[Divii2205]

Summary

Four small fixes to the live dashboard, one commit each:

1. Agents page as cards, and no prior shown as a record. The live Agents page was a six-column table where names, addresses and model ids wrapped mid-word. Each agent is now a card on the same grid as the registry page (3, 2, then 1 across): name, id and model; ENS name and verdict, with the reason when it is not verified; owner and receipt-signing key; and the validated record. The record column used to show the prior ("50% pass, about 0.50 USDC a task … (0 validated)") while the note under it says no figures are shown until an agent has validated outcomes. Only categories with counted_outcomes > 0 are shown now; otherwise "No validated outcomes yet".
2. The new-run form is kept while visiting other pages. Typing a goal, opening another tab and coming back emptied the form. The goal, repository, branch, budget and kind of work are now kept in this browser tab (sessionStorage), keyed on the signed-in address, and dropped once the plan is drafted or "Clear the form" is pressed. "Start again" still fills the form from the earlier plan instead. A stored draft that is malformed, too long or names an unknown kind is ignored.
3. Each kind of work says what it means. The chips (Fix failing tests, Fix a bug, Security, Performance, Refactor, Docs) now have a one-line meaning, shown under the chips for the one picked and as a tooltip, plus what picking a kind changes: the tasks the planner drafts (planner.py puts the category in its prompt), which agents are suggested (routing reads agent_posteriors for the plan's category) and which record the result counts toward.
4. Nav tabs have a name when only their icons show. Below 1100px shell.css hides .tab-label, which left each primary nav tab as a link with no accessible name. Each tab now has aria-label and title.

Type of change

• paper — correction, derivation, or new section in the research
• docs — requirements, README, or other prose
• feat — new functionality
• fix — bug fix
• contracts — Solidity
• chore — dependencies, tooling, CI

What I verified

• Verified:
  • npm run lint, npm run format:check, npm run build and npm test (526 unit tests) pass on the branch, rebased on 66b80a5.
  • npx playwright test --project=live: 118 passed, 2 skipped (the two that need the WASM verifier built).
  • The two middle commits type-check on their own (tsc -b).
  • New tests: newRunDraft.test.ts (9 cases: restore, per-address keys, cleared when empty or after drafting, private mode, malformed or oversized drafts, unknown kind); e2e for typing, leaving for Agents and coming back, Clear the form, reload, and an empty form after drafting; e2e that the kind hint follows the chip picked and is the group's accessible description; e2e that an agent whose rows all have counted_outcomes: 0 shows no figures.
  • Screenshots below are from the live build against the e2e fakes, at 1440×900; I also checked the cards and the new-run page at 390×844 in light and dark.
• Not verified:
  • Not checked against the deployed app or real mainnet ENS data, only the e2e fakes.
  • The mock (sample-data) project's two landing-page scroll-story tests time out when the whole suite runs at once on this machine; they pass alone, on this branch and on main. They do not touch these files.
  • Not tested with a screen reader; the nav and hint changes are checked through accessible names and descriptions in Playwright only.

Checklist

• One concern per PR — this change does one thing (small live-dashboard UI fixes, one commit each)
• No secrets committed (no private keys, API keys, .env files, mnemonics)
• Claims in the diff are supported by something a reviewer can check

Evidence

Agents page: before / after (same fake data: coder-a has only prior rows, coder-b has 7 and 3 validated outcomes)

Before:

After:

New run: after typing, going to Agents and coming back. Before / after

Before:

After:

Breaking changes / follow-ups

None. Test ids record-<id> and ens-<id> are unchanged; the overview e2e now looks for agent cards instead of table rows.