tidyllm

Leave it unset to use the model's default. Supported levels vary by model;

gemini-3.8-flash accepts "low", "medium" and "high", and Gemini itself returns a

clear error for a level the model does not support.

.thinking_budget still works for the current models but now warns that it is

deprecated. Setting both arguments is an error, because Gemini rejects a request that

contains both.

.temperature, .top_p and .top_k stay in gemini(), but Gemini 3.6 Flash and later

ignore them. Google plans to reject them on upcoming models; their help text now says so.

Mistral Large 4

mistral() works with Mistral Large 4 (mistral-large-4, also mistral-large-4-0),

Mistral's new reasoning model with a 512K-token context, image input and tool calls.

It reasons by default; .reasoning_effort now also accepts "none", which turns

reasoning off. Large 4 supports only "high" and "none". Batches sent with

send_batch(mistral()) take .reasoning_effort too.

```r

llm_message("Summarise this contract clause.") |>

chat(mistral(.model = "mistral-large-4", .reasoning_effort = "none"))

```

The default Mistral model does not change, and mistral-large-latest still points to

the previous Large model.

codex_cli() and vibe_cli(): two more command-line providers

Like claude_cli(), these run a command line tool installed on your own machine and

read its output back. codex_cli() runs OpenAI's Codex CLI and vibe_cli() runs

Mistral's Vibe CLI:

```r

llm_message("Explain what a tibble is in one sentence.") |>

chat(codex_cli())

llm_message("Explain what a tibble is in one sentence.") |>

chat(vibe_cli(.model = "mistral-large-4"))

```

codex_cli() uses the login Codex already has. With .use_api_key = TRUE it uses

your OPENAI_API_KEY instead. It reports token counts, takes .reasoning_effort

and .json_schema, and streams, although Codex sends whole messages, not single

words. Codex adds about 12,000 tokens of its own instructions to every call.

vibe_cli() uses your MISTRAL_API_KEY. Vibe reports no token counts or cost and

sends its reply in one piece. Vibe has no system prompt setting of its own, so

vibe_cli() puts the system prompt in front of the prompt text; set

.system_prompt_mode = "ignore" to leave it out.

Both are agents that can run commands and change files. tidyllm turns that off

unless you set .cli_tools = TRUE. Both can keep a conversation on their side with

.stateful = TRUE, both work with send_chat(), and .dry_run = TRUE shows the

command they would run.

Bug fixes

• claude_cli() with its tools off (the default) could still read files, because the Claude CLI kept its read-only tools when the allowed list was empty. It now removes every built-in tool and leaves out your MCP servers, so a plain chat() can no longer read files in the working directory. .cli_tools = c("Read", "Glob") now makes exactly those built-in tools available.
• parallel_chat() with claude_cli() now stops with a message that the provider does not support it. Before, it failed with an error from httr2. The same holds for codex_cli() and vibe_cli(); use send_chat() or chat() in a loop instead.
• Printing the command from a .dry_run of a command-line provider now quotes each argument, so it can be pasted into a shell.
• An image attached with img() is now an error for providers that cannot take images (claude_cli(), codex_cli(), vibe_cli() and deepseek()), in chat(), send_chat() and parallel_chat(). Before, the command-line providers sent the text alone and dropped the image without a word, and deepseek() sent the image to an API that does not take images. provider_capabilities(.what = "media") now also lists images for ellmer(), which has always sent them.
• A streamed request that the provider rejects now shows the provider's error message, for every provider that streams over HTTP. Before, it showed only the HTTP status, such as "HTTP 400 Bad Request.".
• mistral() now shows Mistral's own error message, for example for an unsupported .reasoning_effort or an invalid API key, instead of "Received empty response from Mistral". The same holds for Mistral embeddings, list_models() and the batch functions; checking a batch id that does not exist is now an error instead of an empty table.
• Deprecation warnings raised inside a chat request, such as for .fileid in gemini() and .file_ids in claude(), no longer tell you to report a bug in tidyllm when you call them through chat(), %>%, purrr::map() or tryCatch().
• The deprecated .fileid argument of gemini() no longer also warns about gemini_file_metadata(), a function you did not call.