@qlucent/fishi-governance

Governance plugin for Claude Code — blocks destructive agent operations and logs an audit trail.

Works alongside any framework: Ruflo, BMAD, Oh-My-ClaudeCode, or raw Claude Code.

Why?

AI agents are deleting production files, force-pushing to main, and overwriting .env files — right now, today. This plugin stops that.

Before fishi-governance:

Agent: "Let me clean up the project..."
> rm -rf migrations/    ← your database migrations, gone
> git push --force origin main    ← your team's work, overwritten

After fishi-governance:

Agent: "Let me clean up the project..."
> rm -rf migrations/
[fishi-governance] BLOCKED: Recursive delete of all files in cwd
> git push --force origin main
[fishi-governance] BLOCKED: Force push to main/master

Quick Start (2 minutes)

# 1. Install the plugin
npm install -g @qlucent/fishi-governance

# 2. That's it. Start a Claude Code session.
# Destructive operations are now blocked. Audit trail is logging.

Zero configuration required. Works immediately.

What It Blocks

| Category | Examples | |----------|---------| | Filesystem destruction | rm -rf /, rm -rf *, mkfs, dd if= | | Dangerous git operations | git push --force main, git reset --hard, git clean -fd | | Database destruction | DROP TABLE, TRUNCATE TABLE, DELETE FROM (no WHERE) | | Privilege escalation | sudo, su root, chmod 777 | | Remote code execution | curl ... \| bash, wget ... \| sh | | Environment exposure | .env file overwrites, env > dumps | | System commands | shutdown, reboot, kill -9 |

Audit Trail

Every agent action is logged to .fishi/audit-log.jsonl:

{"ts":"2026-04-02T12:00:00Z","event":"blocked","tool":"Bash","command":"rm -rf migrations/","reason":"Recursive delete of all files in cwd","override":false}
{"ts":"2026-04-02T12:00:01Z","event":"allowed","tool":"Write","command":"src/index.ts","reason":"","override":false}

• Append-only — existing entries can never be modified
• Persists across sessions
• Structured JSONL — easy to parse and review

Works With

• Ruflo — governance hooks layer on top, no conflicts
• BMAD Method — complementary: BMAD plans, FISHI governs
• Oh-My-ClaudeCode — additive hooks, independent operation
• Raw Claude Code — standalone protection for any project
• Full FISHI framework — coexists with v0.20.6+

How It Works

Three Claude Code hooks, zero dependencies:

1. SessionStart — confirms governance is active, creates .fishi/ directory
2. PreToolUse (Bash) — checks commands against 30+ destructive patterns, blocks matches
3. PostToolUse (Bash|Write|Edit) — logs every action to audit trail

All processing is local. Zero network calls. Zero telemetry. Zero tracking.

License

MIT