[junixlabs]

Eleven commits on top of main (5fffe80). They get the gate passing again and implement the roadmap set on 2026-10-03: knowledge, process, evidence and versions, and experience validation at the centre. The last part is an experiment that scores agents on the walk.

Stabilise

• Map contrast fixed. The map page's muted ink was 4.22:1 against its background, which failed the gate's own-pages check on main.
• English output. Every sentence the instrument prints is now English, and the example data matches the code.

Knowledge

• Six topic pools, 64 viewpoints, from the source repositories of GOV.UK, USWDS, Carbon, Primer, Polaris, Mailchimp and the W3C. Every quote was fetched and checked verbatim.
  • The pools: colour (10), writing (10), data display (12), forms (12), navigation and IA (10), motion and feedback (10).
  • The lenses now carry 187 viewpoints.
• uxcli template for workspace, shop, landing, content and mobile-app. Every actor line and research line is a question. apply only ever creates files and records which template was used.
• context show now names the template, the lens for each kind of screen, and which lens each screen was reviewed against. It also shows, per screen, whether it has been drawn and picked yet.
• references/research.md: how to research a product's field before designing for it.

Experience validation

uxcli experience [--page] reports, per workflow:

• steps, clicks, characters typed and scrolls;
• response time, as the first visible change after the action;
• a keystroke-level estimate;
• a wait with no progress signal;
• a control the step needs that is below the fold;
• an answer asked for twice;
• a step that did not reach its state;
• recovery after an error (typed answers kept, error announced);
• navigation consistency between steps.

Every item is a finding with the source of its threshold, pinned on the step's screenshot.

Observer change: the observer now measures the first visible change after an action. If nothing has changed, it waits up to 2.5 s for one instead of calling the page settled at 0 ms. During a wait longer than 1 s it records whether a progress signal was shown.

Privacy: a typed value is stored only as its length and hash. Schemas trace and experience enforce this.

Evidence and versions

• uxcli version save <journey> <name> names one walk. It creates a file and never rewrites one, and the walk is pinned so pruning keeps it.
• uxcli experience --from=<a> --to=<b> compares two versions step by step.
• Proposals accept a new kind, redesign, citing the walk as evidence.

Walkthrough

uxcli walkthrough <journey> [--as=<actor>] | check runs the four questions of Wharton, Rieman, Lewis & Polson (1994) at every step:

• Each answer is given from the step's screenshot, optionally as an actor the project wrote.
• check refuses a yes that the walk contradicts.
• Every no and unsure is a finding.

Studio

uxcli studio [--serve] [--shot] shows the project as a board laid out like a design file:

• Each step's screen as drawn, as built (with findings pinned) and in each named version.
• Pan and zoom, an inspector, and Context and Library tabs.
• In serve mode, a person can pick a variant, ask for a redraw, or leave a note that is written as a redesign proposal. The page refreshes as files change and serves nothing outside .uxcli/.
• Two versions can be compared on the board.

The experiment

Ten fresh Sonnet sessions per arm, scored with one instrument:

Arm	Passed	Median findings
ticket alone	0/10	2
with the skill	8/10	1

• The two failing skill sessions never read the journey's card. uxcli run <url> now names the journey a screen belongs to and points to context show.
• The draw arm removes the lead screens' drawings, so the agent should design first:
  • Of 20 sessions, 0 drew a variant and 20 built the page anyway; 19 said they had skipped the design step.
  • This held before and after context show started printing BEFORE YOU BUILD.
  • The library's exit criterion is measured and not met. The owner needs to decide whether the instrument should hold the work to a pick and a review.

Validation

• uxcli gate passes locally.
• New suites: template, experience, version, walkthrough, studio (this one also runs the page probes on the board), plus additions to brief.
• CI was green on every earlier commit in this PR; the run for the last commit was still pending when this description was written.

Known limits

• Unloaded URLs. The public URLs in the pools were derived from each repository's routing, not loaded, because the design-system sites were not reachable from the sandbox.
• Unvalidated metrics. The experience metrics have not been validated on unseen journeys, so they report findings only.
• Small experiment. One model and one ticket.

🤖 Generated with Claude Code

https://claude.ai/code/session_011TUCUpbzmTpBE6ND7eHfzx