[BUG] Claude stops working between TaskCreate items instead of continuing autonomously

Resolved 💬 3 comments Opened Mar 28, 2026 by dstolts Closed Mar 31, 2026

Description

When Claude has multiple tasks in TaskCreate/TodoWrite and completes one, it frequently stops and asks the user "Want me to continue?" or "Should I kick off the next task?" instead of proceeding to the next pending task autonomously.

This happens even when:

  • The tasks were created by Claude itself as part of an approved plan
  • CLAUDE.md rules explicitly say "keep working autonomously"
  • No RED-zone or blocking action is required for the next task
  • The user has given no signal to stop

Expected Behavior

When a task completes and TaskList shows remaining pending tasks that are not blocked, Claude should immediately start the next one without asking for permission. The user already approved the work when the tasks were created.

Actual Behavior

Claude presents a summary of the completed task and asks variants of:

  • "Want me to kick that off now?"
  • "Should I continue with the next task?"
  • "Ready when you are"

This forces the user to manually confirm each task transition, defeating the purpose of task-based planning.

Reproduction

  1. Ask Claude to plan a multi-step task (e.g., "fix all 77 blog posts")
  2. Claude creates 5 tasks in TaskCreate
  3. Claude completes task 1-4 autonomously
  4. On task 4 completion, instead of starting task 5, Claude stops and asks

This is intermittent -- sometimes it continues, sometimes it stops. The tendency increases at "phase boundaries" (e.g., moving from a build phase to a deploy phase).

Impact

For users running autonomous workflows (AFK, long pipelines, multi-agent orchestration), this is a significant blocker. The user steps away expecting work to complete, returns to find Claude waiting for a confirmation that was never needed.

Possible Root Cause

The model appears to treat task completion as a "decision point" that requires user input, even when the next task is already defined and unblocked. This may conflict with the system prompt instruction to only stop for user input when genuinely blocked.

Environment

  • Claude Code CLI on Windows (Git Bash)
  • Claude Opus 4.6 (1M context)
  • Tasks created via TaskCreate tool
  • CLAUDE.md contains autonomous execution rules

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗