server UI component library, not a hosted
runtime. Keep host auth, persistence, routing, process lifecycle, deployment,
billing, workspace isolation, upload storage, and workflow state outside core.
- Do not plan direct edits inside the vendored Codex submodule except through
upstream sync.
- Do not plan hand edits to auto-created schema files or compiled artifacts.
- Public API plans must include package exports, docs, examples, tests, API
snapshots or package-resolution evidence, and release impact.
- Browser-visible plans must include Playwright and/or
agent-browser
evidence.
Completion
Before reporting ready, confirm the four artifact files exist, the freshness
check is recorded, the analysis pass is reflected in the artifacts, the
validator passed, todo.md is phase-first, and the generated goal-prompt.md
names exact absolute paths, the planning branch, the planning commit, and
repo-specific forbidden edits/checks. Report the artifact directory, selected
branch, planning commit, remote/push result, questions asked or assumptions
made, freshness result, research lanes used, whether web/current research was
used or intentionally skipped, validation selected, artifact validation result,
and remaining risks.