Admin operations

These pages are for Memo administrators. Access depends on the administrator role; a read-only administrator can inspect a page without receiving permission to change settings or approve work. Workspace ownership alone does not grant product-wide administration. The admin sidebar keeps your expanded or collapsed preference on reload, including while the page is still loading. Use its toggle to change the saved preference; Open app returns to the product.

​

Product and customer records

• Users opens a person’s timeline: account, workspace, generation, usage, beta and support events in time order. Follow an event to its related record; missing history is not evidence that nothing happened.
• Workspaces shows the workspace and its members, plan and activity.
• Generations starts with the failures that need attention. Open a row for the stage, model, user-facing result, diagnostic cause, money and related workspace. Create ticket carries that evidence to the support inbox. Counts use the selected period; a tiny sample does not establish a reliable failure rate.
• Usage, Storage, Costs, Treasury and Reconcile answer different questions: recorded use, stored files, cost assumptions, supplier balances and payment-to-credit consistency. An estimate or a balance is not an invoice. Check the date and scope before making a money decision.
• Treasury keeps an unknown supplier balance separate from a measured zero. If a balance is missing or unreadable, its funding amount and days of cover remain unknown. The total funding plan reads Incomplete, and the float shows only the known balances. Check the supplier account before deciding how much to add; a recorded zero still produces a funding warning when there is spending to cover.
• Refresh balances asks the connected suppliers for new readings immediately and saves them before reloading Treasury. Scheduled checks remain hourly; simply opening the page reads the saved snapshots. Each wallet shows its own attempt time in your browser’s timezone. Readings older than 75 minutes are marked Stale balance and excluded from current funding advice. Failed supplier reads stay unknown with an error; a partial refresh still updates the suppliers that answered. Higgsfield has no automatic balance reading yet. Only an admin owner can refresh; this action does not buy credits or change pricing.

​

Support

The Support → Inbox separates Queued, Being worked on, Blocked and Checking repair, while Errors shows failure evidence. Open a ticket to read its Current outcome, Next action, history and any investigation receipt. A rerun requests a new check; it does not by itself prove recovery or close the ticket. Approval, a tested local repair, a deployment and verification of the original problem are separate steps. For automated-check tickets, verified means the original check passed on current main; it does not claim a customer-facing deployment. Automatic repair and check-related Telegram updates are off. Choose work in Support and ask Codex or Claude Code to handle it. See Support and feedback.

​

Start with Today

Open Admin → Today for the daily view: customer health, tickets needing a decision, and work in progress. It reads the same evidence as Health and Support. An unavailable reading stays unavailable; a quiet page is not proof that every customer flow works. Review customer alerts first, then choose a small batch of tickets to work on with Codex or Claude Code. The business numbers and explanations of the detailed pages start collapsed. The top navigation is Today, Tickets, and Plans. There is no new background agent or polling schedule.

​

One workflow across the admin pages

Support holds the work: customer reports, failed checks and Roadmap tickets. Roadmap orders improvements and reads their ticket statuses; it does not execute work. Health shows live operating measurements. Confidence shows test coverage, failures and missing evidence. Schedule lists background jobs and their costs; a scheduled test is not an AI agent. Work on active customer failures first, especially access, billing, lost work and generation failures. Then address reproducible regressions and blocked customer journeys, followed by planned improvements. Choose Fix it to queue a ticket, ask Codex or Claude Code to start it, and keep its owner, current outcome and next action current. Release a tested fix, verify the original problem, then resolve the same ticket. Keep unverified or blocked work open with a concrete next action. Approval alone never starts a background repair. Test failure reports remove recognized credentials, complete cookie and authentication headers, email addresses and URL query strings before sending diagnostic text to Support. The server repeats this filtering for older reporters. Error codes, source locations and test identities remain available. Existing ticket history and private test artifacts are not rewritten by this protection.

​

Find the next action

Use Today as the starting page, Tickets for the work queue, and Plans for the Roadmap. Health, Confidence and Schedule remain available as supporting evidence under Operations details and from Today. You do not need to visit every page during every work session.

​

Health, Schedule and Evals

Health compares measured values with their thresholds. An unknown or low-volume result needs evidence; it is not a green result. The Community queue reading includes the oldest held item. Failed enrichment jobs says why the jobs failed; Retry failed jobs re-queues them for the next enrichment run. Hours since all full suites reported uses the oldest complete full report across browser, database and extension suites. It warns when coverage is older than eight days and turns red after nine. Full runs are manual, so this means old evidence, not an overdue scheduled job. A targeted run, skipped workflow or incomplete report cannot refresh this clock. A full report can contain failures: freshness is not a passing verdict. Schedule shows each configured job once: purpose, cadence, next slot in Cairo, cost basis, latest evidence and pause control where supported. The top calls out paused or overdue jobs and links to the actual GitHub and Vercel usage meters. Not connected means the infrastructure bill is unavailable here. Planning estimate — not your bill is an optional scenario, not measured spend or promised savings. Execution unverified here and No recorded success describe missing evidence, not proof a job never ran. Full regression runs only on request. Core change-time customer and money gates remain. Balance, deployment drift, watched-brand polling and ranking run hourly; Support runs every fifteen minutes; storage auditing runs weekly. Customer recovery, indexing, billing reconciliation and the daily Health digest retain their timing. Lower cadence delays detection or background updates. Pausing skips work but still invokes the function and reads its pause setting; it does not guarantee zero cost. A read-only administrator cannot pause a job. Evals explains what each AI quality check measures, when it is useful and whether it calls a paid provider. These run on request when models, prompts, seeds or embeddings change. Studio library checks are deterministic and make no provider calls. Read the version, environment, sample, threshold and cost basis alongside the result. No recorded result here means this database has no receipt; local evaluations are not automatically published to production. A failed history read is shown separately. Existing scripts retain their own guards; Studio context now requires explicit opt-in and a dollar ledger, with a $1 ceiling per run. No evaluation starts when you open the page.

​

Confidence and Roadmap

Confidence starts with Needs attention. Open Health for a live problem and Support for test failures. Coverage says a check exists; it does not establish a recent pass or production acceptance. The tiles show machine coverage, current problems and the last full-sweep evidence. The old future score is removed. Failures count once per UTC day per promise over thirty days, even when several tests fail that day. A failure stays unresolved until explicit matching tests pass; elapsed time, unrelated passing tests and skipped work cannot clear it. Multiple failed tests require recovery evidence for each. Older receipts without named test identities cannot establish recovery. The promise tree retains the detailed checks and known limitations, including presentation save/open-time gaps. How to read these checks explains the labels. Roadmap puts Needs your decision first, then the next action in each product area. Open Readiness and all work for its acceptance criteria and ranked to-do list. Both views use current Support ticket statuses; if they cannot be read, the page labels its saved fallback. The criteria remain human-reviewed, and a Bar not re-checked note identifies stale readiness. The ticket carries progress and approval; this view does not approve a price, database change or release. How readiness is decided holds the north-star definition and shared rules.

​

Beta and go-to-market

Beta follows the funnel from invitation through activation and feedback, with commitments, the waitlist and a participant timeline. Delivery and bounce receipts attach to the email they describe. Internal/test workspaces stay separate from the customer metrics. A recorded automatic first touch is not necessarily a human reply. Its Emails tab shows every email Memo can send, in three groups (Beta, Account, Sign-in), rendered by the exact template and wording the sender uses, with sample data; it reads the deployed code on every load, so what it shows is what goes out. Contacts opens the person’s file in a drawer. Forms carries the agreed feedback questions. Content and Social track planned work; a calendar entry does not mean a post was published. Keep customer details in the authorized admin workspace and never paste them into a public report.

​

When access or evidence is missing

If a control is absent or a request is refused, check the signed-in administrator role before retrying. Do not work around a permission boundary. If a card has no current receipt, inspect its linked job or support ticket and leave the status unknown until the required evidence exists.