Skip to main content

deep-planning-codex

Deprecated for GPT-5.6 Sol. Use native Codex Plan Mode, with focused verification or adversarial review skills when needed.

ソース情報

リポジトリ
dachent/skills
ソースの最終更新活動
2026年7月20日 12:04
検出された SKILL.md の言語
英語
スター
35
フォーク
2

インストール方法

デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。

ソースファイルを確認

インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。

ファイルエクスプローラー
9 ファイル

SKILL.md を表示中

SKILL.md
ソースの指示 · 読み取り専用プレビュー
name
deep-planning-codex
description
Deprecated for GPT-5.6 Sol. Use native Codex Plan Mode, with focused verification or adversarial review skills when needed.
# Codex Deep Planning Orchestrator > **Deprecated for GPT-5.6 Sol.** Use native Codex Plan Mode, with focused verification or adversarial review skills when needed. Do not use this package for new work. ## Overview Run a Codex-native deep planning protocol that adapts to software with Git, software without Git, business artifact work, and mixed business-coding projects. Produce durable planning artifacts, proof criteria, and a reviewed execution strategy before implementation. ## Core Contract - Treat the user's invocation text as the planning objective. - Create or reuse the target project's `.deep-planning/` workfolder. - During planning, do not edit implementation files. Allowed planning writes are `.deep-planning/`, `.ultraplan/plan.md`, approved design/plan docs, `CONTEXT.md`, and accepted ADRs. - Every phase updates `.deep-planning/state.md`. - Every phase ends with exactly one status: `READY_FOR_PROCEED`, `BLOCKED_NEEDS_USER_DECISION`, `BLOCKED_BY_MISSING_EVIDENCE`, or `FAILED_VALIDATION`. - Stop at explicit proceed gates. Continue only after the user says `PROCEED` or gives equivalent explicit approval. - Use companion skills when available: `$repo-map-codex`, `$verification-plan-codex`, `$adversarial-plan-review-codex`, `$grill-with-docs-codex`, `$grill-me-codex`, `$handoff-codex`, `$ultraplan-codex`, `superpowers:brainstorming`, `superpowers:systematic-debugging`, `superpowers:writing-plans`, and `superpowers:verification-before-completion`. If unavailable, perform the same behavior locally. - Use Codex subagents only when explicitly allowed by the user or current session instructions. Otherwise perform bounded local exploration and review. ## Companion Skill Contract When a companion skill is invoked inside this workflow, require a Deep Planning Delta and merge it into `.deep-planning/state.md` before any proceed gate: ```markdown ## Deep Planning Delta Phase: ... Status: READY_FOR_PROCEED | BLOCKED_NEEDS_USER_DECISION | BLOCKED_BY_MISSING_EVIDENCE | FAILED_VALIDATION Updated artifacts: - ... Blockers: - ... Open assumptions: - ... Next action: ... ``` If the companion skill writes a native artifact outside `.deep-planning/`, such as `.ultraplan/plan.md`, keep that native artifact and also update or link the corresponding `.deep-planning/` artifact so the validator can reason over the full planning packet. ## References Read references only when needed: - `references/project-modes.md` when selecting or revising project mode. - `references/artifact-contracts.md` before writing or validating `.deep-planning/` artifacts. - `references/phase-contracts.md` before running the full workflow or resuming a workflow. - `references/proceed-gates.md` before stopping for approval. - `references/subagent-strategy.md` before delegating planning or execution work. ## Phase Flow 1. **Harness preflight**: record workspace, writable roots, Git state if present, available skills, subagent authorization, allowed artifact paths, and selected project mode. 2. **Intake and criteria**: sharpen success criteria, failure criteria, out-of-scope boundaries, owner decisions, risk tolerance, and documentation permissions. 3. **Project map and evidence catalog**: invoke `$repo-map-codex` if available or produce equivalent `repo-map.md` and `evidence-catalog.md`. 4. **Failure autopsy**: if prior failures exist, use systematic debugging or diagnosis to build `failure-autopsy.md` and `dead-ends-registry.md`. 5. **Assumption attack**: challenge criteria, evidence, dead ends, and unresolved claims; update `assumption-ledger.md`. 6. **Targeted probes**: run low-cost probes mapped to assumptions. Do not implement features. 7. **Design synthesis**: use brainstorming only if real design ambiguity remains; otherwise write the selected design directly. 8. **Grounded implementation plan**: use `$ultraplan-codex` for software-heavy work; for artifact-heavy work write `.deep-planning/implementation-plan.md`. When `$ultraplan-codex` writes `.ultraplan/plan.md`, also write or update `.deep-planning/implementation-plan.md` with the execution sequence, verification command, and a pointer to the native plan. 9. **Verification plan**: invoke `$verification-plan-codex` if available or produce equivalent proof criteria. 10. **Adversarial review**: invoke `$adversarial-plan-review-codex` if available or locally red-team the plan and fix blockers. 11. **Final execution plan**: use `superpowers:writing-plans` for code-heavy work; otherwise write a decision-complete artifact execution plan. 12. **Execution strategy**: choose subagent-driven, parallel, or inline execution only after explicit approval. 13. **Final verification and archive**: use verification-before-completion behavior, then write final validation and handoff records. ## Project Modes Classify the project in preflight: | Mode | Use when | Delivery primitive | | --- | --- | --- | | `software-git` | Git repository and commits/PRs are allowed | branch, commit, PR, CI | | `software-no-git` | Code exists but no commit workflow is allowed | file snapshots, milestone state, validation log | | `business-artifact` | Work is mostly docs, reports, spreadsheets, decisions, or analysis | approval packet, evidence log, deliverable folder | | `mixed-business-coding` | Business deliverables include code, scripts, data, or automation | artifact snapshots plus code and business validation | Default to `mixed-business-coding` when uncertain. ## Required Artifacts Maintain these in the target project's `.deep-planning/` folder as applicable: - `state.md` - `harness-preflight.md` - `project-mode.md` - `success-criteria.md` - `failure-criteria.md` - `out-of-scope.md` - `stakeholder-map.md` - `repo-map.md` - `evidence-catalog.md` - `assumption-ledger.md` - `decision-log.md` - `failure-autopsy.md` - `dead-ends-registry.md` - `probe-results.md` - `design.md` - `implementation-plan.md` - `verification-plan.md` - `adversarial-review.md` - `execution-strategy.md` - `final-validation.md` - `handoffs/` Use `.ultraplan/plan.md` when `$ultraplan-codex` is part of the selected software-heavy workflow, but do not treat it as a replacement for `.deep-planning/implementation-plan.md`. ## Handoff Policy Do not run full handoff after every small phase by default. Always update `.deep-planning/state.md`. Run `$handoff-codex` at major context gates: after Phase 0, after failure autopsy, after final plan approval, before execution handoff, and at final archive. In this workflow, store handoffs under `.deep-planning/handoffs/<gate>.md` or record a state pointer to the handoff path. ## Validation Before presenting a final execution plan, run `scripts/validate_deep_planning_artifacts.py` against the target project when possible. Treat missing required artifacts, incomplete state fields, unresolved proceed gates, failed adversarial verdicts, blocking review findings, or unreviewed verification criteria as blockers.
GitHubで見る