[BUG] Plan mode creates `/.claude/plans` and is over optimizing on simple prompt

[jackstine]

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Not sure if this is a bug, but this is very different behavior.

plan mode now seems like it is over optimizing for simple things. It created a several hundred line plan, and wanted to generate a file in ~/.claude/plans This is new, I do not like the several hundred line plan though. Usually for something like this, it would generate maybe 80 - 120 lines for the plan overview, but now it is reaching 800 or more lines, and full of a lot more detail and over engineering than i requested. Not sure if it is because of the prompt i used or not.

Here is the prompt

Looking at @src/cmd/eod/eodhd/daily_eod_ingestion.py And it's pipeline I would like to make an
enhancement. I believe we are not doing a bulk insert into the database for the historical EOD pricing
table and I believe we are not getting from the database the ticker history rows at a bulk request either.

What Should Happen?

Probably would like to have this type of plan, but it should be activated with a keyword.

Error Messages/Logs

Steps to Reproduce

Here are two prompts that triggered it

Looking at @src/cmd/eod/eodhd/daily_eod_ingestion.py And it's pipeline I would like to make an
enhancement. I believe we are not doing a bulk insert into the database for the historical EOD pricing
table and I believe we are not getting from the database the ticker history rows at a bulk request either.

I believe we fix this issue,  but - no repository should have
two or more tables in it.   There should be one repo dedicated to
 one table.
    - need to fix `get_active_ticker_history_symbols` and
`get_active_ticker_histories` in
`src/repos/equities/tickers/ticker_history_repository.py`
   please check and confirm that all the other files in
@src/repos is not using joins at all.

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.0.51

Claude Code Version

2.0.53

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

No response

[github-actions[bot]]

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/12337
2. https://github.com/anthropics/claude-code/issues/12288
3. https://github.com/anthropics/claude-code/issues/9665

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

[jsauvain]

Plans should be stored in project .claude dir IMHO

[jackstine]

Plans should be stored in project .claude dir IMHO

just new behavior that i did not see on the Change Log, very different than before. I'd suggest a keyword to trigger the heavy plan mode operation. I do not have an objection with where Claude places memory, as it seems like that is what these plans are going to become for the agents.

[v3ceban]

+1. This makes plan mode pretty much unusable for me. The time it takes to plan is doubled, if not trippled, comparing to previous behavior. New behavior is absurd. After generating the plan with this Plan tool it reads the generated file, reads the files mentioned in it (even though they were already in the context/discovered by Explore tool), and then, wait for it, writes another plan.

[edsonbraz]

Terrible experience as well, in terminal mode I cannot see the contents of the plan. Am I supposed to trust the summary? The Plan subagent just waste time, simple tasks now go to Explore and Plan tools. If the goal is to increase Token usage, well done. But I prefer managing my own Context, it worked better before they've introduced these tools.

[jackstine]

I am using version 2.0.47 because this feature on the planning exists. I do not want to use the latest version of claude code, because of this new planning feature. The plan is to verbose, literally had a plan of 800+ lines for making 50 lines of code changes. The time to read the plan is to long, compared to the code generated.

[uhs-robert]

This new plan mode is ineffective, over-complicated, and burns through tokens like no tomorrow.

My Claude Code just burned over 100k tokens on a single prompt within 5 seconds. This is not just unacceptable but also unsustainable. 5 seconds is all it takes to destroy your entire 5-hour window.

I just asked the following question (which, before this update, would have been a 1-5k token back-and-forth exchange):

I want to make a yazi theme generator. I copied the alacritty generator from extras into @extras/yazi/ as a template generator base. Now we need to update that base to generate the toml file for yazi. Here is
 an example toml output @extras/yazi/example.toml that we need to generate. This is a large file, do you think we should we make a template doc or use code to build it line-by-line?

And the result...

● I'll help you create a Yazi theme generator. Let me enter plan mode to
  explore the existing setup and design the best approach for this complex
  generator.
                                                                   
● Entered plan mode

  Claude is now exploring and designing an implementation approach.

● I'll explore the existing generator patterns and the Yazi directory to
  design the best approach for this complex theme generator.

● 2 Explore agents finished (ctrl+o to expand)
   ├─ Explore existing theme generators · 21 tool uses · 52.2k tokens
   │  ⎿ Done
   └─ Examine Yazi directory structure · 9 tool uses · 43.1k tokens
      ⎿ Done

And now I have a useless, overcomplicated, and garbage plan that consumed my entire 5 hour window within an instant. Plan mode is essentially burn money mode while this issue persists.

[rgwood-dd]

I believe this is caused by the "Plan Mode now builds more precise plans and executes more thoroughly" change in v2.0.51.

I've found that plan mode is much harder to use now because the plans are so much longer and more difficult to review (even simple changes generate plans that are 300+ lines).

To work around this, I've pinned Claude Code to version 2.0.50 for now.

[compae]

I’ve had the same problem. I’ve been working on several projects, designing and implementing multiple improvements for months. The other day, after updating to version 2.0.51, everything fell apart. It’s now almost impossible to work the way I used to, planning small, incremental improvements on a stable project. This new “plan mode” seems entirely oriented toward creating projects from scratch and prototyping or making large-scale changes—yet that’s not even 10% of real software development work.

For me, it was extremely disappointing to see something like this happen in a patch release—truly unacceptable. For now, I’ve downgraded back to version 50. I really hope this gets reconsidered and that we can go back to planning with the main agent. I’ve even noticed major issues, like speaking in one language and getting the plan in English, whereas before the plan always matched the language of the conversation.

I hope measures are taken and that some kind of solution is offered.

[jackstine]

if anthropic wants to introduce a new type of plan mode, they should have two types. The new extensive plan mode and the previous plan mode as well.

[github-actions[bot]]

This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.

[uhs-robert]

This issue is still occurring.