feat: language-orthography plugin — enforce diacritical marks for non-English languages

plugins/language-orthography/.claude-plugin/plugin.json

@@ -0,0 +1,5 @@
+{
+  "name": "language-orthography",
+  "version": "1.0.0",
+  "description": "Enforces full orthographic correctness (diacritics, accents, cedillas) when the language setting targets a non-ASCII language"
+}

plugins/language-orthography/README.md

@@ -0,0 +1,58 @@
+# language-orthography
+
+Enforces full orthographic correctness when the `language` setting targets a non-ASCII language (Portuguese, French, Vietnamese, Czech, Turkish, Spanish, German, etc.).
+
+## The problem
+
+The built-in `language` setting injects a generic instruction like *"Always respond in pt-BR"*, which the model often interprets loosely — producing text without proper accents, cedillas, or other diacritical marks. This is especially noticeable in long sessions and after context compaction.
+
+For example, with `language: "pt-BR"`:
+
+| Expected | What you get |
+|----------|-------------|
+| informação | informacao |
+| não | nao |
+| código | codigo |
+| você | voce |
+
+## How it works
+
+A `SessionStart` hook reads your `language` setting from `~/.claude/settings.json` (or `settings.local.json`) and injects an explicit orthographic enforcement instruction into the session context. This instruction frames diacritic omission as an orthographic error rather than a style choice, which makes the model treat it with the same weight as a spelling mistake in English.
+
+The hook is a no-op if:
+- No `language` setting is configured
+- The language is English (`en`, `en-US`, `en-GB`, etc.)
+
+## Installation
+
+```bash
+claude plugin add language-orthography
+```
+
+Or install from the repository:
+
+```bash
+/install-plugin plugins/language-orthography
+```
+
+## Configuration
+
+No configuration needed beyond the standard `language` setting in your Claude Code settings:
+
+```json
+{
+  "language": "pt-BR"
+}
+```
+
+The plugin reads this value automatically.
+
+## Limitations
+
+This plugin works around the issue at the prompt level. It doesn't fix the underlying causes in the core CLI:
+
+1. The built-in language instruction template should explicitly mention diacritical marks
+2. CLAUDE.md instructions are wrapped in a disclaimer that weakens their authority
+3. Context compaction doesn't carry language rules to the summarization step
+
+See [#32886](https://github.com/anthropics/claude-code/issues/32886) for the full root cause analysis.

plugins/language-orthography/hooks-handlers/session-start.sh

@@ -0,0 +1,41 @@
+#!/usr/bin/env bash
+
+# Reads the user's language setting and injects orthographic enforcement
+# instructions so the model preserves diacritical marks in its output.
+
+SETTINGS_FILE="${CLAUDE_CONFIG_DIR:-$HOME/.claude}/settings.json"
+SETTINGS_LOCAL="${CLAUDE_CONFIG_DIR:-$HOME/.claude}/settings.local.json"
+
+# Try local settings first, then global
+LANG_VALUE=""
+for f in "$SETTINGS_LOCAL" "$SETTINGS_FILE"; do
+  if [ -f "$f" ]; then
+    # Extract language field — handles both quoted and unquoted values
+    LANG_VALUE=$(grep -o '"language"[[:space:]]*:[[:space:]]*"[^"]*"' "$f" 2>/dev/null | head -1 | sed 's/.*:[[:space:]]*"\(.*\)"/\1/')
+    [ -n "$LANG_VALUE" ] && break
+  fi
+done
+
+# If no language setting found, nothing to enforce
+if [ -z "$LANG_VALUE" ]; then
+  exit 0
+fi
+
+# Languages that require diacritical marks (non-exhaustive, covers most common cases)
+# If the language is plain English, skip enforcement
+case "$LANG_VALUE" in
+  en|english|English|en-US|en-GB|en-AU)
+    exit 0
+    ;;
+esac
+
+cat << EOF
+{
+  "hookSpecificOutput": {
+    "hookEventName": "SessionStart",
+    "additionalContext": "ORTHOGRAPHIC ENFORCEMENT (language: ${LANG_VALUE})\n\nThe user's language setting is '${LANG_VALUE}'. All text output in this language MUST maintain full orthographic correctness:\n\n- NEVER omit diacritical marks, accents, cedillas, umlauts, circumflexes, tildes, or any other characters required by the language's writing system.\n- Dropping diacritics (e.g. writing 'informacao' instead of 'informação', or 'voce' instead of 'você') is an orthographic error equivalent to a typo or misspelling in English.\n- This applies to ALL text: messages, code comments, strings, commit messages, file contents, and any other output in ${LANG_VALUE}.\n- This rule persists through the entire session, including after context compaction.\n\nTreat this as a hard constraint, not a style preference."
+  }
+}
+EOF
+
+exit 0

plugins/language-orthography/hooks/hooks.json

@@ -0,0 +1,15 @@
+{
+  "description": "Injects orthographic enforcement instructions at session start based on the user's language setting",
+  "hooks": {
+    "SessionStart": [
+      {
+        "hooks": [
+          {
+            "type": "command",
+            "command": "${CLAUDE_PLUGIN_ROOT}/hooks-handlers/session-start.sh"
+          }
+        ]
+      }
+    ]
+  }
+}