dsh-galgame-localization-reverse

Galgame Doctor for DeepSeek Harness

New in v0.9.1: review import errors identify the physical CSV record start line, including multiline records and blank lines. Encoding failures do not echo the translation. Provider registration and skill-body loading were also verified with DSH 0.2.1-rc.3.

Preflight reviewed translations with --dry-run before importing: validate hashes, placeholders, encoding, and strict CSV syntax without writing a catalog.

Keep reviewed translations when a game updates. Catch broken control codes before shipping.

New in v0.9.0: bring human review back into the workflow

1. Run translation-memory migration with --review-csv review.csv.
2. Open the CSV in Excel/WPS and fill only reviewed_target for rows you approve.
3. Import against the original current catalog used by migration:

python -B -X utf8 skill/scripts/vn_review.py --current current.jsonl --review-csv review.csv --output reviewed.jsonl --encoding gbk

The importer rejects a changed catalog, duplicate or unknown IDs, broken placeholders, encoding failures, and conflicting existing translations. It writes a new file only after every row validates. Blank reviewed targets remain pending. Default segment_id, source, and target fields are supported.

Run migration again with reviewed.jsonl as current to combine human decisions with historical exact reuse, then run strict QA before script writeback. Source game files are never changed by the review importer.

Preflight before import

python -B -X utf8 skill/scripts/vn_review.py --current current.jsonl --review-csv review.csv --dry-run --encoding gbk

Preflight rejects --output and writes no files. In its JSON result, ready counts filled, validated review rows; pending counts blank review rows; applied is always zero. These are review-queue counts, not whole-game completion metrics. Any failed row produces a nonzero exit status. After validation, replace --dry-run with --output reviewed.jsonl to import into a new file.

中文 | English

[!IMPORTANT] Unofficial community plugin. Independently developed and maintained; not reviewed or endorsed by DeepSeek.

Audit first, reuse safely, localize second. Give DeepSeek Harness a visual-novel folder and get an evidence-backed engine candidate, safe translation migration, encoding clues, strict QA, and a reversible next step. The audit is offline and read-only; only report paths outside the audited tree are written.

Start in 30 seconds

dsh plugin --profile web add github:julescules/dsh-galgame-localization-reverse#v0.9.1

Restart DSH, then ask:

Use galgame-localization-reverse to audit D:\Games\Example read-only.
Show the engine evidence, localization risks, and the safest next step.

No API key is required. The plugin contains no game assets, decrypted archives, credentials, or title-specific keys.

What you get

One-command localization plan

vn_project_audit.py plan inventories a bounded set of files and reports:

• engine candidates with confidence, signal groups, and concrete path evidence;
• script, archive, image, font, and executable candidates;
• bounded encoding samples and CJK-path-safe relative paths;
• a cautious classification of candidate, weak-candidate, or unknown;
• an evidence status, recommended route, exact next commands, risks, and release gates.

The classification is routing evidence, not a claim that static inspection proves runtime behavior.

python -X utf8 skill\scripts\vn_project_audit.py plan D:\Games\Example `
  --json D:\project\logs\project-audit.json `
  --markdown D:\project\logs\project-audit.md

Lossless source-script adapters

vn_script_adapter.py extracts plain KiriKiri .ks, Ren'Py .rpy, and NScripter 0.txt text into UTF-8 JSONL while preserving source hash, line hash, path, line number, encoding, BOM, newline, prefix, and suffix. Apply always writes to a separate output root and fails if the source changed.

python -X utf8 skill\scripts\vn_script_adapter.py extract D:\Games\Example `
  --engine kirikiri --output D:\project\translations.jsonl

python -X utf8 skill\scripts\vn_script_adapter.py apply D:\Games\Example `
  D:\project\translations.jsonl --output-root D:\project\rebuilt

It does not open XP3/NSA archives or claim compiled-script support.

Safe translation migration after game updates

vn_translation_memory.py carries reviewed translations from one or more prior releases into a newly extracted catalog without silently guessing. Repeat --previous to combine histories. Exact conflicts across releases stay empty and return a non-zero status; fuzzy matches remain review-only. Every input remains unchanged and every catalog SHA-256 is recorded. The optional UTF-8 BOM CSV opens cleanly in Excel/WPS and keeps blank reviewer columns for decisions.

python -X utf8 skill\scripts\vn_translation_memory.py migrate `
  --current D:\project\v2\translations.jsonl `
  --previous D:\project\v1\translations.jsonl `
  --previous D:\project\v1-hotfix\translations.jsonl `
  --output D:\project\v2\migrated.jsonl `
  --report D:\project\logs\translation-memory.json `
  --markdown D:\project\logs\translation-memory.md `
  --review-csv D:\project\review\translation-memory.csv

Strict translation preflight

vn_qa.py rejects missing, added, or reordered control tokens by default. v3 adds glossary and forbidden-term gates, repeated-source consistency, character/line budgets, punctuation balance, empty translations, residual Japanese, and target-encoding checks. Generic JSON/JSONL, script-adapter JSONL, and GalTransl cache fields are supported without third-party packages.

python -X utf8 skill\scripts\vn_qa.py check-jsonl D:\project\translations.jsonl `
  --encoding gbk --require-complete --require-consistent `
  --glossary D:\project\glossary.json --max-chars 80 --max-lines 3 --check-punctuation `
  --report D:\project\logs\translation-qa.json `
  --markdown D:\project\logs\translation-qa.md

An intentionally broken synthetic fixture produces a concise failure report:

FAIL  2 blockers · 3 segments
E control-token-added  scene01:2  added [wait]
E empty-translation    scene01:3  translation required

Release evidence, not another translator

The bundled workflow also provides:

• vn_patch.py — hash-bound file replacement, verification, backup, and rollback;
• vn_image_qa.py — PNG/APNG, JPEG, GIF, and BMP geometry/alpha/frame comparison;
• vn_slot_qa.py — fixed-slot terminator, padding, marker, and outside-region checks;
• focused routes for AI6WIN, BGI/Ethornell, Director, RealLive, Delphi/DirectDraw, PSP/Vita, Ren'Py, KiriKiri, and custom engines.

GalTransl, VNTextPatch, GARbro, LunaTranslator, and engine-specific tools can handle translation or extraction. Galgame Doctor complements them by deciding what the project appears to be, checking their text output, and preserving release evidence.

Try the synthetic examples

python -X utf8 skill\scripts\vn_project_audit.py plan examples\synthetic-project\game `
  --json test_outputs\synthetic-audit.json `
  --markdown test_outputs\synthetic-audit.md

python -X utf8 skill\scripts\vn_qa.py check-jsonl examples\translations.jsonl `
  --encoding gbk --require-complete --glossary examples\glossary.json `
  --report test_outputs\translation-qa.json `
  --markdown test_outputs\translation-qa.md

All example text is synthetic and redistributable.

DSH integration

The bundle uses the official packaged Skill Provider shape:

• inject = ['skills'];
• ctx.skills.registerProvider(...);
• directory resourceBase for on-demand references and scripts;
• the official dsh.bundle.patch install path.

The registered name remains galgame-localization-reverse, so existing prompts continue to work.

Verify the package

npm run check
npm pack --dry-run
node scripts/build-release-metadata.mjs .\dsh-galgame-localization-reverse-0.9.1.tgz .\builds\v0.9.1
.\scripts\verify-release.ps1 -PackagePath .\dsh-galgame-localization-reverse-0.9.1.tgz -ChecksumsPath .\builds\v0.9.1\SHA256SUMS

The check runs provider tests plus deterministic self-tests for planning, script adapters, translation-memory migration, translation QA, and all bundled patch/asset utilities. Each release also publishes SHA-256 and CycloneDX SBOM evidence.

Limits

• Engine formats and binary constants remain title/version specific until proven against exact hashes.
• Script adapters support plain source files only and never imply archive/compiled-script compatibility.
• Static audit and QA do not replace a zero-mutation round trip or in-game UI, save/load, audio, and rollback tests.
• Residual Japanese can be intentional; review each finding or maintain an explicit exclusion instead of silently weakening the gate.
• DeepSeek Harness is in developer preview; pin a reviewed plugin release.

See CONTRIBUTING.md and SECURITY.md for contribution and reporting guidance.

License

MIT

Compatibility: the previous full-host baseline is DSH 0.1.2-rc.1. On 2026-10-18, provider registration and skill loading passed against the real DSH 0.2.1-rc.2 skill registry and Cordis packages. This is a registry contract test, not a full Web UI or model-session test.