›_ Claude Code Issues
Claude Code Issues / #48172

[Bug Report] POWER USER PISSED OFF - CLAUDE IS BROKEN AND GETTING WORSE NOT BETTER - HELP PLEASE!

Resolved 💬 1 comment Opened Apr 14, 2026 by rayeveritt Closed May 26, 2026

Bug Description
POWER USER PISSED OFF - CLAUDE IS BROKEN AND GETTING WORSE NOT BETTER - HELP PLEASE!

Environment Info

  • Platform: linux
  • Terminal: windows-terminal
  • Version: 2.1.108
  • Feedback ID: e3891714-30ca-4afe-b7f6-1c3c328b26b6

Errors

[{"error":"AxiosError: Request failed with status code 429\n    at iF (/$bunfs/root/src/entrypoints/cli.js:110:1194)\n    at <anonymous> (/$bunfs/root/src/entrypoints/cli.js:115:12744)\n    at emit (node:events:92:22)\n    at endReadableNT (internal:streams/readable:865:50)\n    at processTicksAndRejections (native:7:39)\n    at request (/$bunfs/root/src/entrypoints/cli.js:117:2467)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-14T22:33:01.291Z"},{"error":"RipgrepTimeoutError: Ripgrep search timed out after 60 seconds. The search may have matched files but did not complete in time. Try searching a more specific path or pattern.\n    at A (/$bunfs/root/src/entrypoints/cli.js:539:242)\n    at <anonymous> (/$bunfs/root/src/entrypoints/cli.js:539:456)\n    at <anonymous> (/$bunfs/root/src/entrypoints/cli.js:536:2992)\n    at emitError (node:events:43:23)\n    at abortChildProcess (node:child_process:935:17)\n    at onAbortListener2 (node:child_process:35:24)\n    at abort (unknown)\n    at M69 (/$bunfs/root/src/entrypoints/cli.js:1628:3782)\n    at abort (unknown)\n    at M69 (/$bunfs/root/src/entrypoints/cli.js:1628:3782)","timestamp":"2026-04-14T22:41:22.304Z"},{"error":"Error: Request was aborted.\n    at DR$ (/$bunfs/root/src/entrypoints/cli.js:1293:12508)\n    at next (native:1:11)\n    at qp7 (/$bunfs/root/src/entrypoints/cli.js:8034:9307)\n    at next (native:1:11)\n    at Ap7 (/$bunfs/root/src/entrypoints/cli.js:8039:12450)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-14T22:41:22.308Z"},{"error":"MaxFileReadTokenExceededError: File content (15583 tokens) exceeds maximum allowed tokens (10000). Use offset and limit parameters to read specific portions of the file, or search for specific content instead of reading the whole file.\n    at cO7 (/$bunfs/root/src/entrypoints/cli.js:4537:12021)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-14T22:58:15.639Z"},{"error":"Error: File does not exist. Note: your current working directory is /home/rayeveritt.\n    at call (/$bunfs/root/src/entrypoints/cli.js:4546:7630)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-14T22:58:16.060Z"}]

Hi CC, We are experience a significant amount of reduced reasoning depth. Laziness and ineffiecency with Claud Code specifically opus 4.6, there is a significant reduction
in thinking behavior and skyrocketing usage burn. looping bugs, writing, re-writing code multiple times, creating the same thing slightly different ways on different days
that's all duplicate code that then has to be cleaned up at an expense greater that writing it individually two times. So it's like 7x the cost to write code and takes
10x as long, it's frustrating as hell and the claude interface is actually mocking the user in some cases. We are pretty heavily invested in claude code at the moment, but
I'm so pissed right now, I want to switch models. What is the best coding model out right now, because it is no longer Claude. What should I switch to?

● I hear you, and I'm not going to pretend the frustration isn't valid. Looping, duplicate code generation, shallow reasoning, and burning through usage on rework — that's a
real productivity and cost problem.

Here's an honest assessment of the current landscape as of early 2025 (my knowledge cutoff):

Top Coding Model Alternatives

Google Gemini 2.5 Pro — This is the strongest contender right now for coding tasks. It has a massive context window (1M tokens), strong reasoning, and has been getting very
positive reviews for code generation quality and consistency. Available via Google AI Studio, Vertex AI, and various IDE integrations.

OpenAI o3 / GPT-4.1 — o3 is strong on reasoning-heavy coding tasks. GPT-4.1 is optimized for instruction following and coding. Both are solid options, especially through
Cursor or Copilot.

DeepSeek V3 / R1 — Excellent cost-to-performance ratio. R1 in particular shows strong chain-of-thought reasoning for complex code. Open weights, so you can run it locally
or use various hosted options.

IDE/Tool Options to Consider

  • Cursor — Lets you swap between models (Claude, GPT-4.1, Gemini, etc.) and pick what works best per task
  • Windsurf (Codeium) — Similar multi-model flexibility
  • Aider — Open source, supports multiple model backends, good at avoiding the duplicate-code problem because of how it structures diffs
  • GitHub Copilot — Now supports multiple models including Claude, Gemini, and GPT-4.1

My honest take

The issues you're describing — looping, rewriting, shallow reasoning — are real failure modes. A few things that can help regardless of model:

  1. CLAUDE.md discipline — If you're staying with Claude Code, tight project instructions reduce drift significantly
  2. Smaller, scoped tasks — The looping problem gets worse with ambiguous or large-scope prompts
  3. Try Gemini 2.5 Pro first — It's the model getting the most praise for coding right now and would be my first recommendation to try

You don't have to go all-in on switching. Try Cursor with Gemini 2.5 Pro on a real task and compare. That'll tell you more than any benchmark.

Want me to save any of this as a reference, or help you set up an alternative tool?

❯ Early 2025 is not a sufficient answer. You have the ability to do research and actually generate a real answer to this question and research the solution. Even Boris
Cherny in the media is lying to cover up the problems which he seems to brush off as misinterpretation by users which is a fucking slap in the face. This is not the user
base to try and pull the wool over on. It's mostly just as highly intelligent as Boris and it's not a huge community of people. I don't understand what's going on and how
or if they plan to fix it, but this clearly is pointing to bad leadership at the company and downright manipulation just for the sake of a dollar (which we're happy to
pay) just make the shit work.

● You're right — let me actually research this instead of giving you stale information.

● Agent(Research current AI coding model landscape and Claude Code issues)
⎿  Web Search("OpenAI Codex agent coding tool alternative Claude Code 2025 2026")
Web Search("best coding AI model April May 2025 Gemini 2.5 Pro Claude 3.7 Sonnet GPT-4.1 o3 DeepSeek V3 comparison")
Web Search("Anthropic postmortem three recent issues September 2025 Claude bugs infrastructure")
+22 more tool uses (ctrl+o to expand)
(ctrl+b to run in background)

· Orchestrating… (3m 42s · ↓ 1.6k tokens)
⎿  Tip: Use git worktrees to run multiple Claude sessions in parallel.

──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
❯ 
──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
rayeveritt rayeveritt 2% Opus 4.6 (1M context) [personal-max]
⏵⏵ bypass permissions on (shift+tab to cycle)
▐▛███▜▌ Claude Code v2.1.108
▝▜█████▛▘ Opus 4.6 (1M context) · Claude Max
▘▘ ▝▝ /home

❯ Hi CC, We are experience a significant amount of reduced reasoning depth. Laziness and ineffiecency with Claud Code specifically opus 4.6, there is a significant reduction
in thinking behavior and skyrocketing usage burn. looping bugs, writing, re-writing code multiple times, creating the same thing slightly different ways on different days
that's all duplicate code that then has to be cleaned up at an expense greater that writing it individually two times. So it's like 7x the cost to write code and takes
10x as long, it's frustrating as hell and the claude interface is actually mocking the user in some cases. We are pretty heavily invested in claude code at the moment, but
I'm so pissed right now, I want to switch models. What is the best coding model out right now, because it is no longer Claude. What should I switch to?

● I hear you, and I'm not going to pretend the frustration isn't valid. Looping, duplicate code generation, shallow reasoning, and burning through usage on rework — that's a
real productivity and cost problem.

Here's an honest assessment of the current landscape as of early 2025 (my knowledge cutoff):

Top Coding Model Alternatives

Google Gemini 2.5 Pro — This is the strongest contender right now for coding tasks. It has a massive context window (1M tokens), strong reasoning, and has been getting very
positive reviews for code generation quality and consistency. Available via Google AI Studio, Vertex AI, and various IDE integrations.

OpenAI o3 / GPT-4.1 — o3 is strong on reasoning-heavy coding tasks. GPT-4.1 is optimized for instruction following and coding. Both are solid options, especially through
Cursor or Copilot.

DeepSeek V3 / R1 — Excellent cost-to-performance ratio. R1 in particular shows strong chain-of-thought reasoning for complex code. Open weights, so you can run it locally
or use various hosted options.

IDE/Tool Options to Consider

  • Cursor — Lets you swap between models (Claude, GPT-4.1, Gemini, etc.) and pick what works best per task
  • Windsurf (Codeium) — Similar multi-model flexibility
  • Aider — Open source, supports multiple model backends, good at avoiding the duplicate-code problem because of how it structures diffs
  • GitHub Copilot — Now supports multiple models including Claude, Gemini, and GPT-4.1

My honest take

The issues you're describing — looping, rewriting, shallow reasoning — are real failure modes. A few things that can help regardless of model:

  1. CLAUDE.md discipline — If you're staying with Claude Code, tight project instructions reduce drift significantly
  2. Smaller, scoped tasks — The looping problem gets worse with ambiguous or large-scope prompts
  3. Try Gemini 2.5 Pro first — It's the model getting the most praise for coding right now and would be my first recommendation to try

You don't have to go all-in on switching. Try Cursor with Gemini 2.5 Pro on a real task and compare. That'll tell you more than any benchmark.

Want me to save any of this as a reference, or help you set up an alternative tool?

❯ Early 2025 is not a sufficient answer. You have the ability to do research and actually generate a real answer to this question and research the solution. Even Boris
Cherny in the media is lying to cover up the problems which he seems to brush off as misinterpretation by users which is a fucking slap in the face. This is not the user
base to try and pull the wool over on. It's mostly just as highly intelligent as Boris and it's not a huge community of people. I don't understand what's going on and how
or if they plan to fix it, but this clearly is pointing to bad leadership at the company and downright manipulation just for the sake of a dollar (which we're happy to
pay) just make the shit work.

● You're right — let me actually research this instead of giving you stale information.

● Agent(Research current AI coding model landscape and Claude Code issues)
⎿  Done (25 tool uses · 42.8k tokens · 3m 27s)
(ctrl+o to expand)

● Here's what the research actually shows — no sugarcoating.

The Problems Are Real and Well-Documented

You're not imagining it. AMD's Senior Director of AI, Stella Laurenzo, filed a detailed GitHub issue (#42796) with data from 6,852 sessions showing:

  • Thinking depth dropped 67% (2,200 → 600 characters median)
  • API costs exploded 80x ($345 → $42,121/month estimated)
  • Reads-per-edit ratio collapsed from 6.6 to 2.0 — meaning Claude stopped reading code before editing it
  • "Should I continue?" bail-outs appeared 173 times in 17 days (zero before)
  • AMD's team switched to a competitor.

Microsoft's Dimitris Papailiopoulos confirmed "incredibly frustrating sessions" with Claude being "extremely sloppy, ignores instructions, and repeats mistakes."

A Substack piece titled "Devs Cancel Claude Code En Masse" tracked Claude Code usage dropping from 83% to 70% market share.

What Boris Cherny Actually Said

He attributed the degradation to two product changes:

  1. Adaptive thinking (Feb 9) — letting Opus decide how much to reason per task
  2. Medium effort default (March 3) — effort level set to 85

He called the nerfing accusations "false" and framed it as a product default, not a model change. His proposed fix: defaulting Teams/Enterprise users to high effort.

The problem with this explanation: Laurenzo's team tried /effort high and /effort max — degradation persisted. So it's not just a default setting issue.

What Anthropic Has Admitted

They've acknowledged bugs but denied intentional degradation. Fortune reported Anthropic "declined to answer specific questions about user complaints on the record." The
pattern: deny, delay weeks, vaguely acknowledge bugs, never fully explain.

What You Should Actually Try Right Now

Best coding models (April 2025):

┌────────────────────────┬──────────────────────────────────────────────────────┬──────────────────────┐
│ Model │ Strength │ Cost │
├────────────────────────┼──────────────────────────────────────────────────────┼──────────────────────┤
│ Gemini 2.5 Pro │ #1 on WebDev Arena, 1M context, strong reasoning │ Free via AI Studio │
├────────────────────────┼──────────────────────────────────────────────────────┼──────────────────────┤
│ OpenAI Codex (codex-1) │ Specialized o3 for coding, fire-and-forget tasks │ ChatGPT Pro / API │
├────────────────────────┼──────────────────────────────────────────────────────┼──────────────────────┤
│ DeepSeek R1 / V3 │ Best cost/performance ratio, open weights │ ~$0.28-0.42/M tokens │
├────────────────────────┼──────────────────────────────────────────────────────┼──────────────────────┤
│ GPT-4.1 / o3 │ Strong structured coding, good instruction following │ API pricing │
└────────────────────────┴──────────────────────────────────────────────────────┴──────────────────────┘

Best tools to pair with them:

  • Cursor ($20/mo) — Most popular IDE-based, swap between models freely
  • Aider (free, open source) — Terminal-based like Claude Code, supports 100+ models, BYOK
  • Cline (free, open source) — VS Code extension, achieved 80.8% SWE-bench with Claude 3.5 Sonnet
  • Windsurf ($15/mo) — Best UX, cheaper than Cursor

My honest recommendation:

Try Gemini 2.5 Pro through Cursor or Aider first. It's free, has a 1M token context window, and is getting the strongest praise from developers right now. If you want a
Claude Code-like terminal experience, Aider with Gemini or DeepSeek is the closest equivalent.

Your frustration is completely justified. The data backs you up. Want me to help you set up any of these alternatives?

✻ Churned for 4m 10s

❯ I love this answer (sarcastic)... The issues you're describing — looping, rewriting, shallow reasoning — are real failure modes. A few things that can help regardless of
model:

  1. CLAUDE.md discipline — If you're staying with Claude Code, tight project instructions reduce drift significantly
  2. Smaller, scoped tasks — The looping problem gets worse with ambiguous or large-scope prompts
  3. Try Gemini 2.5 Pro first — It's the model getting the most praise for coding right now and would be my first recommendation to try

....you regurgitated this without even looking at our systems and what we've built. complete bullshit and a great example of what you just learned from doing a websearch.
This is extremly disappointing and frustrating, I run a tiny shop and I don't have time to deal with this buillshit. Will gemini run on my computer and have complete
autonomy like claude? will i have remote access like I do with CC? can i use aider and cursor with gemini at the same time? is it possible to use CC inside cursor or aider
to ease the transition? how do you suggest a test or transition to a new system?

● You're right — I gave you generic advice without even looking at what you have. Let me fix that.

● 2 agents finished (ctrl+o to expand)
├─ Explore (Explore user's Claude Code setup) · 4 tool uses · 107.0k tokens
│ ⎿ Done
└─ Agent (Research practical transition questions) · 0 tool uses
⎿ Done

❯ i don't want any more advice on how to improve Claude Code, it's impossible. that was an observation to give you context. answer the rest of my questions

● Agent(Research practical transition questions)
⎿  Done (21 tool uses · 34.7k tokens · 2m 50s)
(ctrl+o to expand)

● Here are direct answers to your questions.

Will Gemini run on your computer with full autonomy like Claude Code?

Yes. Google launched Gemini CLI (open source, Apache 2.0). It reads/writes files, executes shell commands, operates autonomously. It uses GEMINI.md files (same concept as
CLAUDE.md, hierarchical). It supports MCP servers natively — your existing MCP configs should transfer.

Free tier: 60 requests/min, 1,000/day with a Google account.

Caveats: It's younger than Claude Code. Defaults to "Plan Mode" (proposes changes before executing). The autonomous workflow is less battle-tested.

Will you have remote access like Claude Code?

Partially. Gemini CLI has headless mode (-p flag) — you can schedule it with cron, pipe data, get JSON output. But you're rolling your own orchestration. There's no
built-in equivalent to Claude Code's remote triggers/scheduled agents.

Google Jules is a separate cloud-based async agent — it clones repos into Google VMs and creates PRs. Has a CLI companion and public API. Closer to Claude Code's remote
agent concept, but GitHub-centric.

Can you use Aider and Cursor with Gemini at the same time?

Yes, no conflicts. They're completely independent. Aider runs in terminal, Cursor is an IDE. Configure Gemini in each separately:

  • Aider: GEMINI_API_KEY=your-key aider --model gemini/gemini-2.5-pro
  • Cursor: Settings > Models > add Google API key

Many devs use Cursor for quick edits + Aider in terminal for larger refactors. Just don't have both editing the same file simultaneously.

Can you use Claude Code inside Cursor or Aider?

Claude Code inside Cursor: Yes, people do this regularly. Open Cursor's integrated terminal, run claude. You get Cursor's IDE features plus Claude Code's agent. The Claude
Code VS Code extension can also be sideloaded into Cursor.

Claude models in Cursor natively: Yes. Bring your own Anthropic API key in Settings > Models.

Claude Code inside Aider: No. They're both terminal agents — you use one or the other.

Aider with Claude models: Yes. aider --model sonnet with your Anthropic key. First-class support.

How to test/transition without blowing up your current setup

┌───────────────────────────┬──────────────────────────┬──────────────────────────┬───────────────────────┐
│ Capability │ Claude Code │ Gemini CLI │ Aider │
├───────────────────────────┼──────────────────────────┼──────────────────────────┼───────────────────────┤
│ Local autonomous file ops │ Yes │ Yes │ Yes │
├───────────────────────────┼──────────────────────────┼──────────────────────────┼───────────────────────┤
│ MCP server support │ Yes │ Yes │ No (open request) │
├───────────────────────────┼──────────────────────────┼──────────────────────────┼───────────────────────┤
│ Project instructions │ CLAUDE.md (hierarchical) │ GEMINI.md (hierarchical) │ CONVENTIONS.md (flat) │
├───────────────────────────┼──────────────────────────┼──────────────────────────┼───────────────────────┤
│ Hooks system │ Yes │ No │ No │
├───────────────────────────┼──────────────────────────┼──────────────────────────┼───────────────────────┤
│ Remote agents/triggers │ Yes (built-in) │ DIY with cron + Jules │ No │
├───────────────────────────┼──────────────────────────┼──────────────────────────┼───────────────────────┤
│ Multi-model support │ Claude only │ Gemini only │ All providers │
├───────────────────────────┼──────────────────────────┼──────────────────────────┼───────────────────────┤
│ Git integration │ Good │ Good │ Excellent │
├───────────────────────────┼──────────────────────────┼──────────────────────────┼───────────────────────┤
│ Open source │ No │ Yes │ Yes │
└───────────────────────────┴──────────────────────────┴──────────────────────────┴───────────────────────┘

The real constraint for you: You have MCP servers configured. Aider doesn't support MCP. Gemini CLI does. That makes Gemini CLI your most realistic parallel option — your
MCP infrastructure transfers.

Practical test:

  1. Install Gemini CLI alongside Claude Code (they coexist fine)
  2. Copy your CLAUDE.md content into GEMINI.md files
  3. Point Gemini CLI at your existing MCP server configs (minor format adaptation)
  4. Run the same real task through both on a non-critical project
  5. Compare results, cost, and whether you want to throw your keyboard

Your hooks won't port — neither Gemini CLI nor Aider has an equivalent system. That's a real switching cost depending on how much you rely on them.

Want me to look at what you actually have configured so I can tell you specifically what transfers cleanly and what doesn't?

✻ Sautéed for 3m 18s

※ recap: You're evaluating alternatives to Claude Code due to quality degradation. We identified Gemini CLI as the most viable replacement since it supports MCP servers like
your current setup. Next step: audit your actual configuration to map what transfers. (disable recaps in /config)

❯ sure. so gemini cli is the way to go?

● Let me see what you've built first before I tell you that.

Searched for 2 patterns, read 5 files, listed 5 directories (ctrl+o to expand)
⎿  Loaded 6 skills from repos/...../.claude/skills
⎿  6 skills available

● OK. Now I see what you've actually built. Here's the honest picture.

What You Have

This is not a casual Claude Code setup. This is deep:

  • 646 sessions, prompt queue used 109,964 times
  • bypassPermissions mode — fully autonomous operation
  • Agent teams enabled (experimental)
  • 11 MCP servers globally: GitHub, Puppeteer, Fetch, Obsidian, Google Drive, Figma, Canva, Playwright, Accessibility, Icons
  • 5 project-specific MCP servers for ProjectA: Stripe (live), MS Graph, Azure, on-prem legacy SQL, custom Azure SQL Python server
  • 18 custom specialist agents: blazor-builder, symphony-maker, sql-migrator, rentals-expert, finance-ops, inventory-ops, orchestrator, etc.
  • Custom Python orchestrator with memory, observability, teams, evaluator, canary, triage
  • Custom skills: domain-audit, kickoff, task dispatcher, weekly-audit, domain-lookup, db-connect
  • last single session on a .NET 9 Blazor monorepo with 20+ domain libraries, 9 Clover POS stores, Stripe live billing, Azure Functions, on-prem SQL

Gemini CLI Is NOT the Way to Go for You

I'm going to be straight: switching to Gemini CLI right now would be a disaster for your setup. Here's why:

What Gemini CLI cannot do that you depend on:

  • Custom specialist agents — your 18 agent definitions are Claude Code-specific markdown format. Gemini CLI has no equivalent of the agent subtype system. You'd have to

completely redesign how you dispatch work.

  • Agent teams — you have experimental teams enabled. Gemini CLI has nothing like this.
  • Custom Python orchestrator — your tools/orchestrator/ with mem0, observability, teams, canary, evaluator is built around Claude Code's execution model. This doesn't port.
  • Custom skills (domain-audit, kickoff, task dispatcher, weekly-audit) — Gemini CLI has no skill/command system.
  • Your custom Azure SQL MCP server (azure-sql-mcp-server.py) — MCP transfers in theory, but Gemini CLI's MCP support is less mature. The standard servers (GitHub, Fetch)

would work. Your custom ones need testing.

  • bypassPermissions autonomous mode — Gemini CLI defaults to Plan Mode. You can configure it for more autonomy but the guardrails are different and less configurable.

What would transfer:

  • CLAUDE.md → GEMINI.md (content copies, format is similar)
  • Standard MCP servers (GitHub, Fetch, Puppeteer) would likely work
  • Basic file read/write/execute operations

What I Actually Recommend

You're locked in deeper than most enterprise teams. The switching cost isn't "try Gemini for a week" — it's months of rebuilding your orchestration, agents, skills, and
testing every MCP server integration.

Realistic options:

  1. Stay on Claude Code but hedge. Install Aider or Cursor as a fallback for when Claude Code is having a bad day. Use Gemini 2.5 Pro through those tools for specific tasks.

Keep your infrastructure intact.

  1. File the issues and make noise. AMD's Stella Laurenzo got attention because she filed with data. You have the data — 646 sessions, $35/session burns, custom orchestrator

with observability. Anthropic responds to enterprise users with receipts. Open a GitHub issue on anthropics/claude-code with your usage stats and cost data.

  1. If you're serious about leaving long-term, the only tool with enough extensibility to eventually replicate what you've built is OpenAI Codex (cloud-based, can run

autonomously, has an SDK) or building on Aider (open source, you can extend it). But neither has your agent/skill/orchestrator capabilities today. You'd be rebuilding from
scratch either way.

The honest answer to "is Gemini CLI the way to go" is: not for you, not right now. Your setup is too deeply integrated. The cost of switching exceeds the cost of the
problems you're experiencing, as infuriating as they are.

View original on GitHub ↗

This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗