Releases: CosineAI/cli
2.0.7
2.0.6
2.0.5
2.0.4
Improvements in Workflow and Remote Runner
AI Modes & Agent Workflow • Swarm mode got a major upgrade: semi-async subagents, better nudge/query behavior, optional QA agent control, and more reliable parent/child tracking in the agents panel. • Sessions are much easier to pick back up, with auto-resume on startup, improved resumed transcript rendering, and fixes for remote-first resume flow. • Starting the CLI is more flexible: --prompt-file , --mode , and --reasoning flags were added, and start-command flags can now be persisted. • Plan mode is smoother, with implementation plans shown in the timeline, plan-file context injection, optional auto-approve, and fixes for re-entering plan mode without UI corruption. • Training workflows got deeper model control, including training-mode routing, --equivalent-model , and model-specific parallel tool calling. Search, Editing & Tooling • Native tool search is now on by default, with a persistent setting and duplicate-tool controls for power users. • Search got faster and more robust: async file search while typing, less flaky @mention file lookup, optional subprocess grep execution, configurable ripgrep backends, and chunked grep output to avoid large-file crashes. • The edit flow is more reliable, with improved fuzzy matching, better indentation preservation, stronger structural matching rules, and stricter exact-match handling where needed. • Text prediction landed in the prompt composer, along with alignment fixes and better filtering/streaming behavior. • A new multistep automation tool expands what the CLI can orchestrate in one run. Remote Runtime, Git & Environments • Remote sessions gained real workflow upgrades: remote PTY sessions, auto-tunnel URLs, session-based subdomains, Portr TCP support, and faster environment bootup . • Git workflows inside remote environments improved substantially, including on-demand credentials, automatic git setup, better auth/identity handling, push-to-PR wiring, a rebase quick action, and an in-app diff viewer. • Environment management is easier to navigate, with the Environments menu moved under Workspace, configurable environment and pod sizing, and better remote bootstrap and validation. • Session IDs were shortened to 16 characters, making remote/session references easier to handle. Desktop, Browser & Attachments • Desktop mode saw a broad polish pass, including onboarding improvements, split-view work, timeline improvements, homepage updates, and general reliability fixes . • You can now attach files directly to chat prompts. • Browser-driven workflows expanded with Safari and VNC support, plus removal of the same-origin restriction in the dashboard browser. • Desktop/browser UX got a long list of smaller fixes around streaming indicators, dock badges, PATH handling, rate-limit/error states, and image path handling. MCP, Terminal & Model Controls • MCP support is more transparent and stable: the CLI now shows loaded tool counts for MCP servers, resyncs tool inventory correctly, and includes fixes for OAuth flows and Vercel MCP endpoints. • Terminal integration improved with Open Terminal actions, the ability to send terminal output back to the model, better command history handling, interactive streaming, and stall nudges. • Model controls expanded with a new none reasoning level, x-high for GPT-5.4, better Gemini handling and native streaming, clearer provider model IDs, and refreshed model/config behavior. • LiteLLM-specific workarounds were removed as the CLI’s native tool-call handling became more robust. Bug Fixes, Stability & Performance • Windows and Linux builds got a serious hardening pass, including MinGW DLL fixes, static libgcc linking, and external-linker changes for Linux CGO builds. • TUI rendering is more stable: fixes landed for ANSI layout corruption, copy/paste behavior, theme persistence, todo wrapping, sidebar overflow labels, agent- list wrapping, and overview keyboard selection drift. • Large or awkward inputs are handled better, with fixes for big pastes, malformed tool call arguments, repeated streamed argument chunks, no-output tool failures , and stale todo/tool-result edge cases. • Performance improved in several hot spots, including terminal overlays, scroll behavior, file search responsiveness, startup/reconnect paths, and grep on large files. • Update flows are smoother too, with better auto-update restart behavior and automatic installer runs for non-Homebrew installs.