Lane 1 · Web channel

Workflows / Runs UI & UX

Finish the new Workflows/Runs experience: keep the simplification and progressive disclosure that landed in Program 3, fix the one deterministic UX bug that UAT will trip over in the first five minutes, and decide the mobile/PWA scope explicitly instead of implicitly.

5/7

session claims verified in code

~400ms

empty-card flash on every warm prompt

PWA assets (manifest / SW)

5×

React #185 recurrences fixed point-wise

What actually landed (verified against origin/main@734eff41)

Claim	Status	Evidence
IntentSuggestion explicit UI phases `pending / unavailable / quiet / suggested / run_summary` + trace hooks (`data-workflow-suggestion-state`, `window.__winWorkflowSuggestionTrace`)	LANDED	`IntentSuggestion.tsx:114,139–154,587–604`; 8 state tests
Suggester conversational contract: `decision: suggest \| quiet \| clarify_only` + `quietReason`, confidence floors 0.6 / 0.65	LANDED	`tmux-suggester.ts:38–40, 1120–1179`; `types.ts:327`
Layout grammar: 720px centered thread column, `--gutter` token	LANDED	`WorkflowCanvasPage.tsx:721,736,832`; `globals.css:127,244`
Frozen “Applied receipt” setup during a run (no Edit-setup escape hatch), richer run-detail hero (phase, status, quick actions)	LANDED	`WorkflowRunSummaryCard.tsx:84,126,207`; `JobDetailUnified.tsx:271–442`
Responsive layer: files drawer below `md`, jobs grid reflow, dvh-relative artifact iframe, reduced-motion provider	LANDED	`files/index.tsx:147`; `JobsTable.tsx:37,76`; `LiveArtifactShell.tsx:178`; `MotionProvider.tsx`
“Monitor run” label + command rail (claimed by PR #655’s completion message)	NOT FOUND	No “Monitor” string or rail component in `packages/ui/src`; only “Open live run” link + hero quick-action links
PWA / installability	NOT FOUND	Viewport meta only (`ui/index.html:5`). No manifest.json, no service worker, no theme-color

Nothing above is hosted-validated Everything marked LANDED is merged but running nowhere the user can see: staging/beta still serve runtime 0d5cbb6a while main is at 734eff41. Every UI claim on this page needs its hosted proof row after the next deploy (see the Deploy tab).

The suggestion-card flash — root cause confirmed

Your report (“empty suggestion card instead of loading indicators”) is a deterministic design artifact, not a flaky race. On the common path the worker is already warm before you finish typing, so on commit:

status="ready", bundle=null, decision=null — and isPendingSuggestionStatus() (IntentSuggestion.tsx:135–137) only treats warming/busy/retrying as pending.
The request-key debounce (useComposerSuggestion.ts:962–968, default 400ms, unoverridden at the only call site useWorkflowCanvasPage.ts:1118–1132) hasn’t queued yet.
So suggestionUiPhase() falls through to "quiet" and renders the “No suggestion — looks specific enough to run as-is” card for ~400ms, then flips pending → suggested.

Worse: IntentSuggestion.test.tsx:73–79 asserts this quiet-card render as intended for ready + null bundle, so the test suite actively defends the bug. No test exercises the ready→busy transition timing.

Plan of attack

Implementation worksheet0/8

RUX-1Fix the flash: treat ready + bundle=null + decision=null (or settledRequestKey !== currentRequestKey) as pending, not quiet

Update the intent test that currently locks in quiet-card behavior; add a fake-timer transition test (warm → commit → advance < debounceMs → assert pending chrome). Files: IntentSuggestion.tsx, useComposerSuggestion.ts + both test files.

RUX-2Reconcile the run-card affordance: implement “Monitor run” + command rail as #655 claimed, or normalize on “Open live run” and correct the docs

The session’s completion message and the code disagree — decide the label once, grep-verify, and update PR-655 work record.

RUX-3Decide PWA scope explicitly: minimum viable = manifest.json + icons + theme-color (installable shell); defer service-worker/offline

Greenfield either way. If deferred, strike “PWA” from UAT checklists so it isn’t rediscovered as a regression. A manifest + icons is a <1-day slice; offline caching of an SSE/job app is a real project — don’t conflate.

RUX-4Guard the React #185 pattern generically (5 prior incidents): lint rule or shared test helper for poll-fed effect deps without stable identity

Pattern: poll returns fresh array/object → useEffect dep → setState loop. Point fixes exist (useWorkflowCanvasPage.ts:1961, useWorkflowSessionState.ts:725) but nothing prevents the sixth recurrence.

RUX-5Boundary-test the suggester decision contract: table-driven cases at 0.59/0.60/0.61 and 0.64/0.65/0.66 per decision type

tmux-suggester.ts is 3,502 lines with only ~4 direct decision-value assertions. This is the brain of “field any query” — it needs contract-grade tests.

RUX-6Hosted-validate the full landed set on beta after deploy: suggestion trace order, applied receipt, run hero, responsive checkpoints (375/768/1280)

Use the suggestion trace hooks (__winWorkflowSuggestionTrace) that #645 added — they exist precisely for this. Capture ledger rows per hosted-validation-runbook.md.

RUX-7Add responsive/viewport regression tests (none exist): scrollWidth ≤ innerWidth at 375, drawer vs rail visibility, card-list vs grid

Today the responsive layer is CSS-class-driven with zero automated coverage; the “pass at 375” definition already exists in the GOAL doc — encode it.

RUX-8Progressive-disclosure audit pass: walk hero → suggestion → running → delivered as a first-time user; every advanced control must be one disclosure deep, nothing mission-critical hidden

Use the frontend review exercises already embedded in docs/plans/pre-uat-and-prod-deploy/plan.html (#worksheet) — simplicity, disclosure, hierarchy, polish, safety, mobile.

Query-coverage matrix (can users field anything?)

Entry-point truth, corrected from the session’s own description: /api/workflows/:slug/test-run is for draft workflow testing only. Ad-hoc runs and applied suggestions both go through /api/me/jobs (employee) or /api/run (admin console), distinguished by whether workflowSlug is set — see useWorkflowLaunch.ts:199–259. Don’t build UAT scripts around the wrong split.

Query shape	Path today	State
Simple ad-hoc task, no config	Composer → run-as-is → `/api/me/jobs` (ad-hoc fallback preset)	LANDED — hosted spawn proven on beta pre-drift
Task matching a published workflow	Composer → daemon suggester (`decision=suggest`) → apply → run	LANDED — suggester retrieval moved server-side (#555); hosted quality spot-checks never run
Ambiguous ask	`decision=clarify_only` → clarification UI state	LANDED — needs hosted proof + copy review
Specific ask, no workflow fits	`decision=quiet` → run-as-is encouragement	OPEN — currently doubles as the flash bug’s wrong state (RUX-1)
Daemon unavailable	`/warm` soft-degrade 200 + `unavailable:true` → retry chrome	LANDED (#554) — hosted /warm proof still an open gate

Agent handoff prompts

prompt · RUX-1 flash fix

Fix the deterministic empty-suggestion-card flash in the WIN workflow canvas.

Context: on prompt commit with a warm suggester worker, useComposerSuggestion has
status="ready", bundle=null, decision=null for the ~400ms request-key debounce window
(useComposerSuggestion.ts:962-968; debounceMs default 400, call site
useWorkflowCanvasPage.ts:1118-1132). IntentSuggestion.isPendingSuggestionStatus
(IntentSuggestion.tsx:135-137) treats only warming/busy/retrying as pending, so
suggestionUiPhase() renders the "quiet" no-suggestion card during the window, then flips
pending → suggested. IntentSuggestion.test.tsx:73-79 currently asserts the wrong behavior.

Do: (1) make the debounce window render pending chrome — gate on unsettled request key or
add the ready+null-bundle+null-decision case; (2) update the locked-in test; (3) add a
fake-timer transition test (warm → commit → advance < debounceMs → assert pending, advance
past resolve → assert suggested); (4) run bun run --filter @win/ui test + typecheck; keep
the change surgical — no refactor of the debounce itself. PR per repo rails (specific git
add, PR record docs/github/prs/PR-<n>.md, work-record untested list includes hosted proof).

prompt · RUX-3 minimal PWA shell

Add a minimal installable-app shell to packages/ui (WIN dashboard). Scope strictly:
manifest.json (name WIN, short_name WIN, display standalone, theme/background colors
matching globals.css tokens, maskable icons 192/512 generated from the existing logo),
<link rel="manifest">, theme-color meta, apple-touch-icon. NO service worker / offline
caching in this slice (SSE-heavy app; defer intentionally and say so in the PR body).
Verify with a Lighthouse installability pass locally and document the deferred offline
story in the PR record. Wire nothing else.

WIN launch workbook · generated 2026-07-06 by Claude (Fable) from session mining + repo audit of origin/main@734eff41 · worksheet state persists in your browser (localStorage, key prefix winwb:runs-ux)