hyper
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/omarhafez1--hydra)HTML
<a href="/listings/omarhafez1--hydra"><img src="/api/listings/omarhafez1--hydra/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
Install it from this repository, then run setup once:
/plugin install hyper --marketplace OmarHafez1/Hyper /hyper:setup
From a shell, the same in two commands: `claude plugin marketplace add OmarHafez1/Hyper`, then `claude plugin install hyper@hyper`.
Setup checks your machine, asks which models you have, researches what each is good at, and writes your use-case file. To try Hyper without the research, copy an example roster and adjust it:
python3 <plugin folder>/scripts/hyper.py init --example claude-only
Install it for one project (`--scope project`) if you want Hyper only there.
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/marketplace.json · Manifest · 17 lines
- {
- "name": "hyper",
- "owner": {
- "name": "OmarHafez1"
- },
- "metadata": {
- "description": "Hyper: the right model and effort for every step, from an idea to a finished branch."
- },
- "plugins": [
- {
- "name": "hyper",
- "source": "./",
- "description": "One job from an idea to a finished branch: ask first, design with EGM, build with Superpowers, and give every step the model and effort that fit it.",
- "version": "0.2.1"
- }
- ]
- }
README.md · README · 225 lines
- # Hyper
- One job from an idea to a finished branch, with every step on the model and effort that fit it.
- Hyper is a Claude Code plugin. It does not replace the tools you already use. It puts two of them in order and adds
- the one thing none of them has: a table that says which of *your* models does which kind of work.
- ```
- ASK Hyper the interview: one round of questions, each with a recommended answer
- DESIGN Google's EGM the design document, in EGM's template, approved by you once
- BUILD Superpowers the plan, test-first builds in parallel, reviews, the finished branch
- ROUTE Hyper the size of the job, the record, the model and effort of every agent
- ```
- ## Why
- - **Different work suits different models.** A strong reasoner should write the design. A cheaper, careful model
- should build from it. A model from another vendor should review it, because it misses different things.
- - **Small work should stay small.** A rename needs no agents. A new subsystem needs a reviewed design. Hyper sizes the
- work first and runs only the steps that size needs.
- - **Asking first is cheaper than rebuilding.** Nothing is designed or built until you have confirmed, in a written
- record, what was understood.
- - **Time is spent where it pays.** One reviewer per gate, independent tasks built at once, and per-task reviews only
- while the job's own history says they still find something. `hyper time` shows where a job's time went.
- ## Requirements
- - Claude Code 2.1.271 or later, and Python 3.9 or later on your PATH as `python3`.
- - For the full flow: [EGM](https://github.com/google/elephant-goldfish-model) and
- [Superpowers](https://github.com/obra/superpowers), both enabled where you use Hyper.
- - Recommended: [Matt Pocock's skills](https://github.com/mattpocock/skills), for the `domain-modeling` skill that
- writes the words of the domain into a glossary after the interview. They only need to be installed.
- - Optional: the Codex command-line tool, when you want a second vendor to review. (The Gemini tool also works, at a
- basic level.)
- ```
- claude plugin marketplace add google/elephant-goldfish-model
- claude plugin install egm@egm
- claude plugin enable egm@egm
- claude plugin install superpowers@claude-plugins-official
- claude plugin install mattpocock-skills@claude-plugins-official
- ```
- ## Install and set up
- Install it from this repository, then run setup once:
- ```
- /plugin install hyper --marketplace OmarHafez1/Hyper
- /hyper:setup
- ```
- From a shell, the same in two commands: `claude plugin marketplace add OmarHafez1/Hyper`, then
- `claude plugin install hyper@hyper`.
- Setup checks your machine, asks which models you have, researches what each is good at, and writes your use-case
- file. To try Hyper without the research, copy an example roster and adjust it:
- ```
- python3 <plugin folder>/scripts/hyper.py init --example claude-only
- ```
- Install it for one project (`--scope project`) if you want Hyper only there.
- ## Install the Codex skill
- Run these commands at the root of a clone of this repository.
- ```sh
- skill_dir="${CODEX_HOME:-$HOME/.codex}/skills/hyper"
- mkdir -p "$skill_dir"
- cp codex/skills/hyper/SKILL.md "$skill_dir/SKILL.md"
- ln -s "$PWD/scripts/hyper.py" "$skill_dir/hyper.py"
- ```
- Use `$hyper <idea>` or `$hyper resume` in Codex. Both hosts use the same model settings and jobs.
- ## Upgrading from Hydra
- Install the new plugin ID, then disable the previous plugin in Claude Code so its hooks run only once.
- Hyper uses existing settings and jobs in place. When an old data directory exists alone, Hyper links
- the new name to it so saved paths, Git worktrees and concurrent sessions keep using the same files.
- On filesystems that cannot create symlinks it reads the old directory directly. `HYPER_HOME` and
- `HYPER_OFF` take precedence over the previous environment variable names.
- ## Use
- ```
- /hyper:feature <idea> the whole flow, from the first question to the finished branch
- /hyper:feature build <design.md> the design exists and is reviewed: plan and build it
- /hyper:feature resume continue a job in a later session
- /hyper:setup refresh your models changed
- ```
- Two skills also work by themselves, without the flow: `hyper:asking-first` runs whenever you ask for something new, and
- `hyper:routing` gives the model and effort for any single agent.
- ## How it decides
- Three inputs, three settings:
- | Input | Setting |
- | --- | --- |
- | The kind of work (the *seat*) | which model |
- | How hard the step is (the *level*: easy, normal, hard) | how much effort |
- | How big the job is (the *lane*) | which steps run |
- **Lanes.** The ask-first step looks for four facts: the work is costly to undo, it is wide, its requirements conflict,
- or it touches a risk path (money, authentication, user data, concurrency).
- | Lane | When | Steps |
- | --- | --- | --- |
- | trivial | a mechanical edit, none of the four facts, at most 2 files | just do it |
- | small | none of the four facts, and new work of at most 5 files (or a larger mechanical edit) | ask, plan, build, final review, close |
- | feature | one or two facts, or new work of more than 5 files | adds a design document (approved once) and one review of the design and the plan together |
- | critical | three or more facts, or costly to undo on a risk path | the same steps at harder efforts, with a second reviewer from another vendor at the final review |
- A concept review (EGM's brainstorm lenses, before any design) runs only when you ask for one. In any lane, work on a
- risk path gets a security checklist in the final review, and the final review also lists where the code differs
- from the design.
- **Builds.** Up to `max_parallel` plan tasks (default 3) build at once, each in its own git worktree, when their
- dependencies are merged and their files do not overlap. Easy tasks get no review of their own; hard tasks always do;
- normal tasks are reviewed until `trust_after_clean` reviews in a row (default 2) came back clean, and again as soon as
- one finds a Critical or Important problem. The final review always runs. Every builder and reviewer writes its full
- report to a file and replies in at most 10 lines.
- **Roles.** The table inside the plugin never names a model. It names five roles, and setup binds each to one of your
- models:
- | Role | Does | Chosen for |
- | --- | --- | --- |
- | designer | architecture, the design document, hard plans | strength; nothing later catches a design miss cheaply |
- | builder | code, tests, fixes, ordinary plans | build quality for its cost |
- | critic | the first review of everything | being different from the builder: another vendor when you have one |
- | second | a second review, only beside a critic from another vendor | low cost |
- | scout | search, fetches, mechanical edits | lowest cost |
- **Effort.** Each model has three efforts: a floor, a normal one, and a hard one. Hyper never asks for an effort
- below a model's floor. Search and
- mechanical work, and builds whose code is already given, run at the floor. A step runs at the hard effort when it is
- hard by the facts, when it is the last gate before the work is finished, or when it is a redo after a miss. The rest
- runs at normal.
- ## Your files
- Everything about you lives outside the plugin, in `~/.config/hyper` (or `$XDG_CONFIG_HOME/hyper`, or `$HYPER_HOME`).
- `hyper doctor` prints the folder.
- | File | What it is |
- | --- | --- |
- | `models.json` | your models, the role each plays, each model's three efforts |
- | `USE-CASES.md` | the same as a readable page: who does what and why |
- | `evidence/` | what the choices rest on, each claim with its source, date and confidence |
- | `settings.json` | your changes: a role, an effort, a single cell of the table, a setting |
- | `ledger.jsonl` | one row for every agent Hyper checked |
- Each project keeps its own in `.hyper/`: the review bar (`bar.md`), an optional `settings.json` that wins over yours,
- and one folder per job with its record, its state and EGM's working files. Hyper writes a `.hyper/.gitignore` that
- keeps the job folders out of git; the bar and the project settings are yours to commit.
- Useful commands (`hyper` means `python3 <plugin folder>/scripts/hyper.py`):
- ```
- hyper pick build --level hard which model and effort for this work
- hyper model off <id> a model hit its limit; its stand-in runs
- hyper config set fix_round_cap 3 fix rounds per task before what is left is settled
- hyper config set max_parallel 2 plan tasks built at once
- hyper job list the jobs of this project and where each stands
- hyper time <job> where a job's time and tokens went, per step and per seat
- hyper gate normal --task 3 whether a task gets its own review (the build loop runs it)
- hyper ledger --summary what ran, on which model
- hyper doctor what is installed, what is missing
- ```
- ## What it changes in the other tools
- Nothing in their files. Where their rules meet, Hyper chooses as follows, and says so in `docs/DESIGN.md`:
- - The interview happens once, in the ask-first step, and it is written down by Hyper. Superpowers' brainstorming is
- not run again. EGM still asks about design decisions the record leaves open.
- - The design is written in EGM's template in one pass and approved by you once, without egm-draft's section-by-section
- gates. EGM's staged review is replaced by one review of the design and the plan together; EGM's brainstorm lenses
- run only for a concept review you ask for.
- - Superpowers' advice on which model to use is replaced by the table. Its build runs several tasks at once in
- worktrees, and its implementers and reviewers read its own prompt templates. Its fix loop keeps its rules, with a
- lower cap on rounds (default 1; `fix_round_cap` 5 gives its own length), and a fix is re-reviewed only when it
- answers a Critical finding.
- - The approved design is committed, the build's branch carries it, and decisions made during the build are written
- back into it. The final review lists where the code differs from the design, and you decide each difference.
- - A decision the build made that you reject sends the code back to be fixed.
- How a job ends is yours to choose in Superpowers' last step: merge, pull request, or keep the branch.
- ## Turning it off
- `touch .hyper/off` in a project stops Hyper's session note there. `HYPER_OFF=1` does the same for one session. The
- hook only checks agents in a session that is running a Hyper job, or agents whose prompt starts with a Hyper tag; it
- is silent everywhere else, and it never stops a session when something goes wrong inside it. It checks the model of
- every such agent, and the effort whenever Claude Code reports it.
- ## Version 0.2
- 0.2.0 is the speed pass: fewer steps (a critical job has 8 instead of 15), one reviewer per gate, parallel builds,
- per-task reviews that adapt to the job's quality, short replies, and `hyper time`. Jobs started under 0.1 keep
- working: a step done under an old name counts under its new one (`approve` as `design`, `plan_check` as `review`,
- `finish` as `close`), and a settings override of a removed seat is ignored with a warning from `hyper validate`.
- Not here yet: a calibration run that measures your models on planted defects, a script that proposes table changes
- from the ledger, automatic reading of usage meters, a saved account of a half-finished interview, and the final
- result of a background agent in the log. Windows support is partial: the hooks need a `python3` command, and two
- commands at once are not protected by file locks there. The Gemini tool is supported at a basic level.
- The 0.2.0 lanes have not yet carried a real feature from start to finish. Treat the first job as a trial.
- Known limits. Hyper is built for one user, one project and one job at a time, with the design inside the
- repository. It has not been hardened for: job folders that are symbolic links, job files edited by hand, a design
- kept in a worktree of another branch than the one being built, and several sessions driving one job at the same
- moment (a second session is refused, not merged).
- ## Credits
- Hyper drives, and does not copy, [EGM](https://github.com/google/elephant-goldfish-model) (Google, Apache-2.0),
- [Superpowers](https://github.com/obra/superpowers) (Jesse Vincent, MIT) and
- [Matt Pocock's skills](https://github.com/mattpocock/skills) (MIT). Hyper itself is MIT licensed.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.