jig

README

jig

A jig holds workpieces in place — jig holds your repos in place.

Jig manages a local workspace made of many related Git repositories, driven by a single shared schema. Your organization describes its repositories once — layout, dependencies, tags, generated support files — in a schema hosted in its own Git repository. Everyone then materializes just the parts they need, and jig keeps the checkouts converged, restored, and up to date.

jig clone services/checkout   # clones the service and everything it depends on

Why

• One shared map. Hundreds of repos, one versioned schema: where everything lives, what depends on what, what is archived, how it is tagged.
• Partial by design. Nobody needs the whole org. Clone one service, a group, or a tag — dependencies come along automatically. Ideal for spinning up task-scoped workspaces with all the context an AI agent needs.
• Safe by default. Jig never deletes work: dirty checkouts, unpushed commits, and locally modified files are always detected and preserved.
• Fast. Git operations run in parallel, and a machine-wide clone cache makes repeat clones nearly instant.

Install

go install github.com/fvdsn/jig@latest

Or grab a prebuilt binary for macOS or Linux from the releases page.

Quick start

jig init git@github.com:acme/jig-schema.git ~/Code/acme   # create a workspace from a shared schema
cd ~/Code/acme                  # (or start from scratch in an empty directory: jig init)
jig list                        # browse the catalog
jig clone services/checkout     # install a service + its dependencies
jig status                      # branches, dirty state, ahead/behind
jig fetch && jig status         # what changed across the workspace?
jig pull                        # fast-forward everything installed
jig rm services/checkout        # uninstall

Commands

Command	Description
init [<git-url|file> [dir]]	Create a workspace from a schema repository or a local draft file; no args starts a fresh starter schema
list [path]	List the catalog: groups, repos, files, dirs
info <path>	Show one entry's metadata
deps <path>	Show a repo's recursive dependencies
clone [path]	Install repos/files matching a path, plus dependencies (--no-deps to skip them)
sync [path]	Converge the workspace: moves, origins, file updates, restores (--prune deletes what left the schema)
pull [path]	git pull --ff-only across installed repos, in parallel
fetch [path]	git fetch across installed repos, in parallel
checkout [-b] <branch> [path]	Switch installed repos to a branch (-b creates it); never discards local changes
status [path]	One line per installed entry, plus a summary
rm <path>...	Uninstall: delete the checkout and stop tracking it
update [--sync]	Fast-forward the schema from its remote (then sync)
validate [file]	Validate the schema — also usable in the schema repo's CI
cache [clean]	Inspect or prune the clone cache

Most commands accept --tags a,b (entries carrying all listed tags), --archived, and paths that address a single entry or a whole subtree.

Commands are position-aware: run from a subdirectory, pathless commands scope to that subtree (inside a checkout, they address that one repo), and path arguments resolve like filesystem paths — ., .., and a leading / for the workspace root all work. jig status in services/ shows just your services; jig pull inside a checkout pulls just it.

The schema

A JSON tree where paths are the directory layout. Repos, files, and dirs are leaves; any level can carry group metadata that children inherit.

{
  "version": 1,
  "tree": {
    "platform": {
      "$group": { "description": "Shared platform services", "tags": ["backend"] },
      "auth":   { "$repo": { "id": "auth-service", "git": "git@github.com:acme/auth.git" } },
      "dev.sh": { "$file": { "src": "git@github.com:acme/config.git#scripts/dev.sh", "executable": true } }
    },
    "services/checkout": {
      "$repo": {
        "id": "checkout-service",
        "git": "git@github.com:acme/checkout.git",
        "tags": ["go", "api"],
        "dependsOn": [{ "path": "platform", "reason": "uses platform services" }]
      }
    },
    "tools/ci": {
      "$dir": { "src": "git@github.com:acme/config.git#scripts/ci" }
    }
  }
}

• $repo — a Git checkout. id is a stable identity that survives renames: move the entry in the schema and jig sync moves the checkout on disk.
• $file — a single generated file fetched from a source repo (or a symlink to another file via link). Updated on sync when the source changes; never overwritten if locally modified. src is <clone-url>#<path>, or simply a file URL pasted from the forge web UI (https://github.com/o/r/blob/main/…, GitLab/Bitbucket/Gitea equivalents; default branch only). src may also be a list of sources concatenated in order — e.g. one AGENTS.md assembled from a base section plus sections that follow the installed repositories; list entries can be { "src": ..., "onlyWhen": ... } objects gating individual sections, and when every source is gated off no file is generated.
• $dir — a whole subtree materialized from a source repo (<clone-url>#<subtree>, or a pasted …/tree/main/… web URL). src may also be a list of sources merged in order (first wins on conflicts) — e.g. one .agents/skills assembled from several skill repositories; list entries can be { "src": ..., "onlyWhen": ... } objects to gate individual sources. A $dir can instead declare link to become a relative symlink to another $dir — one real skills directory, symlinked into every harness path. Jig tracks a manifest of what it wrote, so updates touch only unmodified files and user files inside are never touched.
• $group — metadata on a directory: description, tags, dependsOn, archived, onlyWhen are inherited by everything beneath it.
• dependsOn — cloning a repo brings its dependency closure along (optional: true deps only with --with-optional-deps).
• onlyWhen — conditional entries, active only when some active or installed repository matches a path, carries all listed tags, or both — e.g. API skills materialize whenever anything tagged api is installed.
• archived — hidden and skipped by default, kept synced if already installed.

Files and dirs follow the repositories around them: a support file placed inside a group is materialized whenever any repo of that group is installed; root-level files follow the workspace as a whole.

Agent skills

$dir is a natural fit for assembling one agent-skills directory for the whole workspace from several skill repositories. Jig itself ships a skill at .agents/skills/jig/SKILL.md that teaches agents how to drive it — use it as a source:

{
  "tree": {
    ".agents/skills": {
      "$dir": {
        "id": "agent-skills",
        "src": [
          "https://github.com/fvdsn/jig/tree/master/.agents/skills",
          { "src": "git@github.com:acme/billing-skills.git#skills",
            "onlyWhen": { "path": "billing" } }
        ]
      }
    },
    ".claude/skills": { "$dir": { "id": "claude-skills", "link": ".agents/skills" } }
  }
}

jig sync merges the sources into .agents/skills (first wins on conflicts, onlyWhen gates per source) and symlinks it into harness-specific paths — every agent working in the workspace then knows how to use jig. Note the first source is a directory URL pasted straight from the GitHub UI; jig resolves it to the clone URL and subtree path.

How it works

Everything jig manages lives under .jig/ at the workspace root:

.jig/source/       Git checkout of the schema repository — jig reads the schema live from it
.jig/config.json   which file inside the checkout is the schema
.jig/state.json    what is installed, and where

• State records intent. Deleting a checkout by hand doesn't uninstall it — jig sync restores it. jig rm is the uninstall verb, with rm-like ergonomics (-r for groups, -f to override the dirty/unpushed safety checks). jig sync --prune batch-deletes entries that were removed from the schema, with the same safety checks — anything dirty, unpushed, or locally modified is kept and reported.
• The schema is a working copy. Edit .jig/source/<schema>, test immediately with jig sync, then commit and push it like any repo. Teammates pick it up with jig update --sync. Conflicts are plain Git conflicts.
• Clone cache. Jig keeps a bare mirror per remote in the user cache directory; clones hardlink from it, so history transfers over the network once per machine. Checkouts stay fully independent — deleting the cache is always safe. JIG_CACHE_DIR relocates it (empty disables), jig cache clean --unused 30 prunes it.
• Terminal-aware output. list and status align and truncate on a terminal; piped output stays full and tab-separated for scripts.

Compatibility

• Schemas are stable. A version: 1 schema keeps working across jig releases; schema and CLI behavior follow semver from v1.0.0.
• Versioned formats fail loudly. When jig meets a schema, workspace config, or state file with a version it does not understand, it stops with an "upgrade jig" error instead of guessing (or silently dropping fields a newer jig wrote).
• Concurrent-safe. Commands that mutate the workspace take a lock (.jig/lock); state files are written atomically.
• Exit codes tell the truth. Commands exit non-zero when any entry could not be brought to its desired state, so scripts and agents never mistake a partial run for success.
• Platforms. macOS and Linux are supported and tested in CI. Windows is not supported.

Validating the schema in CI

In the schema repository:

validate:
  image: golang:1.26
  script:
    - go install github.com/fvdsn/jig@latest
    - jig validate jig.json