⚡ Claude Terminal v1.2.16
Claude can now run your Automations - and permission prompts stopped lying to you.
The headline
🤖 Automations, scriptableAsk Claude to create, edit, pause or delete an Automation and it just does it. Six dedicated MCP tools, no graph editing required. |
🔁 One answer, one runAn Automation set to fire when Claude replies ran several times per answer. It now runs once, when Claude has actually finished. |
🔐 Your ask rules, respectedWhen one of your own |
🌿 Subagents, transparentA subagent's summary now shows the git worktree branch its work landed on, and flags when the model was swapped mid-run. |
🤖 Automations Claude can actually drive
Automations used to be read-only from Claude's side. It could list them, but not create one - and editing one through the workflow node tools was silently overwritten on your next save, because an Automation's graph is generated from its settings.
| What you can ask for | |
|---|---|
| Create | "Every Friday at 6pm, summarise the week's commits" - schedule or event, in one sentence |
| React to events | Fire on a git commit, a file change, a failed command, a session ending, or a reply that mentions something specific |
| Update | Change the prompt without touching the schedule, or the schedule without touching the prompt |
| Pause & resume | Arm or pause an Automation without losing its settings |
| Delete | Removes the Automation and its run history |
| No more silent loss | The workflow node tools now refuse an Automation and point at the right tool, instead of writing edits that get overwritten |
Everything else
- Automation MCP tools: list, view, create, update, enable/disable, delete
DirectoryAddedhook - fires when a directory joins the session via/add-dir- RefreshMcpTools, SendFeedback and ProposeSkills now render as proper tool cards
- Bumped
claude-agent-sdkto 0.3.226
- Forking from an earlier message used to silently discard a prompt you had queued mid-turn - it now warns you instead of losing it
- Permission cards use the real tool title and description rather than reconstructing them from the tool name
- A session with skip-permissions no longer auto-approves a prompt your own ask rule demanded
- The "when Claude replies" trigger fired on every chunk of text between tool calls, so an Automation ran three or four times while Claude was still typing. It now fires once, on the complete answer - and its text filter tests the whole reply instead of a fragment
📜 Full changelog — 10 commits
feat(mcp) add dedicated tools for automations
feat(hooks) add the DirectoryAdded hook
feat(chat) refuse a fork that would drop a queued turn
feat(chat) show which ask rule forced a permission prompt
feat(chat) surface a subagent's worktree branch and model swaps
feat(chat) render RefreshMcpTools, SendFeedback and ProposeSkills
fix(chat) fire the Claude-replied trigger once per turn
chore(deps) bump the Agent SDK to 0.3.226
chore(agents) add the release agent definition
chore bump version to 1.2.16