Named Agent Skillsets

skillset manages multiple named global agent skill collections under ~/.agents/skillsets and activates one collection through stable relative symlinks. Collections can be upstream-managed or explicitly hand-managed. The Linux-only CLI provides lifecycle, inspection, and diagnostic commands plus managed delegation to the upstream skills CLI.

Requirements and PATH

The CLI requires Linux and Python 3 and uses only the Python standard library. Add its directory to your shell startup file using the syntax for your shell, then start a new shell:

# Bash and Zsh
export PATH="$HOME/.agents/bin:$PATH"

# Fish
fish_add_path "$HOME/.agents/bin"

Use an alternate home directory

skillset stores its managed data under $HOME/.agents. To operate on a different home directory for a single command, override HOME for that command:

HOME=/path/to/alternate-home skillset list

This reads the managed layout from /path/to/alternate-home/.agents. The alternate home directory must already exist. The override also applies to any upstream skills command delegated through skillset skills.

Store skillsets in another directory

By default, named collections live in the real ~/.agents/skillsets directory. To keep them in a repository or another durable location, create that directory first and add a real ~/.agents/config.json:

{
  "version": 1,
  "skillsets_directory": "/home/alice/Repos/dotfiles/skills"
}

The path must be absolute and normalized, outside the managed ~/.agents root, and it must name an existing real directory rather than a symlink. The configured directory contains the named set directories (default/, work/, and so on), not one flat collection of skills. Configuration does not permit symlinked sets, skills/ directories, or individual skill directories. Use a dedicated directory inside the Git repository, not the repository root; otherwise repository metadata such as .git is treated as a named skillset and makes the configured layout invalid.

On a fresh installation, write the config before initialization:

mkdir -p /home/alice/Repos/dotfiles/skills
skillset init default

Initialization creates the exact absolute ~/.agents/skillsets -> /home/alice/Repos/dotfiles/skills link and creates the initial named set in the configured directory. Other valid named sets may already be present, but the initial name must be unused. Runtime aliases and locks remain under ~/.agents.

Before migration, close any skillset or delegated upstream command that is currently using the installation. Move the real ~/.agents/skillsets directory to the desired destination, write config.json, and replace the old location with the exact absolute symlink to that destination. One ~/.agents installation should own a configured directory; sharing it between installations is unsupported. Then run:

skillset doctor
skillset list
skillset current

An alternate HOME continues to select both its own .agents/config.json and its own runtime state.

Shell completions

Load completions for the current shell session:

# Bash
source <(skillset completions bash)

# Zsh
source <(skillset completions zsh)

# Fish
skillset completions fish | source

For persistent completion, redirect the generated script to a file in your shell's normal completion directory and restart the shell. The destination depends on the shell and distribution. For Zsh, use an underscore-prefixed filename such as _skillset in a directory on fpath; compinit will discover and autoload it.

Script generation does not require an initialized or healthy managed layout. Command and option completion is always available after loading the script; existing skillset-name completion calls skillset list and therefore requires a healthy managed layout at completion time. Arguments following skillset skills are left to the upstream CLI.

Initialize the current installation

Choose a lowercase name for the current global collection:

skillset init default

This adopts existing real ~/.agents/skills and ~/.agents/.skill-lock.json entries. Missing entries are created. Existing symlinks or a partial managed layout are refused rather than guessed at.

Create skillsets

Create an empty inactive set:

skillset create experiment

Create a hand-managed collection without upstream lock metadata:

skillset create --manual personal

Manual collections contain a real skills/ directory and an empty .skillset-manual marker, but never a local .skill-lock.json. Cloning a collection preserves its mode. --manual and --from cannot be combined.

Or clone both the complete skills tree and lock metadata from another set:

skillset create experiment --from default

Create and activate an empty set in one command:

skillset create --use experiment

Clone and activate in one command:

skillset create --use --from default experiment

Options may appear before or after NAME. If creation succeeds but activation fails before replacement, the new set remains and the previous set stays active; inspect it and retry skillset use NAME.

Names use lowercase letters, digits, underscores, and hyphens, and must start with a letter or digit.

Activate a skillset

skillset use experiment

Activation retargets ~/.agents/active. Managed collections keep the root lock alias pointed at active/.skill-lock.json; a manual collection instead uses the read-only shared empty-lock sentinel ~/.skillset-manual-empty-lock.json. Keeping this runtime-owned file outside the skillset repository prevents version-control tools from trying to rewrite it during commits. skillset use recreates it when needed; skillset doctor --fix can also restore a missing sentinel after confirmation. An interrupted activation may leave a recovery record for skillset doctor --fix.

Install and maintain skills

After activating the intended set, run upstream commands through the wrapper:

skillset skills add SOURCE
skillset skills list
skillset skills remove SKILL
skillset skills update

While a manual collection is active, lock-aware upstream commands (add, list, remove, and update, including their aliases) are refused before npx starts. Maintain those skills by editing their files directly. Scope-free upstream commands such as find and use remain available.

For add, list/ls, remove/rm, and update, the wrapper injects --global unless exact -g or --global is already present before an option terminator. It inserts the flag before --, or appends it when no terminator is present. Exact -p and --project options are rejected for those commands because managed skillsets contain global state; tokens after -- remain literals. Scope-free find, use, and upstream init commands, as well as an invocation with no upstream arguments, pass through unchanged:

skillset skills find formatter
skillset skills use SKILL

Unknown upstream commands also pass through unchanged, but the wrapper warns on stderr that global scope was not injected. It does not reinterpret other upstream arguments or output. The delegated process inherits terminal streams, signals, and exit status; only its copied environment has XDG_STATE_HOME removed so lock metadata resolves through the active managed alias.

skillset uses advisory locks to reduce accidental overlap between its own operations. This is best effort: do not run conflicting skillset or direct upstream mutations at the same time. Managed inspection remains read-only, and skillset augment skills list can inspect discovery roots without requiring a managed layout. After initialization, prefer skillset skills ... for global installation and maintenance.

Inspect skillsets

List sets in sorted order. The active set is the only one prefixed with * :

$ skillset list
* default
experiment

Hand-managed sets are annotated with [m] (yellow on a color-enabled TTY):

$ skillset list
* default
personal [m]

Verbose output with either -v or --verbose uses an aligned inventory table. Valid entries use declared skill names; malformed entries use their directory name followed by [invalid: REASON]. A set with no inspectable entries shows (no skills):

$ skillset list --verbose
  SKILLSET   | SKILLS
  -----------|---------------------------------------------
* default    | alpha, broken [invalid: missing description]
  experiment | (no skills)

Print only the active name and its terminating newline for use in scripts:

$ skillset current
default

Show the active set's direct skill directories, sorted by the displayed name:

$ skillset show
SKILL            | DESCRIPTION
-----------------|-------------------------------
alpha            | First skill
broken-directory | [invalid: missing description]

NAME is optional for skillset show [NAME]: omitting it inspects the active skillset, while an explicit name inspects that set whether it is active or inactive. Non-empty sets use the headered, aligned SKILL | DESCRIPTION table shown above. An empty set prints No skills installed. without table headers. Manual sets first print Manual skillset [m]: no upstream lock metadata..

When stdout is a TTY, NO_COLOR is absent, and TERM is not dumb, verbose list and show make both header cells bold and valid skill names cyan. They color [invalid: missing description] annotations yellow and every other invalid annotation red, and dim every row's | delimiter and the complete separator. Verbose list also makes the active marker and skillset name bold cyan and dims (no skills); show dims its empty-set message. All non-TTY output, including pipes and redirects, is plain text; NO_COLOR and TERM=dumb also suppress ANSI styling.

show reads UTF-8 SKILL.md files whose first line is exactly --- and that have a later exact --- closing line. Before that close, top-level name and description must each appear exactly once and be nonempty. The focused, dependency-free reader supports plain scalars, YAML-style single quotes with doubled apostrophes, JSON-compatible double-quoted escapes, and | or > block markers with optional + or - chomping indicators (such as >-), followed by indented block content. Plain and quoted values may continue on indented physical lines, and trailing comments are ignored outside quotes. It ignores extra keys and the body, and collapses whitespace in both displayed values.

Malformed, non-UTF-8, or invalid decoded-Unicode candidates remain visible in the show table as <directory> | [invalid: <reason>] and in verbose list as <directory> [invalid: <reason>] inventory entries. Verbose list uses (no skills) only when inspection finds no valid or invalid skill-directory entries; ignored direct regular files do not prevent this state. Direct skill-directory symlinks and SKILL.md symlinks are never followed. Declared names and descriptions preserve ordinary printable Unicode, while terminal controls, invisible format controls, and line/paragraph separators are rendered as deterministic \xNN, \uNNNN, or \UNNNNNNNN escapes. Wrapper-created table delimiters and line endings remain unchanged.

Rename skillsets

Rename an inactive set without changing the active aliases:

skillset rename experiment trial

The active set can be renamed with the same command:

skillset rename default baseline

An active rename moves the complete set first and then atomically retargets ~/.agents/active; the stable root aliases remain unchanged. Rename refuses any existing destination, including files and symlinks, without overwriting it. If active retargeting fails before it commits, the directory rename is rolled back. If rollback cannot restore a valid layout, the error names the old, new, and active paths and identifies the remaining data location. Preserve that copy and inspect the reported state before either moving the set back or retargeting active; then run skillset doctor.

Remove skillsets

Remove an inactive set with an interactive confirmation:

$ skillset remove trial
Remove skillset 'trial'? [y/N] yes

Only y or yes, ignoring case and surrounding whitespace, confirms removal. Use --yes for noninteractive operation:

skillset remove trial --yes

The active set is always refused, even with --yes. Activate another set with skillset use before removing it.

Use skillsets in Codex

Codex recursively discovers skills below ~/.codex/skills and a project's .codex/skills directory. Each enabled entry is an absolute link to a managed skillset's skills/ directory. The commands below never replace unrelated Codex entries.

Enable

Enable a skillset globally:

skillset codex enable personal

This creates ~/.codex/skills/personal pointing to ~/.agents/skillsets/personal/skills. enable is idempotent for that exact link and refuses to replace any other entry. It creates only missing real .codex and skills directories.

Use -l or --local to manage a .codex/skills directory beneath the current working directory instead. The target remains an absolute path derived from the current user's home directory:

skillset codex enable personal --local

-g/--global explicitly selects ~/.codex and is the default for enable and disable. The scope flags are mutually exclusive and are available on enable, disable, and list.

Disable

Disable a canonical global link:

skillset codex disable personal

Use --local to disable the corresponding link in the current project instead. disable removes only the expected absolute target; it refuses a missing or unrelated entry.

List

List every skillset Codex can discover from the global and current-project locations:

$ skillset codex list
[g] always-on
[g] personal [m]
[l] obra

[g] identifies a global link and [l] a local link. They are cyan and green, respectively, on a color-enabled terminal. A skillset enabled in both locations appears twice. There is no * marker: all listed skillsets are available to Codex.

Use -g/--global or -l/--local to inspect one location only; the scope marker is then omitted because it would be redundant:

skillset codex list --global
skillset codex list --local

Add -v or --verbose to either the combined or filtered listing to show the same skill inventory table used by skillset list:

skillset codex list --verbose
skillset codex list --local --verbose

Rename and removal refuse a globally Codex-enabled skillset; disable it first so the link cannot become stale. Local links are intentionally not tracked outside their project directory, so disable them before renaming or removing their referenced skillset.

Use skillsets in Claude Code

Claude Code discovers direct skill directories rather than collection directories. skillset claude therefore creates one flat link per direct skill and records the owning collection separately:

~/.claude/
├── .skillsets/
│   └── personal -> ~/.agents/skillsets/personal/skills
└── skills/
    ├── feature-status -> ~/.agents/skillsets/personal/skills/feature-status
    └── handoff -> ~/.agents/skillsets/personal/skills/handoff

The actual targets are absolute paths. Existing unrelated entries beneath .claude are preserved.

Enable a collection globally, or beneath the exact current working directory:

skillset claude enable personal
skillset claude enable personal --local

Global scope is the default for enable and disable; an unscoped list combines global and exact-current-directory registrations. -g/--global and -l/--local are mutually exclusive and work with all three commands. More than one collection can be enabled when their direct skill-directory names do not overlap. Enable also checks the effective global and exact-current-directory scopes against each other. A different-target name collision is refused before registration or projected links are changed; an identical canonical target in both scopes is allowed.

Rerun enable after adding or removing direct source skills. It adds missing links, removes only stale links owned by that registration, and leaves unrelated files, directories, and links intact. A partial operation retains its registration as recovery intent: rerun the same enable to complete the current projection, or run disable to remove its owned links.

Disable a global or exact-current-directory registration:

skillset claude disable personal
skillset claude disable personal --local

The registration is removed last. The .claude, .claude/skills, and .claude/.skillsets containers remain even when empty.

List global and exact-current-directory registrations:

$ skillset claude list
[g] always-on
[g] personal [m]
[l] obra

Use --global or --local to omit the redundant scope label, and --verbose to include the normal skill inventory. This reports synchronized registrations, not every skill Claude Code can discover from ancestor, descendant, plugin, bundled, or enterprise locations. Claude Code gives a personal skill precedence over a project skill with the same name, so listing refuses different-target collisions across the effective global and local scopes instead of presenting the masked local registration as available.

Global enable can preflight only the exact current project; local registrations in other projects are not globally indexed. If a global skill is later enabled from elsewhere, running claude list or claude enable inside an affected project reports the collision. Disable one of the conflicting registrations before using that local projection.

Local registrations and links contain machine-specific home paths. Keep this generated state out of version control without ignoring any hand-authored .claude/skills content that the project intentionally commits. The command does not edit .gitignore.

Claude Code may need a session restart when a top-level skills directory is created after the session starts. Changes inside an already watched skill directory are normally visible through the live link.

Rename and removal operations refuse globally registered Claude Code skillsets, including malformed registration state that cannot safely be classified. Local registrations are not indexed globally, so disable every known local registration before renaming or removing its source set.

Use skillsets in Auggie

Auggie, like Claude Code, discovers direct skill directories. skillset augment therefore projects a managed collection as flat, absolute links and keeps its registration outside the public skills directory:

~/.augment/
├── .skillsets/
│   └── personal -> ~/.agents/skillsets/personal/skills
└── skills/
    ├── feature-status -> ~/.agents/skillsets/personal/skills/feature-status
    └── handoff -> ~/.agents/skillsets/personal/skills/handoff

Enable or disable a managed collection globally, or in the exact current directory with --local:

skillset augment enable personal
skillset augment enable personal --local
skillset augment disable personal
skillset augment disable personal --local

Global scope is the default for enable and disable; an unscoped skillset augment list combines global and exact-current-directory managed registrations. -g/--global and -l/--local are mutually exclusive for all three commands. --verbose adds the usual managed-skill inventory.

enable never replaces unrelated entries. It rejects different-target direct skill-name collisions between the effective global and local .augment projections, while allowing the identical canonical target in both scopes. Rerun enable after changing direct source skills to reconcile missing and stale links; disable removes owned projected links before its registration. If a normal operation is interrupted, rerun enable to reconcile the projection or use disable to remove its owned links. Empty containers remain in place.

Global Auggie registrations, including malformed registration state, block rename and removal of their source skillset. Local registrations are not globally indexed, so disable every known local registration before renaming or removing its source set. Local .augment links and registrations contain machine-specific absolute paths: do not commit this generated state. The command does not edit .gitignore or unrelated entries.

Inspect effective Auggie discovery

skillset augment skills list is separate from skillset augment list: it is a read-only view of effective individual skills, not the managed registration ledger. It scans all six documented discovery roots in this precedence order and prints each winning skill in a SKILL | SOURCE | DIRECTORY table. SOURCE uses concise, unbracketed labels: user roots are ~/.augment, ~/.claude, or ~/.agents; workspace roots are ./.augment, ./.claude, or ./.agents. DIRECTORY is the resolved skill-directory path. Paths under the configured home directory use ~; paths outside it remain absolute:

Order	Precedence root	Skill root
1	User/.augment	~/.augment/skills
2	Workspace/.augment	$PWD/.augment/skills
3	User/.claude	~/.claude/skills
4	Workspace/.claude	$PWD/.claude/skills
5	User/.agents	~/.agents/skills
6	Workspace/.agents	$PWD/.agents/skills

Only direct child directories containing SKILL.md are candidates. The first candidate for a skill-directory name wins; lower-precedence duplicates are omitted. Effective winners print first in lexicographic concise source-label order, then in canonical resolved skill-directory order. On eligible TTY output, headers are bold, dividers are dim, skill names and ~/... sources are cyan, and ./... sources are green. Captured output and TTY output with NO_COLOR set or TERM=dumb remain plain. --verbose instead retains its non-tabular diagnostic blocks: its first line uses bracketed concise source labels, followed by unchanged resolved-directory and SKILL.md detail lines. The command resolves all six roots before --global or --local filters the winners, so a lower-precedence local skill shadowed by a global one is not reported as effective.

The .claude/skills and .agents/skills roots are compatible, read-only external inputs to this view. Auggie management never creates, validates, modifies, registers, or removes entries there; a same-name compatible skill does not block a managed .augment projection because documented precedence selects the effective skill.

Diagnose the managed layout

Run the read-only diagnostic command after initialization, when another command reports an invalid layout, or before attempting manual recovery:

skillset doctor

doctor inspects the managed layout without creating it and reports all discovered layout problems. A missing .agents root or managed lock is reported alongside missing aliases, active, and skillsets. It checks all three canonical aliases, the active target grammar and existence, direct skillset directory names and shapes, stale use/create staging markers, readable version-3 lockfiles, and each direct skill directory's SKILL.md metadata.

Every finding is written to stderr on one line. Structural, alias, lockfile, metadata, symlink, and staging problems are skillset: error: findings. A lock entry without a corresponding real direct skill directory, or an installed directory absent from the lockfile, is a skillset: warning: finding. Direct regular files under skills are not installed skill directories and are otherwise ignored. Finding text uses the same visible escaping for controls and line separators, so each finding remains one physical line. Errors are printed before warnings. Any error produces exit status 1; warnings alone and a healthy layout produce status 0. A healthy run may print nothing.

Recovery

Operations refuse invalid layouts and stale staging paths without deleting them. When an error names recovery paths:

1. Preserve the only remaining copy of every skill directory and lockfile.
2. Run skillset doctor and review every error and warning before changing anything.
3. Inspect the concrete original, staged, old, new, and active paths named by the failed operation.
4. Restore a missing original only from its staged counterpart, or move a set back/retarget active only after confirming which path contains the data.
5. Run skillset doctor again after manual recovery.

doctor --fix can repair a deliberately small set of lossless omissions after an interactive y or yes confirmation. It may create a missing advisory lock and a missing version-3 empty lockfile only when that skillset's skills directory is empty. It does not overwrite existing files. Each created file is reported as skillset: repaired:. Resolve every other finding manually after preserving the data it identifies.