Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++
4.3 KB
Raw Blame

Strata UI fixes — handoff for Claude Code

This package comes from the Strata UI audit of 7 October 2026: 177 findings across 210 screen states. Here they are split into 19 tasks. Each task file lists the findings it covers, the files to change, and how to check the result.

What's inside

  • tasks/ has one file per task, ready to give to a subagent.
  • findings.json has all 177 findings with severity, state, issue, suggested fix and assigned task.
  • designs/ holds the reference designs for T09 (Model) and T10 (Engine v3). They are design-component source files: read them for structure, copy, colours and behaviour. They won't run on their own.
  • .claude/agents/strata-fixer.md is a subagent definition (Sonnet, focused on one task).
  • PROMPTS.md has the prompts to paste into Claude Code.
  • REVIEW.md is the checklist for the review after each phase.
  • PROGRESS.md is the log that each task appends to.

Setup

  1. Copy this folder into the repo root as audit-fixes/ (so paths read audit-fixes/tasks/T01-…).
  2. Copy audit-fixes/.claude/agents/strata-fixer.md to .claude/agents/strata-fixer.md in the repo.
  3. Make sure docs/evidence/audit-2026-10-07/ (the screenshots) is in the repo, because the task files point to it.
  4. Run npm install, then npm run check and node --test tests/ once. Note anything that already fails, so it isn't blamed on a task.

Order

Phase 1 fixes shared components, so later tasks get smaller. Tasks in the same phase with no dependency between them can run in parallel on separate branches.

Phase 1 — root causes (shared components)

Task Title Findings High / Med / Low Depends on
T01 One model, one name 11 6 / 2 / 3 —
T02 Overlay layering and dialogs 5 2 / 3 / 0 —
T03 Primary, disabled and on states 11 1 / 7 / 3 —
T04 One error component, no fixture text 10 0 / 9 / 1 —
T05 Show numbers when they matter 6 2 / 2 / 2 T01
T06 Fix the response hover layout 1 1 / 0 / 0 —
T07 Hide what is not built 2 0 / 1 / 1 —
T08 Engine and Model in the main nav 1 1 / 0 / 0 T01

Phase 2 — rebuild Model and Engine

Task Title Findings High / Med / Low Depends on
T09 Build the single-model Model page 17 0 / 9 / 8 T01, T03, T04
T10 Build Engine v3 17 2 / 9 / 6 T01, T02, T03

Phase 3 — remaining findings per screen

Task Title Findings High / Med / Low Depends on
T11 Chat: remaining findings 38 0 / 20 / 18 T01, T05, T06, T07
T12 Time-skip: remaining findings 6 0 / 3 / 3 —
T13 Serving: remaining findings 19 0 / 6 / 13 T04
T14 Monitor: remaining findings 2 0 / 1 / 1 T05
T15 Skills and Tasks: remaining findings 9 0 / 2 / 7 T02, T03, T04
T16 Settings and Setup: remaining findings 7 0 / 2 / 5 T02, T09
T17 Navigation and overlays: remaining findings 4 0 / 2 / 2 T02, T08
T18 Preview panel: remaining findings 7 0 / 3 / 4 —
T19 Dark mode: remaining findings 4 0 / 1 / 3 T02, T03

Running it cheaply

  • Do high and medium first. Every task file puts low findings under "Optional polish". Skip them in the first pass. That cuts about a third of the work.
  • Use a fresh session per task. Each subagent starts clean with only its task file, so context stays small.
  • Check with tests first, and use screenshots only for the states listed in the task.
  • After each phase, run one review (see REVIEW.md) instead of reviewing after every task.

Rough cost at API list prices (Sonnet 5.5, medium effort, for the subagents; one review per phase):

  • Phase 1 (8 tasks): $10–20
  • Phase 2 (2 large tasks): $8–16
  • Phase 3 (9 tasks): $10–20, or about half that if you skip the low findings
  • Reviews: $8–15

Ground rules for every task

  • Stay within the task's scope. If a fix needs a file that another task owns, make the smallest change and note it in PROGRESS.md.
  • Don't change the design kit tokens in src/styles/kit/ unless the task says so. Run npm run sync:kit if you do.
  • Keep both themes working. Check light and dark.
  • Existing tests must stay green. Add tests for new logic (formatters, clamping, fit rules).
  • A finding can be closed as "already fixed by Txx" or "not reproducible", with a one-line reason.