Changelog
What this prototype changes against the dashboard, written as proposals for Martin and Stefan. Each row says what is there today and what the change would take.
2026-11-06 · compared with augapro at 29831baa, branch feat/agent-health
Proposed9ready to look at
To decide25a product or design decision first
Prototype only2nothing to do in the dashboard
Built in the dashboard0nothing here is merged
Naming
| No. | The change, and why | In the dashboard today | What it would take | Status |
|---|---|---|---|---|
| 1 | Compliance in the menu and the breadcrumbSeveral laws apply, and the EU AI Act is one of them. The place should not carry the name of one law. | The rail says AI Act, the wide rail and the breadcrumb say EU AI Act. | In copy/nav.ts, set both short and label of aiact to Compliance, and update docs-site/docs/dashboard/navigation.md. Keep the route /ai-act: the tenant's Caddy sends /compliance to the engine (LENS_PATHS in auga/tenant/caddy.py). Keep AI Act everywhere else: the page title, the article labels, the palette's hint and the Mandates panel. The rail needs room for the word: at --sh-rail-w: 6.5rem with a scrollbar showing, Compliance broke into two lines here. The prototype uses 7rem and a thin scrollbar, and lets a word stay whole (Shell.css). | Proposed |
| 2 | Coach in the menu, Coach your agents in the breadcrumbCoach names the area as a whole, and reviewing a decision is a task in it. It is a working name, not settled yet. | The rail says Review, the wide rail and the breadcrumb say Decision review. | In copy/nav.ts, set short of review to Coach and label to Coach your agents, and update docs-site/docs/dashboard/navigation.md. The route /review can stay, and nothing stops /coach either. Keep decision review as the name of the task: the page title, the facts and the links use it, and it pairs with claim review, the other inbox. | Proposed |
| 3 | The menu follows a manager's questionsFirst the agents, then what they may do, then what can be shown, then what should change. Agents moves to Inspect, Reports stays with Compliance, and What-If and Improve get the new group Optimise. | The groups in copy/nav.ts are Inspect, Govern, Comply and Operate. Agents is in another group there, and Mandate Simulator has no entry. | Decide whether the dashboard takes the same order and the group Optimise. | To decide |
The dataset's page
| No. | The change, and why | In the dashboard today | What it would take | Status |
|---|---|---|---|---|
| 1 | The dataset in the header opens the dataset's pageThere is no way today to look at a dataset itself. An eye says the fragment can be opened. | The header shows a picker for the dataset, or the name alone when there is only one. Nothing opens. | In Shell/Header.tsx, make the eye, the label and the name one link, with the chevron beside it as the picker. The page needs a route of its own. | Proposed |
| 2 | A new page for the dataset, with Traces & SpansWhat is in the file: traffic, traces, services, attributes, whether there is enough data. Auga and Exchange are not on this page: they have places of their own. | No such page. The Dataset Overview prototype has it, with four tabs. | Build the page from the prototype, without the Auga and Exchange tabs and the settings popover. The prototype computes its figures in the browser, and a summary route does not exist yet. | Proposed |
| 3 | Bundle Analysis only when a bundle was uploadedThe training ground and fine-tuning area: run a simulated bundle with a root cause hidden in it and see whether Auga finds it. The answer stays sealed until the person asks for it, so they can look at Auga's findings first and then compare. Open, the page leads with the root cause and what was found, and it does not repeat the panels of Traces & Spans. It replaces the name Proving Ground, which the docs already use for a hosted tenant. | No such tab. The Inbox reads a bundle and shows its report, and the dataset list carries a score. | Show the tab when the dataset came from a bundle. score in the dataset list is not enough: it only says there is an answer key, and a bundle's ground-truth.json may hold none (Places/ValidationReport.tsx). Then the page must say so. The upload's source needs to be kept with the dataset. Sealed means the whole answer: the root cause, the Planted and Verdict columns, the chart markings, the scenario and the score. In the product the answer should not reach the browser until it is asked for. Here it sits in the page source, which a static page cannot avoid. | Proposed |
| 4 | The page opens after a successful uploadThe upload ends where the person wants to look next, with a notice saying what was imported. | The Inbox shows the state, a toast says it is done, and the actions are Clear or Reject. | In Places/Inbox.tsx, open the page when an entry reaches done, and add Open the dataset to the row. Add Open to the table in Places/Datasets.tsx. | Proposed |
New places
| No. | The change, and why | In the dashboard today | What it would take | Status |
|---|---|---|---|---|
| 1 | Improve in the menu, under GovernWhat changed in the agents, whether it helped, and what to change next. | Not in nav/registry.ts. | A product decision first. Much of what the page stands on is marked Not built or Partly on the page itself. | To decide |
| 2 | Intelligence Exchange in the menu, under OperateThe first screen for what a tenant shares. | Not in nav/registry.ts. The prototype notes that the route exists, and that no registry is running. | A product decision first. | To decide |
| 3 | Connections in the menu, moved out of SettingsHow Auga is connected and the ways in: terminal, API and agents. | A tab in Settings, /settings/integrations, with the table of parts only. | A new place in nav/registry.ts, and the tab removed from Places/Settings.tsx. | To decide |
| 4 | Agent Health reached from an agent's profile, Preview a stage as a view of MandatesBoth belong to a place that exists, so neither takes a place in the menu. Preview a stage answers the question Mandates raises, what the next stage would do, and it sets nothing. It is not part of What-If: What-If estimates money from assumptions, the preview counts calls in a real file. What-If links to it for the number of reviews. | Agent Health is not reached from the dashboard. The Mandate Simulator is at /simulator with its own entry under Govern, in PR 122, not merged. | Done in PR 122: the route /mandates-preview (not /mandates/preview, where that segment is an agent), no rail entry, a view switch Stages and Preview a stage on both pages, and the agent from Mandates chosen when the preview opens. | Proposed |
| 5 | Plans, a tab under SettingsA page for what a tenant has and can add: Starter, Platform, Enterprise and the modules. The plans, prices and what runs at the customer's side come from the draft plan for the hosted service, v0.3 of 2026-11-09. The rest of the wording comes from the earlier pricing page. | No such page, and nothing in the dashboard knows about a plan. | A product decision first. The prices are a draft. To work for real it needs the tenant's plan and modules from somewhere the dashboard can read, which the draft plan puts in a control plane with billing. | To decide |
| 6 | Upgrade on Bundle AnalysisA way from the training ground to the plans. It opens Plans with the module marked. | No such button. | Decide what Upgrade means here: the page is shown in full, so the button sells nothing that is locked. If the module is paid, decide what a tenant without it sees. | To decide |
| 7 | What-If, a new place and moduleA calculator for all of the organisation's agents, because several agents can together harm or help the company. You adjust the assumptions, simulate the outcome and take out reports for top management. It overlaps Cost Optimization, so the two may become one. | Nothing at the level of the organisation. On feat/agent-health, the Value sheet in the engine's health report (auga/health_ui/value.js, auga/health_value/) prices one report at a time. | A running prototype exists, with its own model, and is frozen here in Huvudbok with a board report. Decide whether it is a place in the menu, a module, or both, and whether its model and the Value calculations should become one. | To decide |
| 8 | Agent Health with the Value sheetWhat the report's quantities cost and put at stake: cost, value at stake, value lost per switch with its interval, and every assumption with its source. | In the report on feat/agent-health, in its own blue look. Not reached from the dashboard. | Decide whether the dashboard shows the report in Huvudbok, as here. The figures here use illustrative prices. | To decide |
| 9 | The board report as a spreadsheet, a presentation, a document, Power BI and for a board portalThe group that prepares the meeting works in its own tools, so the report should reach them there. "Spreadsheet" replaces "Excel", since it is not the only tool. | The What-If prototype writes a spreadsheet, and its report prints to PDF. On feat/agent-health, Agent Health has a Power Query and a connection file (lens_service/health_api.py). What-If has no server side, and nothing writes a presentation. | Power BI needs the scenarios saved on the tenant and a route for them. /api/whatif/scenarios on the page is a proposal. A presentation and a document need a .pptx and a .docx writer, and a decision on whether they follow your own templates. A board portal takes the PDF as it is, which is not tried in any portal, so a direct integration can wait. | To decide |
| 10 | The globe beside the verdict, with EnlargeThe verdict and the globe are seen together when the page opens, and Enlarge gives the globe the whole window. The globe covers one trace file, with its agents on the southern half, so it is one globe per report and not one per agent. | On feat/agent-health, the report opens on the globe at full window height in its own blue look, with the report under it. Not reached from the dashboard. | The layout, a dialog and Huvudbok's colours. At half width the mouse wheel scrolls the page, and it turns the globe only when enlarged. | To decide |
| 11 | Systems that receive evidence from AugaAuga should be the place for the risks and incidents of agents, and the company's GRC system keeps the full list. Connections now says so, with an honest status per receiver. | Mandate events go out by webhook: stage changes, revoked mandates and risk acceptances. AUGA_PLAN_TOKEN_CLAIMS.md plans signed JSON and an OSCAL export for GRC. No OSCAL code exists. | Incidents and risks as places, events for them, and a signed evidence package that can be checked without Auga. | To decide |
| 12 | Agents as the agent register, with rolesGovernance starts with a person who answers for each agent. The register shows the roles, the stage now and the sources of every agent, and its history. Only register sources create agents, never uploaded files or bundles, and the register compares declared agents with the agents that Auga sees. An agent page shows the roles, the classification and the history. | Agents are found in the telemetry per dataset, by auga.agent.id. The only owner in the engine is the owner of a risk acceptance. The ledger records decisions, reviews, amendments and actions, not agents. | Register sources, declared agents, roles per agent, and owner changes signed into the ledger. | To decide |
| 13 | IncidentsSomething that went wrong needs an owner, a level of certainty and a way to close it. The three levels keep a change apart from a confirmed harm, which removes the contradiction about wrongful denials. | The engine writes a deviation analysis per run (auga/incident.py), and Coach can mark an outcome as harmful. There is no workflow and no page. The AI Act module spec plans an incident manager. | An incident with owner, level and status, made from the evidence. Events for the ticketing system and a signed evidence package. | To decide |
| 14 | RisksA risk about one agent and a field of work, which Auga can measure, belongs in Auga. Other risks stay in the GRC system. | Risk acceptances exist per agent and field of work, signed and with an owner and an end date. There is no list of risks. | A risk with mechanism, cost, decision, owner, check date and the signal that opens an incident. | To decide |
| 15 | The agent's page as its profile, with Health as a sectionAgent Health covers one time window of a whole workflow, with several agents, so it is not the health of one agent. The agent's page becomes its profile: roles, health, sources and history. Its Health section gives the workflow's status, the strongest change and the agent's own figures, and leads into the report. Agents leads to the section, and the report's breadcrumb goes through the agent. A portrait shows five real runs of the workflow with three.js, using the techniques of the portrait drafts: the floor dips where a step took time, the debt-to-income rule is a ring, and a failure is a red peak. | An agent has no page of its own. Agent Health is a report per trace file on feat/agent-health, not reached from the dashboard. | An agent page first (row 12). Then a Health section that reads the agent's rows from the latest report of the workflow it takes part in. three.js in the dashboard for the portrait, served from the dashboard itself. Decide the report's name: it now covers value, vitals and changes for a whole workflow, so a name such as Health report may say more than Agent Health. The engine and the docs say Agent Health. | To decide |
Signing up, signing in and the first day
| No. | The change, and why | In the dashboard today | What it would take | Status |
|---|---|---|---|---|
| 1 | Sign-up, from choosing a planA person who wants to try Auga starts alone: choose a plan, create an account, check the e-mail, set up the tenant. The draft plan v0.3 makes Starter self-service. | The dashboard only signs in on a tenant that already exists (HostedGate). Nothing creates a tenant. | A service outside the tenant that makes the tenant, its address and its first admin token, and takes the plan. The draft plan puts this in a control plane with billing. The pages for the terms of use are not drawn. | To decide |
| 2 | A code to the work e-mailBefore a tenant exists, the code shows that the address is the person's. | No such step. | A mail sender and a code that works for a short time. The 15 minutes on the page are a proposal. | To decide |
| 3 | The data path is read before the tenant is madeOn Starter, the telemetry does not pass the edge first, so content can reach Auga. Plans says so already. Here the person accepts it before any data arrives. | No such step. The Plans page is new too (row 5 under New places). | Record the acceptance with the tenant. Choose the wording with whoever owns data protection at Auga Labs. | To decide |
| 4 | The admin token is shown once, at the end of sign-upThe first admin signs in with an access token, as today, so sign-up ends by showing the token once. | Tokens are made under People and tokens and shown once. The first token has to come from somewhere. | The service that makes the tenant also makes the first admin token. | To decide |
| 5 | Home on the first dayNo agents, the steps still to do, and where to start: upload a file, or connect the Collector. The end of the trial is shown. | Home already shows the setup steps until all are done, with a count of the steps to go. The token, the trial and the pointer to a file upload are new. | Add the empty state and the pointer in Places/Home.tsx. The trial needs the plan from the service in row 1. | To decide |
| 6 | Sign-in with a way back for a lost tokenThe sign-in stays an access token on the tenant's address. A person who lost the token has no way back, so an admin can get a link by e-mail. | HostedGate asks for an access token and answers unknown, revoked or suspended. Nothing helps a person who lost it. Single sign-on is planned. | A route that sends a one-time link to the admin e-mail and signs the person in once. Decide what a person who is not an admin does. | To decide |
| 7 | A request to talk to us, for Platform and EnterpriseThese plans are sold with a conversation. The request says what the person runs and needs. | No such page. On Plans, Talk to us is a button with no page behind it. | A form that reaches a person at Auga Labs, and a decision on who answers. | To decide |
Huvudbok
| No. | The change, and why | In the dashboard today | What it would take | Status |
|---|---|---|---|---|
| 1 | Move the marked CSS into huvudbok.cssEvery page marks what the design system lacks: 13 named sections, among them links in an island, the select field, chart roles and the stage ladder. | Much of it lives in component CSS, for example Shell.css and Places.css. | Take each section name once. The README says which global rules to leave out, because tokens.css and base.css already set them. | Proposed |
| 2 | Quiet text in the rail reaches 7:1Huvudbok's rule is 7:1 for every text. The quiet rail text measures 6.51:1 in light mode. | --hb-cover-ink-2 has alpha 0.72. | Alpha 0.76 gives 7.04:1. One value in huvudbok.css. It is left unchanged in this prototype, so the copy stays identical. | Proposed |
| 3 | Red on an inked panel head stays over 7:1Red on the slate head measures 5.93:1 in dark mode. The prototype sets a stamp in a head on paper. | The ledger and the boundary set a whole heading in red when the chain breaks or the allowlist is broken. Not measured in the dashboard. | Add the stamp rule from the page section, and look at those two headings. | Proposed |
| 4 | The Coach page in panelsThe other places group their content in panels, so the page would read the same way. | Headed sections directly on the sheet (.mo-section). | A small change in MandateOversight. | To decide |
| 5 | Reports in Huvudbok's panelsThe prototype draws key figures, exports and risks like the other pages. | A frame with the engine's Compliance Lens, in its own toned cards. | A change in compliance_ui.html, or a page of its own in the dashboard. | To decide |
| 6 | The evidence graph in inkHuvudbok has no accent colour but red. The prototype tells grades apart by line weight. | The lens draws edges in green by grade, and lens.css says grade colours keep their meaning. | A design decision: the graph may need its colours. | To decide |
Only in this prototype
| No. | The change, and why | In the dashboard today | What it would take | Status |
|---|---|---|---|---|
| 1 | Static pages with fictional figuresTo show and discuss a screen with no engine running. Buttons and fields do nothing. | Not applicable. | Nothing. | Prototype only |
| 2 | A frozen preview of a stageThe playback needs a script, so one state is shown. The charts in Agent Health are frozen in the same way. | A prototype in Drive, in a blue look of its own. | Nothing, unless they are built. | Prototype only |
Nothing on this page is agreed. The rows describe the dashboard's source at the commit above, not a running tenant.