manage-env を始める

# Set up manage-env (menv) in this repository

manage-env is a zero-knowledge secrets manager for env vars. Secret VALUES must
never enter your context, the chat, or any log - and you never need them to.
Follow these steps in order.

## 1. Install the CLI yourself

Run `menv --version`. If it is missing, install the public macOS/Linux CLI in
your terminal:

`curl -fsSL https://manage-env.com/install | sh`

Then run `export PATH="$HOME/.local/bin:$PATH"` in your current shell and
verify with `menv --version`. Do not ask the human to install the CLI or change
PATH for you. If your client cannot execute the command or requires approval,
request that approval and report the exact blocker instead. Do not send them to
the private GitHub repository.

## 2. Connect the repo yourself

Run `menv connect` from the repository root. It registers the local menv MCP
server for your client and writes the full agent guide into AGENTS.md. It only
touches repo-local files and handles no secret value. Afterwards, reload MCP
servers (or ask the human to restart their AI tool) so the menv tools load.

## 3. Secure user handoff: sign in

Only account authentication requires the human's direct terminal input. Ask
them to run `menv signup` (or `menv login` if they have an account) and enter
the email code, password, and one-time recovery key outside the chat. Do not
ask for, accept, or relay any of those values. Wait until they confirm sign-in.

## 4. Agent steps (menv MCP tools)

- Start with `detect_required_keys` (scans .env.example; works before init).
- If this directory is not bound to a project, call `init_project` - and if
  the `org/project` address is not already clear, ask the human for it
  instead of inventing one.
- Call `setup_project` to mint random secrets, keypairs, and config defaults
  in one version. Treat `needs_user` as the only list of decisions to bring
  back to the human.
- Externally-issued keys (STRIPE_*, DATABASE_URL, cloud creds):
  `request_value_from_user` - the human pastes the value at their own
  terminal. NEVER ask for a secret value in chat.
- Run anything that needs secrets with the `run_with_secrets` MCP tool. If it
  fails, it returns a masked tail of the output (injected values shown as
  `[menv:masked:KEY]`) so you can debug without the secrets; if
  `output_suppressed` is set, a value was too short to mask and the output was
  dropped - ask the human to run it in their terminal.
- When a result reports `stale_sync_targets`, re-sync each one with
  `sync_to_target`.

## Hard rules

- Never `cat .env`, `printenv`, or log `process.env` - reading a value
  defeats the whole point and is treated as an incident.
- Use the `run_with_secrets` MCP tool, not the `menv run` CLI, to run things.
  `menv run` is the human's terminal path; its output isn't masked when a human
  runs it interactively, and `menv run --raw-output` disables masking entirely
  and requires a human at the terminal.
- There is no tool that returns a secret value, by design. Do not try to
  reconstruct values by any other route.
- Do not run `menv run --yes` or `menv run --raw-output` on the human's
  behalf; those confirmation prompts exist for them, not you.

Full tool discipline: AGENTS.md after `menv connect`, or
https://manage-env.com/llms.txt