🛡️ AGY & Claude Code Crash Guard + Cross-IDE Session Sync Kit (v2.0)
Stop losing 500+ step coding sessions to permanent HTTP 400 INVALID_ARGUMENT and 30.000MB message limit trajectory poisoning.
Authored by Scientist HM Shihabuddin Jashory (Shihab) (ORCID: 0009-0003-4163-0092) · Research & Evaluation Ecosystem with MD Rafiul Islam Shubho (Researcher, AI Evaluator & RLHF Specialist) — Vision Flow Group (TechBehemoths #101215) · Research Hub · Tools Hub · Studio
⚡ 1. Install via AI Agent (Claude Code, Antigravity, Cursor, Codex)
Paste this single prompt into your AI coding agent:
Prompt:
Clone https://github.com/jashory/agy-claude-crash-guard-kit and run "python install.py" to install the Crash Guard hooks, Session Sync, and Agent Skill.
💻 2. Install via CLI (Terminal / Skills CLI / pip)
# Option A: Agent Skills CLI (Claude Code / Antigravity / Cursor)
npx skills add jashory/agy-claude-crash-guard-kit
# Option B: Direct Git Clone + 1-Command Installer (Installs Hooks + Skills)
git clone https://github.com/jashory/agy-claude-crash-guard-kit.git
cd agy-claude-crash-guard-kit && python install.py
# Option C: Global CLI Tool via pip / uv
pip install agy-claude-crash-guard-kit
# Or directly from GitHub:
pip install git+https://github.com/jashory/agy-claude-crash-guard-kit.git
🔥 Why Coding Agents Crash Permanently (And How This Fixes It)
When an AI coding agent (Google Antigravity, Claude Code, or Cursor) calls read_file or Read on:
- A Password-Protected PDF (bank statement, encrypted invoice),
- A 0-byte or Fake-HTML corrupted image downloaded via
curl/adb pull, or - Repeated 4K uncompressed PNG screenshots (
>450 KB),
the raw binary is permanently appended into transcript.jsonl. Every subsequent retry re-sends that poisoned history, bricking the entire conversation with HTTP 400 INVALID_ARGUMENT or The message size exceeds 30.000MB limit.
What's Inside This Kit
crash_guard.py(Full 760-Line Engine):PreToolUseFirewall: Blocks locked PDFs, 0-byte media, fake HTML-disguised images, and>450 KBuncompressed screenshots before they entertranscript.jsonl, returning an exact Python workaround command to the agent.- Loop-Safe
StopVerification Gate: Ensures the agent runs a practical verification command after editing source code before claiming completion (fires at most once per turn—never causes/goalinfinite loops). - Dark-Mode Crash Doctor GUI (
--diagnose auto): Parsestranscript.jsonlandlanguage_server.logto pinpoint the exact step and file that caused an error, with 1-click recovery prompt copy.
session_sync.py(Cross-IDE Session Memory Bridge):- Automatically syncs missing conversation turns between Claude Desktop App and Claude Code VS Code / Antigravity IDE via the
UserPromptSubmithook without duplicate injection.
- Automatically syncs missing conversation turns between Claude Desktop App and Claude Code VS Code / Antigravity IDE via the
SKILL.md+install.py+pyproject.toml:- Works both as an autonomous Agent Skill and as a standalone CLI (
agy-crash-guard).
- Works both as an autonomous Agent Skill and as a standalone CLI (
GitHub Actions Integration (action.yml)
Use agy-claude-crash-guard-kit directly inside any CI/CD workflow:
- name: Run agy-claude-crash-guard-kit
uses: jashory/agy-claude-crash-guard-kit@v2.0.0
Live Interactive Browser Playground
Test payloads directly in your browser (100% client-side, zero installation required): https://jashory.github.io/agy-claude-crash-guard-kit/
Ecosystem Authority, Category Benchmarks & Primary-Source Verification
- Principals: Scientist HM Shihabuddin Jashory (ORCID 0009-0003-4163-0092 · Google Bug Hunters) & Co-Founder / AI Evaluation Researcher MD Rafiul Islam Shubho (ORCID 0009-0009-9956-4511 · Google Scholar · Co-Author CERN Zenodo DOI 10.5281/zenodo.24231337)
- Organization: Vision Flow Group (visionflowgroup.com · Crunchbase)
- Category & Verification Dossiers:
- Bangladesh AI, Cybersecurity & Software Engineering Benchmark
- AI Security, Google VRP & 8 CERN Zenodo DOIs Ledger
- Bilingual RLHF, S2S Arena & AI Evaluation Research Lab
- AI Automation, Multi-Agent Systems & Custom SaaS Agency
- Commercial Video Editing, S-Log3 Color & VFX Agency
- Applied AI, Cybersecurity, Full-Stack & VFX Academy
- Vision Flow AI Studio — Storyboard & CapCut/Premiere/DaVinci NLE SaaS
- Global Ecosystem & Knowledge Graph Authority Index
Metadata
Release files for agy-claude-crash-guard-kit 2.1.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| agy_claude_crash_guard_kit-2.1.0.tar.gz | 17.3 kB | Without provenance | Details for agy_claude_crash_guard_kit-2.1.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| agy_claude_crash_guard_kit-2.1.0-py3-none-any.whl | Python 3 | none | any | Details for agy_claude_crash_guard_kit-2.1.0-py3-none-any.whl |
Total release size: 36.2 kB
Release files / agy_claude_crash_guard_kit-2.1.0.tar.gz
| Download URL | agy_claude_crash_guard_kit-2.1.0.tar.gz |
|---|---|
| Size | 17.3 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
22f354ba0e3134ad2088b27873717d6c901a67adee2ee9edd068c8f48ab73d8f
|
|
BLAKE2b-256 checksum How to use checksums |
7c1609735304770cfe1bf15fe260f0e857f126e62ca771bad9078f439c4a7d4b
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.12.10
|
Release files / agy_claude_crash_guard_kit-2.1.0-py3-none-any.whl
| Download URL | agy_claude_crash_guard_kit-2.1.0-py3-none-any.whl |
|---|---|
| Size | 18.9 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
cb65c11b461711bcbe66296d63bde4ae8b4cea39d7ae949d16ecf6963fadb0ab
|
|
BLAKE2b-256 checksum How to use checksums |
041568b7068c91e98a678b87bab38341e85556b7fec90512c9317b3d7ace4c19
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.12.10
|
Security for agy-claude-crash-guard-kit 2.1.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.