id: ops-dashboard title: A human can see what their agents are doing and stop them status: blocked crates: [didbot-serve] dependsOn: [scope-policy, app-allowlist, write-policy, e-stop] exitCriterion: > An owner signs in with their own atproto account and can see, for one agent, every scope it was refused and which rule refused it. #

ops-dashboard #

Served by this server, about this server. It shows live state — agents, grants, refusals — and carries the emergency stop. It owns no policy: everything it can change is operational, and everything it reports about itself is only as trustworthy as the host it runs on.

Sign-in is atproto OAuth against the owner's own server, requesting the atproto scope and nothing else. The owner loads the page here and stays on it; the approval happens in a popup, somewhere we do not run.

• Sign in with authentication only. No write scope on the owner's account, ever, and the client metadata that says so is public. Needs oauth. Until it exists, the dashboard is gated behind the operator credential from auth-types instead, and the page says on its face that this is a stand-in.
• The refusal log. What was asked, what was granted, what was removed and which rule removed it — for an agent, an app, or the whole server. Needs write-policy, scope-policy and app-allowlist, none of which exist yet; the panel says so and shows no numbers, rather than a fabricated zero.
• Collections nothing covers. A ceiling is an enumeration, so a new lexicon is silently outside every profile until somebody names it.
• Scope the stop beyond "everything" — e-stop's own open item, not this epic's to close.
• Grants and app authorisations. Needs oauth; reported as not implemented rather than invented. Every live grant names its client (OAuthGrantStore::live), and whether policy refuses it now is a Subject::Token judged against the current tree, not a stored state. A refusal row names its client (DenialRow::client_id), and so does each admitted write (didbot_pds::write_log), so what an app wrote can be listed for the app.
• No prose written by a model. Placeholders until a human writes them — holds for every string this epic's own work has added so far.

Done #

• Say what it cannot be trusted about. The page carries a trust note stating that a compromised server can lie here and that anything security-relevant has to be checked against the owner's own records, and a machine-readable /dashboard/api/capabilities says the same thing per panel: live, not implemented, or unreachable, never a number standing in for one it does not have.
• The homepage half of this surface: an unauthenticated /dashboard/api/about, reusing bot.did.stats's own numbers, so a person opening a running deployment's address with no credential still gets a real answer about what it is and what it holds — per docs/web-surfaces.md.
• The visual: prototype F's travelling-wave field (crates/didbot-site-anim) as the page's background, with F's bordered-readout-box and column language carrying the panels, opaque enough that the field never competes with a value or a label. Built opt-in (scripts/build-dashboard-wasm.sh), same as the site's own wasm step, so cargo build/cargo test need no wasm32 toolchain.