[Resource]: Claude Capsule Kit - Persistent memory + dependency intelligence for Claude Code

[arpitnath]

Display Name

Claude Capsule Kit

Category

Tooling

Sub-Category

None

Primary Link

https://github.com/arpitnath/super-claude-kit

Secondary Link

No response

Author Name

arpitnath

Author Link

https://github.com/arpitnath

License

MIT

Other License

No response

Description

Adds persistent context memory and dependency analysis to Claude Code through hook-based automation. Includes session resume with 24-hour persistence, dependency intelligence tools (circular detection, impact analysis, dead code identification), and progressive file reading with semantic AST chunking.
Built using SessionStart, PreToolUse, and PostToolUse hooks to track files, tasks, and discoveries across sessions.

Validate Claims

No response

Additional Comments

Production-tested over 2 weeks by 6 developers (frontend, backend, full-stack, DevOps) across TypeScript, Python, Go, and JavaScript codebases (200-2,000 files).

Built this after getting frustrated re-explaining the same context to Claude every session.

Unique features not in other Claude Code tools:

• Go module import resolution for dependency tracking
• XML-based hook outputs (60% token reduction vs emoji/box-drawing)
• Works with or without git (file-based change detection fallback)
• Subdirectory support (hooks work from any folder)

Credits: Anthropic's hook system made this possible.

Recommendation Checklist

• I have checked that this resource hasn't already been submitted
• My resource provides genuine value to Claude Code users, and any risks are clearly stated
• All provided links are working and publicly accessible
• I am submitting only ONE resource in this issue
• I understand that low-quality or duplicate submissions may be rejected

[hesreallyhim]

@arpitnath wow really nice thank you for sharing 🙏🏽 . i just started poking around and just a few quick comments: (i) are you familiar with "SuperClaude" framework? it's really popular so just a little worried about "brand confusion" but that's totally up to you of course, as is everything else i have to say; (ii) we have already a few persistent memory solutions on the list, so what really stands out to me is the intelligent file-reading / AST-based approach, i haven't seen that before in this context, as well as TOON - do you have an example/dummy/test repo (could even be some other open source project you don't own, doesn't matter) where I could clone it and use this capability and see for myself the results in token usage? (like "go to this folder and send this prompt" - sort of what you do in the side-by-side gifs) - that would be really helpful in validating. again, only took a brief look, so just my initial thoughts, but looks really really nice (and nice README) - regarding the name, it seems to me your project is young enough you could maybe avoid a naming collision with SuperClaude, but that's not my business, just my POV - e.g. SuperClaude is so popular i think there's plugins/etc built on top of SuperClaude - so if this isn't using SuperClaude I would have to say like "btw this is not the other SuperClaude" - anyway names are tough, please don't make any important decisions on my advice, either way i think this looks great and will take a closer look, but wanted to send a quick note asking if you could provide or suggest a test repo for the intelligent/efficient parsing/reader. Thanks.

[arpitnath]

@hesreallyhim Apologies for the late reply, I was traveling. For the name, I didn't realize "SuperClaude" was that established. Will update by today, probably something like "Claude Kit".

For testing it out, I used React's codebase, ReactFiberWorkLoop.js - 187KB, roughly around 45K tokens.
For token counting I added a small tool that uses Claude's tokenizer - https://huggingface.co/Xenova/claude-tokenizer

Start a claude with debug flag so that we can check the detailed tool call logs.
Send this prompt:
Summarize all functions in ReactFiberWorkLoop.js, organized by category (scheduling, error handling, rendering, commits, etc.). List each function with a one-line description.

Without the kit: Claude hits MaxFileReadTokenExceededError, then makes 8+ Read calls with offsets and limits basically guessing 800 lines at a time until it covers the file. Works perfectly, but inefficient.

With the SC kit: Same error, but Claude recognizes the file is large and switches to progressive-reader. One --list call gives the complete function structure (all ~100 functions with signatures). If it needs details on specific functions, it reads individual chunks. Total: 2-3 operations instead of 8+ read calls.

The core value isn't just "fewer operations" here, it's that progressive-reader chunks by AST boundaries (functions, classes), so Claude gets complete code units instead of arbitrarily cut lines. When we're trying to understand a thousand lines file, getting "lines 800-1200" is less useful than getting "here are all the scheduling functions.

And also smaller context (claude code context window) get populated means better attention. When Claude reads 8K tokens of exactly what it needs vs 45K tokens of everything, the responses are more focused.

Steps to test it yourself:

1. Setup
  git clone https://github.com/facebook/react.git
  cd react
  curl -fsSL https://raw.githubusercontent.com/arpitnath/super-claude-kit/master/install.sh | bash

  2. Install token counter dependencies (one-time)
  python .claude/tools/token-counter/token-counter.py --install

  3. Check full file token cost
  python .claude/tools/token-counter/token-counter.py packages/react-reconciler/src/ReactFiberWorkLoop.js
  # Should show ~44,980 tokens

  4. Run Claude in debug mode
  claude --debug

  5. Send the test prompt
  Summarize all functions in ReactFiberWorkLoop.js, organized by category (scheduling, error handling, rendering, commits, etc.). List each function with a one-line description.

  6. Observe and check the debug log for any details
    - You'll see MaxFileReadTokenExceededError in the output
    - Then Claude switches to progressive-reader, 
    Check debug log for details (open the latest one and look for)
    - "MaxFileReadTokenExceededError"
    - "executePreToolHooks called for tool: Bash" (progressive-reader calls)

  8. Validate token savings with the chunks Claude used
  python3 .claude/tools/token-counter/token-counter.py packages/react-reconciler/src/ReactFiberWorkLoop.js 19,20,21

[hesreallyhim]

@arpitnath thanks, we have a ClaudeKit already actually lol. what about "Claude Mastermind" or something 🤷🏽

i don't know how busy you are, but if this is something easy for you to set up and demo and you can record a screencast (you can share it privately even if it shows your directory names or whatever) that would be really helpful, we're hopefully going to get some good traffic this week 🤞🏽 so i'm focused on getting that together.

[arpitnath]

Sure, i will add a screencast in the same thread.

[arpitnath]

@hesreallyhim Rebranding to Claude Capsule Kit. Here's the screencast, will push the updates shortly.

full_final.mp4

[hesreallyhim]

@arpitnath

@hesreallyhim Rebranding to Claude Capsule Kit. Here's the screencast, will push the updates shortly.

hey is this ready to share or are you renaming it or no?

[arpitnath]

@hesreallyhim Yes, ready to share! Just released with rebrand complete (Super Claude Kit to Claude Capsule Kit).
https://github.com/arpitnath/super-claude-kit/releases/tag/v2.0.0
The repo URL stays the same (super-claude-kit) to preserve existing links.

[github-actions]

🤖 Validation Results

✅ All validation checks passed!

Your submission is ready for review by a maintainer.

Validated Data:

{
  "display_name": "Claude Capsule Kit",
  "category": "Tooling",
  "subcategory": "General",
  "primary_link": "https://github.com/arpitnath/super-claude-kit",
  "secondary_link": "",
  "author_name": "arpitnath",
  "author_link": "https://github.com/arpitnath",
  "license": "MIT",
  "description": "Adds persistent context memory and dependency analysis to Claude Code through hook-based automation. Includes session resume with 24-hour persistence, dependency intelligence tools (circular detection, impact analysis, dead code identification), and progressive file reading with semantic AST chunking. \nBuilt using SessionStart, PreToolUse, and PostToolUse hooks to track files, tasks, and discoveries across sessions.",
  "last_modified": "2025-12-21:18-37-22",
  "active": "TRUE",
  "last_checked": "2025-12-22:17-23-08"
}

---

_{This comment is automatically updated when you edit the issue.}