NEWS

llmjoin 0.3.0 (2026-06-16)

Removed check_joint() — LLM-based joint validation was unreliable (circular LLM self-validation) and its parsing logic was fragile.
Removed magrittr dependency. The %>% re-export is no longer provided. Use the native R pipe |> (available since R 4.1.0).

Changed cat() to message() for user-facing output in set_llm() and chat_llm(), complying with CRAN best practices.
Removed unconditional cat() preview from parse_joint().
Fixed joint_prompt() @returns tag (incorrectly stated data.frame; now correctly states character string).
Added missing @returns tag for set_llm().

remove thinking mode support, the author does not have the energy to set up individual thinking adaptation interfaces for the countless LLMs.
build_joint() now takes a raw LLM response string instead of data.frames. Signature changed from build_joint(x, y, key1, key2, ...) to build_joint(llm_response, key1, key2). The LLM call and prompt construction are now orchestrated by llm_join().

llm_join() explicitly orchestrates the full pipeline: joint_prompt() → chat_llm() → build_joint().

Config file moved from ~/.LLMJOIN.yml to tools::R_user_dir("llmjoin", "config"), complying with CRAN policy on writing to the user home directory.
Added missing @examples to set_llm() and check_joint().
Set LazyData: false in DESCRIPTION to avoid CRAN NOTE about missing data directory.
Removed deprecated %>% re-export, validate_llm_config(), and test_llm_service_minimal().

Thinking/reasoning support for all providers. chat_llm() enables max reasoning intensity by default. Claude: thinking: {type: enabled, budget_tokens: 16000}. OpenAI / Gemini: reasoning_effort: "high".
Provider layer (R/providers.R) with pluggable .providers registry supporting OpenAI, Claude (Anthropic), and Gemini. Four internal helpers handle per-provider auth, request body, response parsing, and URL construction.
Config caching: validate_llm_config() writes VERIFIED: true into the YAML after a successful probe, skipping redundant network checks.
test_llm_service_minimal() for one-token connectivity probing.
%||% null-coalescing operator.

Replaced tidy-R dependency calls with native R functions, removing magrittr from Imports.
Fixed chat_llm() input validation: missing .message now properly errors.
Config validation errors now include the file path and suggest set_llm().
Modularized auth header construction per provider (provider_headers()).

Renamed R/network-utils.R → R/connection.R, R/main.R → R/llmjoin.R.
Added R/utils.R with pipe re-export, %||%, httr/jsonlite imports.
Added comprehensive test infrastructure using testthat (>= 3.0.0) with local_mocked_bindings mock patterns for offline testing.

check_joint() for LLM-based joint validation. Asks the LLM to identify problematic mappings and filters them from the result.
joint_prompt() for generating structured matching prompts from two key columns.
tbl2md() for converting data.frames to markdown tables.
build_joint() for constructing fuzzy-join mapping tables via LLM.
llm_join() for end-to-end fuzzy join: build_joint + optional check_joint + merge.
set_llm() for configuring LLM service credentials (provider, URL, key, model) stored in ~/.LLMJOIN.yml.
chat_llm() as the single entry point for LLM API calls, supporting OpenAI, Claude, and Gemini providers.