Changelog

All notable changes to Translation Workbench are recorded here.

[0.2.3] - 2026-09-29

Changed

• Finalization may condense the user’s recorded reasons when they are verbose, as voice dictation often is: fillers, repetitions, false starts, mis-transcriptions, and unrelated asides are removed while reasoning steps, examples, and judgment words are kept. A condensed version is labeled and shown to the user; once confirmed, it counts as the user’s own words. This keeps review notes compact for later distillation.

[0.2.2] - 2026-09-28

Changed

• Finalization scopes the model’s agreement analysis to what changed. Kept drafts need none; revised passages explain what was wrong with the draft and how the final wording solves it; when the user already gave a reason, the model adds only what they left out; mechanical fixes get a one-line note without confirmation. A rewritten analysis needs fresh confirmation.
• Distillation now uses user-confirmed model analysis as a source of reasons, weighted below the user’s own words, and checks reasons that rest on it alone with the user before they enter a reference.

[0.2.1] - 2026-09-28

Changed

• Finalization now presents the model’s reason for agreeing with proposed wording and waits for the user to confirm that understanding before marking the passage confirmed or advancing. The user need not supply a separate explanation.
• Review notes distinguish pending wording, model-written analysis, the user’s own reasons when provided, and explicit user confirmation. Updated the template and both READMEs to match.

[0.2.0] - 2026-09-25

Changed

• Drafting notes now list actual translation choices, source passages for the user to inspect, and target-language passages that still read awkwardly. The translation stage checks subject carryover, broken causal links, and repetition across sentences where relevant to the target language.
• Translation preserves a separate initial-draft copy. Finalization creates its own passage-by-passage decision record and no longer requires an independent-review stage or review checksum markers.

Added

• Added a separately invoked translation-distillation skill for user-started cross-unit reference updates, with a Chinese Claude command shortcut. Publishing and dashboard work remain outside both skills.

[0.1.2] - 2026-09-18

Added

• .claude-plugin/marketplace.json: the repository now publishes itself as a Claude Code plugin marketplace holding one plugin, so /plugin marketplace add Alexu0317-FATHER/translation-workbench works. Before this the command failed with Marketplace file not found, because the repository only ever carried a bare skill and Claude Code had nothing to read. The skills.sh route is unchanged and remains the only one that covers Codex.
• CONTRIBUTING.md: how to report a problem, what a workflow change needs to argue, and the checks a pull request has to pass.

Changed

• scripts/validate_repository.py also checks the marketplace manifest: that the plugin entry exists, points at ./skills/translation-workbench, and cites the same version SKILL.md declares. The manifest is a third place the version number appears, so CI holds it to the single source rather than letting it drift.
• scripts/validate_repository.py no longer hardcodes the version number. It reads metadata.version out of SKILL.md, checks that it is a semantic version, and then checks that both READMEs cite the same one. A release bump now has a single source, and CI names whichever document was left behind instead of only reporting that the number is wrong.
• The bundled showcase was replaced by a single screenshot in the README. The translated chapters it demonstrated now live in their own repository, franz-lohners-chronicle-zh, so this repository stays generic and small.

Removed

• showcase/ and the validation that checked its HTML dependencies.

Documentation

• Both READMEs rebuild the installation section into four labelled routes, easiest first: asking your agent to do it from a chat box, the Claude Code plugin, skills.sh, and copying the folder by hand. The chat-box route spells out what project scope and global scope mean for someone who has not met the distinction before.
• Invocation is its own top-level section instead of a trailing paragraph under one installation route, since which agent you invoke the skill from has nothing to do with how you installed it.
• Both READMEs now carry the skills.sh install-count badge, linking to the skill’s page in that directory. The directory ranks skills by installs reported from the skills CLI, which is the same command the READMEs already give for installation, so the badge is the one entry point a reader can use to check the project’s standing there.
• The README screenshot was replaced. It used to show a bilingual reading page, which separate build scripts produce rather than this skill. In its place is a table taken from chapter 4 of the same project: one source line, what the independent review flagged, the AI’s draft, and the finalized wording. That is what the skill actually produces, and it shows why the person has to be in the loop.
• Both READMEs were rewritten around what the workflow is for: the AI supplies the reasoning behind a rendering, not just the rendering, so that a translator whose command of the source language is uneven still keeps every substantive decision. Added a worked example from chapter 1, a stage table of what the person and the AI each do, an example screenshot, and links to the published bilingual chapters.
• Both READMEs now state the installation scope. npx skills add installs into the current project by default, and -g installs into the user account instead; neither was mentioned before. The directories a project install writes to and the scope flags for npx skills update are documented alongside them.

[0.1.1] - 2026-09-04

Added

• Simplified Chinese README with bidirectional language links.
• A standalone showcase containing the translated chapter index, five bilingual chapter pages, and their referenced local images.
• Automated validation for local HTML links and image dependencies in the showcase.

Changed

• Public-content validation now permits project-specific material only inside the explicitly separated showcase while keeping the installable skill generic.

[0.1.0] - 2026-09-04

Added

• Cross-runtime Agent Skill for Codex and Claude Code.
• Adaptive project initialization and existing-material intake.
• Source preparation, translation, independent-review, and finalization guidance.
• Templates for project documentation, terminology, context, drafting, and review.
• Read-only context and stage checkers with JSON output.
• Unit and command-line tests for context states, glossary validation, stage gates, review-file protection, and draft checksums.
• MIT License.