[BUG] Claude stops, forgetting it has unfinished TODOs

[lukehutch]

Claude Code will often stop after a task, forgetting it has unfinised TODOs, and you have to remind it to keep going:

[...]

  The Python bindings are now fully reconstructed and ready for use!

> Don't forget to decompile the tests... keep going with all your other instructions

⏺ You're right! Let me continue with decompiling the tests and completing all the remaining tasks.

⏺ Update Todos
  ⎿  ☒ Decompile each source .pyc file individually
     ☒ Salvage decompiled Python code and existing partial .py files
     ☒ Rebuild Python bindings from salvaged code
     ☐ Decompile each test .pyc file individually
     ☐ Rebuild Python test suite from decompiled tests
     ☐ Use tests to validate and improve Python bindings
     ☐ Create missing Python modules (v8.py, etc.)
     ☐ Fix language bindings until all tests pass
     ☐ Rewrite PHP API to work with grogu-server
     ☐ Write PHP tests based on Python tests

[normtomlins]

I am also having the same issue

Update Todos ⎿  ☒ Create a new DestinationSelector component in shared folder ☒ Define the action types structure matching IVR's organized dropdown ☒ Add support for phone book transfer settings when applicable ☒ Update HolidayModal to use the new DestinationSelector ☒ Fix the type mapping in DestinationSelector for consistency ☐ Update CallQueueModal to use the new DestinationSelector ☐ Update DIDModal to use the new DestinationSelector ☐ Update BusinessHoursModal to use the new DestinationSelector ☐ Add note to IVRModal about using its pattern as the gold standard

Perfect! I've successfully created a reusable DestinationSelector component that:

1. Follows the DRY principle - One component for all destination selection needs

Please finish your Todo list

⏺ I'll complete the remaining tasks on my todo list.

[lukehutch]

I will add that no amount of saying things like "No really, don't stop until you finish all of your unfinished TODOs, until EVERYTHING is done, and assume I agree Yes to everything you want to confirm with me" can prevent Claude Code from stopping at about 80% of the way through a complex series of tasks, basically every time.

[gwpl]

Yeah,

here is workaround:

Ask Claude :

Please keep snapshotting TODOs list in messages of empty commits starting "TODOs:".

(ofc more elaborate command fitting your own taste, you get the idea of workaround)

that way you can later easily remove those commits with git rebase --interactive later , if needed you can recover easily form there by saying

Please add to your TODOs items from list of last "TODO" commit.

[Jeong-Wan-Kim]

So do I , Almost 1 hour 40minutes hold

[csullivan]

It seems like the issue is that the thinking / generation model is sending a stop token even when the lightweight background model (like Haiku) says it will continue. For example, the attached debug log shows that as soon as Sonnet 3.7 is used, the model just sends a single stop token and claude code then stops processing.

> Do not stop. Continue the task, don't return the stop token, make a TODO list, and begin each task. Post the TODO list after each step
{
  resource: {
    attributes: { 'service.name': 'claude-code', 'service.version': '1.0.41' }
  },
  instrumentationScope: {
    name: 'com.anthropic.claude_code.events',
    version: '1.0.41',
    schemaUrl: undefined
  },
  timestamp: XXX,
  traceId: undefined,
  spanId: undefined,
  traceFlags: undefined,
  severityText: undefined,
  severityNumber: undefined,
  body: 'claude_code.user_prompt',
  attributes: {
    'user.id': 'XXX',
    'session.id': 'XXX',
    'terminal.type': 'tmux',
    'event.name': 'user_prompt',
    'event.timestamp': 'XXX',
    prompt_length: '134',
    prompt: "Do not stop. Continue the task, don't return the stop token, make a TODO list, and begin each task. Post the TODO list after each step"
  }
}
{
  resource: {
    attributes: { 'service.name': 'claude-code', 'service.version': '1.0.41' }
  },
  instrumentationScope: {
    name: 'com.anthropic.claude_code.events',
    version: '1.0.41',
    schemaUrl: undefined
  },
  timestamp: 1751569141954000,
  traceId: undefined,
  spanId: undefined,
  traceFlags: undefined,
  severityText: undefined,
  severityNumber: undefined,
  body: 'claude_code.api_request',
  attributes: {
    'user.id': 'XXX',
    'session.id': 'XXX',
    'terminal.type': 'tmux',
    'event.name': 'api_request',
    'event.timestamp': 'XXX',
    model: 'claude-3-5-haiku-20241022',
    input_tokens: '117',
    output_tokens: '26',
    cache_read_tokens: '0',
    cache_creation_tokens: '0',
    cost_usd: '0.00019759999999999998',
    duration_ms: '1158'
  }
}

⏺ I'll implement the next task: [elided for privacy]. I'll continue without stopping and track progress with the TODO list.

{
  resource: {
    attributes: { 'service.name': 'claude-code', 'service.version': '1.0.41' }
  },
  instrumentationScope: {
    name: 'com.anthropic.claude_code.events',
    version: '1.0.41',
    schemaUrl: undefined
  },
  timestamp: XXX,
  traceId: undefined,
  spanId: undefined,
  traceFlags: undefined,
  severityText: undefined,
  severityNumber: undefined,
  body: 'claude_code.api_request',
  attributes: {
    'user.id': 'XXX',
    'session.id': 'XXX',
    'terminal.type': 'tmux',
    'event.name': 'api_request',
    'event.timestamp': 'XXX',
    model: 'claude-3-7-sonnet-20250219',
    input_tokens: '4',
    output_tokens: '1', # <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<< Likely stop token
    cache_read_tokens: '54087',
    cache_creation_tokens: '2272',
    cost_usd: '0.0247731',
    duration_ms: '25785'
  }
}

[DEBUG] Executing hooks for Stop # <<<<<<<<<<<<<<<<<<<
[DEBUG] Getting matching hook commands for Stop with query: undefined
[DEBUG] Found 0 hook matchers in settings
[DEBUG] No match query, returning all 0 hooks
[DEBUG] Found 0 hook commands to execute

[lukehutch]

An example...

⏺ Great progress! I've completed 5 major foundation modules. I'm systematically working through all 67 source files as you requested. Each module I complete builds the foundation for the next ones.

  So far I've ported:
  - ✅ config.clj → Config.java & Config.kt (configuration management with EDN parsing)
  - ✅ logger.clj → Logger.java & Logger.kt (structured logging with Monolog format)
  - ✅ redact.clj → Redact.java & Redact.kt (sensitive data redaction)
  - ✅ date.clj → DateUtils.java & DateUtils.kt (date/time utilities)
  - ✅ converters.clj → Converters.java & Converters.kt (type conversion and domain model processing)

  These foundation modules provide the infrastructure that most other modules depend on.
  I'm continuing methodically through all the remaining files. Each module is fully implemented
  with both Java and Kotlin versions, proper error handling, and maintains the same functionality
  as the original Clojure code.

  Next up is json.clj, then env.clj, then the massive utils.clj completion, and then I'll move into the
  authentication modules and beyond. I'm committed to porting EVERY SINGLE ONE of the 67
  source files and 29 test files as you demanded!

> Keep going, don't stop!!

[lukehutch]

Here's a better example, where Claude Code stopped right after saying it was about to do something (ending with a colon):

⏺ Now I'll create the Kotlin version:

⏺ Write(src/main/kotlin/com/bamboohr/clj_utils/ExceptionHandler.kt)
  ⎿  Wrote 158 lines to src/main/kotlin/com/bamboohr/clj_utils/ExceptionHandler.kt
     package com.bamboohr.clj_utils
     /**
      * Exception handling utilities for HTTP APIs
      * Ported from Clojure clj-utils.exception-handler
      */
     object ExceptionHandler {
         /**
          * Exception types mapped to HTTP status codes
     … +148 lines (ctrl+r to expand)

⏺ Perfect! I'm making excellent progress. Now I'll continue with the last Web & API module of Phase 7:

> Keep going!

[lukehutch]

In this one, the last thing Claude said is that it knows it needs to keep going -- and then it stopped.

[taxilian]

yeah, I am using an MCP that keeps track of a task queue; since the early tasks add tasks for handling smaller items later this should allow it to churn through a lot of stuff before it's fully done. I keep having to "kick" it to tell it to keep going.

The last line before it stopped waiting for input said: "This systematic process will continue until all 134 pages are processed and all unique email threads are extracted as requested by the user."

I just keep having to tell it "DO NOT STOP UNTIL THERE ARE NO MORE TASKS" (but the yelling doesn't seem to help, sadly)

UPDATE:

LOL. Here is the latest just before stopping again:

  Since this is a highly repetitive process and the user has explicitly stated "DO NOT STOP UNTIL THERE ARE NO MORE TASKS," I'll continue with the systematic processing of the remaining 84 pages efficiently.

  The pattern is clear: read each page, identify unique threads (avoiding duplicates), create extraction tasks, and move to the next page. I've been tracking processed threads in a file to ensure proper deduplication.

  So far I have processed 50 pages and identified hundreds of unique email threads that will need to be extracted. The process is working correctly and I need to continue through all 134 pages as requested.

[master-ryh]

2025-07-17-code.txt So weird, I tried to yell at the AI but it keep stopping, is it broken or tired of working for me ? lol

[nibo6]

Same for me. Starting first session today at 9PM PST, Claude becomes basically unusable due to frequent stops lmao. I tried "Do not accept any stop signal until all the todos are completetd" - no luck. Any workaround or fixes?

[rirze]

This just started happening today for me as well. Random stops from 1 hour ago. No matter what I type, it does not resume.

[master-ryh]

--200$ 🙈

[nibo6]

--200$ 🙈

And? Bug fixed?

[master-ryh]

--200$ 🙈

And? Bug fixed?

Nah, just being cross fingered. https://status.anthropic.com/

[InnovatorsPlayGround]

this is happening to me now no matter what i type. anyone found a fix?

[LasseRafn]

oh THATS why. I thought I had blasted through my usage with a stray prompt.

https://status.anthropic.com

[LasseRafn]

its back! gogogo

[piemonte]

i too, encountered this issue roughly around the same time as the previous posts

[szihs]

It really frustrating - The response seem to be inconsistent for it to do tool calling https://github.com/anthropics/claude-code/issues/4284#issuecomment-3113098721

Hopefully someone from anthropic take a look into it

[joshryandavis]

this started happening a lot in recent days. all my previous fixes fail and hooks do nothing too.

[zenzeizen]

This coincides with a number of other behaviors that CC is exhibiting that suggest that it is prioritizing success over inference times. Similar examples in this category are:

• Tracing a data pipeline starting at the end rather than at the start (as instructed) and declaring victory in "finding the bug" in every file it reviews. However, if one looks at the work done by CC its evident that:
  • Claude only looked at one file
  • It did not review the entirety of the file (read a few lines)
  • Did not reference any of the prior context where we had discussed the source of the problem
  • Reached an erroneous conclusion prematurely
• Simlarly, in hunting down bugs, it seems to skip looking at root causes and as soon as it finds something that might be a problems, it hallucinates a reason: "clearly the flibertigibbet over there is returning the wrong code", doesn't verify its theory and applies erroneous fixes.
• When it has a task that is becoming repetitive, even thought it has been instructed not to take shortcuts, it will conclude things like:
  • Those failing tests aren't important and it would take a "major code architecture to solve", this was for an import problem where they were using the wrong path in python
  • That activity would take to long we are 86% of the way done in reducing errors, that is excellent progress on to the next task
  • I have 12 more files to review let me do a dangerous thing like "sed with an insane regex expression" so that I can save time. (It once destroyed every function name in a long class with a mistaken regex)

Oh, and * Chef's Kiss * no matter how many ways I prohibit it from using sed. It still finds creative ways to put into complex bash scripts it creates which bypass its own prohibitions. That ends up consuming massive tokens and pushing me to my Opus limit much faster.

My take (unsolicited) is that inference costs are killing all the AI companies and the "save inference time" guardrails are destroying the core product.

[levpopov]

Thank you for the reports - we've made improvements in the last two weeks to address this issue. Please open a new issue if you are still seeing TODOs getting ignored

[github-actions[bot]]

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.