claude-code-patches
A simple patch that toggles claude-code thinking traces to be fully expanded by default without having to press ctrl + o
Star
Fork
Watch
77 6 0
Files
patch-subagent-models.js
patch-thinking.js
README.md
README.md

Claude Code Patches #

Enhance Claude Code with custom patches for thinking display and subagent model configuration.

Available Patches #

  1. Thinking Display Patch - Make thinking blocks visible by default
  2. Subagent Model Configuration - Configure which models subagents use

Thinking Display Patch #

Make Claude Code's thinking blocks visible by default without pressing ctrl+o.

The Problem #

Claude Code collapses thinking blocks by default, showing only:

∴ Thought for 3s (ctrl+o to show thinking)

You have to press ctrl+o every time to see the actual thinking content. This patch makes thinking blocks visible inline automatically.

Current Version: Claude Code 2.1.86 (Updated 2026-03-28)

Required Setting (v2.1.64+) #

Starting in v2.1.64 (briefly reverted, permanent from v2.1.69), Claude Code sends a redact-thinking-2026-02-12 beta header with every API request. This tells the API to strip thinking text from the response — thinking blocks still arrive with valid cryptographic signatures, but the thinking property is an empty string. The rendering patch is correct but has nothing to display.

To disable this redaction, you must add the following to ~/.claude/settings.json:

{
  "showThinkingSummaries": true
}

This setting is not documented in the official Claude Code settings docs. Its internal schema description is "Show thinking summaries in the transcript view (ctrl+o). Default: false." — but what it actually controls is whether the API request includes the redact-thinking beta header. Without it, the API never sends thinking content to the client, and no client-side patch can make it visible.

Without this setting, the patch will apply cleanly but you will see no thinking content. The patch script will warn you if the setting is missing.

See also: anthropics/claude-code#31326 — upstream bug report confirming thinking content is empty since v2.1.69.

Quick Start #

# 1. Enable thinking content from the API (required since v2.1.64)
# Add "showThinkingSummaries": true to ~/.claude/settings.json

# 2. Clone or download this repository
cd claude-code-thinking

# 3. Run the patch script (automatically detects your installation)
node patch-thinking.js

# 4. Restart Claude Code

That's it! Thinking blocks now display inline without ctrl+o.

Works with:

  • ✅ Local installations (~/.claude/local)
  • ✅ Global npm installations (npm install -g @anthropic-ai/claude-code)
  • ✅ All Node version managers (NVM, nodenv, asdf, etc.)

What This Patch Does #

Before:

∴ Thought for 3s (ctrl+o to show thinking)
[thinking content hidden]

After:

∴ Thinking…

  [thinking content displayed inline]
  The actual thinking process is now visible
  without any keyboard shortcuts needed

How It Works #

This patch modifies two locations in Claude Code's compiled JavaScript:

Patch 1: Remove the Banner (v2.0.30) #

Before:

function GkQ({streamMode:A}){
  // ... displays "Thought for Xs (ctrl+o to show thinking)"
}

After:

function GkQ({streamMode:A}){return null}

Effect: Removes the collapsed thinking banner entirely.

Version History:

  • v2.0.9: Function named Mr2
  • v2.0.10: Renamed to br2, used PE.createElement
  • v2.0.11: Renamed to er2, uses _E.createElement
  • v2.0.13: Renamed to hGB, uses TL.createElement
  • v2.0.14: Renamed to pGB, uses TL.createElement, TX1.useState
  • v2.0.15: Renamed to KYB, uses xL.createElement, mX1.useState
  • v2.0.19: Renamed to aFB, uses ZM.createElement, BV1.useState
  • v2.0.21: Renamed to wVB, uses XM.createElement, DV1.useState
  • v2.0.22: Renamed to YOB, uses NM.createElement, zK1.useState
  • v2.0.24: Renamed to GSB, uses oM.createElement, kD1.useState
  • v2.0.25: Renamed to YSB, uses tM.createElement, xD1.useState
  • v2.0.26: Renamed to KjQ, uses QO.createElement, gKA.useState
  • v2.0.28: Renamed to RjQ, uses IO.createElement, iKA.useState
  • v2.0.29: Unchanged from v2.0.28 (RjQ, IO.createElement, iKA.useState)
  • v2.0.30: Renamed to GkQ, uses NO.createElement, dDA.useState
  • v2.0.31: Renamed to _kQ, uses MO.createElement, nDA.useState
  • v2.0.32: Renamed to wkQ, uses LO.createElement, oDA.useState
  • v2.0.37: Renamed to nR2, uses CR.createElement, AwA.useState
  • v2.0.42: Renamed to cR2, uses RR.createElement, UwA.useState
  • v2.0.46: Renamed to Et2, uses xP.createElement, HTA.useState
  • v2.0.53: Renamed to hq2, uses QP.createElement, dMA.useState
  • v2.0.55: Renamed to nM2, uses UP.createElement, bOA.useState
  • v2.0.56: Renamed to CL2, uses HP.createElement, xOA.useState
  • v2.0.57: Renamed to HM2, uses EP.createElement, oOA.useState
  • v2.0.58: Renamed to SM2, uses $P.createElement, GRA.useState
  • v2.0.59: Renamed to DO2, uses MP.createElement, CRA.useState
  • v2.0.61: Renamed to RR2, uses rj.createElement, vTA.useState, P container
  • v2.0.62: Renamed to ZT2, uses GP.createElement, rTA.useState, P container

Patch 2: Force Thinking Visibility (v2.0.46) #

Before:

case"thinking":if(!K&&!Z)return null;
  return H7.createElement(T32,{addMargin:Q,param:A,isTranscriptMode:K,verbose:Z});

After:

case"thinking":
  return H7.createElement(T32,{addMargin:Q,param:A,isTranscriptMode:!0,verbose:Z});

Effect: Forces thinking content to render as if in transcript mode (visible).

Version History:

  • v2.0.9: Used S2B component
  • v2.0.10: Changed to DOB component, z→H variable
  • v2.0.11: Changed to SOB component, H→z variable
  • v2.0.13: Changed to xlB component, K→D variable swap
  • v2.0.14: Changed to dlB component, maintains K and D variable pattern
  • v2.0.15: Changed to FpB component, z3→C3 variable
  • v2.0.19: Changed to NoB component, C3→B7 variable
  • v2.0.21: Changed to H8Q component, B7→G7 variable, checks K and D
  • v2.0.22: Changed to nNB component, G7→e3 variable, checks K and D
  • v2.0.24: Changed to nTB component, e3→Y7 variable, checks K and D
  • v2.0.25: Changed to aTB component, maintains Y7 variable, checks K and D
  • v2.0.26: Changed to CTQ component, Y7→Y3 variable, checks only V
  • v2.0.28: Changed to LTQ component, Y3→C3 variable, checks only V
  • v2.0.29: Unchanged from v2.0.28 (LTQ component, C3 variable, checks V)
  • v2.0.30: Changed to sjQ component, C3→D3 variable, checks V and I, added verbose parameter
  • v2.0.31: Changed to MSQ component, D3→E3 variable, checks V and I
  • v2.0.32: Changed to ljQ component, E3→F3 variable, checks V and I
  • v2.0.37: Changed to n$Q component, F3→K3 variable, checks V and I
  • v2.0.42: Changed to xLQ component, K3→w3 variable, checks V and I
  • v2.0.46: Changed to T32 component, w3→H7 variable, checks K and Z
  • v2.0.53: Changed to o09 component, H7→L3 variable, checks K and G
  • v2.0.55: Changed to J29 component, L3→y3 variable, checks K and G
  • v2.0.56: Changed to b29 component, y3→v3 variable, checks K and G
  • v2.0.57: Changed to K49 component, v3→b3 variable, checks K and G
  • v2.0.58: Changed to k49 component (lowercase k), b3 variable, checks K and G
  • v2.0.59: Changed to F89 component, u3 variable, checks K and G
  • v2.0.61: Changed to T69 component, A3 variable, checks F and G
  • v2.0.62: Changed to X59 component, J3 variable, checks F and G
  • v2.1.31+: Structural change to memo cache block with q[] indices
  • v2.1.80: rE8 component, S3 namespace, guard if(!X&&!w), two-layer gating
  • v2.1.81: Same patterns as v2.1.80
  • v2.1.84: uL8 component, C5 namespace, guard if(!P&&!w), memo q[31-36]

Installation #

Prerequisites #

  • Claude Code v2.1.84 installed
  • Node.js (comes with Claude Code installation)

Install Steps #

  1. Download the patcher:

    # Clone this repository
git clone <repository-url>
cd claude-code-thinking

  2. Run the patcher:

    node patch-thinking.js

  3. Restart Claude Code for changes to take effect.

Command-Line Options #

The script supports several options:

# Apply patches (default)
node patch-thinking.js

# Preview changes without applying
node patch-thinking.js --dry-run

# Restore original behavior from backup
node patch-thinking.js --restore

# Show help
node patch-thinking.js --help

Installation Detection #

The script automatically detects Claude Code installations using a robust 4-tier detection strategy:

Detection Methods (Priority Order) #

  1. Local Installations (Priority 1)

    • ~/.claude/local/node_modules/@anthropic-ai/claude-code/cli.js
    • ~/.config/claude/local/node_modules/@anthropic-ai/claude-code/cli.js

  2. Global npm Installation (Priority 2)

    • Dynamically detected via npm root -g
    • Works with all Node version managers (NVM, nodenv, asdf, nvm-windows)
    • Automatically resolves symlinks

  3. Derived from Node.js Binary (Priority 3)

    • Falls back to process.execPath derivation
    • Works when npm command is unavailable

  4. Unix Binary Location (Priority 4)

    • Uses which claude on macOS/Linux
    • Traces binary back to installation directory

Key Features #

  • ✅ Works with any Node version - No hardcoded paths
  • ✅ Supports all version managers - NVM, nodenv, asdf, etc.
  • ✅ Cross-platform - macOS, Linux, Windows
  • ✅ Automatic symlink resolution - Handles npm global symlinks
  • ✅ Comprehensive error messages - Shows all attempted paths

Backup Created:

  • cli.js.backup (in the same directory as cli.js)

Important: After Claude Code Updates #

When you run claude update, the patches will be overwritten. You must re-apply them:

cd claude-code-thinking
node patch-thinking.js
# Restart Claude Code

The patch script automatically:

  • Detects your Claude Code installation
  • Creates a backup before patching (if it doesn't exist)
  • Applies both patches atomically
  • Reports success or failure
  • Safe to run multiple times

Rollback #

To restore the original behavior:

Option 1: Using the script

node patch-thinking.js --restore

Option 2: Manual restore

# The backup is created in the same directory as cli.js
cp ~/.claude/local/node_modules/@anthropic-ai/claude-code/cli.js.backup \
   ~/.claude/local/node_modules/@anthropic-ai/claude-code/cli.js

Then restart Claude Code.

Verification #

Check if patch is applied (for v2.1.84):

# Check thinking visibility patch (should show if(0) instead of if(!P&&!w))
grep -o 'case"thinking":{if(0)' $(npm root -g)/@anthropic-ai/claude-code/cli.js

# Should output: case"thinking":{if(0)

Troubleshooting #

"Could not find Claude Code installation" #

Cause: The script cannot locate your Claude Code installation.

The script will display all attempted detection methods and paths, for example:

❌ Error: Could not find Claude Code installation

Searched using the following methods:

  [local installation]
    - ~/.claude/local/node_modules/@anthropic-ai/claude-code/cli.js
    - ~/.config/claude/local/node_modules/@anthropic-ai/claude-code/cli.js
  [npm root -g]
    - /opt/homebrew/lib/node_modules/@anthropic-ai/claude-code/cli.js
  [derived from process.execPath]
    - /opt/homebrew/bin/../lib/node_modules/@anthropic-ai/claude-code/cli.js
  [which claude]
    - /opt/homebrew/bin/../lib/node_modules/@anthropic-ai/claude-code/cli.js

💡 Troubleshooting:
  1. Verify Claude Code is installed: claude --version
  2. For local install: Check ~/.claude/local or ~/.config/claude/local
  3. For global install: Ensure "npm install -g @anthropic-ai/claude-code" succeeded
  4. Check that npm is in your PATH if using global install

Solutions:

  1. Verify installation: Run claude --version to confirm Claude Code is installed
  2. For local installations: Check that cli.js exists in ~/.claude/local or ~/.config/claude/local
  3. For global installations:
    • Ensure npm install -g @anthropic-ai/claude-code completed successfully
    • Verify npm is in your PATH: npm --version
    • If using NVM: Ensure you've activated the correct Node version
  4. Check file permissions: Ensure the script has read access to the installation directory

"Pattern not found" #

Cause: This means:

  1. Claude Code has been updated to a newer version
  2. The patches are already applied
  3. The file structure has changed

Solution:

  1. Run node patch-thinking.js --dry-run to check status
  2. If already applied, you're good!
  3. If version changed, the patterns may need updating for the new version

Thinking Still Collapsed After Patching #

Solution: You must restart Claude Code for changes to take effect.

Backup File Missing #

The patch script creates a backup automatically on first run. The --restore command will fail if the backup doesn't exist.

Cross-Platform Support #

The script works on:

  • macOS ✅ Full support (all 4 detection methods)
  • Linux ✅ Full support (all 4 detection methods)
  • Windows ✅ Supported (3 detection methods - excludes which claude)

Platform-Specific Notes #

macOS & Linux:

  • All 4 detection methods available
  • Includes which claude binary resolution
  • Automatic symlink resolution for Homebrew, NVM, etc.

Windows:

  • Uses first 3 detection methods
  • Works with nvm-windows and system Node
  • Skips Unix-specific which command

Path detection is fully automatic using Node.js built-in modules.

Version Manager Support #

The script automatically works with all Node.js version managers:

Version Manager Support Notes
NVM (Node Version Manager) ✅ Full Detects via npm root -g
nodenv ✅ Full Detects via process.execPath
asdf ✅ Full Detects via npm root -g
nvm-windows ✅ Full Works with all 3 Windows methods
System Node ✅ Full Standard installation detection
Homebrew (macOS) ✅ Full Symlink resolution included

How it works:

  • npm root -g dynamically finds the global node_modules directory regardless of version manager
  • process.execPath derives the path from the current Node.js binary location
  • No hardcoded paths means it works with any Node setup

Technical Details #

File Structure #

  • cli.js: ~3,600+ lines, ~9+ MB (heavily minified)
  • Version: Claude Code 2.1.84
  • Patches: Non-invasive, minimal changes

Installation Detection System #

The patcher uses a sophisticated detection system with multiple fallback methods:

// Priority 1: Local installations
~/.claude/local/node_modules/@anthropic-ai/claude-code/cli.js
~/.config/claude/local/node_modules/@anthropic-ai/claude-code/cli.js

// Priority 2: Global npm (dynamic)
$(npm root -g)/@anthropic-ai/claude-code/cli.js

// Priority 3: Derived from Node binary
$(dirname process.execPath)/../lib/node_modules/@anthropic-ai/claude-code/cli.js

// Priority 4: Unix binary resolution
$(which claude) → resolve symlinks → find cli.js

Key implementation details:

  • Uses child_process.execSync() with proper error handling
  • Automatically resolves symlinks via fs.realpathSync()
  • Tracks all attempted paths for detailed error reporting
  • Cross-platform compatible (Windows, macOS, Linux)

How It Works (Two-Layer Gating) #

The case"thinking" handler has two independent layers that suppress thinking output:

  1. Layer 1 — Early return guard: if(!P&&!w)return null Returns null when not in transcript mode AND not verbose. The patch changes this to if(0)return null (dead code).

  2. Layer 2 — Component prop: isTranscriptMode:P Controls whether the component shows content or collapses it. The patch changes this to isTranscriptMode:!0.

Both layers must be fixed or thinking stays invisible. The banner function (ZT2/vo4 etc.) was deprecated since v2.0.71.

Pattern Evolution Across Versions #

The minified code patterns change with each Claude Code update:

Version Banner Function Component Variables
2.0.9 Mr2 S2B Various
2.0.10 br2 DOB H check
2.0.11 er2 SOB z check
2.0.13 hGB xlB K check
2.0.14 pGB dlB K check
2.0.15 KYB FpB K check
2.0.19 aFB NoB K check
2.0.21 wVB H8Q K,D check
2.0.22 YOB nNB K,D check
2.0.24 GSB nTB K,D check
2.0.25 YSB aTB K,D check
2.0.26 KjQ CTQ V check
2.0.28 RjQ LTQ V check
2.0.29 RjQ LTQ V check
2.0.30 GkQ sjQ V,I check
2.0.31 _kQ MSQ V,I check
2.0.32 wkQ ljQ V,I check
2.0.37 nR2 n$Q V,I check
2.0.42 cR2 xLQ V,I check
2.0.46 Et2 T32 K,Z check
2.0.53 hq2 o09 K,G check
2.0.55 nM2 J29 K,G check
2.0.56 CL2 b29 K,G check
2.0.57 HM2 K49 K,G check
2.0.58 SM2 k49 K,G check
2.0.59 DO2 F89 K,G check
2.0.61 RR2 T69 F,G check
2.0.62 ZT2 X59 F,G check

When Claude Code updates, function names and component identifiers are regenerated during minification. In some cases (like v2.0.29), the patterns remain unchanged.

Limitations #

  1. Breaks on updates: Must re-run after claude update
  2. Minified code: Fragile, patterns may change with version updates
  3. No official config: This is a workaround until Anthropic adds a native setting
  4. Version-specific: Patterns are specific to v2.1.84

Feature Request #

Consider requesting this as an official feature from Anthropic:

  • Configuration option to always show thinking
  • User preference in settings (~/.claude/config.json)
  • Toggle command like /thinking show or /thinking hide
  • Environment variable like CLAUDE_SHOW_THINKING=true

Submit feedback: https://github.com/anthropics/claude-code/issues

Contributing #

If Claude Code updates and the patches stop working:

  1. Locate the new patterns in cli.js:

    • Search for the thinking banner function (look for "Thought for" text)
    • Search for case"thinking" to find the visibility check

  2. Update the script with new patterns

  3. Test thoroughly before committing

  4. Update this README with the new version information

Pull requests welcome!

Subagent Model Configuration #

Configure which AI models Claude Code uses for different subagent types (Plan, Explore, general-purpose).

The Problem #

By default, Claude Code hardcodes the models used by subagents:

  • Plan subagent: Uses Sonnet (for planning tasks)
  • Explore subagent: Uses Haiku (for code exploration)
  • general-purpose subagent: Inherits from main loop model

You cannot change these defaults without modifying the source code.

The Solution #

This patch allows you to configure subagent models via a configuration file (~/.claude/subagent-models.json).

Current Version: Claude Code 2.0.46

Quick Start #

# 1. Create configuration file
cat > ~/.claude/subagent-models.json <<EOF
{
  "Plan": "sonnet",
  "Explore": "sonnet",
  "general-purpose": "sonnet"
}
EOF

# 2. Run the patch script
node patch-subagent-models.js

# 3. Restart Claude Code

Configuration #

Create ~/.claude/subagent-models.json with your preferred models:

{
  "Plan": "sonnet",
  "Explore": "haiku",
  "general-purpose": "sonnet"
}

Valid model values:

  • "haiku" - Fast and efficient (Claude Haiku 4.5)
  • "sonnet" - Balanced performance (Claude Sonnet 4.5)
  • "opus" - Most capable (Claude Opus 4.1)

Subagent Types:

  • Plan - Used in planning mode for breaking down tasks
  • Explore - Used for codebase exploration and file searching
  • general-purpose - Used for general multi-step tasks

Command-Line Options #

# Apply patches
node patch-subagent-models.js

# Preview changes without applying
node patch-subagent-models.js --dry-run

# Restore original behavior
node patch-subagent-models.js --restore

# Show help
node patch-subagent-models.js --help

How It Works #

The patch modifies Claude Code's cli.js to change the hardcoded model assignments:

Before (v2.0.37):

// Plan subagent
R3A={agentType:"Plan",...,model:"sonnet"}

// Explore subagent
model:"haiku"}});var R3A;

After (with config: Plan="haiku", Explore="sonnet"):

// Plan subagent
R3A={agentType:"Plan",...,model:"haiku"}

// Explore subagent
model:"sonnet"}});var R3A;

Important Notes #

  1. After Claude Code Updates: Must re-run after claude update:

    node patch-subagent-models.js

  2. Backup: Created automatically at:

    ~/.claude/local/node_modules/@anthropic-ai/claude-code/cli.js.subagent-models.backup

  3. Version-Specific: Patterns are specific to v2.0.37. May need updates for newer versions.

Restoration #

Using the script:

node patch-subagent-models.js --restore

Manual restore:

cp ~/.claude/local/node_modules/@anthropic-ai/claude-code/cli.js.subagent-models.backup \
   ~/.claude/local/node_modules/@anthropic-ai/claude-code/cli.js

Example Use Cases #

Use Sonnet everywhere (avoid Haiku):

{
  "Plan": "sonnet",
  "Explore": "sonnet",
  "general-purpose": "sonnet"
}

Use Haiku for speed:

{
  "Plan": "haiku",
  "Explore": "haiku",
  "general-purpose": "haiku"
}

Balanced approach:

{
  "Plan": "sonnet",
  "Explore": "haiku",
  "general-purpose": "sonnet"
}

Troubleshooting #

"No model configuration found"

  • Create ~/.claude/subagent-models.json with your desired configuration
  • Ensure the JSON syntax is valid

"Pattern not found"

  • Claude Code version may have changed
  • Run with --dry-run to see which patterns are detected
  • Check if patches are already applied

Changes not taking effect

  • Restart Claude Code after applying patches
  • Verify patches were applied with --dry-run

Version History #

Version Plan Default Explore Default Notes
2.0.31 sonnet haiku Previous
2.0.32 sonnet haiku Previous
2.0.37 sonnet haiku Current

License #

This patch is provided as-is for educational purposes. Use at your own risk.

Credits #

Developed through analysis of Claude Code's compiled JavaScript. Special thanks to the community for identifying the thinking display issue.

Last Updated: 2026-03-26 Claude Code Version: 2.1.84 Status: ✅ Working

Quick Reference #

# Install
node patch-thinking.js

# Preview
node patch-thinking.js --dry-run

# Restore
node patch-thinking.js --restore

# Help
node patch-thinking.js --help