Taskdeck-first adapter sketch

T5 #286, T1 #282, product #2901. Proposal only; browser execution is NOT RUN.

Verified source map

Inspected Taskdeck commit 622d9d820f48e68ac7fa77d94c9ca2536155ddf8:

frontend/taskdeck-web/tests/e2e/workspace-overhaul.spec.ts: keeps Home capture and saved thinking across all experience combinations; retains separate approval and apply gates while switching experiences.
tests/e2e/support/authSession and boardHelpers are used by that specification for synthetic auth and board setup. Its local setup and seedCard helpers are not exported adapters.
frontend/taskdeck-web/playwright.config.ts provides TASKDECK_E2E_DB, TASKDECK_E2E_WORKERS, API/frontend origin resolution and project chromium; default traces retain failures only.
Taskdeck AGENTS.md names an available project browser controller for interactive reproduction (default Chrome DevTools) and Playwright for durable regression. Do not impose a second controller on one session.

Source links: specification, configuration. Revalidate on the actual candidate; current source is not an execution receipt.

First local deterministic baseline

From an authorized isolated Taskdeck checkout, after its documented dependencies/browser/backend setup, use a run-owned synthetic database and verify no live provider credentials or personal services are reused. An example command, not executed here:

Set-Location frontend/taskdeck-web
$env:TASKDECK_E2E_DB = Join-Path $env:TEMP ('taskdeck-ux-' + [guid]::NewGuid().ToString() + '.db')
$env:TASKDECK_E2E_WORKERS = '1'
npx.cmd playwright test tests/e2e/workspace-overhaul.spec.ts --project=chromium --grep 'keeps Home capture and saved thinking across all experience combinations' --trace on --workers=1

Inspect the repository’s server-reuse and provider setup before running; a new database environment variable does not prove an already-running server uses it. Refuse an unknown/non-isolated server. Trace-on is a local evidence override so successful controls are retained, not a proposed CI change. The command exercises an existing test, not the new JSON scenario runner. Do not imply it executes every assertion in the draft pack.

Proposed adapter boundary

Future harness-owned module home: ux_evaluation/adapters/taskdeck.py, after T1 semantics are accepted. No such module is shipped here. Interfaces are a design sketch, not callable APIs:

bind(pack, product_checkout, capability_receipt) -> BoundScenario: resolve exact source/fixture/controller/environment; no execution on mismatch.
observe(bound_scenario, run_root) -> ObservationManifest: call only reviewed project operations and collect per-step evidence.
validate_observation(manifest, run_root) -> EvidenceValidation: deterministic completeness/digest/identity checks, no quality judgment.
judge(validated_evidence, rubric, budget) -> AdvisoryFindings: model-independent result contract, no publication.
prepare_issue(findings, existing_issue_index) -> DraftIssue: deduplication and sanitized Markdown/JSON, no automatic filing.

Harness owns binding/evidence/adapters. Taskdeck owns actual fixture helpers, product invariants and any later test changes. claude-config owns invocation recipes. Do not export product internals or rewrite app code just to satisfy this sketch.

Coverage and limits

The example binds UX-01 from #2901: preserve the unsaved Home thought while switching experiences, then explicitly save. Existing source asserts preserved text and later durable thinking; the draft adds explicit counter requirements for zero unwanted capture/proposal/provider/approval/apply effects. Those counter bindings are missing work, not proved coverage.

A follow-up covers the review/apply journey using the second existing test. UX-02 moderated comparison, the full 4×3×2 shell invariant matrix, renderer/Auto resolution and UX-04 observation import/export remain owned by #2901; one pilot does not complete them. Rotate experience order for preference studies, retain raw outcomes and never declare a winning experience from this synthetic pack.

Before T2 runs, record capability probes, exact fixture/reset recipe and counter sources. Collect a known-good control plus one seeded defect in a disposable fixture; do not deliberately mutate product main. After a real finding is reproduced, the product repo gets the smallest deterministic regression PR. Keep .github, branch protections and subjective CI gates untouched.