Kilo Code JetBrains 7.0.12-rc.3: Agent Manager terminals, multi-project, shell-permission fix
Large release candidate spanning the JetBrains, VS Code and CLI surfaces. Headline items are embedded side-panel terminals and a multi-project foundation for Agent Manager, plus a fix that enforces permissions on shell commands the parser cannot scan. Relevant to anyone running Kilo Code agents against multiple repos or relying on command approval.
PUBLISHED2026-07-31
OBSERVED2026-08-11
AGE11d
SOURCES1
Added
- Agent Manager: embedded side-panel terminal destination, multiple side-panel terminals, project and setup scripts run in the selected terminal, diff scope selector, worktree name on hover, jump-shortcut badges
- VS Code: multi-project Agent Manager, prompt navigator rail in the chat transcript
- Permission prompt now shows verbatim skill shell commands and the skill name (VS Code and JetBrains)
- TUI: collapsible Context/Token Usage sidebar sections,
/auto-approveslash command, commands execute in skill context - Web search availability configurable for all providers; opencode websearch via Exa routed through the Kilo proxy
- Persian (Farsi) UI language; telemetry now includes host OS properties
Fixed
- CLI: permissions enforced on shell commands the parser fails to scan
- CLI: agent loop no longer freezes when a provider stalls after sending headers
- CLI: session reverts kept atomic; unset config keys removed from every layered config file; startup time reduced by deferring Kilo module loading and telemetry
- CLI:
gpt-5.6excluded from ChatGPT subscriptions; credentials included in console URLs printed for headless users; honest exit codes for headless runs - VS Code: MCP server toggle state persisted, faster embedded terminal startup, past chats listed across the worktree family, eager worktree watchers avoided
- Windows: snapshot diffs made resilient
_Source release notes are truncated mid-list upstream; the fixed section above is partial._
COMMUNITY
No curated reactions recorded for this event. Facts and takes are kept in separate layers — community context is added by hand, never blended into the record above.