crabshell

Crabshell

Claude Code plugin that makes Claude remember, verify, and self-correct.

Three pillars:

1. Session memory — Auto-saves context across sessions. Delta extraction, Haiku summarization, token-based rotation. No manual setup.
2. Behavioral correction — Injects verification-first rules and interference pattern detection every prompt. Twelve guard hooks block sycophancy, scope reduction, overcorrection, and shortcuts at runtime. Pressure System: three pressure counters (feedbackPressure.level, feedbackPressure.oscillationCount, tooGoodSkepticism.retryCount) with user-prompt-driven level (W021: profanity-only NEG patterns) plus assistant-side oscillation and too-good skepticism counters. UNLEASH keyword (renamed from BAILOUT in v21.79.0 / W021) resets all three.
3. Structured workflows — D/P/T/I/H/W document system with 21 skills for planning, investigating, iterative improvement (regressing), hotfix recording, and light-workflow tracing.

All plugin output lives under .crabshell/ — gitignored, clean project root.

Installation

/plugin marketplace add ZipperBagCoffee/crabshell
/plugin install crabshell

After installation, you don't need to do anything. It works automatically.

Codex Compatibility

This repository is a dual-runtime plugin repo:

• Claude Code uses .claude-plugin/plugin.json, hooks/hooks.json, commands/, and skills/.
• Codex uses .codex-plugin/plugin.json, codex-skills/, and the scripts/codex-*.js wrappers.

Installing the plugin in one runtime does not automatically activate the other runtime. The files can ship in the same GitHub repository, but each product only reads its own manifest and entrypoints.

After installing through Claude, run the manual bridge command when you want the same checkout available in Codex:

/crabshell:install-codex

The shared state is .crabshell/. Claude and Codex can both read and write the same memory/document store when they are used in the same project:

node scripts/codex-memory.js load
node scripts/codex-memory.js save --title="Codex session note" --message="..."
node scripts/codex-memory.js search "query"
node scripts/claude-to-agents.js
node scripts/codex-docs.js investigation "research topic"
node scripts/codex-docs.js worklog "task title"

Codex compatibility is explicit skill/script based. Claude-style automatic hooks such as SessionStart, PostToolUse, PreToolUse, and Stop are not activated by Codex.

How It Works

1. Session start - Loads saved content from previous sessions into Claude's context
2. During work - Auto-save triggers every 15 tool uses (configurable), Claude records decisions/patterns/issues directly
3. Session end - Full conversation backup + final save

What Gets Saved

Automatic (No action needed)

• logbook.md - Session summaries accumulate here (auto-rotates at 23,750 tokens)
• logbook_*.md - Rotated archives (L2)
• *.summary.json - L3 summaries (Haiku-generated)
• sessions/*.l1.jsonl - Detailed session transcripts (L1)

Manual Setup (Optional)

If there's information you want Claude to know every session, directly edit the files:

# Create/edit files in your project's .crabshell/memory/ folder
echo "React + TypeScript web app." > .crabshell/project.md

Or just ask Claude: "Save the project info to project.md"

With this setup, Claude starts every new session knowing this information.

Slash Commands

Works in any project where the plugin is installed: