v0.10.7
Highlights
- A read-only plan mode: toggle /planmode and arterm explores and proposes without changing a single file
- Five new tools — lsp for definitions, references and hover; github for issues, pull requests and comments; diagnostics for compiler and linter errors; repo_map for a repo-wide tree with symbol outlines; and monitor to watch a command or websocket and feed matches back into the turn
- Every file-changing tool can now be undone, and can commit what it touched on its own
Improvements
- Permission rules are now per tool and per argument, written as deny, ask or allow, with deny outranking ask and ask outranking allow
- Named config profiles: keep one set of settings per kind of work and pick it with --profile or ARTERM_PROFILE
- Subagents can be given personas of your own, written as markdown files in .arterm/agents
- Agents spawned into a swarm can each work in their own git worktree instead of sharing one checkout
- Lifecycle hooks can now POST to an HTTP endpoint, or gate a turn by asking a model
- Sessions extract what they learned at the end of a turn, so it is there for the next one
- Tool descriptions are trimmed to their prompt budget, leaving more of the context window for your work
- Windows ARM64 builds ship with embeddings again, and a missing platform can no longer leave a release quietly
Fixes
- A permission rule set to ask now refuses when nobody can be asked, instead of quietly allowing the call
- The session list from another machine is no longer cut short by the handshake, so remote machines stop looking idle when they are not
- device forget now accepts the fingerprint that device list printed
- Updated h2 for RUSTSEC-2026-0258, where a peer could send empty data frames without bound
Full changelog: v0.10.6...v0.10.7
Platform availability
- Linux x86_64: available
- Linux aarch64: available
- macOS Apple Silicon: available
- macOS Intel: available
- Windows x86_64: available
- Windows ARM64: available
- FreeBSD x86_64: available