Release v2.5.0 · panjie/mods

v2.5.0 - 2026-06-27

Interactive setup with mods --config, plus first-run configuration when no settings or conversations exist.
Provider and model setup wizard entries, including web search configuration and clearer base URL handling.
Continuous chat mode for longer interactive sessions.
Built-in Mods identity guidance so models understand the CLI, config, tools, and expected behavior.
Safe workspace filesystem tools (fs_read, fs_write, fs_list, fs_stat) with scoped access.
Plan mode improvements with multi-proposal selection and local shell command classification.
Configurable web-search-api-key-env, defaulting to TAVILY_API_KEY.
Clipboard image shorthand flag.
Nightly release automation using GoReleaser.

Config precedence is now explicitly CLI flags > mods.yml > MODS_ env > defaults.
Custom OpenAI-compatible endpoints now favor explicit config via mods --config or mods.yml instead of base URL environment auto-discovery.
Tool approval rules now use directory-scoped grants instead of shell command prefix matching.
Reasoning controls now actively enable or disable provider-specific thinking behavior, including DeepSeek.
The build system moved from Mage to Taskfile.
Streaming, request handling, and tool policy internals were centralized and optimized.
Thinking and review UI output was softened and simplified.

Respect -T reasoning settings in plan mode and provider-specific reasoning paths.
Suppress loading animation in debug mode and preserve unsolicited reasoning output.
Separate model responses cleanly after tool calls.
Flush cached output and config path updates reliably.
Validate CLI inputs before provider calls.
Improve config wizard navigation, provider model setup, base URL screens, and Shift+Tab behavior.
Fix always-allow directory rule handling and trust LLM shell review analysis.
Clarify LLM prompt guidance and update README test expectations.

Full comparison: v2.0.0...v2.5.0