Add dynamic English translation mod

Please make me a mod for Amazing Cultivation Simulator that translates all text to English dynamically.

I’m treating this as a usable, installable mod—not just a code sketch. I’ll first load your project directives plus the `apocky-loop` and `critical-analysis` skills, then inspect the local game/runtime surfaces and current modding interfaces. I’ll keep translation credentials optional, add persistent caching and fail-safe UI behavior, and verify what I can without assuming the game is installed.

The first load-bearing fork is now clear: the game’s supported Workshop API may only cover authored data, while truly dynamic coverage likely requires a runtime loader that can observe rendered strings. Before locking that in, I’m using the loop’s decision, delegation, and security references—and the CSL reference required by your global instructions—then I’ll verify the installed game (if present) and primary-source mod APIs. No files are being changed yet.

✓ The game is installed locally as Steam app `955900`, so we can replace guesswork with direct evidence from its managed assemblies and shipped mod examples. I’m going to inspect only the game’s files; installation remains untouched. If its runtime confirms a stable hook surface, that will determine the blueprint.

The install includes the developer’s own `APIDoc.rar`, English mod README, and a dedicated “Mod Translation” document. I’m using the `documents` skill because that official `.docx` is likely the best evidence for the native path; this is still read-only research and won’t alter the game.

The document renderer hit a missing local conversion binary (`WinError 2` before any page output). Per the engineering loop, I’m loading the debug pipeline, checking the bundled binary paths once, then falling back to structural DOCX extraction if LibreOffice truly isn’t present. This affects only visual layout review of the official guide, not its text content.

The native route is now directly evidenced: ACS ships a Mono/Unity 2017.4 runtime, exposes every FairyGUI object’s virtual `text` property plus `GRoot` traversal to Lua, calls `OnStep`, and ships asynchronous `UnityWebRequest`. That lets the mod stay inside the supported Workshop format—no DLL injector. Blueprint now locked to the request: scan rendered FairyGUI text incrementally; translate only CJK-bearing, non-editable text; preserve markup/placeholders; deduplicate and cache; use bounded asynchronous requests; never replace text after its source changes; retain originals on every error. I’ll support a local LibreTranslate endpoint and an optional OpenAI endpoint so privacy/cost remains an explicit user control. I’m loading `openai-docs` only to verify that optional provider’s current wire contract. Rollback is deletion of one mod folder; I will not install it into the live game without separate permission.

A fresh primary-source pass changed the implementation choice: current XUnity.AutoTranslator already ships a FairyGUI adapter, async translation, batching, and disk caching. Building a second translator pipeline would duplicate the hardest logic and cover fewer text frameworks. I’m rotating the blueprint to a tailored ACS distribution: pinned BepInEx 5 + pinned XUnity 5.6.1 + an ACS-safe English configuration, install/uninstall tooling, provenance hashes, and a coverage/limitations ledger. I’ll verify the adapter and package layout directly before downloading anything. The custom native Lua path remains the rollback option if the existing FairyGUI hook fails ACS-specific acceptance.

The exact package lane is now verified: the installed ACS executable is x64 Mono, so the compatible pair is BepInEx `[redacted:ip-address]` x64 and XUnity `5.6.1` BepInEx. I need network permission only to download those two official GitHub release archives into `work/`; I’ll hash and inspect them before packaging, with no game-directory writes.

Permission is back and the research lanes have converged. One important denominator correction: ACS also uses a small number of Unity UI/TextMesh paths, so “all” cannot honestly mean baked image text or every custom renderer. The XUnity package still wins because it covers FairyGUI plus UGUI/TextMesh frameworks, while the tailored config will enable only the observed safe surfaces. I’m downloading the two pinned archives now, then I’ll verify their hashes and inspect every payload file before assembly.

The first download attempt failed at PowerShell’s HTTP/TLS authentication layer before creating either file; no partial archive exists. I’m changing one variable only—using Windows `curl` against the same official URLs—then I’ll compare the XUnity hash to the maintainer-published SHA-256.

✓ Both archives now match the maintainers’ published SHA-256 digests. Before assembly I’m checking for path traversal, unexpected executables/install scripts, duplicate target paths, and the exact XUnity configuration surface; then I’ll build the tailored layer around unchanged upstream bytes.

The reusable path is now settled: I’m packaging the current BepInEx/XUnity stack with ACS-specific framework coverage, explicit network consent, and ownership-aware rollback. Next I’m validating the exact config schema, then I’ll exercise install/uninstall against an isolated fake game root before producing the ZIP.

Two PowerShell one-liners exposed shell-parsing edge cases, so I’ve paused the fragile inline approach and narrowed the probe. The archives themselves extracted cleanly; I’m switching the schema inspection to a deterministic Cecil reader and will keep the package source testable rather than relying on ad hoc commands.

The package source is assembled, including a non-network passthrough mode and an explicit-consent Google mode. A schema probe also caught and neutralized an upstream security-sensitive default: certificate-validation bypass is now explicitly forced off in the ACS config. I’m entering the lifecycle tests now—fresh install, shared-loader coexistence, provider reconfiguration, and rollback.

The independent security pass found a real junction-based containment bypass in the installer path primitive, plus a rollback edge case for pre-existing identical configs. I’m holding the release, replacing lexical containment with reparse-point rejection on every existing ancestor, and adding regression cases for both before I package anything.

I’m closing the verified release blockers now: manifest authority, reparse-safe mutation, replacement consent, rollback ownership, and Windows PowerShell test reliability. After that I’ll rerun the package through both shells, adversarial lifecycle tests, assembly loading, checksum validation, and malware scanning before producing the ZIP.

The hardening also tightens the consent boundary: changing providers will rebuild the translator configuration from the package’s known-safe template, clear fallback/custom service URLs, keep certificate validation enabled, and verify the installed translator binaries before Google translation can be enabled. Existing settings are backed up first.

The expanded adversarial suite is running now. It covers both normal lifecycle behavior and the exact failure modes the independent reviewers reproduced, including junction redirection, manifest tampering, directory collisions, provider-binary tampering, and byte-identical rollback ownership.

I’m also correcting the redistribution bundle itself: the upstream loader includes transitive components with licenses beyond the two top-level MIT notices. I’m adding the exact tagged license texts and source pointers so the final package is not merely functional, but properly attributable and redistributable.

The final audit caught a real redistribution blocker: two missing transitive licenses and an unresolved SharpZipLib source-version mismatch. I’m holding the release archive, fixing what can be fixed from exact upstream evidence, and will only ship once the source/license chain and both-host lifecycle tests are clean.

The SharpZip mismatch now has a promising exact provenance path: the embedded assembly is Mono’s .NET 2.0 build of SharpZipLib 0.84—the source explicitly emits assembly version `[redacted:ip-address]`. I’m comparing it against Mono’s archived 2.6.7 binary, then I’ll include the matching source/build tree rather than the incorrect standalone 0.84 archive.