Skip to content

v0.3.0 — Native tool rows, provider workers and binding plan mode

Latest
Latest

Choose a tag to compare

View all tags
@greenpolo greenpolo released this 28 Sep 23:17
v0.3.0
e50b0bd
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Multi 0.3.0 draws native Cursor, Antigravity and Grok actions as Claude Code tool rows, replaces per-model workers with one worker type per provider, and makes plan mode binding across providers and workers. It also adds an end-to-end test suite that runs a real Claude Code binary in CI on Linux, macOS and Windows.

  • Native tool rows: Cursor, Antigravity and Grok actions render the way Claude Code draws its own tools, inside the worker session with their native tool names. Unconfirmed native actions are no longer shown as successes.
  • One worker type per provider: pick the model with the Agent tool's model parameter instead of choosing among per-model, per-effort worker types. Agent rows show the resolved model. Native harness workers run with isolation: "worktree" and survive Claude-loop snapshots.
  • Plan mode enforcement: OpenAI automatic review now honours plan mode, including plan mode entered or left mid-turn. A planning parent keeps every native worker in plan, and workers whose permissions are unknown fail closed. Antigravity plan mode allows only read-only native tools.
  • Auto mode: subscription auto mode no longer switches to billed classifier requests after a non-Claude reply.
  • Fixes: Claude responses are no longer cut at three minutes, and tool definitions are deferred the way Claude Code does natively. Grok settings are admitted with Grok's own rules, and Zen models that no longer serve requests are gone. On Windows, empty arguments survive when Multi launches npm's claude.cmd.

Update the installed Multi plugins together: provider plugins now require multi-core ~0.3.0.

Validation: npm run check (567 tests) and the hermetic E2E suite (106 scenarios with a real claude binary, fake providers and fake native CLIs) pass in GitHub CI on Linux, macOS and Windows. All 70 local Claude Mods tests pass. Live provider checks ran on Linux during development but not against the final release commit.

Known outstanding work: the Auto-mode classifier report (#20), WSL2 reproduction and session reliability (#21), upstream Workflow spawn context (#32), and accurate context sizing for other providers (#37). Remote Control (#19) still depends on upstream custom-gateway support.

See the full changelog and changes since v0.2.1.

Assets 2
Loading