Skip to content
Back to search

insane-review

SkillWorks with Claude CodeGitHub

GPT Pro(웹 전용·API 없음 — 최신 플래그십 모델의 Pro 추론)를 Claude Code 안에서 활용 — repomix로 관련 코드만 정밀 패킹 → 구독 ChatGPT Pro에 투입 → 분석/의견 회수. agent-council 웹 전용 멤버로도 통합.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/fivetaku--insane-review/badge.svg)](/listings/fivetaku--insane-review)

HTML

<a href="/listings/fivetaku--insane-review"><img src="/api/listings/fivetaku--insane-review/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

1. Add the marketplace (once)

/plugin marketplace add https://github.com/fivetaku/gptaku_plugins.git

2. Install

/plugin install insane-review

3. Restart Claude Code

Required for the plugin to load.

4. Prepare the browser bridge (once per machine)

Pro is web-only, so insane-review needs a real, logged-in browser on a debug port:

# launch Comet (or Chrome) with the CDP port, then log into chatgpt.com and pick the Pro reasoning tier
open -a Comet --args --remote-debugging-port=9222

# verify everything is wired up (node/repomix, playwright, pyperclip, CDP browser)
python3 bin/pack_and_ask.py --check-env

5. Run it

/insane-review review the auth flow in src/auth

Or just say "have Pro review this" / "ask GPT Pro about this design" — Claude figures out the target and packs it.

---

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

.claude-plugin/plugin.json · Manifest · 21 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. {
  2. "name": "insane-review",
  3. "version": "0.6.15",
  4. "description": "GPT Pro(웹 전용·API 없음 — 최신 플래그십 모델의 Pro 추론)를 Claude Code 안에서 활용 — repomix로 관련 코드만 정밀 패킹 → 구독 ChatGPT Pro에 투입 → 분석/의견 회수. agent-council 웹 전용 멤버로도 통합.",
  5. "author": {
  6. "name": "fivetaku",
  7. "url": "https://github.com/fivetaku"
  8. },
  9. "keywords": [
  10. "chatgpt",
  11. "gpt-pro",
  12. "repomix",
  13. "browser-automation",
  14. "agent-council",
  15. "subscription"
  16. ],
  17. "license": "MIT",
  18. "dependencies": [
  19. "gptaku-core"
  20. ]
  21. }
README.md · README · 200 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. English | [한국어](README.ko.md) | [中文](README.zh.md) | [日本語](README.ja.md) | [Español](README.es.md)
  2. # insane-review
  3. <div align="center">
  4. <img src="assets/hero.png" width="860" alt="insane-review cinematic hero">
  5. </div>
  6. > **GPT Pro (the Pro reasoning tier of the current flagship — currently GPT-5.6 Sol) has no API. This plugin uses it from inside Claude Code anyway.**
  7. GPT Pro lives only in the ChatGPT web app (subscription) — there is no official API, so you can't reach it from Codex CLI, `omc ask`, or an agent-council API provider. insane-review drives your **logged-in ChatGPT web session over CDP**: it packs the relevant code with repomix, sends it to Pro, and harvests the review. **Zero API cost** — it runs on your existing ChatGPT plan.
  8. [Quick Start](#quick-start) • [Why insane-review?](#why-insane-review) • [How it works](#how-it-works) • [Features](#features) • [Tuning & timeouts](#tuning--timeouts) • [Requirements](#requirements)
  9. ---
  10. ## Quick Start
  11. ### 1. Add the marketplace (once)
  12. ```
  13. /plugin marketplace add https://github.com/fivetaku/gptaku_plugins.git
  14. ```
  15. ### 2. Install
  16. ```
  17. /plugin install insane-review
  18. ```
  19. ### 3. Restart Claude Code
  20. Required for the plugin to load.
  21. ### 4. Prepare the browser bridge (once per machine)
  22. Pro is web-only, so insane-review needs a real, logged-in browser on a debug port:
  23. ```bash
  24. # launch Comet (or Chrome) with the CDP port, then log into chatgpt.com and pick the Pro reasoning tier
  25. open -a Comet --args --remote-debugging-port=9222
  26. # verify everything is wired up (node/repomix, playwright, pyperclip, CDP browser)
  27. python3 bin/pack_and_ask.py --check-env
  28. ```
  29. ### 5. Run it
  30. ```
  31. /insane-review review the auth flow in src/auth
  32. ```
  33. Or just say "have Pro review this" / "ask GPT Pro about this design" — Claude figures out the target and packs it.
  34. ---
  35. ## Why insane-review?
  36. - **The only way to reach Pro programmatically** — no API exists. A logged-in web session driven over CDP is the bridge, and it costs nothing beyond your subscription.
  37. - **Claude picks the complete relevant set** — you don't have to hand-list files. For reviews it sends **full code** (no `--compress`, which strips function bodies and produces false "looks fine" verdicts) and audits the packed file list so nothing is silently missing.
  38. - **fail-closed by design** — wrong model, unverified login, truncated prompt, an empty pack, or a previous turn's answer are all refused rather than silently sent or saved. Hardened across four rounds of Pro's own self-review (6 → 0 P0s).
  39. - **Two roles, one engine** — a standalone reviewer when you ask for a fix/review, or a web-only member of [agent-council](references/council-setup.md) so Pro can debate alongside Codex/Gemini/others.
  40. - **Citations you can follow** — line numbers are packed in, so Pro's findings come back as `file:line` you can jump to.
  41. ---
  42. ## How it works
  43. ```
  44. "have Pro review this" / council member call
  45. ↓
  46. Claude selects the COMPLETE relevant file set (full code — no --compress for reviews)
  47. ↓
  48. repomix pack (line numbers · secretlint · packed-file-list audit · token count)
  49. ↓
  50. CDP-attach the logged-in ChatGPT session
  51. Select Pro effort (flagship auto-follows; currently GPT-5.6 Sol) → re-open menu and VERIFY (mismatch = abort, fail-closed)
  52. ↓
  53. Attach pack + prompt → confirm the prompt actually landed in the composer → send
  54. ↓
  55. Wait for THIS turn to complete (turn-scoped: new assistant node + new copy button)
  56. Optionally cut long reasoning early with --force-answer-after
  57. ↓
  58. Harvest the answer → save to ./.insane-review/response_*.md (atomic write)
  59. ```
  60. Output lands in the **current project's** `.insane-review/` folder (like kkirikkiri's `.kkirikkiri/`), never inside the plugin:
  61. ```
  62. .insane-review/
  63. ├── pack_<target>_<ts>.md # what was sent (chmod 600)
  64. └── response_<target>_<ts>.md # Pro's answer + verified model header
  65. ```
  66. ---
  67. ## Features
  68. ### Commands
  69. | Command | What it does |
  70. |---------|-------------|
  71. | `/insane-review [target/question]` | Pack the relevant code and send it to GPT Pro for review |
  72. | natural language | "have Pro review this", "ask GPT Pro about X" — same flow |
  73. ### Two modes
  74. 1. **Standalone reviewer** — you ask for a fix/review → Claude scopes the target → repomix pack → Pro analysis → applied back.
  75. 2. **agent-council member** — register Pro as a web-only council member so it debates with other models. See [`references/council-setup.md`](references/council-setup.md).
  76. ### Key flags
  77. | Flag | Purpose |
  78. |------|---------|
  79. | `--target <dir>` | Folder to pack (omit for a prompt-only opinion) |
  80. | `--include <glob>` / `--ignore <glob>` | Narrow the packed set |
  81. | `--effort <step>` (alias `--model`) | Reasoning effort — default `pro`; `instant`/`medium`/`high`/`xhigh` are applied exactly, mismatch aborts |
  82. | `--require-model "<name>"` | Pin the model name — abort on mismatch. **The current UI shows the model name only at the Pro step (e.g. `6`) and it is unreadable at other steps, so pinning there always aborts.** Omit it to follow the flagship |
  83. | `--prompt "..."` / `--prompt-file` | The question |
  84. | `--pack-only` | Just pack (inspect token count), don't send |
  85. | `--council` | Council mode — response on stdout, logs on stderr |
  86. | `--compress` | tree-sitter skeleton only — **don't use for reviews** (drops function bodies) |
  87. | `--check-env` / `--install` | Diagnose / install the local toolchain |
  88. ---
  89. ## Tuning & timeouts
  90. Response waiting and pack timeouts are adjustable from both the CLI and the environment — useful because Pro's full reasoning can take 10–15 minutes.
  91. | Control | Default | What it does |
  92. |---------|---------|-------------|
  93. | `--max-wait <sec>` | `1200` (20 min) | Max time to wait for Pro's response before giving up (fail-closed, no partial save) |
  94. | `INSANE_REVIEW_MAX_WAIT` | `1200` | Same as `--max-wait`, via environment |
  95. | `--force-answer-after <sec>` | off | Soft cut: if Pro is still reasoning after N sec, click **"Get answer now"** so it answers **from the reasoning it has done so far** — a complete, saved answer (see below) |
  96. | `INSANE_REVIEW_REPOMIX_TIMEOUT` | `300` | Max seconds for the repomix packing step |
  97. | `--retries <n>` | `1` | Re-attempts if a send/harvest fails |
  98. **Two different "timeouts" — don't confuse them:**
  99. - **`--force-answer-after N` (soft cut, recommended for bounding cost).** Pro reasons for a long time; this clicks ChatGPT's *"Get answer now"* at N seconds, so Pro stops reasoning and replies based on **what it has reasoned up to that point**. That reply is a normal, complete turn — it's harvested and saved like any other. Use it to cap a council member at, say, 120s instead of waiting 10+ minutes.
  100. - **`--max-wait N` (hard ceiling, fail-closed).** If the turn never completes within N seconds *and* no answer was forced, insane-review gives up **without saving** the half-streamed text — an incomplete answer is treated as a failure, not a result. This is intentional: it never hands you a truncated review pretending to be done.
  101. Other environment overrides:
  102. | Variable | Default | What it does |
  103. |----------|---------|-------------|
  104. | `INSANE_REVIEW_CDP_PORT` | `9222` | Browser remote-debugging port |
  105. | `INSANE_REVIEW_COMET` / `INSANE_REVIEW_CHROME` | app default path | Browser executable path |
  106. | `INSANE_REVIEW_REPOMIX_VERSION` | `1.15.0` | Pinned repomix version (reproducibility) |
  107. | `INSANE_REVIEW_OUT` | `./.insane-review` | Output directory (also `--out-dir`) |
  108. ```bash
  109. # example: give Pro up to 25 minutes, but cut reasoning at 5 minutes if it's still thinking
  110. INSANE_REVIEW_MAX_WAIT=1500 python3 bin/pack_and_ask.py \
  111. --target . --include "src/**" --model pro \
  112. --force-answer-after 300 --prompt "Where are the concurrency bugs?"
  113. ```
  114. ---
  115. ## Requirements
  116. ### Required
  117. - [Claude Code](https://docs.anthropic.com/claude-code)
  118. - Python 3.11+ with `playwright` and `pyperclip`
  119. - Node.js / `npx`
  120. - **A subscription ChatGPT account with GPT Pro**, logged in inside Comet/Chrome launched on the debug port (`--remote-debugging-port=9222`)
  121. ### What's auto-handled vs. what you do
  122. | Dependency | First-run behavior |
  123. |------------|-------------------|
  124. | **repomix** | **Fully automatic** — pulled on demand via `npx -y repomix@<pinned>`, never needs manual install |
  125. | **playwright / pyperclip** | Checked on first use by `--check-env`; install them with `--install` (runs `pip install`). A normal run without them stops with a clear instruction (fail-closed) rather than failing mid-way |
  126. | **Browser login + GPT Pro** | **Manual** — can't be automated; you log into `chatgpt.com` and select Pro once |
  127. ```bash
  128. # one shot: checks node/repomix, playwright, pyperclip, CDP browser — and installs the pip deps if missing
  129. python3 bin/pack_and_ask.py --check-env --install
  130. ```
  131. ### Note
  132. Web-UI automation is not endorsed by OpenAI's ToS and selectors may need maintenance when the ChatGPT DOM changes. Intended for personal subscription use only.
  133. ---
  134. ## License
  135. MIT
  136. ---
  137. <div align="center">
  138. **No API. Still Pro.**
  139. </div>

Version history

  • v0.6.1511/5/2026

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.