Harnesses & adapters

A harness is a target AI coding agent. Each one loads skills from its own path with its own schema. adept sync renders every canonical skill into every enabled harness, and adept sync-from adopts harness-side edits back into canonical.

Enable and inspect harnesses:

adept harness add claude-code     # enable a harness
adept harness remove cursor       # disable a harness
adept harness list                # registered harnesses + enabled state

Enabled harness ids are recorded in .adeptability/config.json under harnesses[].

Specialized renderers

Five harnesses ship dedicated renderers that get the details right — correct frontmatter, activation translation, sidecar handling, and size budgets:

Harness	id	Output	Global
Claude Code	`claude-code`	`.claude/skills/<id>/SKILL.md` (full sidecars, `allowed-tools`)	`~/.claude/skills/<id>/SKILL.md`
Cursor	`cursor`	`.cursor/rules/<id>.mdc` (activation → `alwaysApply`/`globs`/`description`)	not global-capable
OpenCode	`opencode`	`.opencode/skill/<id>/SKILL.md` (full sidecars)	`~/.config/opencode/skill/<id>/SKILL.md`
Codex	`codex`	`AGENTS.md` (aggregated, 32 KiB cap, lowest-priority dropped first)	`~/.codex/AGENTS.md`
GitHub Copilot	`copilot`	`.github/instructions/<bucket>.instructions.md` (aggregated per-glob)	not global-capable

See the Harness Comparison for exactly what each renderer emits and why it differs, and Scopes for what --global changes and why Cursor and Copilot are excluded.

Generic per-skill adapters

Beyond the five specialized renderers, adept ships built-in generic per-skill adapters for 50+ more agents — each writing SKILL.md into that agent’s convention. A sample:

amp, cline, continue, gemini-cli, github-copilot (agent-skills mode), goose, junie, kilo, openhands, qwen-code, roo, warp, windsurf, zencoder, and many more that write into a shared .agents/skills/<id>/SKILL.md.

Run adept harness list for the authoritative, live registry — it always reflects the version you have installed.

Config-driven adapters (no recompile)

You can teach adept about a new harness by dropping a YAML adapter file into $ADEPT_LIBRARY/adapters/ (or a library’s adapters/ dir). It’s loaded at startup. Example:

id: jetbrains-junie
name: JetBrains Junie
kind: per-skill
output: ".junie/guidelines/{id}.md"
needs-directory: false
size-budget-bytes: 0

frontmatter:
  include:
    - description
    - tags
  rename:
    description: summary

body:
  prefix: |
    > Auto-generated by adeptability. Do not edit by hand.

detect:
  - ".junie/"
  - ".junie/guidelines/"

Adapter files are validated against an embedded schema on load. Key fields:

kind — per-skill (one file per skill), aggregator-single, or aggregator-per-glob.
output — the on-disk path template; {id} expands to the skill id.
frontmatter.include / frontmatter.rename — which canonical fields to emit, and any key renames.
body.prefix / body.suffix / body.replace — body transforms (with {{.Project.Name}} templating).
detect — path globs used to auto-adopt existing files on adept init.
import.rename — optional reverse-render hints; the default sync-from flow auto-derives the inverse of frontmatter.rename, so this is only needed for non-bijective renames.

See examples/adapters/jetbrains-junie.yaml in the repo for the fully annotated version.

Global support

Two optional keys give a config-driven adapter a home-level target, same as the built-in renderers:

id: jetbrains-junie
name: JetBrains Junie
kind: per-skill
output: ".junie/guidelines/{id}.md"
global-output: ".junie/guidelines/{id}.md"    # path relative to the global render root ($HOME)
global-base-dir: ".junie"                     # detection root when scoped globally

Leave both out and the adapter simply isn’t global-capable: adept --global harness add <id> fails with harness has no global config location, same as the built-in Cursor and Copilot adapters. See Scopes for the full global-scope model.