ChatCLI was designed to be global. The user interface, including menus, hints, and status messages, is fully internationalized.
Supported Languages
| Language | Code | Status |
|---|
| Portuguese (Brazil) | pt-BR | Complete |
| English | en | Complete (default fallback) |
Automatic Detection
The language is automatically detected from system environment variables, in the following priority order:
| Priority | Variable | Example |
|---|
| 1 (highest) | CHATCLI_LANG | pt-BR |
| 2 | LANG | pt_BR.UTF-8 |
| 3 | LC_ALL | pt_BR.UTF-8 |
Force a Language
To force a specific language, set CHATCLI_LANG in the .env file or in the environment:
# No .env
CHATCLI_LANG=pt-BR
# Ou via export
export CHATCLI_LANG=en
Coverage
The i18n system holds 1925 translation keys covering all user-facing strings (parity across en.json, en-US.json, and pt-BR.json). Every new user-facing string introduced in code goes through i18n.T() wrapping — a mandatory rule for any command handler, prompt.Suggest autocomplete entry, or user-visible error message.
As of the April 2026 i18n sweep, every slash command (/hooks, /mcp, /worktree, /cost, /channel, /websearch, /config + subsections, /skill, /agent, /memory, /session, /switch, /context, /auth, /plugin) and the CLI autocomplete are 100% routed through i18n.T. Hardcoded PT/EN strings inside fmt.Println or colorize(...) count as a conformance bug.
What Is Translated
- Interactive menus and prompts
- Status and progress messages (including history-compactor pipeline phases: trim, summarize, emergency)
- Command help (
/help)
- Error messages and warnings
- Agent mode UI (execution plan, actions)
- Coder mode feedback
- Pre-flight and payload-recovery messages (413/WAF/EOF) with specific labels per failure type
- Server and remote connection messages
- Session, context, and plugin commands
- K8s Watcher and Operator notifications
Main namespaces
| Key prefix | Scope |
|---|
welcome.*, help.* | Onboarding, welcome, help |
agent.* | Agent mode, turns, action plan |
agent.microcompact.* | Progressive compaction of old tool results (no LLM) |
agent.preflight.* | Proactive history-size vs proxy-cap warnings |
agent.recovery.* | Context overflow + proxy 413/WAF/EOF recovery |
coder.* | Coder mode, cards, policy decisions |
compact.* | /compact command (manual and guided) |
compact.status.* | Live feedback during each compaction pipeline phase |
context.* | /context attach, /context show, etc. |
cmd.* | Legacy CLI command keys (/auth, /connect, /switch, …) |
cmd.core.* | /session fork and similar under the core command_handler |
cfg.* | Hierarchical /config — section titles, kv labels, placeholders (cfg.val.*, cfg.kv.*, cfg.sub.*, cfg.panorama.*, cfg.section.*, cfg.msg.*) |
hooks.cmd.* / mcp.cmd.* / wt.cmd.* / cost.cmd.* / chan.cmd.* | One namespace per manager slash-command (/hooks, /mcp, /worktree, /cost, /channel) |
skill.cmd.* / persona.cmd.* / mem.cmd.* | skill_handler, persona_handler, memory_command handlers |
sw.cmd.* / sess.cmd.* / ctx.cmd.* | Fills in cli_llm (handleSwitchCommand), cli_session, context_handler |
ws.cmd.* | /websearch command (status, list, provider, reset) |
complete.* | Autocomplete descriptions — complete.root.* for top-level slash commands, complete.{connect,watch,context,session,plugin,agent,skill,switch,auth,generic}.* for sub/flags |
llm.* | Streaming, provider errors |
server.* | Server mode and authentication |
Special pattern for AI instructions
The ai.response_language key is a meta-instruction written in English across all locales (including pt-BR), which tells the AI to respond in the user’s language. Intentional: LLMs follow English instructions with much higher fidelity, so the prompt that asks for “respond in pt-BR” is itself in English. The response content, of course, comes out in pt-BR.
This instruction is appended to the system prompt across all three modes (chat, agent, coder), together with a [ACTIVE MODE: ...] block that tells the AI which mode it’s operating in. See cli/prompts.go and cli/cli_llm.go.
AI responses are in the language you use in the conversation — i18n controls both the ChatCLI interface and the meta-prompt that instructs the AI. The actual content the LLM generates depends on the model following the instruction (which 4.x models do very well).
