Copilot CLI v1.0.6: Autopilot Fixes & Memory Improvements
Copilot CLI 1.0.6 fixes autopilot error recovery, plugs major memory leaks, and stabilizes Windows. Tool search for Claude, human-readable subagent IDs, and expanded plugin spec support included.
TL;DR
- Autopilot no longer gets permanently blocked after errors — sessions can now recover
- Major memory leak fixes: tree-sitter WASM cleanup, environment variable duplication, streaming optimization
- Windows stability improvements: auto-update race conditions, module prebuilds, integrated terminal double-triggers fixed
New
- Tool search for Claude models — Model can now discover and use tools dynamically instead of relying on static tool definitions
- Subagent human-readable IDs — Task-launched agents now get names like
math-helper-0instead of genericagent-0identifiers - PR URL in pull request tool output —
create_pull_requestnow includes the direct link so agents can share it - Open Plugin spec support — Expanded compatibility for
.lsp.json, PascalCase hook names, exclusive path mode, and namespace separators - Extension tool permissions — Tools now integrate with the permissions system; use
skipPermissionper-tool to bypass prompts
Fixed
- Autopilot error recovery — Continuation no longer gets permanently blocked after an error in a previous turn
- Memory leak in tree-sitter — WASM objects now freed after shell command parsing
- Environment variable duplication — Eliminated redundant copies per child process spawn, reducing overall memory footprint
- Streaming and tool-output memory — Improved memory efficiency during agent execution
- Windows auto-update race conditions — Auto-update now correctly recovers; CLI loads latest version after update
- Windows module prebuilds — Native modules (e.g., conpty.node on ARM64) load reliably on first launch
- VS Code integrated terminal double-triggers — Links and right-click paste no longer trigger twice
- HTTP/2 connection pool crashes — Resolved session crashes when sub-agents are active
- Command output loss — Command output no longer lost when a command causes the shell to exit
- Kill command validation — No longer incorrectly blocks legitimate commands like
p.kill()in Python scripts - Screen reader accessibility — Input placeholder text no longer read aloud on every prompt submission
- Help dialog scrolling — /help dialog now starts scrolled to the top in alt-screen mode
- Prompt input truncation — All lines now render without truncation in alt-screen mode
- Shift+Enter handling — Fixed handling in VS Code with old /terminal-setup config
- Hook configuration cross-platform — Hook config files now work across VS Code, Claude Code, and CLI without modification by accepting both PascalCase and camelCase event names
- Custom instruction file paths —
COPILOT_CUSTOM_INSTRUCTIONS_DIRSnow loads correctly - Plugin discovery — Plugins using
.claude-plugin/plugin.jsonare discovered when loaded via--plugin-dir - Quota widget accuracy — Remaining requests widget no longer shows inaccurate data for Copilot Free users
- Agent creation wizard — Shows correct user agents directory path
- Subagent elapsed time — Freezes when idle and resumes when active in /tasks view
- SDK flag support — Flags
--enable-all-github-mcp-tools,--add-github-mcp-toolset, and--add-github-mcp-toolnow take effect in SDK (ACP mode) - Task completion summary — Now required in autopilot and renders as markdown
- Read agent output — Includes inbound messages that triggered each turn in multi-turn agents
- Hook config structure — Now supports Claude Code's nested matcher/hooks structure and optional type field
- Error messages and browser links — Friendlier error messages with keyboard shortcut to open event links in browsers
- Instruction file frontmatter —
applyTofield now accepts both string and array values - Session resumption hooks — Hooks now fire correctly when resuming a previous session
Update via: npm install -g @github/copilot-cli@latest or check the release page for platform-specific instructions.
Source: Copilot CLI