claudish-to-english

A Claude Code plugin that shows a plain-language rewrite of every assistant message directly under the original. The rewriter is a model you choose with one environment variable: a local ollama model by default, or the codex CLI, the Anthropic API or any OpenAI-compatible endpoint. It is display-only — the saved transcript and what Claude itself sees keep the original text, only what is rendered on screen changes — so the hook cannot alter or swallow an answer, and a failure of any kind leaves the original on screen. An optional second hook rewrites Markdown files as an agent writes them, either in place or to a NAME.plain.md sibling. The /claudish command switches rewrites on and off and changes the display mode, the style preset, the output language and the model mid-session, and a dashboard prints each setting with the place its value came from, flagging every override that will outlive the session.

Who built itThirty-seven commits and nine contributors, but this is one person’s project: 21 commits carry his own GitHub account and six more are his name from two different Macs. He signs his review comments “Human Mike”, and 24 of the 37 commits carry a co-author trailer — 21 of those name a Claude model, with Opus 5 on a one-million-token context appearing eleven times. The rest of the history is other people’s: the provider layer, an oauth login path, a caveman style preset and a slash command all arrived as pull requests.

1. 01

   Eighteen days, twelve versions, and 2,707 stars

   The repository was created on 2026-08-10T18:32:23Z and its first commit landed twelve minutes later, titled “Initial release of claudish-to-english v0.1.0”. Everything the README describes was built in the eighteen days that followed: 37 commits, all of them in August 2026, the last on 2026-08-28. The changelog records twelve versions — 0.1.0 and 0.1.1 on the first day, 0.2.0 and 0.3.0 on 2026-08-13, 0.4.0 on 2026-08-14, 0.5.0 on 2026-08-19, 0.5.1 and 0.6.0 on 2026-08-20, 0.7.0 on 2026-08-26, 0.7.1 and 0.8.0 on 2026-08-27, and 0.9.0 on 2026-08-28 — while tags number thirteen, those twelve plus one called assets, and the GitHub releases list holds only three entries: v0.9.0, v0.8.0 and the assets release carrying the README comparison image. It finished the month with 2,707 stars, 129 forks, ten open issues and thirty issues and pull requests in total. Nine people have contributed, and 24 of the 37 commits carry a co-author trailer, 21 of them naming a Claude model — Opus 5 on a one-million-token context eleven times, Opus 4.8 nine times across two context sizes, and Fable 5 once.

2. 02

   The hook fires once per streamed chunk, not once per message

   The central fact about the display hook is that Claude Code fires MessageDisplay once per streamed chunk, not once per message, and each firing is a separate process carrying only that chunk’s delta. So the plugin appends every fragment to a temp file keyed by session and message id, emits nothing, and calls the model exactly once — on the chunk flagged final, when the whole message is known. That call also reads the original user question out of the transcript and passes it in as context only, with an instruction never to answer or repeat it. Two display modes follow: append, the default, streams the original and adds the rewrite underneath, and replace, marked experimental, suppresses the chunks and shows only the rewrite after the model answers. Being display-only is what makes the design safe, and the repository keeps a long argument about it: issue 12 opened with a report that the hook fired, emitted schema-correct output and rendered nothing; the reporter narrowed that to an interactive-terminal bug, then retracted it entirely — the cause was "verbose": true in his own settings file, because verbose mode deliberately shows the original text and bypasses the replacement.

3. 03

   Four providers, and a default that never leaves the machine

   The first thing users asked for was not to need ollama: issue 9 is three lines — “Would love to try this if it didn’t need Ollama” — plus a reply asking for a model-agnostic version. nircoh-onyx answered with a pull request that lifted the model call out of both hooks into a shared providers.sh, keeping ollama the default and putting any OpenAI-compatible endpoint and the Anthropic Messages API behind CLAUDISH_PROVIDER. The maintainer’s review added two things the contributor had not: a rewrite cut off by an output-token cap is discarded rather than shown, because in the Markdown hook’s overwrite mode it would replace a real document with half a rewrite; and the HTTP status is captured, so a failure can be explained. A codex CLI provider followed, contributed by datvo06 and bounded by a poll-and-kill loop because stock macOS ships no timeout(1). Then JackBhanded’s oauth mode, which rides the Claude Code login instead of an API key: a tester found the token in the macOS login Keychain rather than the credentials file and contributed that read with an end-to-end http=[200] proof, after which the maintainer made the mode refuse to run when the base URL is overridden, so a subscription token cannot reach a proxy. The README still prints the warning it earned: unofficial, use at your own risk, repeated on screen at the first rewrite of each session.

4. 04

   A slash command that arrived twice, and a pronoun bug that arrived once

   MakhBeth had built a /claudish command in a fork called claudish-tldr; the maintainer was building the same thing, replied “I was already working on the integration of your slash command into the plugin”, merged his own version as a separate pull request, credited him under the wrong handle, apologised, and ported the one piece he had missed. The command interpolated $ARGUMENTS unquoted into the shell line, and the 0.5.1 fix replaced the quoting with a quoted here-doc on stdin, because Claude Code substitutes the argument textually before the shell parses the line, so quotes inside the command could not neutralise $(…), backticks or a bare *. A second outsider contribution became the caveman style preset, and arrived with coloured labels the maintainer tested live, found working, and replaced anyway: the 16-colour codes are palette indices a terminal theme may remap onto grey, and the escapes leak as literal bytes into claude -p output piped to a file. The labels became an emoji and a bold word. The last community round is the sharpest: issue 24 shows two rewrites with the roles inverted, “Found the root cause.” coming back as “You’ve found the root cause.” The maintainer asked which model was in use; the reporter answered with the prompt paragraph he had written himself, and 0.9.0 put a version of it in every prompt.

5. 05

   The rules were written down after they had already been broken

   The last release is documentation. 0.9.0 adds CONTRIBUTING.md, a pull-request template, a CLAUDE.md that Claude Code picks up automatically, and a /release skill that cuts a version in two phases — and it exists because the release process and the repository’s traps lived, in the maintainer’s own words, “only in review comments and in whoever happened to remember them”. What those documents record is what had already gone wrong. A new style preset is read in four places, and the caveman preset shipped with one of them — the session-start notice — missing, so a style could persist across sessions with nothing on screen to say so. The allowed-tools line in the slash command has to close its quote after the path, and it had regressed twice, once through a merge-conflict resolution, leaving /claudish unusable in every release from 0.5.0 to 0.7.0. A language value in a project’s settings file — a file that travels with a repository — could carry a terminal escape sequence onto the screen, because the sanitiser folded whitespace and an escape is not whitespace. The rule that outranks all of them is stated first and repeated in every document: every hook must fail open. There is no test suite; a change is verified by piping synthetic payloads into the hook and checking that three malformed inputs print nothing and exit 0.