> ## Documentation Index > Fetch the complete documentation index at: https://reversi.mintlify.site/llms.txt > Use this file to discover all available pages before exploring further. # Operator console > See what an agent changed, decide approvals, run rollbacks, edit manifests and policy, and download evidence in a browser. The console is a set of server-rendered pages at `/console`, with no client-side JavaScript. The server, the gateway (on its own port) and `reversi console` can all serve it. ## Modes The console is **off by default** and serves nothing until you choose a mode. | Mode | Login | Who may use it | Turn on with | | - | - | - | - | | `off` (default) | none | Nobody. `/console` answers 404 with a hint | nothing | | `open` | none | Your own browser on this machine. Fixed role (default `operator`, or `auditor` / `approver`), **never admin** | `REVERSI_CONSOLE=open`, `reversi console`, gateway `--console open` | | `login` | paste an API key or an OIDC access token | Anyone with an `auditor`, `approver`, `operator` or `admin` credential. Agent credentials are refused | `REVERSI_CONSOLE=login`, gateway `--console login` | ```bash theme={null} reversi console # open mode on http://localhost:7421/console as operator reversi console --role auditor --port 7500 # read-only REVERSI_CONSOLE=login REVERSI_KEYS=./keys.json node packages/server/dist/main.js ``` The same setting can live in `reversi.config.yaml` (`console: { mode: login }`; for `open`, `role:`), and for the server `REVERSI_CONSOLE_ROLE=auditor|approver|operator` picks the open-mode role. **Open mode is for a laptop.** It has no login, so it is fenced: * Only requests whose `Host` is `localhost`, `127.0.0.1` or `::1` are answered, any `Origin` must equal the host, and `Sec-Fetch-Site` must be `same-origin` or `none`. Anything else gets a 403 page (tests cover foreign hosts and origins). * The server and the gateway **refuse to start** in open mode if they listen on a non-loopback address (`HOST` for the server, `listen.host` for the gateway; the Docker image sets `HOST=0.0.0.0`, so use `login` there). They also refuse the `admin` role. * It is labelled "open console, no login" in the header. **Login mode** takes an API key or OIDC *access token* pasted into a form. There is no browser redirect to an identity provider (OIDC redirect login is not built). The session is a signed cookie, re-checked against the key ring on every request, so disabling or demoting a key ends the session at once. An OIDC session never outlives its token. ## Pages | Page | For | | - | - | | **Runs** | Filter by actor, tenant, tool; "needs attention" shows runs with unresolved, failed-compensation or unrecoverable steps. Link to compare two runs | | **Run detail** | Steps and statuses with risk scores, the rollback plan (dry run), the full event trail, pending approvals, verification of the undo, report and bundle links, and the actions below | | **Approvals** | Pending queue and recent decisions, with approve, deny, break-glass and delegation | | **Alerts** | Live list of what needs attention: severity, what, since when, which run | | **Manifests** | Enforced manifests with version, owner, data class, undo and lint findings; a manifest editor | | **Policy** | The active policy, its own tests, a rule builder, an editor with simulation, and versions | | **Impact** | Preview a call to a tool: decision now, risk score, approvers, rollback outlook | | **Tools** | Inventory: manifests, tools the upstream offers now, use counts, which are unclassified | | **Connectors** | Each upstream: kind, health, tools, how many have manifests. Filled by the gateway | | **Compare** | Two runs side by side | | **Evidence** | Ledger verification, signed whole-ledger bundle, kill switches | | **Settings** | Roles, API keys, legal holds, retention, notification channels, integrations | ### Run detail * **Auditor report** (HTML and PDF) and a **signed evidence bundle** (needs a signing key). **Replay with a change** appears when capture is on: change one argument and see which decisions would differ, in a sandbox that executes nothing. * **Resolve** (operator): for a step stuck at `intended` or `uncertain`, choose `succeeded` or `failed` and write how you checked. Recorded with your name. * **Rollback plan**: each action, the compensation tool, whether it needs approval, residue, and whether the undo is read back. * **Run the rollback / Retry the rollback** (operator): shown when a compensator is configured. It runs in the background, the page refreshes every 2 s, and the result lists compensated, failed and unverified steps. The checkbox "also approve compensations that need approval" works only if you hold the approver role too. Without a compensator the page says how to run it elsewhere. * **Escalate to on-call** (operator): shown when a compensation failed or a step is unrecoverable. Writes an `escalated` record and sends a critical notification. Where a compensator comes from: | Console served by | Executes rollbacks? | | - | - | | Server | Yes, if `REVERSI_COMPENSATOR_CMD` is set (a command that gets the call as JSON on stdin; exit 0 = undone). Declared read-backs run through `REVERSI_READER_CMD` (a command that gets `{"tool","arguments"}` as JSON on stdin and prints the result as JSON). Without it those steps stay `unverified` and the report says incomplete | | Gateway | Yes, always: compensations and read-backs are called on the upstream | | `reversi console` | No. It shows the plan; use `reversi rollback` | ### Approvals Each row shows why approval is needed, the **who may decide** column (the required roles or groups, "any approver", the quorum progress such as `1/2`, an `escalated` badge) and argument **names** only, never values. Approve and deny take a note. A requester cannot approve their own request, and the server applies the same role, group and quorum rules as the API. **Break-glass** (admin) overrides a pending request with a reason of at least 10 characters and an evidence reference such as a ticket; it is marked `break-glass` on the approvals page and in the decision note. **Delegation** hands your approval authority to someone for 1 to 2160 hours; the delegate decides as themselves on your behalf. It appears only when the engine services are present. ### Manifests Needs a **registry folder** (created by `reversi manifests publish`, or by `reversi apply`) as the `manifests:` setting. With plain YAML files the console shows manifests but they are **read-only**, and the page says so. With a registry: | Action | Role | | - | - | | New manifest (form builder or YAML), validate, see the diff against the current version | operator | | Save as **draft** (not enforced) | operator | | Draft from MCP tools (paste a `tools/list`, or the connected upstream's tools are prefilled) | operator | | Run contract tests (the `reversi test` format) | any console role | | **Publish as approved**, **approve** a draft (optionally "the author may not approve it"), **deprecate**, **restore** an older version | admin | Every save is a new version; history shows the author, approver and signature. An approve or restore reloads the running configuration, and if the new configuration does not load the page says so and the previous one stays active. Drafts from MCP tools come with a confidence level and notes; annotations are the server's claims and only ever produce drafts. A manifest's `approval:` object (roles, quorum, timeout, `escalate_to`) is kept as written by save, approve, deprecate and restore. ### Policy Needs a **policy versions folder** (`reversi policy publish`) as the `policy:` setting; a single policy file is shown read-only. The rule builder (tool, effect, environment, tenant, data class, one argument condition, approver groups, quorum) adds a rule to the editor. Any console role can validate (runs the policy's own tests, shows a diff against the active version) and **simulate** against the last 7 days of captured calls (needs capture on): newly blocked, newly needing approval, newly allowed. **Publish and activate**, publish without activating, and **activate** an older version (which is how you roll back) need admin. Versions list the author, tests passed and note, with diffs between neighbours. ### Settings | Section | Notes | | - | - | | Roles | What each role may do | | Users and API keys | Needs the server started with `REVERSI_KEYS` pointing at a writable file. Admin can create (the key is shown once), disable, enable and remove. Not available in the gateway console or with OIDC-only sign in | | Legal holds | List; admin can place and release ([retention](/operations/retention)) | | Retention | Needs a `retention:` section in the policy. Admin can dry-run or apply | | Notification channels | Names; admin can send a test ([notifications](/operations/notifications)) | | Integrations | Slack approvals, OpenTelemetry, anchoring, evidence signing and egress status | ### What needs which role | Role | Gets | | - | - | | `auditor` | Everything read-only, contract tests, policy validate and simulate, report and bundle downloads | | `approver` | Plus approve, deny, delegate | | `operator` | Plus resolve, run or retry a rollback, escalate, replay, draft and save manifests, import from MCP tools | | `admin` | Plus publish approved manifests, approve or deprecate or restore, publish or activate policy, kill switches, break-glass, keys, holds, retention, test notifications | Buttons the role cannot use are hidden and the POST is refused anyway. ## Security * The session cookie is signed (HMAC-SHA-256), `HttpOnly`, `SameSite=Strict`, scoped to `/console`, and `Secure` with `REVERSI_SECURE_COOKIES=1`. It lasts 8 hours and ends on restart unless you set `REVERSI_SESSION_SECRET`. * Every action is a POST with a per-session CSRF token. The login form body is capped at 16 KB. * All values are HTML-escaped and responses carry a strict `Content-Security-Policy` (`default-src 'none'`, `form-action 'self'`). * Roles are checked on every page. Terminate TLS in front of the server. Do not expose the console to the internet without a reverse proxy and network controls. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.