NEWS

agenticr 0.3.3 (2026-07-28)

CRAN compliance: added missing \value tags to 6 Rd files
CRAN compliance: switched from ~/.agenticr/ to tools::R_user_dir() for data storage
CRAN compliance: removed assign(.GlobalEnv) monkey-patching (reviewer confirmed eval(.GlobalEnv) is acceptable for interactive tools)
CRAN compliance: added SystemRequirements and cran-comments.md
Fixed agentic_skills() showing NA bytes
Fixed agentic_enable() cli formatting crash when no API key configured
Removed dead code: write_turn_history(), orphan roxygen, unused MCP fields, todos
Replaced getExportedValue() (always errored) with getFromNamespace()
Pure-R grep fallback when neither rg nor grep is available
Windows cross-platform fixes: Sys.chmod(), rc.settings(), path regex, findstr fallback

Default provider is now local (Ollama Qwen3-1.7B) — zero config needed
First-run auto-detection: if no API key, offers to install Ollama automatically with platform-specific commands (brew/curl|sh/winget)
/provider and /model slash commands for switching providers and models mid-session
agentic_config() renamed params: api_base → base_url, api_model → model (old names accepted for backward compat)
Per-provider max_tokens defaults with fallback in setup wizard
Bundled config-api skill with provider reference data and doc URLs
LLM tests: 100% pass rate against local Qwen3-1.7B model
Test improvements: agent-loop tests for multi-step queries, explicit fail guards on all if/else branches, removed unreasonable "fix mtcar typo" test

Per-provider max_tokens defaults (auto-set when switching providers)
agentic_setup() wizard now asks for max output tokens
Custom model and max_tokens preserved when re-selecting the same provider
reasoning_effort and api_key no longer leak across provider switches
Eliminated CRAN NOTE for global assignments by using options()-only pager control
Fixed MCP server connection (processx API compatibility, JSON-RPC id type)
MCP capabilities now sent as {} instead of []
Fixed get_function_help tool blocking on interactive pager
Fixed help() and ? blocking in tool_execute_r_code via pager option
Exclude editor/IDE folders from tarball and git
Updated config template and documentation

Initial release
AI-powered R console assistant with natural language interface
Supports DeepSeek, OpenAI, and compatible LLM providers
REPL mode and error-interceptor mode
Tool framework: read_file, list_files, execute_r_code, memory_write, task management
Skill system with anthropic-style frontmatter loading
MCP server support
CRAN-ready