# @agentic-kit/cli

`agent` — a local, secure-by-default coding agent that builds your frontend and your backend. Skills, a backend harness, and offline-capable skill releases, batteries included. ```bash npm install -g @agentic-kit/cli agent # interactive TUI session in the current project agent -p "explain this repo" # one-shot print mode ``` ## What it does On startup, `agent` assembles the harness skill tree and drops you into an interactive terminal session: 1. **Base layer** — a pinned release of [`constructive-io/constructive-skills`](https://github.com/constructive-io/constructive-skills), fetched as a git tag/branch/SHA tarball into `~/.constructive/data/skills//` (offline fallback: newest cached release). 2. **Local overlay** — `~/.constructive/config/skills-overlay//SKILL.md` (agentskills.io layout), highest precedence: a same-name skill replaces the base skill wholesale. 3. **Typed db tools** — the [`@agentic-kit/pi`](https://www.npmjs.com/package/@agentic-kit/pi) extension (provision_database, provision_blueprint, describe_schema, add_records, run_codegen, … with confirm gating on pi's built-in y/n dialog) is materialized into the agent dir and loaded by pi automatically. The merged tree is materialized into the appstash-owned agent dir, so the session picks the skills up natively. All harness state lives in the [appstash](https://github.com/constructive-io/appstash) `~/.constructive/` layout; project directories are untouched. ## Commands | Command | Description | |---------|-------------| | `agent [options...]` | Start an interactive session with the harness skills | | `agent -p "prompt"` | One-shot print mode | | `agent init` | Configure the skills source (repo + pin) interactively | | `agent login` | Sign in to the Constructive platform (interactive) | | `agent logout` | Revoke the API key and clear the stored session | | `agent whoami` | Show the signed-in account, backend, and masked API key | | `agent skills list` | Resolve and list the effective skill set per layer | | `agent skills update` | Re-fetch the base release and re-materialize | | `agent help` | Usage | ## Configuration `~/.constructive/config/skills-manifest.json`: ```json { "repo": "constructive-io/constructive-skills", "pin": "^1.0.0", "manifest": { "sources": [ { "name": "constructive-skills" }, { "name": "local-overlay" } ] } } ``` Env overrides: `AGENT_SKILLS_REPO`, `AGENT_SKILLS_PIN`, `AGENT_HOME` (appstash base dir), `GITHUB_TOKEN` (private skills repos). ## Authentication `agent login` signs in to the Constructive platform with email and password. The flow asks for a backend first: `localnet` (a local backend on `*.localhost:3000`), `devnet` (`*.launchql.dev`), or a custom API URL. After sign-in, the CLI mints a 5-year API key for the db tools and checks it on each startup. `agent logout` revokes the key and deletes the session. Credential files (mode `0600`, plaintext): - `~/.constructive/config/agent/account.json` — the signed-in session (user, access token, API key) - `~/.constructive/config/agent/backend-config.json` — the selected backend endpoints The db tools read these files on each tool call, so a login mid-session is picked up without a restart. Tokens never enter the environment of child processes. Env overrides for CI and headless use (these win over the stored session; read per tool call): `CONSTRUCTIVE_USER_ID`, `CONSTRUCTIVE_ACCESS_TOKEN`, `CONSTRUCTIVE_API_KEY` (optional), `CONSTRUCTIVE_API_ENDPOINT`, `CONSTRUCTIVE_MODULES_ENDPOINT` (default to the local backend). ## Roadmap - Harness system-prompt section + templates/prompts materialization, shared with constructive-desktop via a common adapter package.