Replace unsafe string concatenation patterns using strcpy/strcat with explicit memcpy calls and proper null-termination in bash_repair.c, context_manager.c, context_summarizer.c, and db.c. Add missing null-pointer checks for strdup results in bash_executor.c. Introduce atomic_int for spinner_running flag in http_client.c. Add consecutive_tool_error_iterations field to agent struct and refactor agent heartbeat/token tracking to use messages_get_db instead of db_open. Remove r_config_get_max_tokens declaration from r_config.h. Add json_repair.c to Makefile build sources and include db.h in messages.h with new messages_get_db function declaration.
Implement spawn_tracker module to enforce recursive agent spawning limits via R_MAX_SPAWN_DEPTH and R_MAX_TOTAL_SPAWNS environment variables, replacing the previous max_subagents parameter in tool_agent.c. Add tool_deepsearch.c implementing an iterative multi-query research algorithm with configurable system message, and register the new source files in Makefile build targets.
Introduce `include/markdown.h` with retoor copyright header and add `r_process_result_t` struct plus `r_bash_execute_ext` / `r_process_result_free` declarations to `include/bash_executor.h` for advanced async process handling.
- Add agent_set_is_subagent and agent_set_tool_registry to agent API for subagent support
- Add timeout_seconds parameter to r_bash_execute to prevent hanging processes
- Refactor context_manager truncation logic with dynamic ratio-based truncation and total size calculation
- Add messages_to_json_string helper and r_config_get_max_tokens accessor
- Introduce tool_registry_get_specialized for type-filtered tool registries
- Add completion_phrases and passive_phrases detection to agent completion logic
- Suppress spinner in http_client when stderr is not a TTY
- Add comprehensive agent_benchmark.py with TestCase framework and logging
This commit deletes the complete implementation of the agent loop, authentication module, web browsing utilities, chat prompt builder, database helper, and HTTP client. All corresponding header files and test harnesses are also removed. The diff shows the deletion of 6 source files and 6 header files totaling over 1000 lines of code, including the agent state machine, API key resolution logic, curl-based HTTP functions, SQLite database initialization and CRUD operations, and the JSON chat prompt construction.
Implement pthread-based parallel execution in tool_registry_execute to run tool calls concurrently, replacing the previous sequential loop. Add tool_thread_args_t struct and tool_thread_func wrapper for thread-safe execution. Introduce MAX_PARALLEL_PROBES constant and probe_t struct in network_port_scan_execute to enable concurrent socket probing with non-blocking connect patterns, replacing the single-threaded sequential port scan loop.
Implement context manager integration that detects API errors indicating context length exceeded, then automatically shrinks message history and retries the request. Add new messages API functions (remove_range, get_object, replace_at) and a new R_ERROR_CONTEXT_TOO_LONG error code to support this functionality.
Add use_tools() helper that reads R_USE_TOOLS environment variable to
conditionally include tools descriptions in chat JSON output, defaulting
to true when unset or unrecognized.
Implement full AppImage packaging support including AppRun entry point, .desktop launcher, icon asset, and recursive shared library collection via collect_so_files.sh. The Makefile build target now generates a portable Linux executable by assembling the directory structure, copying the binary to AppImage/usr/bin/r, running the dependency collector, and invoking appimagetool to produce a self-contained AppImage binary. Also update .gitignore to exclude the AppImage build directory and add shell scripts to the ignore list.
The diff adds a tip about creating bash files for environment variables and symlinking them for easy access, along with new configuration sections for OpenAI (using gpt-4o-mini) and Grok (using grok-2 with strict mode disabled) API setups in the README documentation.
The fix changes the glob pattern from a literal dot to a wildcard asterisk when the target directory is ".", enabling recursive matching of files in subdirectories during the build process.
The change adds a conditional check in `tool_function_directory_rglob` that converts a target_dir of "." to "*", enabling the recursive glob to match files in nested subdirectories rather than only the current directory's immediate contents.
Replace hardcoded global API URL variables with getter functions that check the R_BASE_URL environment variable, falling back to the static defaults. Introduce a joinpath utility in utils.h to safely concatenate a base URL with a path segment, handling trailing slashes and leading path slashes. Remove conditional compilation blocks for RD and OLLAMA backends, consolidating all API URL logic into the new get_completions_api_url and get_models_api_url functions. Update all call sites in openai.h to use the new getter functions instead of direct global variable access.
- Remove the initial commit placeholder progress message from Makefile
- Add `!tools` command to REPL that prints available tools descriptions
- Include tools.h header and integrate tools_descriptions() function
- Switch default RD backend from DeepInfra/Llama to Anthropic/Claude models
- Update advanced_model and fast_model to claude-3-5-haiku-20241022
- Change completions_api_url to Anthropic API endpoint
- Enable verbose mode by default (is_verbose = true)
- Update README to reflect new Claude/Ollama focus and remove old download links
Introduce a new `file_version_history` table to the SQLite schema and implement the `db_store_file_version` function that reads a file's current content and inserts a version record before overwriting it. The `tool_function_write_file` now calls this backup function prior to opening the file for writing, ensuring every write operation creates a recoverable snapshot.
Add build_mingw target using x86_64-w64-mingw32-gcc cross-compiler with
dedicated MinGW flags and publish step. Introduce Dockerfile with Ubuntu
base image installing all required build dependencies (gcc, make, readline,
ncurses, curl, openssl, json-c, sqlite3, python3-dev). Add compose.yml for
convenient Docker-based development shell with mounted working directory.
Extend Makefile with docker, docker_make, docker_run, run_mingw targets and
suppress error output on publish commands for build and build_free targets.
Introduce conditional compilation via OLLAMA preprocessor flag to switch between OpenAI and Ollama API endpoints, model names, and URLs. Add `build_free` and `run_free` Makefile targets for the Ollama variant. Enhance REPL with `!models` command to list available models, improve error handling in `openai_process_chat_message`, and fix history line addition logic.
The change inserts a `printf("%s", output)` call inside the `tool_function_linux_terminal` function, immediately after reallocating the output buffer and before appending new data. This enables real-time display of captured process output to the console as chunks are received, rather than only returning the final aggregated string. The modification affects the core execution loop in `tools.h` where command output is accumulated from the process stream.
Add explicit system("clear") call and newline in handle_sigint to visually reset the terminal display before showing the double-Ctrl+C exit message, ensuring clean output regardless of prior screen state.
Add file path completion support alongside existing command completion in the readline-based input system. Include <glob.h> header and implement line_file_generator function that uses glob pattern matching to suggest file paths based on partial input. Modify line_command_completion to detect file path context (non-command input after start position) and delegate to the new file generator, enabling tab-completion of filenames and directories in the shell-like interface.
- Add SIGINT handler that exits on two Ctrl+C presses within 1 second
- Introduce --free flag to switch auth type to AUTH_TYPE_FREE via auth_free()
- Implement "model" REPL command to print current prompt model name
- Register index_source_directory tool description and function in tools.h
Add a new `message_add_object` function in messages.h that appends a json_object to the global message list. Modify openai_chat in openai.h to call this function after processing a chat completion, ensuring the assistant's response object is stored. Also comment out the previous `message_add_tool_call` call to avoid duplicate message storage when tool calls are present, and add a missing `continue` in main.c's repl loop after the dump command to prevent fallthrough.
- Added rpylib_prompt and rpylib_system functions to rpylib.c for user and system role OpenAI chat
- Removed build_free target from Makefile and added publish step for rpylib.so build
Implement auth_init() in auth.h to detect API keys from environment variables (R_KEY or OPENAI_API_KEY) and set auth_type accordingly. In main.c, call auth_init() during init() and replace the conditional commercial/free version startup messages with a simple carriage return clear.
Add AUTH_TYPE enum and global auth_type variable to auth.h, replacing the FREE_VERSION preprocessor conditional in chat.h with a dynamic get_prompt_model() function that returns "gpt-3.5-turbo" for free users and "gpt-4o-mini" for API key holders. Update all references to prompt_model to use the new function, enabling runtime model switching without recompilation.
The tool_function_directory_glob function previously returned NULL when the glob
pattern failed to match, causing a null pointer dereference in the user profile
avatar loader when it attempted to process the result. Changed the return value
to an empty string to ensure callers always receive a valid pointer.
The diff removes a diagnostic fprintf to stderr that warned about a missing
environment variable, and replaces the null return path with a hardcoded
fallback API key string in auth.h's resolve_api_key function.
retoor