Files
genshin-assistant/docs/ARCHITECTURE.md
T
AzuTear 639b0b7f59 feat(scanner): add native artifact pipeline
Add native IK-style capture processing, Artifact Inventory, explicit promotion and single-result review. Confirm the three live OCR corrections in the eval corpus and preserve extraction/value separation.
2026-07-09 23:30:42 +02:00

334 lines
21 KiB
Markdown

# Architecture
This document describes the structure, boundaries, flows, and technical rules of the system.
## High-Level Overview
Genshin Artifact Assistant is a local desktop application. Electron owns OS integration, screen capture, IPC, and overlay windows. React owns the interactive UI. Domain logic for OCR parsing, scoring, scanner state, and recommendations lives in TypeScript modules under `src/lib`.
The scanner performance direction is now native-first: the C# input helper owns
the fast capture/click/scroll loop and uses vendored Inventory Kamera
`inventorylists` as the scanner dictionary source. Electron is the process,
IPC, packaging, hotkey, and dev-control shell. React is only the visual control
and status surface for this path.
```mermaid
flowchart LR
User["User"]
Genshin["Genshin Impact Window"]
Electron["Electron Main Process"]
Native["C# Input Helper / Native Scanner"]
React["React Renderer"]
Parser["OCR Parser and Scoring"]
LocalData["Local Snapshot / Future SQLite"]
User --> React
React --> Electron
Electron --> Native
Native --> Genshin
Native --> Electron
Electron --> React
React --> Parser
Parser --> React
React --> LocalData
```
## Architecture Style
| Field | Value |
| --- | --- |
| Style | Desktop modular monolith |
| Primary reason | One local app with tight UI, capture, and parsing feedback loops |
| Main trade-off | Faster iteration now, but scanner heuristics must stay isolated to avoid UI coupling |
## Folder Structure
```text
/
electron/ Electron main process and preload bridge
src/ React app, domain types, parser/scoring logic
data/ Seed or package data
docs/ Project documentation and decisions
dist/ Generated renderer build
dist-electron/ Generated Electron build
outputs/ Packaged app output
```
## Module Responsibilities
| Module | Responsibility |
| --- | --- |
| `electron/main.ts` | Electron app composition, dependency wiring, app lifecycle, hotkeys, and IPC/dev-control registration |
| `electron/appWindowManager.ts` | Main window and overlay window lifecycle, menu-bar removal, dashboard focus behavior, and renderer window command delivery |
| `electron/services/inputHelper.ts` | Stable JSON protocol client for the compiled C# input/capture sidecar plus PowerShell fallback startup |
| `electron/services/inputHelperPowerShellFallback.ts` | PowerShell fallback script body for environments where the compiled helper is unavailable |
| `electron/services/nativeScannerProcessingService.ts` | Post-capture processor for native IK runs; reads crop jobs, OCRs/parses card crops, matches IK metadata, writes processing outputs, and safely loads native crop previews |
| `electron/services/nativeScannerResultWorkflowService.ts` | Authoritative promotion and review workflow for native results; reloads run state, validates edits, updates durable results/logs, and only writes the artifact store after explicit confirmation |
| `electron/services/goodFileService.ts` | Local GOOD export file writing and GOOD import file dialog/read handling |
| `electron/preload.cjs` | Safe renderer bridge exposed as `window.assistantApi` |
| `native/input-helper/IkInventoryLists.cs` | Loads and validates the vendored IK `inventorylists` feature catalog |
| `native/input-helper/NativeScannerFiles.cs` | Writes native scanner manifest, status, and JSONL crop-job files |
| `src/lib/artifactStore.ts` | Pure signature/id/record helpers for the persistent artifact store |
| `src/lib/ikArtifactMatcher.ts` | Pure IK inventorylist matcher for native artifact set/piece/slot validation and GOOD key metadata |
| `src/lib/ikCatalogMatcher.ts` | Pure IK inventorylist matcher for simple weapon, character, and material names plus compact GOOD-key samples |
| `src/lib/ikScanCapabilities.ts` | Pure capability summary that separates IK catalog coverage from implemented native capture support |
| `src/lib/scanResultEntry.ts` | Pure durable scan-result entry/status helpers that keep extraction confidence separate from deferred artifact value evaluation |
| `src/App.tsx` | Thin React entry that renders the app page |
| `src/pages/AppPage.tsx` and `src/pages/app/*` | App page composition and high-level layout routing |
| `src/features/scan/hooks/useScanViewController.ts` | Scan feature state composition and view-controller assembly |
| `src/features/scan/hooks/scanViewScanActions.ts` | Manual scan and visible-grid scan orchestration |
| `src/features/scan/hooks/scanViewEntryActions.ts` | Guided auto-entry choreography for visible inventory, direct inventory, and ESC/B fallback paths |
| `src/features/scan/hooks/useScanGoodInterop.ts` | Scan-page GOOD import/export actions against renderer repository ports |
| `src/features/inventory/*` | Scanned-artifact inventory browser for native result entries, stored artifacts, filters, sorting, compact detail state, native crop preview display, and IK catalog status |
| `src/lib/artifactOcrParser.ts` | Converts OCR output into a parsed artifact candidate with confidence and notes |
| `src/lib/fuzzyMatch.ts` | Generic fuzzy string matching for OCR text against known game data |
| `src/lib/genshinLookup.ts` | Pure lookup and validation API for generated Genshin data |
| `src/lib/autoScanEntry.ts` | Pure entry-mode planning and auto-scan preflight validation |
| `src/lib/cardReadyGate.ts` | Detail/page fingerprint readiness gate for scan timing |
| `src/lib/artifactEvaluation.ts` | Planned deterministic artifact value evaluation with score reasons and review-safe output |
| `src/lib/upgradeProjection.ts` | Planned best/middle/worst upgrade projection for under-leveled artifacts |
| `src/lib/scoring.ts` | Recommendation and build scoring logic |
| `src/lib/demoData.ts` | Temporary local demo snapshot |
| `data/ik-inventorylists/*` | 1:1 vendored Inventory Kamera inventory lists used by the native scanner data preflight and future matching |
| `src/data/genshinGameData.json` | Generated local dictionary of characters, artifact sets, slots, and stats |
| `scripts/generate-genshin-data.cjs` | Regenerates the local Genshin dictionary from `genshin-db` |
| `src/types/*` | Shared app, capture, and domain contracts |
| `src/styles/global.css` | Stylesheet entrypoint importing split style modules |
| `src/styles/base.css` | Shared application, layout, scanner workspace, modal, triage, build, and overlay styles |
| `src/styles/diagnostics.css` | Diagnose/dev-view specific styles |
## Dependency Rules
- Renderer code calls Electron only through the preload bridge.
- Electron main process must not import React renderer modules.
- Pure parsing and scoring modules must not depend on Electron APIs.
- OCR uncertainty must be represented in data, not hidden in UI only.
- Generated folders must not be treated as source of truth.
## Smart Capture Flow
```mermaid
sequenceDiagram
participant UI as React Scan UI
participant Bridge as Preload Bridge
participant Main as Electron Main
participant Game as Genshin Window
participant OCR as OCR Worker
UI->>Bridge: captureSource(sourceId, 0, focusGenshin=true)
Bridge->>Main: IPC capture:captureSource
Main->>Main: Hide app window
Main->>Game: Focus Genshin window
Main->>Main: Capture primary screen via GDI
Main->>Main: Detect artifact detail panel
Main->>OCR: OCR focused crops
OCR-->>Main: Text and confidence
Main->>Main: Restore app window
Main-->>UI: Full capture, detail preview, crops, OCR results
UI->>UI: Parse candidate and show result
```
## Data Flow
| Data | Source | Owner | Consumers |
| --- | --- | --- | --- |
| Capture sources | Electron desktopCapturer | Electron main | Scan UI |
| Screenshot | Windows GDI / desktopCapturer | Electron main | Cropper, OCR, UI preview |
| OCR crops | Electron main | Electron main | Details modal, parser |
| Native card crops | C# input helper | Native scanner run directory under app userData | Electron status, inventory preview, later OCR/parse queue |
| IK inventorylists | `data/ik-inventorylists` copied from Inventory Kamera 1.4.4 | Source data package | Native scanner data preflight and future matcher |
| IK simple catalog match | IK weapon/character/material inventorylist catalog | `IkCatalogItemMatch` | Future category scanners, inventory catalog evidence |
| Game dictionary | `genshin-db` generated JSON | `src/data/genshinGameData.json` | OCR parser |
| Parsed artifact candidate | OCR parser | Renderer domain logic | Result panel, post-capture report, future local DB |
| IK artifact match | IK artifact inventorylist catalog | `ScanResultIkMatch` | Native post-capture review gate, scan result entry, inventory detail |
| Parser field confidence | Native post-capture parser | `ScanResultFieldConfidence` | Scan result entry, inventory detail review signal |
| Scan result entry | Scan loop or native post-capture processor | `StoredScanResultEntry` / native run `scan-results.json` | Live scan rail, artifact inventory, summary |
| Artifact evaluation | Deterministic evaluator | `src/lib` | Result pills, inventory sort/filter, detail reasons |
| Upgrade projection | Projection helper | `src/lib` | Artifact detail view only |
| Review samples | User action in Scan UI | Electron userData `review-samples.jsonl` | Future regression tests and OCR training |
| Stored artifacts | Manual/automatic scans | Electron userData `artifact-store.json` (dedupe by content signature) | Future triage, recommendations, SQLite migration |
| Recommendations | Scoring module | Renderer domain logic | Triage and builds views |
## Scan Modes
**Manual scan** is read-only: the user clicks artifacts in Genshin, the app repeatedly runs Smart Capture, deduplicates by content signature, persists new artifacts to the local store, and saves review samples when crops/OCR are missing, total confidence is low, field confidence is low, or parser notes indicate incomplete data.
**Automatic grid scan** is user-triggered input automation limited to clicking detected inventory tiles and wheel-scrolling the inventory. Safety and reliability rules:
- All input goes through the helper service boundary (currently a C# sidecar with
fallback support behind the same JSON protocol). The helper owns focus, cursor
movement, click, scroll, guard-state polling, elevation detection, and GDI
capture. Mouse movement is sent as iterated relative input deltas instead of
relying on a single absolute cursor jump. The helper verifies the cursor
reached the target and refuses to click otherwise.
- `npm run dev:admin` is the validated dev path for automation when elevated
input is required. The elevated PowerShell startup is handled by
`scripts/dev-admin.ps1` and logged to `outputs/admin-start/admin-dev.log`.
The user must approve UAC manually; the app cannot approve the Secure Desktop
prompt itself.
- Failsafe: before every click and scroll the renderer polls cursor position and ESC state. Holding ESC or moving the mouse away from the last automated position aborts the scan immediately; the Stop button also aborts. Only the `GetAsyncKeyState` held-down bit (0x8000) is used - the "pressed since last call" bit fires for stale ESC presses from normal Genshin menu navigation and caused false aborts.
- SendInput's return value is checked: zero injected events (UIPI, e.g. elevated Genshin vs. non-elevated app) aborts with an explicit hint instead of silently clicking into nothing.
- Dev-only probes under `http://127.0.0.1:17317` are used for live validation:
`/automation/probe-click?index=N` tests one read-only tile selection, and
`/scanner/start?limit=N` starts the native capture scan with a temporary
limit payload from an already visible artifact detail view.
The live known-good result on 2026-07-07 is documented in
[AUTOMATION_LIVE_SCAN.md](AUTOMATION_LIVE_SCAN.md).
- Click verification: after each click the parsed detail-panel signature should change. An unchanged signature is a soft miss (it can also mean two OCR-identical neighbor pieces, common among +0 artifacts), so it is retried once with a small offset, logged with the stuck artifact name, and then skipped - never fatal on its own. The scan aborts only when the first ~6 clicks of page 1 produce nothing new (diagnosis hint: elevated Genshin blocks SendInput via UIPI, or grid coordinates are wrong) or a later page yields zero new artifacts.
- Scan stats separate clicked (click attempts), parsed (readable captures), stored (persisted), review (review samples), duplicates, and misses, so "scanned" cannot be mistaken for "successfully read".
- Artifact grid automation uses 32 safe click targets per full page
(`8 x 4`). The apparent lower fifth row sits in the bottom control band on
16:9 captures and is not clicked automatically.
- Guided auto-entry is state gated. The normal scan button first performs a
lightweight no-OCR preflight; OCR/store/review work starts only after the
artifact inventory grid and artifact detail card are visually confirmed.
- The normal scan button does not navigate into inventory when that preflight
fails; it blocks and asks the operator to open the Artifact inventory with a
visible detail card. Explicit Dev-Control entry modes can still test direct
inventory or Paimon-menu choreography, but they are not the merge-ready
default path.
- Item verification uses the artifact OCR capture's own detail fingerprint, so
the loop no longer performs a separate card-ready capture before OCR. Page
waits remain fingerprint based and can proceed as soon as the inventory pane
changes and stabilizes.
- Scrolling sends one wheel notch per grid row with the cursor anchored over the inventory (assumption: roughly one row per notch; overlap is absorbed by dedupe, and a page without new artifacts stops the scan).
- The scan never deletes, enhances, feeds, locks, or spends anything; it only selects tiles to read them.
**Native IK capture scan** is the new high-speed path. It runs inside the C#
sidecar, computes a fixed 16:9 8x4 visible-inventory grid, clicks and scrolls
natively, captures the artifact detail card as PNG crops, and reports progress
through IPC/dev-control. The renderer only starts, stops, and displays status.
OCR, parsing, GOOD persistence, and artifact value evaluation are deliberately
separate follow-up stages so capture speed is not blocked by UI work.
Each native run writes a self-contained run directory under app userData:
- `manifest.json`: run schema, IK data version/categories, Genshin bounds,
grid, detail crop rectangle, explicit scan category, and downstream queue
contract.
- `capture-jobs.jsonl`: one job per captured card crop with page/row/column,
client/screen coordinates, click event count, image path, and downstream
`ocr-parse-store` marker. Jobs include the scan category; today only
`artifacts` produces jobs. Native preflight and post-capture processing also
guard this category boundary so catalog-only weapon, character, and material
entries cannot be accidentally parsed as artifacts.
- `status.json`: latest scanner status snapshot for recovery and dev tooling.
Its `supportedCategories` block separates IK catalog availability from native
capture support: artifacts are the only native-capture category today, while
weapons, characters, and materials are catalog-only until their own capture
flows have evidence.
- `scan-results.json`: durable per-artifact result entries with capture
metadata, parsed artifact identity when available, extraction status, and
value status. Native capture currently writes `deferred` value status for
clean extraction and `review` for uncertain extraction.
- `processing-report.json`: optional post-capture OCR/parse report generated
from `capture-jobs.jsonl`. It records `queueConcurrency` for the bounded
post-capture OCR/parse worker and is non-persisting by default until native
crop OCR has live validation evidence.
Parsed artifacts from manual and renderer auto-scan modes are persisted into
`artifact-store.json` keyed by a content signature that excludes the equipped
character, so re-equipping updates a record instead of duplicating it. Leveling
an artifact currently creates a new record (documented limitation until
rescan-merge exists). Native IK capture currently writes card crops first; a
downstream OCR/parse processor can report parsed artifacts from those crops
without slowing capture. Store promotion remains opt-in until native crop OCR is
validated.
The Inventory surface derives a dry-run promotion summary from native
`scan-results.json` entries and the local artifact store. It can show which
native artifacts are ready for explicit promotion, already stored, review-only,
or blocked, but it does not write store records by itself.
During auto-scan, artifact store writes can be batched and flushed after the
click/capture/OCR loop to avoid per-artifact save/reload churn in the hot path.
Auto-scan artifact captures also bypass Electron source-list enumeration and use
the GDI capture helper directly once the selected source/Genshin state has been
preflighted. Manual captures and source refresh still use `desktopCapturer`.
## Scan Results And Inventory UX
The next product surface is documented in
[scanner-results-inventory-roadmap.md](scanner-results-inventory-roadmap.md).
Architecture rules:
- The active scan view is an operator surface, not the full evaluator. It should
show the screenshot/preview, the right-side live result rail, Stop/status
controls, and review access.
- The live result rail receives completed artifact results only. Intermediate
OCR/debug stats stay in diagnostics or detail.
- A scan result row preserves extraction status and artifact value status as
separate data even when the UI shows one compact pill.
- The artifact inventory view owns browsing, filtering, sorting, opening detail,
and showing the current artifact-only native pipeline state.
- Weapons, materials, and character details stay out of the active Inventory UI
until their values are actually scanned; loaded IK catalog data alone is not a
user-facing scanner capability.
- The artifact detail view owns screenshot/crop inspection, OCR confidence,
parser notes, value score reasons, and upgrade projection.
- Native crop previews are served through the Electron bridge only for PNG paths
inside the selected native scanner run directory.
- Native artifact post-processing uses IK artifact set/piece/slot matching as a
review gate. A conflict is extraction uncertainty, not a weak artifact value.
- Native artifact post-processing stores per-field parser confidence so detail
review can explain which OCR/parser fields are trustworthy.
- Upgrade projection is a local deterministic/probabilistic helper, never a
claim that an artifact will roll a specific way.
Planned result flow:
```mermaid
flowchart LR
ScanLoop["Scan loop"]
Parsed["Parsed artifact"]
ReviewGate["Extraction confidence / review gate"]
Value["Artifact value evaluator"]
Store["Artifact store"]
Rail["Live result rail"]
Inventory["Artifact inventory"]
Detail["Artifact detail"]
ScanLoop --> Parsed
Parsed --> ReviewGate
ReviewGate --> Value
Value --> Store
Value --> Rail
Store --> Inventory
Rail --> Detail
Inventory --> Detail
```
Future queue refactor:
- One capture/game-control worker may click, scroll, focus, and poll failsafes.
- OCR/parse/evaluation may process bounded queued screenshot/crop jobs.
- Queueing must preserve stop behavior, duplicate handling, review decisions,
and the existing read-only safety boundary.
- The native post-capture processor already consumes crop jobs with bounded
parallelism while preserving report order. Further queue work is secondary to
live throughput evidence and result quality.
## Security And Safety
| Area | Policy |
| --- | --- |
| Game access | Screen capture only; no memory reads, hooks, or process injection |
| Automation | Future feature only, opt-in and reversible |
| Data privacy | Local-first; no upload path in scanner MVP |
| Secrets | No cookies or API tokens required for scanner MVP |
| Unsafe actions | Never delete, feed, enhance, or spend resources |
## Error Handling
- Capture failures should surface in the scan status row.
- Missing Electron bridge should explain that browser preview cannot capture Genshin.
- OCR failures should leave the capture available and mark parsed fields as unknown.
- Low-confidence data should go to review instead of silent acceptance.
- Parsed artifact fields carry individual confidence and source metadata so the UI can show uncertainty per field.
## Performance
Current OCR is measured through the eval harness and live scan assessments
rather than assumed good. The app keeps a Tesseract.js worker pool and reports
capture, OCR, card-ready, scroll-ready, active-scan, and projected-100 timings.
For the next product phase, performance work should not displace extraction
quality, result clarity, or review safety unless evidence shows a regression.