Unity Translation Workflow
Operating Principle
Treat each Unity game as a separate format-discovery task. Do not assume that a previous game's extraction boundary, scenario layout, writeback method, or font injection strategy applies to the current game.
- Inspect the build and locate actual display-text sources.
- Select the smallest applicable adapter.
- Extract project-named, one-column text files.
- Ask for matching one-column translation files.
- Validate row alignment and structural tokens; blank only row-addressable failures when the user elects partial writeback.
- Dry-run, prove source-specific serialization/runtime safety, build candidates outside the game root, back up, then write only approved rows.
- Run the same source-specific residual audit on the composed final candidate; resolve every visible candidate and rerun until zero unreviewed visible occurrences remain.
- Build a SHA-verified binary patch only from pristine originals to that final runtime-tested candidate.
Use three gates for every project:
- Discovery gate: report Unity version/layout, Mono versus IL2CPP, Addressables, middleware clues, and all candidate text sources before choosing an adapter.
- Translator gate: expose only project-named one-column CSVs for row-mapped sources; keep broad
extracted_text.csv as audit evidence, not as a writeback source.
- Runtime gate: a clean CSV validation and a tool-level dry run are not enough for runtime-sensitive containers; prove the edited file loads through the same game path before distributing it.
Prefer static occurrence-mapped writeback. Use a runtime text adapter only for confirmed visible strings that static discovery cannot reach, and keep its global key-based output separate from static manifests and writeback artifacts.
Read references/unity-text-map.md when locating candidate Unity text containers or middleware clues.
Read references/runtime-sensitive-adapters.md when handling BansheeGz databases, Addressables bundles/catalogs, ES3 saves, or runtime font compatibility.
Read references/workflow-modes.md before choosing semi-automatic or full-automatic translation.
Read references/ainiee-integration.md before running full-automatic translation through ainiee-translate.
Read references/mono-raw-residual.md when TypeTree reading fails, an embedded Utage book is incomplete, or visible Japanese survives normal writeback.
Read references/font-asset-replacement.md before selecting, converting, or replacing any TMP font asset or font bundle.
Read references/runtime-text-adapters.md before generating or testing an AutoTranslator candidate for IL2CPP, Naninovel, GameCreator, custom-rendered, or runtime-only text.
Read references/version-migration-and-release-audit.md when porting a translation to a newer game version, separating scratch candidates from durable project evidence, selecting a runtime Locale/StringTable, auditing short bytecode choices or residual Japanese, choosing static CJK font coverage or smoke-test depth, or preparing a release-only patch baseline.
Read references/embedded-csv-and-runtime-diff.md when Naninovel-like CSV data is embedded in assets or when comparing runtime AutoTranslator captures.
Public repository release notes belong in GitHub Releases. Do not add CHANGELOG.md to the GitHub repository.
Workflow Modes
Semi-automatic mode extracts Unity text and waits for matching *_translation.csv files before validation and writeback.
Full-automatic mode extracts Unity text, converts each source CSV to an AiNiee cache.json, lets the agent translate through the sibling ainiee-translate skill, converts the cache back to *_translation.csv, then runs the same Unity validation and writeback gates. Full-auto only automates translation; it does not bypass row alignment, structural validation, source hashes, dry runs, backups, or runtime checks.
Work Directory
Keep work products under the game root:
GameFolder/_translation/unity-text-report/
report.md
extraction_manifest.json
text/
GameTitle_text.csv
GameTitle_text.manifest.json
GameTitle_text_translation.csv
GameTitle_ui_text.csv
GameTitle_ui_text.manifest.json
source_text_index.json
validation/
runtime/
autotranslator/
writeback/
backups/
scan_unity_text.py and extract_unity_text.py default to this directory and ignore _translation while scanning. Put distributable patch archives outside the game directory only after verification.
Reports, manifests, and CSVs may live under _translation, but rebuilt bundles, catalogs, font assets, and other runtime binary candidates must be written to a separate candidate root outside the game directory. Applying a candidate is a distinct, backed-up step.
Translator File Contract
Use one source file and one translation file per extracted Unity source. Name translator-facing files with the project folder name followed by a short semantic description, not the internal Unity asset path:
<ProjectFolderName>_<description>.csv
Use <ProjectFolderName>_text.csv for the primary story/dialogue text source when it is clear, and descriptions such as _ui_text, _utage_text, _profiles_text, or _resources_text for additional or generic sources. For example:
M男いじめスクールライフ_text.csv
M男いじめスクールライフ_ui_text.csv
Source file generated by extraction:
original_flat
...
Translation file provided after translation:
zh_cn
...
Name the translation by appending _translation to the source CSV stem:
M男いじめスクールライフ_text.csv
M男いじめスクールライフ_text_translation.csv
Rules:
- Keep one translated cell for every source row; do not reorder or delete rows.
- Keep duplicate source rows. Each row identifies one specific occurrence and may receive a different action.
- Leave
zh_cn blank to preserve that occurrence in the game.
- Keep resource paths,
path_id, offsets, and source hashes in the manifest only; do not expose them in the translator-facing CSV filename.
- Preserve TMP/Unity tags, placeholders, and literal
\n markers unless a verified parser-specific change is required.
- A headerless one-column translation file is accepted. Generated/normalized output uses the
zh_cn header.
- Do not edit
*.manifest.json; it maps CSV rows to source-file occurrences and stores the extracted file hash.
Legacy original_flat,zh_cn translation files remain accepted by validation/writeback for existing projects, but do not create new projects in that format.
Choose An Adapter
| Evidence | Use now | Status |
|---|
level*, resources.assets, sharedassets*.assets, globalgamemanagers, bundle candidates with serialized UTF-8 text | scan_unity_text.py, generated per-source manifests, writeback_unity_text.py | Automated row-mapped workflow |
Broad extracted_text.csv contains mixed visible/internal rows and needs manual narrowing | build_visible_safe_extract.py, then project_translation_to_subset.py if an existing legacy baseline must be projected | Audit/subset workflow; not a row-mapped writeback source |
| Raw profile/encyclopedia MonoBehaviour strings missed by the normal pass | extract_character_profiles.py, writeback_character_profiles.py | Automated one-column workflow |
Utage ALL_Texts.book / ALL_EventTexts.book in scenario assets or embedded assets | extract_utage_scenarios.py, writeback_utage_scenarios.py | Automated one-column workflow after confirming text column/books |
| MonoBehaviour TypeTree fails, embedded Utage data is skipped, or Japanese remains after writeback | extract_mono_raw_utf8.py, occurrence approval, writeback_mono_raw_utf8.py | Two-stage raw-object audit; exact controls; external canary only when object spans resize |
bansheegz_database TextAsset plus BGDatabase.dll/BansheeGz.BGDatabase | Build a one-shot game-runtime export/write/save probe and reinsert the verified payload | Runtime-specific structured adapter |
| Loose CSV/JSON/TXT or custom scenario string parsers | Inspect format and preserve its record/newline contract | Project-specific |
0x06 + uint32 LE + UTF-8 length-prefixed records, often near a DMSL signature | extract_dmsl_text.py, writeback_dmsl_text.py | Fixed-byte occurrence workflow; byte budget is blocking |
TMP/UI m_text, label fields, or Unity Localization m_TableData.*.m_Localized | extract_typetree_ui_text.py, writeback_typetree_ui_text.py | TypeTree occurrence workflow with field-path verification |
Addressables catalogs/bundles or IL2CPP global-metadata.dat | Confirm exact runtime text, integrity metadata, and game-runtime loading before applying candidates | Project-specific |
| Confirmed UGUI/TMP/IMGUI/TextMesh strings exist only at runtime, including IL2CPP, Naninovel, or GameCreator resolution | Validate the row-mapped CSV, then use generate_autotranslator_txt.py; provision any compatible loader/plugin separately | Optional global key-based candidate; no automatic deployment |
| Rebuilt Addressables bundle loads to a black screen or logs CRC mismatch | patch_addressables_catalog_crc.py, then paired bundle/catalog canary | Binary catalog candidate; unique internal-name match only |
| Chinese displayed as boxes/tofu in Mono/TMP games with JSON injection tables | install_tmp_chinese_font_fix.py --dry-run, then apply | Preflighted runtime fallback workflow |
| TMP font bundle candidate for resource-level replacement | trace_tmp_font_usage.py, inspect_tmp_font_bundle.py, then one-font runtime canary | Visible-component identity and coverage gate |
| Correct Unicode renders as another valid Chinese character | Verify CSV/code point, inspect populated TMP tables, then build a matched static set | Glyph-index conflict; not translation substitution |
| IL2CPP game without a verified runtime loader | Inspect candidates or generate an exact-version TMP asset; do not use the Mono injector | Project-specific font workflow |
Embedded CSV signatures in sharedassets*.assets or bundles | extract_embedded_csv.py | Discovery-only occurrence manifest; no generic writeback |
Core Commands
Scan and extract candidates:
python scripts/scan_unity_text.py "path/to/GameFolder"
This keeps the broad extracted_text.csv and extraction_manifest.json for audit, and creates translator-facing, occurrence-preserving files in text/.
For fixed-width DMSL operands, extract from explicit inputs or a constrained candidate root:
python scripts/extract_dmsl_text.py "GameFolder" \
--candidate-root "Game_Data/StreamingAssets/aa/StandaloneWindows64" \
--out-dir "GameFolder/_translation/unity-text-report/text"
After validation, dry-run the fixed byte budgets, then write only to an external candidate root:
python scripts/writeback_dmsl_text.py "GameFolder" \
--manifest "GameFolder/_translation/unity-text-report/text/GameTitle_dmsl_text.manifest.json" \
--translation-csv "GameFolder/_translation/unity-text-report/text/GameTitle_dmsl_text_translation.csv" \
--report "CandidateReports/dmsl_dry_run.json"
Add --base-root "PriorCandidateRoot" --out-dir "NextCandidateRoot" --write only after the dry run is clean. Any over-budget row rejects the whole batch and must be shortened and revalidated; the writer never truncates UTF-8.
Extract component UI fields and Unity Localization StringTables with:
python scripts/extract_typetree_ui_text.py "GameFolder" \
--scan-root "Game_Data/StreamingAssets/aa" \
--out-dir "GameFolder/_translation/unity-text-report/text"
Dry-run writeback_typetree_ui_text.py with its manifest, source CSV, and translation CSV. Use --base-root to compose over a DMSL candidate, and add --out-root "NextCandidateRoot" --write only after every object identity and field path still matches.
For a rebuilt Addressables bundle, generate the paired external catalog candidate from its internal name:
python scripts/patch_addressables_catalog_crc.py \
"GameFolder/Game_Data/StreamingAssets/aa/catalog.bin" \
--bundle "CandidateRoot/Game_Data/StreamingAssets/aa/StandaloneWindows64/changed.bundle" \
--game-root "GameFolder" --out-dir "CatalogCandidateRoot" --write
The catalog tool rejects unknown/duplicate records and byte-compares the result so only the selected CRC field can change. Apply and roll back the bundle/catalog pair together.
When a broad audit CSV must be narrowed for human translation review, produce a conservative visible-safe subset and audit the omissions:
python scripts/build_visible_safe_extract.py \
"GameFolder/_translation/unity-text-report/extracted_text.csv" \
--manifest "GameFolder/_translation/unity-text-report/extraction_manifest.json" \
--out-dir "GameFolder/_translation/unity-text-report/validation/visible-safe"
Use this subset only as a reviewer aid. For automated writeback, prefer the per-source CSV plus its sidecar manifest. If an older original_flat,zh_cn translation baseline must be projected onto an approved subset, run:
python scripts/project_translation_to_subset.py \
--translation-csv "GameFolder/_translation/unity-text-report/validation/legacy/GameTitle_text_translation_utf8.csv" \
--target-source-csv "GameFolder/_translation/unity-text-report/validation/visible-safe/extracted_text_visible_safe.csv" \
--out-dir "GameFolder/_translation/unity-text-report/validation/projection"
Revalidate the projected translation against the exact target source CSV before any writeback decision.
Validate a translated source file:
python scripts/validate_translation_csv.py \
--source-csv "GameFolder/_translation/unity-text-report/text/GameTitle_resources_text.csv" \
--translation-csv "GameFolder/_translation/unity-text-report/text/GameTitle_resources_text_translation.csv" \
--out-dir "GameFolder/_translation/unity-text-report/validation/resources"
Validation blocks row-count, placeholder, tag, and literal \n count mismatches by default. Use --allow-newline-changes only after confirming the game's parser or layout accepts the change.
When runtime evidence proves that supported visible text cannot be reached statically, generate an AutoTranslator audit report from the exact source CSV and its normalized validated translation:
python scripts/generate_autotranslator_txt.py \
--source-csv "GameFolder/_translation/unity-text-report/text/GameTitle_ui_text.csv" \
--translation-csv "GameFolder/_translation/unity-text-report/validation/ui/GameTitle_ui_text_translation_utf8.csv" \
--out-dir "GameFolder/_translation/unity-text-report/runtime/autotranslator"
The default writes only autotranslator_export_report.json. Inspect it, then repeat with --write to create _PreTranslated.txt. Any row-count mismatch, conflicting or partly blank duplicate key, ambiguous plain-text key/value, or serialized-key collision blocks the whole candidate. The skill does not install BepInEx, XUnity.AutoTranslator, configs, or fonts; verify the external loader/plugin contract and complete a runtime canary as described in runtime-text-adapters.md.
For embedded CSV discovery, produce an occurrence-preserving audit before considering a runtime-specific writer:
python scripts/extract_embedded_csv.py "GameFolder" \
--out-dir "GameFolder/_translation/unity-text-report/text/embedded-csv"
For runtime candidate comparison, keep the diff separate from deployment:
python scripts/diff_autotranslator_files.py \
--old "GameFolder/_translation/unity-text-report/runtime/autotranslator/_PreTranslated.previous.txt" \
--new "GameFolder/_translation/unity-text-report/runtime/autotranslator/_PreTranslated.txt" \
--output "GameFolder/_translation/unity-text-report/runtime/autotranslator/runtime_diff.json"
Review removals and broad changes manually, then rerun the runtime canary. These adapters never merge or install runtime files.
For full-automatic translation, convert a Unity source CSV to an AiNiee cache, translate it with ainiee-translate, then convert it back to a Unity translation CSV:
"$AINIEE_PY" scripts/unity_csv_to_ainiee_cache.py \
--source-csv "GameFolder/_translation/unity-text-report/text/GameTitle_text.csv" \
--out-cache "GameFolder/_translation/unity-text-report/ainiee/GameTitle_text/cache.json"
Use the sibling ainiee-translate skill to run the batch translation loop on that cache.json. After translation:
"$AINIEE_PY" scripts/ainiee_cache_to_unity_translation.py \
--cache "GameFolder/_translation/unity-text-report/ainiee/GameTitle_text/cache.json" \
--source-csv "GameFolder/_translation/unity-text-report/text/GameTitle_text.csv" \
--out-csv "GameFolder/_translation/unity-text-report/text/GameTitle_text_translation.csv"
Then run validate_translation_csv.py exactly as in the semi-automatic path.
When the user explicitly chooses to apply only structurally safe rows, filter row-addressable failures and then validate the filtered output again:
python scripts/filter_approved_translation.py \
--source-csv "GameFolder/_translation/unity-text-report/text/GameTitle_text.csv" \
--translation-csv "GameFolder/_translation/unity-text-report/validation/text/GameTitle_text_translation_utf8.csv" \
--validation-report "GameFolder/_translation/unity-text-report/validation/text/translation_validation_report.json" \
--out-csv "GameFolder/_translation/unity-text-report/text/GameTitle_text_translation.csv" \
--report-out "GameFolder/_translation/unity-text-report/writeback/approved_translation_report.json"
This is valid only for issues carrying individual row numbers, such as placeholder, tag, or newline mismatches. Do not filter around row-count or source-alignment failures.
Dry-run and write a supported serialized/raw source:
python scripts/writeback_unity_text.py \
--manifest "GameFolder/_translation/unity-text-report/text/GameTitle_resources_text.manifest.json" \
--source-csv "GameFolder/_translation/unity-text-report/text/GameTitle_resources_text.csv" \
--translation-csv "GameFolder/_translation/unity-text-report/validation/resources/GameTitle_resources_text_translation_utf8.csv" \
--out-dir "GameFolder/_translation/unity-text-report/writeback/resources" \
--dry-run
Run the same command without --dry-run only after inspecting skipped rows and planned occurrences. The writer verifies the source file SHA256 recorded at extraction time before modifying a row-mapped source.
After any successful write to a source asset, re-extract that asset before applying another translation batch to it; the stored source hash intentionally rejects stale manifests.
Extract raw profile records when normal extraction misses visible profile text:
python scripts/extract_character_profiles.py \
--asset "GameFolder/Game_Data/resources.assets" \
--out-csv "GameFolder/_translation/unity-text-report/text/GameTitle_profiles_text.csv" \
--manifest "GameFolder/_translation/unity-text-report/text/GameTitle_profiles_text.manifest.json"
Validate the resulting one-column _translation.csv, then dry-run/write it:
python scripts/writeback_character_profiles.py \
--manifest "GameFolder/_translation/unity-text-report/text/GameTitle_profiles_text.manifest.json" \
--source-csv "GameFolder/_translation/unity-text-report/text/GameTitle_profiles_text.csv" \
--translation-csv "GameFolder/_translation/unity-text-report/validation/profiles/GameTitle_profiles_text_translation_utf8.csv" \
--out-dir "GameFolder/_translation/unity-text-report/writeback/profiles" \
--dry-run
Extract Utage book rows from a scenario or embedded asset after confirming the visible text column:
python scripts/extract_utage_scenarios.py \
--asset "GameFolder/Game_Data/StreamingAssets/Utage_AdvProject/Windows/utage_advproject.scenarios.asset" \
--game-root "GameFolder" \
--out-dir "GameFolder/_translation/unity-text-report/text"
This creates GameTitle_utage_text.csv; its manifest records the ALL_Texts.book and ALL_EventTexts.book row identities. Validate its one-column _translation.csv, then write it with:
python scripts/writeback_utage_scenarios.py \
--manifest "GameFolder/_translation/unity-text-report/text/GameTitle_utage_text.manifest.json" \
--source-csv "GameFolder/_translation/unity-text-report/text/GameTitle_utage_text.csv" \
--translation-csv "GameFolder/_translation/unity-text-report/validation/utage/GameTitle_utage_text_translation_utf8.csv" \
--out-dir "GameFolder/_translation/unity-text-report/writeback/utage" \
--dry-run
The extractor now blocks with coverage_complete=false when a requested book is missing or a skipped MonoBehaviour contains raw Japanese candidates. Audit those objects before continuing:
python scripts/extract_mono_raw_utf8.py \
--asset "CandidateRoot/Game_Data/sharedassets0.assets" \
--path-id 126 \
--out-dir "GameFolder/_translation/unity-text-report/residual/mono_raw"
Review the occurrence-level approval CSV, setting every action to translate or preserve, then rerun with --approval-csv. Follow references/mono-raw-residual.md; never approve by text value alone.
Source Review
Per-source extraction is deliberately occurrence-preserving, not an assertion that every row is safe to translate. Before translating or writing:
- Leave product identity/config strings in
globalgamemanagers blank unless changing the save/log namespace is intended.
- Leave resource names, audio/CG keys, animation parameters, control labels, and sample/tutorial blocks blank unless runtime evidence shows they are visible text.
- For an occurrence that shares the same original text with a visible row, translate only the visible row and leave the internal occurrence blank.
- Keep an audit note for intentionally omitted categories when preparing a distributable patch.
- Treat duplicate originals as separate occurrences when a per-source manifest exists; do not project by text value unless the workflow is explicitly legacy/audit-only and duplicate risk is reported.
- Do not convert occurrence-specific choices into a runtime key when duplicate originals differ or mix selected and blank rows; a key-based hook would apply one result globally.
The older broad cleaners can support exploration, but do not use a deduplicated CSV as a row-mapped writeback source.
Special Structures
For a scenario stored as a serialized CSV-like string, inspect its real parser contract before writing. Preserve record separators, in-cell newlines, command columns, quoting behavior, and control tokens. Byte presence alone does not prove the scenario still parses.
For Utage, check both StreamingAssets/Utage_AdvProject and embedded books in sharedassets*.assets. Confirm that column index 8 is the visible Text column before relying on the default; pass --text-column when a project differs. A TypeTree failure is a coverage failure, not a harmless skipped object. Use the raw-object audit for every skipped MonoBehaviour with Japanese candidates. Preserve the ordered CRLF/CR/LF signature recorded in the v2 manifest; marker count alone is insufficient.
For a BansheeGz/BGDatabase payload embedded in a TextAsset, do not infer its serialized field format from exported bytes. If the game's managed API can load and save it, use a temporary runtime probe to export occurrences, compare each selected source value, write translated field values, and save a verified replacement payload. Remove all probe registration and DLLs before smoke-testing the normal game.
For Addressables or IL2CPP metadata, require a backup and an integrity plan before edits. Locate a binary catalog entry from the bundle's internal AssetBundle.m_Name, never from an opaque disk filename alone. When a uniquely identified CRC must be disabled, use patch_addressables_catalog_crc.py to change only that four-byte field in an external catalog candidate; do not guess catalog.hash or bundleSize. Treat bundle plus catalog as one canary and one rollback unit. A candidate that UnityPy can reopen is not approved unless the game runtime loads it without Addressables/bundle errors.
After an Addressables edit causes a black screen, inspect Player.log before revising translations or fonts. Search for CRC mismatch, Will not load AssetBundle, bundle dependency/load errors, and catalog parse/hash/cache failures. Preserve the log and roll back the bundle/catalog pair before trying another candidate.
Font Diagnosis
Treat remaining Japanese as a translation/extraction gap. Distinguish uniform tofu squares from repeated glyph fragments or boxed atlas snippets: tofu indicates missing glyphs, while fragments indicate a font-table, atlas, stream, material, or platform-texture mismatch. A correct source character rendered as another valid Han character indicates a glyph-index conflict: verify the CSV and Unicode first, then inspect the TMP tables instead of changing the translation. Inspect the runtime log and map the visible component to TMP or legacy UI before applying a fix.
Never trust a font package filename or guess the primary font by name. Use trace_tmp_font_usage.py from a visible failing component to establish the actual FontAsset/Material/CAB/bundle chain. Run inspect_tmp_font_bundle.py to read dynamic/static mode, populated tables, source Font PPtr/family identity, Unity version, glyph coverage, atlas metadata, and SHA-256. If a dynamic asset already has character/glyph tables, never replace only its source TTF. Static replacement must move the matched character table, glyph table, material parameters, and SDF atlas together with build_static_tmp_fontasset.py. Require a one-font, one-screen runtime canary before broad replacement. UnityPy reopening is structural evidence, not visual or GPU-runtime proof. Follow references/font-asset-replacement.md for stop conditions and rollback rules.
Trace from a visible string or component name while indexing dependency bundles:
python scripts/trace_tmp_font_usage.py \
--asset "ComponentBundle" --candidate-root "AddressablesBundleRoot" \
--text-contains "Visible failing text" --out "CandidateReports/tmp_usage.json"
When static replacement is justified, pass the traced target FontAsset, Material, and Texture2D PathIDs explicitly to build_static_tmp_fontasset.py, along with a donor, approved translation root, game root, external output, and report path. Exact Unity versions and real donor coverage are required by default; explicit mismatch or donor-alias opt-ins mark the result canary-only.
For compatible Mono/TMP games with ScriptingAssemblies.json and RuntimeInitializeOnLoads.json, prefer the runtime fallback preflight:
python scripts/install_tmp_chinese_font_fix.py "GameFolder" \
--out-dir "GameFolder/_translation/unity-text-report/font-fix" \
--dry-run
The fixer compiles the DLL and validates injection files during preflight, then applies all planned writes with backups only when run without --dry-run. Its default runtime behavior chooses an installed CJK font file, adds a dynamic TMP fallback, and changes legacy UI fonts only for strings with missing CJK glyphs; it does not replace every TMP font. Use --patch-embedded-fonts --font "Microsoft YaHei" only when a reviewed embedded legacy Font replacement is needed.
After installation, launch the game and inspect Player.log. Require an initialization log and no MissingMethodException, Unable to load font face, or repeated missing-glyph failures. Some stripped Mono builds omit Font.CreateDynamicFontFromOSFont or ParameterInfo.HasDefaultValue; use the shipped file-path/reflection-compatible fixer rather than assuming those APIs exist. If the game directory permits content writes but rejects rename replacement, keep the backup, accept verified copy-overwrite only, and remove or report staged .tmp files.
Older Mono builds without the JSON injection tables still require project-specific work in globalgamemanagers or a managed assembly hook. IL2CPP builds require a separately verified IL2CPP loader/plugin or an exact-version resource asset; do not claim the Mono installer applies.
Save Discovery
When asked for save locations, check the game's persistence middleware before assuming Unity's persistentDataPath. For Easy Save 3 (ES3, ES3Defaults, SaveFile.es3, or AutoSave.es3 clues), inspect the serialized defaults resource and confirm an actual modified save file on disk. ES3 may place a save beside game data rather than under AppData\LocalLow; report the confirmed file and distinguish it from PlayerPrefs/display settings.
Patch Delivery
Build a distributable package only after the composed external candidate passes reopen checks, a second source-specific residual audit with zero unreviewed visible occurrences, font verification, and a game-runtime canary. Build from pristine original hashes directly to that final candidate, never incrementally from an earlier translation candidate:
python scripts/build_binary_patch.py \
--game-root "GameFolder" \
--backup-root "GameFolder/_translation/unity-text-report/writeback/resources/backups/<timestamp>" \
--file "Game_Data/resources.assets" \
--out-dir "GameFolder/_translation/unity-text-report/patch/cn_patch" \
--zip
Repeat --file for each modified game file. The package stores SHA256-protected GZip-compressed binary middle-span payloads, an install_cn_patch.bat launcher backed by apply_cn_patch.ps1.txt, and a Python verifier/applicator for development checks. The installer refuses unknown target files, skips already patched files, and creates _cn_patch_backup_<timestamp> in the target game before applying changes.
For changes spread throughout a large resaved asset, the binary middle-span payload may be large; report payload sizes rather than claiming small deltas.
Install the ZIP into a fresh directory populated from the pristine originals. Require the installed hashes to match the live final candidate, confirm automatic backup creation and repeat-install behavior, and record the ZIP SHA256. Any later translation, residual, or font fix invalidates the package and requires a full rebuild and clean install test.
Completion Cleanup
After the validated game build is working, keep the final project-named source/translation CSVs and manifests, validation reports, applied writeback report, one original backup per modified game file, required runtime font DLL/config registration, and any distributable patch. Remove failed candidates, probe DLLs/registration backups, dry-run payload copies, duplicate backups, compile output, and staged .tmp files unless they are specifically needed to diagnose an unresolved issue.
Required Reporting
Report:
- detected build layout and selected adapters
- generated project-named CSVs and their manifests
- validation failures/warnings and normalized translation paths
- dry-run and applied writeback counts, skipped rows, backups, and source-hash checks
- intentionally preserved identity/config/resource rows
- first and final residual-audit counts, every preserved internal category, and zero-unreviewed-visible evidence
- runtime text candidate/report hashes, duplicate-key decisions, external loader/plugin evidence, canary result, and removal path when a runtime adapter was used
- TMP font inspection report, required-character coverage, canary scope, visual result, and rollback path when fonts were changed
- confirmed save path and persistence evidence when save-location work is requested
- patch manifest, payload sizes, installer verification, and package ZIP when distributing
- whether any outputs were produced from a broad audit subset rather than a row-mapped source manifest
Never claim complete translation coverage when unsupported containers or skipped occurrences remain.