aandrew-me/tgpt v2.16.0

tgpt 2.16.0

on GitHub

Changelog

New AI Providers

• API Route (apiroute): Added OpenAI-compatible gateway support with tool calling (default model gpt-5.5; supports API_ROUTE_API_KEY, API_ROUTE_MODEL, API_ROUTE_URL, API_ROUTE_BASE_URL).
• DemonRoute (demonroute): Added OpenAI-compatible provider (https://api.demonroute.com/v1) with tool calling.
• Runcrate (runcrate): Added OpenAI-compatible provider (https://api.runcrate.ai/v1) with tool calling (default model deepseek-ai/DeepSeek-V4.1-Flash; supports RUNCRATE_API_KEY, RUNCRATE_MODEL, RUNCRATE_URL, RUNCRATE_BASE_URL).
• Obscurify.ai (obscurifyai): Added OpenAI-compatible provider (https://obscurify.ai/v1) with tool calling (default model deepseek/deepseek-v4-flash; supports OBSCURIFY_API_KEY, OBSCURIFY_MODEL, OBSCURIFY_URL, OBSCURIFY_BASE_URL).

New Web Search Providers & Built-in Tools

Added opt-in choices for --search-provider / SEARCH_PROVIDER and corresponding built-in tools (documented in SEARCH_SETUP.md):

• Keenable (keenable / web_search_keenable): Works without an API key.
• Sofya (sofya / web_search_sofya): Uses SOFYA_API_KEY.
• Parallel (parallel / web_search_parallel): Works without an API key via anonymous Search MCP, with idle connection cleanup forwarding.
• SpecStream (specstream / web_search_specstream): Uses SPECSTREAM_API_KEY.

Bug Fixes & Improvements

• OpenCode provider: Fixed it.
• Gemini provider: Replaced the shut-down gemini-2.0-flash default model with gemini-3.6-flash.
• DeepSeek Web provider: Fixed output truncation at both the beginning (handling SET events on fragments and response paths) and end of streams (increased bufio.Scanner buffer limit to 1 MiB in src/helper/helper.go).
• SSE stream parsing: Stopped dropping chunks whose generated text contains "data: " across deepseek, gemini, groq, minimax, openai, opencode, pollinations, and koboldai (and prevented panics on "data:" without a space in koboldai).
• Sampling parameters (--temperature, --top_p): Restored forwarding of --temperature, --top_p, TGPT_TEMPERATURE, and TGPT_TOP_P across OpenAI-compatible providers, obscurifyai, pollinations, ollamacloud, and koboldai.
• MiniMax provider: Included tools in chat completion requests.
• CLI --tools / -t flag parsing: Normalized -t <tool_list> / --tools <tool_list> arguments so subsequent flags (e.g., -t web_search_exa -i) continue to parse properly.
• Windows glob tool: Switched matching from filepath.Match to path.Match so * and ? do not match across / directory boundaries on Windows.
• Piped stdin & error output: Preserved newlines between piped stdin lines (up to 1 MiB per line) and wrote non-interactive HTTP error bodies to stderr instead of stdout.