KingKongRoboticsKingKongRobotics/jumper/AGENTS.md

jumper

AGENTS.md3.1k starsChanged 32 days agoApache-2.0View on GitHub

The working rules for this repository live in CLAUDE.md — commands, the architecture worth knowing before editing, and the conventions that are not visible from any one file. They are not Claude-specific; read that file first whatever agent you are. Jumper is the shared entry point for Train and Design. Read docs/WORKFLOWS.md to select the route. Movement training uses the root repository. Appearance and environment creation use the independent jumper-design repository: read its current AGENTS.md and referenced workflow/protocol first. Read…

Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 31,774 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository7 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- bundle-manualSkill
- controlsSkill
- deploySkill
- new-taskSkill
- setup-envSkill
- trainSkill

The file

We split this AGENTS.md into the rules it gives an agent

2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.

≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.

Common
Distinctive
Very specific

Very specific:
Keep separate Train and Design virtual environments.
line 18
Distinctive:
Use lightweight models for bounded reading/documentation, standard models for routine implementation and checks, and advanced models for …
line 21

≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.

Files

AGENTS.md

The working rules for this repository live in CLAUDE.md — commands, the architecture worth knowing before editing, and the conventions that are not visible from any one file. They are not Claude-specific; read that file first whatever agent you are.

Start here

Jumper is the shared entry point for Train and Design. Read docs/WORKFLOWS.md to select the route. Movement training uses the root repository. Appearance and environment creation use the independent jumper-design repository: read its current AGENTS.md and referenced workflow/protocol first. Read remotely for guidance; when execution needs source or assets, reuse a suitable existing checkout or clone it separately, then fetch its required LFS assets. Record the Design commit used for generated outputs. Do not add it as a submodule or assume its code or binary assets are present in this checkout. Keep separate Train and Design virtual environments. Imported maps are replay-only in Train; there is no Train .skin loader. Do not silently reconcile robot mechanical differences.

Use lightweight models for bounded reading/documentation, standard models for routine implementation and checks, and advanced models for architecture or difficult debugging. Delegate independent tasks and preserve required validation regardless of model choice.

task	source of truth
Installation, documentation navigation and project status	`docs/PROJECT_GUIDE.md`
Any code or documentation change	`CLAUDE.md` — repository rules and architecture
Set up or repair an environment	`docs/AGENT_SETUP.md`, then the `setup-env` skill
Add a task	`docs/USAGE.md`, then the `new-task` skill
Train a policy and report it	`docs/TUTORIAL.md`, then the `train` skill
Change controls	`docs/CONTROLS.md`, then the `controls` skill
Export or deploy a policy	`deploy/README.md`, then the `deploy` skill
Read or write a bundle consumer	`deploy/BUNDLE.md`
Navigate the repository or run checks	`CONTRIBUTING.md`

Skills

A skill is the playbook for one kind of task, kept at .claude/skills/<name>/SKILL.md. Claude Code finds them there by itself; nothing else about them is Claude-specific. Whatever agent you are, before starting a task one of them covers, read its SKILL.md in full and follow it — each exists because the failures on its path are silent, and its steps are the checks that make them loud. The description: at the top of each file decides whether it applies; the table only points at it.

skill	reach for it when
`setup-env`	getting the Python environment and the controller's Rust toolchain running on a new machine, or repairing one that imports the wrong checkout, ignores the GPU or cannot build the controller
`new-task`	creating a training task — its directory, registration, configs and operator controls
`train`	training a policy and reporting it — sizing the run for the machine's backend, reading whether it learns, replaying it with the measurement on, and handing over a checkpoint with measured numbers rather than an assumed success
`controls`	giving a stick, a button or a key a meaning — a task's `controls.yaml`, a mode switch in `deploy/manifests.json`, a recorded motion's `go`
`deploy`	taking a policy from a checkpoint to the robot — export, RKNN, the bundle, the cross-build, the board
`bundle-manual`	translating a built bundle's `manual.en.json` into Chinese and packing it into the `.app` — after every `scripts/deploy.py` build

A new skill gets a row here; tests/test_layout.py fails until it has one.

Their scripts need no agent framework, only Python and this repository:

For bringing an environment up on a new machine, docs/AGENT_SETUP.md is the procedure and .claude/skills/setup-env/scripts/ holds two standard-library scripts that automate the mechanical parts.
For putting a trained policy on the robot, deploy/README.md is the map and .claude/skills/deploy/scripts/check_board.py is a read-only preflight of the board.
For binding a stick or a button to a meaning, docs/CONTROLS.md is the design and .claude/skills/controls/scripts/write_controls.py writes the file from controller/vocabulary.json and refuses a name that is not in it, because an invented control name parses, binds nothing and looks exactly like one that does not work. It arranges sys.path itself.

The bundle

deploy/BUNDLE.md is the format of what scripts/deploy.py writes — one directory and its .app, from which every host takes its own part, file by file — and what a host that loads one has to do. It is the document to work from when writing or changing a consumer, in this repository or outside it.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

It worked
Partly
It didn't

What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free