# Product

## Platform

Web.

## Stack

A DeepSeek Harness Host + Web Client plugin. The current Beta uses native JavaScript, React ModuleLoader, Node.js, and Chrome DevTools Protocol; it adds no runtime dependency.

## Users

Primary users are AI product managers, designers moving into product, independent builders, and small-team leads who use DSH or Codex to generate or change local Web products. When an Agent says the work is complete, they still need to decide whether it is deliverable.

## Product purpose

DSH Visual Acceptance turns scattered browser checks, evidence, state coverage, human decisions, and post-change retests into one traceable local acceptance loop. Success is not producing a report: it is allowing a person to confirm the target and scope, decide what may be changed, and retest under the same conditions.

## Positioning

Vision tools answer “what can an Agent see?” This product answers “what was checked this time, what is unacceptable, where is the evidence, who decides, and did the change pass a retest?” Its core is Target, Matrix, Evidence, Issue, Decision, and Retest—not OCR, a VLM, or a composite score.

## Operating context

- It runs in the DSH session-level `conversation.view`.
- The current session workspace is the storage and path-permission boundary.
- The user chooses Quick, Standard, Page-breakpoint, or Custom coverage. The generated Matrix is visible before a Run and remains capped at eight Checkpoints.
- Custom Checkpoints may use six bounded state actions. Arbitrary JavaScript and inferred states remain forbidden, and the UI explicitly warns against putting credentials in the local Matrix.
- The Host runs a one-off local Chrome check and releases Chrome and temporary directories when the Run ends.
- A visual Provider is optional. Runtime facts remain available without one.
- Confirmed problems can produce a local read-only change package. The user copies it to an Agent; there is no automatic session injection.

## Capabilities and constraints

- V0.1 is limited to local HTML, localhost Web products, and UI prototypes.
- Runtime facts, visual facts, AI experience judgement, and product judgement must remain separate layers.
- `unreached` can never count as passing.
- A target project cannot be changed without human confirmation.
- The product does not recreate undifferentiated OCR, grounding, crop, or color-picking. Pixel comparison is limited to original-Matrix Retest evidence and remains a Candidate.
- The core stays Harness-neutral. DSH is the first delivery surface; the Codex Skill is deferred.
- DSH changes quickly, so the Adapter must stay thin and removable.
- The closed-loop Beta delivers Candidates, project-level Issues, append-only Decisions, read-only change packages, immutable-Matrix Retests, and human retest confirmation. It does not claim the full V0.1 is delivered.
- The `0.1.2` source candidate adds responsive coverage plans, readable-CSS breakpoint discovery, basic accessibility/reachability signals, screenshot-content validation, Retest Diff, and controlled state recipes. CSS viewports are not real devices.

## Brand commitments

- Product name: **DSH Visual Acceptance**; Chinese UI label: **视觉验收**.
- It follows the DSH Operate style: quiet, dense, and task-oriented—not a marketing landing page.
- Copy distinguishes facts, candidate findings, and human decisions.
- A single overall score never hides individual problems.

## Evidence on hand

- `docs/spike-report.md`: DSH Host, Chrome, Provider Adapter, and Issue-retest spike evidence.
- Local `artifacts/`: run evidence and screenshots. They are not published in the public repository because they can expose workspace content.
- Tested DSH installation: `0.1.1-rc.2`.
- The plugin has been installed locally with `link:` and was observed displaying the **Visual Acceptance** tab.
- `0.1.1-closed-loop-beta` completed a fixture closed loop in real DSH Web: one Candidate was explicitly promoted to an Issue and approved for change; a local change package was generated; an original-Matrix Retest was then human-confirmed as `unresolved`.
- The `0.1.2` responsive Spike and linked-source DSH Web session exercise passed on Chrome `151.0.7922.174`, covering breakpoint discovery, quality facts, nonblank evidence, same-Matrix Diff, controlled state replay, redacted action evidence, and 390px layout. This is not marketplace-release approval.
- There is no real user-project dataset, subjective visual-judgement evaluation, or published production case. None may be implied.

## Product principles

1. Deterministic problems are answered by browser and rules, not AI guesses.
2. Candidate results, deterministic facts, AI judgements, and human decisions remain separate.
3. Every conclusion binds to the correct target, coverage, impact, and decision owner.
4. Local first; reuse the host; Providers are replaceable.
5. Retesting under the same conditions matters more than a one-off report.

## Accessibility and inclusion

The workbench must support keyboard focus, light/dark appearance, zoom, and a 390px narrow viewport. Status must not depend on color alone; loading, empty, error, and disabled states require readable text.