v1.0.16

ACP agent

• Added hooks: run custom commands at six lifecycle seams (PreToolUse, PostToolUse, UserPromptSubmit, PreCompact, SessionStart, Stop), configured in the hooks section of settings.yaml. Hooks can block or rewrite tool calls and prompts, inject context, or ask the agent to keep going at the end of a turn.
• Added Auto mode: an opt-in approval mode where a classifier model rates each permission request. Low-risk actions run immediately, medium-risk actions run with a notice in the transcript, and high-risk ones fall back to the approval dialog. Requires a classifier configured in settings.yaml.
• Added mid-turn steering: new prompts can now be folded into a running turn instead of waiting for it to finish.
• Added settings.yaml configuration for named subagents, including descriptions, default selection, and custom agent servers.
• Subagents now reuse the running ACP server instead of spawning a separate process.
• Sessions now continue after a length-truncated model response instead of terminating.
• The plan to build mode switch is now always reviewed by the user, even under "Allow all".
• Fixed a memory leak that caused agent memory to grow in long-running sessions.
• Login with an HTTP API URL now automatically retries over HTTPS on network failure.

ACP client (TUI)

• Pressing enter while the agent is busy now steers the prompt into the running turn. Use ctrl+enter to queue it for the next turn instead.
• Added a startup dialog when a newer pool version is available.
• New /delete command to select and delete previous sessions.
• Subagent token and cost usage is now included in /usage and the usage tooltip.
• Unified the UI for --resume and /resume.
• Copied selections no longer contain soft line wraps.
• Improved mouse support in dialogs and selectors.
• Fixed lag and high CPU usage in long-running sessions.