[BUG] Armenian script generation broken - outputs "delays" instead of Armenian characters
Preflight Checklist
- [x] I have searched existing issues and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code
What's Wrong?
When generating Armenian text, Claude outputs the English word "delays" instead of Armenian characters. The first 1-2 Armenian characters may render correctly, then all subsequent characters are replaced with the literal string "delays".
Example output when asking for "Hello" in Armenian:
Բdelays delays delays!
Instead of:
Բարեւ
This bug affects ONLY Armenian out of 58 tested scripts/languages. Even rare scripts like Cherokee (ᎣᏏᏲ), N'Ko (ߊߟߎ߫), Javanese (ꦲꦭꦺꦴ), and Tibetan (བཀྲ་ཤིས་) work correctly.
The bug is also reproducible in claude.ai web interface, indicating this is a model-level issue, not specific to Claude Code.
What Should Happen?
Claude should output proper Armenian text. For example:
Input: "Write 'Hello, I received a notification' in Armenian"
Expected output:
Բարեւ, ես ծանուցում եմ ստացել
(Full Armenian text with proper Unicode characters from the Armenian block U+0530–U+058F)
Error Messages/Logs
No error messages - the model completes successfully but outputs incorrect text.
Sample corrupted output:
Բdelays delays delays delays delays delays delays delays! Delays delays delays delays Postcard-delays delays delays delays delays delays delays delays delays delays delays delays delays delays delays delays delays delays delays delays delays delays delays delays delays delays delays.
The word "delays" consistently replaces Armenian characters.
Steps to Reproduce
- Open Claude Code CLI (or claude.ai web interface)
- Enter a prompt requesting Armenian text generation, for example:
- "Write 'Hello, how are you?' in Armenian"
- "Translate this to Armenian: I received a notification from the delivery service"
- Observe the output - Armenian characters will be replaced with the word "delays"
Alternative test:
- Ask Claude to write an email in Armenian
- Notice that after the first 1-2 characters, all Armenian text becomes "delays delays delays..."
Verification that this is Armenian-specific:
- Test the same prompts with Georgian: Works ✅ (გამარჯობა!)
- Test with Greek: Works ✅ (Γεια σας!)
- Test with Hebrew: Works ✅ (!שלום)
- Test with Arabic: Works ✅ (!مرحبا)
- Test with Armenian: Broken ❌ (Բdelays delays delays!)
Claude Model
Opus
Is this a regression?
I don't know
Last Working Version
Unknown - not tested on previous versions
Claude Code Version
2.1.3 (Claude Code)
Platform
Anthropic API
Operating System
macOS
Terminal/Shell
Other
Additional Information
Comprehensive Multi-Script Testing
I tested 58 different non-Latin scripts/languages. Results:
| Status | Count | Languages |
|--------|-------|-----------|
| ✅ Working | 57 | Ukrainian, Bulgarian, Serbian, Belarusian, Russian, Macedonian, Mongolian (Cyrillic), Kazakh, Kyrgyz, Tajik, Greek, Georgian, Hebrew, Arabic, Persian, Urdu, Pashto, Kurdish, Uyghur, Sindhi, Yiddish, Thai, Lao, Khmer, Myanmar, Tibetan, Chinese (Simplified), Chinese (Traditional), Japanese, Korean, Hindi, Sanskrit, Nepali, Marathi, Bengali, Gujarati, Punjabi, Tamil, Telugu, Kannada, Malayalam, Sinhala, Odia, Assamese, Amharic, Tigrinya, N'Ko, Cherokee, Inuktitut, Dhivehi, Javanese, Balinese, Sundanese, Uzbek, Bashkir, Chechen, Chuvash |
| ❌ Broken | 1 | Armenian |
What this rules out
- ❌ NOT a general Unicode issue (57 other scripts work)
- ❌ NOT a non-Latin script issue (Georgian, Greek work)
- ❌ NOT a RTL issue (Hebrew, Arabic, Persian, Urdu work)
- ❌ NOT a multi-byte UTF-8 issue (all tested scripts use multi-byte)
- ❌ NOT a low-resource language issue (Cherokee ~2k speakers works, Armenian has 3M)
- ❌ NOT a script complexity issue (Tibetan, Myanmar, Khmer work)
Reproducibility
- ✅ Reproducible in Claude Code CLI
- ✅ Reproducible in claude.ai web interface
- ✅ Consistent replacement with "delays" (not random corruption)
Impact
Armenian is spoken by ~3 million native speakers and is the official language of Armenia. This bug makes Claude unusable for Armenian language tasks.
9 Comments
As a workaround, downgrading Claude Code to 2.0.0 fixes the problem temporary
This can be reproduced in current Claude 4.5 Opus.
I can confirm. As a workaround, I use cursor for Armenian texts, and i18n library, and ask cursor to write json translations there.
Claude Opus 4.6 has the same problem but it now speaks some more tokens in Armenian before
Delays Delayscoming and also now it has some newDefaults Defaultsin the response@chrislloyd , Claude team. Please look at this
The Write and Edit tools corrupt Armenian Unicode characters (U+0530–U+058F range) during file write operations, replacing them with garbled bytes or stripping them entirely or causing the
Delays.The issue appears to stem from an encoding layer that mishandles multi-byte UTF-8 sequences for non-Latin scripts. As a workaround, Armenian text can be generated programmatically via Python chr(), but not a viable one.
Fresh reproduction — Claude Opus 4.6 (1M context), 2026-03-29
Still broken. Here is what just happened in a live Claude Code session:
Context: User asked Claude to interact with an Armenian-language website via Playwright MCP. The page snapshot contained standard Armenian UI labels (e.g. the Login button text). When Claude attempted to reference this Armenian text in a tool call, it corrupted the output by inserting the English word "delays" in place of Armenian characters mid-word.
The model literally cannot write Armenian words it just read from a page snapshot. It inserts "delays" into Armenian Unicode character sequences.
Generic reproduction — ask Claude to write "Hello, how are you?" in Armenian:
This is not a cosmetic issue — it makes Claude Code unusable for any Armenian-language workflow: i18n files, web scraping, content management, UI testing, localization, etc.
Why this keeps getting deprioritized
I inspected the repo's GitHub Actions workflows and scripts. Here's how the automation works against this issue:
claude-issue-triage.yml— Uses Claude (viaclaude-code-action) to auto-label new issues. It labels Armenian bugs asexternal+area:model, pushing them outside the Claude Code team's scope. Ironically, Claude is triaging its own bug and deciding it's someone else's problem.sweep.yml— Runs twice daily. Labels issuesstaleafter 14 days of inactivity, then auto-closes them 14 days later. This issue has been labeledstaletwice already (2026-02-20, 2026-03-21) — not because it's resolved, but because no Anthropic team member ever engaged.auto-close-duplicates.yml— Runs daily. When the triage bot marks new Armenian issues as "possible duplicate" of this one, and nobody responds within 3 days, it auto-closes them. This killed issues #29332, #36969, and #25635.Net effect: New reports get auto-closed as duplicates. The original stays open but goes stale. Nobody at Anthropic ever reads it.
@ThariqS @bcherny @chrislloyd @shawnm-anthropic @dicksontsai @ashwin-ant — tagging directly because this issue has had zero team engagement since it was opened, and the automation layer (stale bot + duplicate closer) keeps suppressing it. A 30-second acknowledgment would go a long way.
Claude's support told me that this issue had been passed to Engineering team 2 months ago, still no response, it's really frustrating, it's the only popular LLM that can't speak a 10M speakers' language
<img width="998" height="1024" alt="Image" src="https://github.com/user-attachments/assets/d13c7865-b875-41be-802a-4ee4bb8a7152" />
<img width="992" height="887" alt="Image" src="https://github.com/user-attachments/assets/0dba1acc-1cdb-4182-93e8-b3f213c06cf9" />