The Context Flood
Exploring a big codebase pours everything into one thread, until the thread loses track of the plan.
No. 0.9.7Oct 2026Free · MIT
Give Claude Code a whole league of superpowers.
SuperDuperPowers is a Claude Code plugin that turns a rough request into a reviewed spec, a bounded plan, cross-checked code and a verified branch, with every model it uses on the record.
›Using SuperDuperPowers, let's add rate limiting to the API
◆brainstorming: spec drafted
◆review panel (Claude + Codex): 3 findings, all fixed
✓you approved the spec
◆plan: 3 slices, 9 tasks, contracts first
⇉lanes 1–3 building in parallel
✗Codex review, task 4: the retry window never resets
◆fixed on the same thread, re-verified
✓42 tests pass (fresh run)
✓ledger: 27 seats recorded, 0 off route
?merge · open a PR · keep · discard
Example session
Our story begins with a big coding job…
Every hero has a weakness
Hand an AI coder a large task and the same three villains show up.
The Faceless Model
Something wrote this code, but nobody can say which model it was or what effort setting it ran at.
The Echo
A reviewer from the same model family shares the author's blind spots and waves the bug through.
Luckily, help is on the way…
Eight powers, one plugin
Every power is a real part of the plugin. Its secret identity is printed on the card, and the file it lives in is at the bottom.
Precognition
Secret identity: spec review panel
Spots the problems before there's any code. Claude and Codex reviewers pick the spec apart, and nothing gets built until you approve it.
skills/brainstorming
Swarm Recon
Secret identity: scouting pyramid
A swarm of small scouts reads your codebase in parallel. Your main thread gets back a short answer with citations and keeps its context clear.
agents/code-explore.md
Multiplicity
Secret identity: parallel lanes
Splits the plan into slices and builds them side by side as headless Claude Code lanes, one wave at a time.
scripts/lane-run.py
Second Sight
Secret identity: cross-family review
Risky work gets reviewed by the other model family. Claude's code goes to Codex and Codex's goes to Claude, so the reviewer doesn't share the author's blind spots.
skills/code-review
Total Recall
Secret identity: routing ledger
Remembers which model ran every step and at what effort. Any work that went off route shows up in the audit.
scripts/routing-ledger.py
Perfect Casting
Secret identity: model catalog
Every role, from builder to judge, is pinned to a named model and effort in one catalog, so nothing runs on a silent default.
scripts/model-catalog.json
Force Field
Secret identity: enforcement hooks
Eleven hooks watch every dispatch, block the ones that break the routing rules, and tell you plainly what they can't enforce.
hooks/hooks.json
Truth Sense
Secret identity: verification before completion
It won't call anything done without proof. Tests run fresh before any claim that they pass, and again before you merge.
skills/verification-before-completion
How it works
From rough idea to verified branch, in four panels
-
1 · Spec
YouLet's add rate limiting to the API.
Brainstorming turns your request into a design. Claude and Codex reviewers pick it apart, and nothing gets built until you approve it.
-
2 · Plan
ClaudeThree slices, nine tasks. Contracts go first.
The work is cut into slices, each with its own files, its own checks and clear contracts with the slices around it.
-
3 · Build
CodexTask 4 never resets its retry window.
Lanes build in parallel. Every task checks its own work, and risky tasks get a review from the other model family.
-
4 · Finish
ClaudeAll 42 tests pass. Merge, open a PR, keep or discard?
A fresh test run and a ledger audit come first. Then you make the call.
No hero works alone
Assemble the league
SuperDuperPowers gives each model the job it does best and has them check each other's work.
Claude
The lead
Runs the show from your main thread. It plans the work, makes the calls and keeps you posted.
Codex
The partner
Builds or reviews, whichever Claude isn't doing. It's the other half of every cross-family check.
Grok & Muse
The sidekicks
Add a second opinion to every review. They can flag anything, but they never block a merge.
The Judge
The tiebreaker
Steps in only when reviewers disagree, and lays out its reasoning so Claude can rule.
You
The hero
You approve the spec, choose who builds and decide whether to merge. Nothing you can't undo happens without your say.
Later, in the changelog…
Powers you can measure
The plugin measures its own routing choices and writes the results up in its changelog. A few of them:
11.91.9min
Median time per judgment scout on Sonnet 5.5 at high effort, against Haiku 5.5 at max, across 129 measured seats.
88vs81
Items found on the same brief by the faster seat, in 7.8 minutes instead of 20.1. It cost $2.17 against $1.19, and the ledger shows both.
8.3min
A scripted two-file Codex build, end to end. Its first self-review caught a real shell-ordering bug.
23releases
Shipped so far, from 0.1.0 to 0.9.7, each one written up in the changelog.
Figures come from the project's changelog and method notes, measured by its author.
To be continued…
Join the league early
SuperDuperPowers is MIT licensed and opens to the public soon. Leave your email and we'll tell you the day it does.