القائمة
Adversarial plan-critique-gate loop using build and deep subagents. Use when the user asks for megaplan, or a task is high-risk, multi-system, or ambiguous.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Build a fast, actionable branch/PR/commit context pack with scope, intent, changed files, hotspots, risks, and next actions. Use before handing work to another agent, reviewing a branch, resuming unfamiliar work, or preparing a PR/code-review prompt.
Create reviewed Codex goal setup packages for long-running /goal work. Use when the user wants to turn an idea, backlog, project mission, or vague objective into durable goal files under a project goals slug folder, with Plannotator review gates for brief, narrative plan with acceptance criteria, verification, blockers, and the final /goal prompt.
User guide for the local squash-safe `stack` CLI for stacked PR repair. Use when someone asks how to inspect, track, sync, merge, repair, document, or undo stacked pull requests in squash-merge repositories. Prefer this tool over GitHub's `gh stack` command for this workflow.
Manage long-running terminal jobs in a dedicated tmux-opencode session using wrapper scripts for run/wait, health checks, and crash-recovery cleanup.
Video understanding and transcription with intelligent multi-provider fallback. Use when: (1) Transcribing video or audio content, (2) Understanding video content including visual elements and scenes, (3) Analyzing YouTube videos by URL, (4) Extracting information from local video files, (5) Getting timestamps, summaries, or answering questions about video content. Automatically selects the best available provider based on configured API keys - prefers full video understanding (Gemini/OpenRouter) over ASR-only providers. Supports model selection per provider.
Designs and refactors React component APIs toward compound composition: fewer boolean mode props, clearer seams, context-backed internals, and JSX-first variant assembly. Use when building or reviewing React components with many props, conditional render branches, render props/slots, shared form/composer/card/table internals, or when the user mentions composition, compound components, Radix-style APIs, prop drilling, boolean props, or "composition is all you need".
| name | megaplan |
| description | Adversarial plan-critique-gate loop using build and deep subagents. Use when the user asks for megaplan, or a task is high-risk, multi-system, or ambiguous. |
High-rigor planning with adversarial critique, explicit gate decisions, and resumable execution.
Skip it for obvious, low-risk edits.
| Role | Subagent | Purpose |
|---|---|---|
| Planner | build | plan, revise, finalize |
| Critics | 3x deep | feasibility, architecture, risk |
| Reviewer | deep | final independent review |
| Orchestrator | you | state, gate, execution loop |
init -> plan -> critique -> gate -> revise/finalize -> execute -> review
plan, revise, finalize delegate to buildcritique runs 3 parallel deep critics and merges flagsgate is computed by you, then decision delegated to buildexecute is owned by the main threadreview delegates to deepUse .megaplan/plans/<name>/runbook.md as the continuity source of truth.
Trust runbook.md over conversational memory, especially after compaction or resume.
During execute, the main thread owns:
Delegate only fresh-eyes work:
simplifyreviewThe main thread must not approve its own diff unaided.
Before each commit:
simplifyreviewDuring execution, set /never-stop using the exact ## Continue section from runbook.md.
Refresh it after meaningful steps. Clear it on complete or blocked.
Keep runbook.md minimal. It should contain only:
StatusStepNextBatchLast commitWatchBlocker## ContinueDo not turn the runbook into a diary. Keep history and evidence in other artifacts.
Use .megaplan/plans/<name>/.
Essential artifacts:
runbook.md — continuity control planeplan_v<N>.md and plan_v<N>.meta.jsonfaults.jsongate.jsonfinalize.jsonreview.jsonAdditional artifacts may be written as needed for critiques, execution evidence, and summaries.
When Step is execute, use this loop:
simplifyreviewrunbook.mdStop only on:
complete — all planned work finished, reviewed, committedblocked — missing requirements, auth, permissions, secrets, or irreconcilable contradiction