English | 한국어

GPT-5.5 Pro has no API. This plugin uses it from inside Claude Code anyway.

GPT-5.5 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.

Quick Start • Why insane-review? • How it works • Features • Tuning & timeouts • Requirements

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

/plugin install insane-review

Required for the plugin to load.

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 GPT-5.5 Pro
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

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

Or just say "have Pro review this" / "ask GPT-5.5 Pro about this design" — Claude figures out the target and packs it.

• 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.
• 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.
• 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).
• Two roles, one engine — a standalone reviewer when you ask for a fix/review, or a web-only member of agent-council so Pro can debate alongside Codex/Gemini/others.
• Citations you can follow — line numbers are packed in, so Pro's findings come back as file:line you can jump to.

"have Pro review this"  /  council member call
  ↓
Claude selects the COMPLETE relevant file set (full code — no --compress for reviews)
  ↓
repomix pack  (line numbers · secretlint · packed-file-list audit · token count)
  ↓
CDP-attach the logged-in ChatGPT session
Select GPT-5.5 + Pro effort  → re-open menu and VERIFY (mismatch = abort, fail-closed)
  ↓
Attach pack + prompt  → confirm the prompt actually landed in the composer  → send
  ↓
Wait for THIS turn to complete (turn-scoped: new assistant node + new copy button)
Optionally cut long reasoning early with --force-answer-after
  ↓
Harvest the answer → save to  ./.insane-review/response_*.md  (atomic write)

Output lands in the current project's .insane-review/ folder (like kkirikkiri's .kkirikkiri/), never inside the plugin:

.insane-review/
├── pack_<target>_<ts>.md        # what was sent (chmod 600)
└── response_<target>_<ts>.md    # Pro's answer + verified model header

1. Standalone reviewer — you ask for a fix/review → Claude scopes the target → repomix pack → Pro analysis → applied back.
2. agent-council member — register Pro as a web-only council member so it debates with other models. See references/council-setup.md.

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.

Two different "timeouts" — don't confuse them:

• --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.
• --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.

Other environment overrides:

# example: give Pro up to 25 minutes, but cut reasoning at 5 minutes if it's still thinking
INSANE_REVIEW_MAX_WAIT=1500 python3 bin/pack_and_ask.py \
  --target . --include "src/**" --model pro --require-model "GPT-5.5" \
  --force-answer-after 300 --prompt "Where are the concurrency bugs?"

• Claude Code
• Python 3.11+ with playwright and pyperclip
• Node.js / npx
• A subscription ChatGPT account with GPT-5.5 Pro, logged in inside Comet/Chrome launched on the debug port (--remote-debugging-port=9222)

# one shot: checks node/repomix, playwright, pyperclip, CDP browser — and installs the pip deps if missing
python3 bin/pack_and_ask.py --check-env --install

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.

MIT