agent-harness-adapter-0.1.0 is not a library.

agent-harness-adapter

agent-harness-adapter: plug a command-line tool, written in any language, into LLM agent harnesses: Claude Code, OpenAI Codex, Gemini CLI, GitHub Copilot, Cursor, Factory Droid, Pi, OpenCode, Kilo Code, Qwen Code, Devin, and any agent that reads AGENTS.md.

Your tool declares its integration once, in a manifest file. agent-harness-adapter install writes it into each harness's own files, puts shared content in one place, and never overwrites what the user changed. agent-harness-adapter hook turns every harness's hook input into one JSON format for your tool, and your tool's answer back into what that harness expects.

Install

npm install -g @six5536/agent-harness-adapter   # prebuilt binaries
cargo install agent-harness-adapter             # from source

Both install the agent-harness-adapter command. Rust tools can use the library agent-harness-adapter-core directly instead.

Quick start

mytool.harness.toml:

version = 1
name = "mytool"
instructions = { file = "agent/instructions.md" }
allow_commands = ["mytool"]

[[skills]]
name = "mytool"
description = "Use mytool to check the project."
body = { file = "agent/skill.md" }

[[skills]]
dir = "agent/skills/review"   # an existing skill directory, copied as it is

[[hooks]]
event = "pre-tool"
tools = "shell"
run = "mytool guard"          # bridged: mytool reads the hook contract

[[mcp_servers]]
name = "mytool"
command = "mytool"
args = ["mcp"]

agent-harness-adapter install --manifest mytool.harness.toml --harness claude,codex
agent-harness-adapter status  --manifest mytool.harness.toml

Commands

agent-harness-adapter install --manifest <file> --harness <ids|all> [--scope project|user|local] [--root <dir>] [--force] [--without <parts>] [--json]
agent-harness-adapter uninstall --manifest <file> --harness <ids|all> [--scope ..] [--root <dir>] [--force] [--json]
agent-harness-adapter status  --manifest <file> [--harness <ids|all>] [--scope ..] [--root <dir>] [--json]
agent-harness-adapter hook [--tool <name>] [--tools <kind>] <harness> <event> -- <command> [args..]
agent-harness-adapter schema <manifest|hook-input|hook-answer|result>

Harness ids: claude, codex, factory, gemini, copilot, cursor, pi, opencode, kilo, qwen, devin, agents; all is every harness of the manifest with files at the scope. status without --harness reports the installed ones.
A part changed by hand is reported edited and left alone; --force overwrites it (install) or removes it (uninstall).
uninstall takes the tool's content back out: its files, its block in the instructions file, its entries in settings files (the user's stay), and files and folders that are then empty. Content another installed harness still reads is kept (kept … used by <harness>). all is the installed harnesses. Run it with the manifest you installed with.
Exit codes: 0, or 2 with error: <message> on stderr.

The manifest

TOML, or JSON when the file name ends in .json. agent-harness-adapter schema manifest prints its JSON Schema.

Key
`version`	`1`
`name`	the tool's name; region markers are `<!-- <name>:harness -->`
`harnesses`	`"all"` (default) or ids in preference order: the earliest writes a file several harnesses share
`adapter`	the command bridged hooks run (default `agent-harness-adapter`), e.g. an absolute path
`record`, `declined`	paths under the scope's root (defaults `.<name>/harness.toml`, `.<name>/harness.local.toml` at local scope, `.<name>/config.toml`)
`instructions`	text placed between the tool's markers in each harness's instructions file
`skills`	`dir` (a skill directory: its `SKILL.md` as it is and every other file in it, hidden ones aside), or `name`, `description`, `body`, `files` (path → text)
`hooks`	`event`, then `run` (bridged) or `command` (a template), `tools`, `timeout` (seconds), `commands` (harness → template), `owned`
`hook_match`	which hook entries are the tool's: `{ prefix }`, `{ contains }` or `{ any = [..] }`
`mcp_servers`	`name`, then `command`, `args`, `env`, or `url`, `headers`
`allow_commands`, `allow_mcp_tools`	what the agent may run without asking
`agents`, `commands`	`name`, `description`, `prompt`
`parts`	raw `file`, `region` or `merge` parts for one harness
`scopes.<scope>`	items, `record` and `declined` that replace the top level's at that scope

Any text is a string or { file = "<path>" }, relative to the manifest.

Events: session-start, session-end, prompt-submit, pre-tool, post-tool, stop, pre-compact. A hook command template may use {harness} and {event} anywhere; {{ and }} are literal braces.

The hook contract

A bridged hook runs agent-harness-adapter hook --tool <name> <harness> <event> -- <run> (with --tools <kind> when the hook gives tools: on harnesses that cannot match tools, such as Copilot, Cursor, Pi, OpenCode and Kilo Code, the bridge then skips your command for other tools). Your command gets one JSON object on stdin:

{"v": 1, "harness": "claude", "event": "pre-tool", "session_id": "…", "cwd": "…",
 "tool": {"name": "Bash", "kind": "shell", "input": {"command": "ls"}}, "raw": {…}}

Other fields, when the harness sends them: transcript_path, prompt, tool_output, source, continuing (a stop hook already asked the agent to go on), last_message. raw is the harness's own JSON.

It writes one answer on stdout:

{"answer": "allow"}
{"answer": "deny", "reason": "…"}        // pre-tool, prompt-submit
{"answer": "continue", "reason": "…"}    // stop: keep the agent going
{"answer": "context", "text": "…"}       // session-start, prompt-submit, post-tool

If your command cannot start, exits non-zero or writes no valid answer, agent-harness-adapter allows and prints <tool>: <reason> on stderr: your tool failing never blocks the agent. An answer a harness cannot express for the event is an error (exit 1). agent-harness-adapter schema hook-input and agent-harness-adapter schema hook-answer print the JSON Schemas.

Versioning

The manifest's version and the contract's v change only in a release that says so. Pre-1.0, minor versions may contain breaking changes.

License

MIT. See LICENSE.

agent-harness-adapter 0.1.0