Boffin adds guardrails to AI coding agents

Boffin is a new control layer for AI coding agents that routes file-specific architectural constraints to the agent before it edits code, then requires verification afterward. Available as the boffinit npm package, it is powered by ParselFire Core and supports Cursor, Claude Code, Codex, and OpenCode.

The project targets a familiar failure mode: a developer requests a 15-line fix, but the coding agent returns a 500-line renovation. Rather than supplying one repository-wide instruction block, Boffin selects the constraints relevant to the file under review.

How Boffin differs from AGENTS.md

Boffin says it is not another AGENTS.md file, prompt pack, linter, CI gate, or speed tool. Its focus is review safety: architectural guidance before the edit and proportional verification after it.

Recommended reading

Weak AI safety rules can raise product risks

Compared with a static rules file, Boffin:

• Routes constraints to the current edit instead of applying one block to the whole repository.
• Requires verification proportional to the change.
• Records evidence through case studies rather than making unsupported proof claims.
• Delivers signed, portable instruction packs powered by ParselFire Core.

The rules are readable, versioned Markdown files stored under packs/ and GPG-signed. Boffin selects the applicable rules at edit time; the project says nothing is hidden during installation, allowing users to inspect each pack before trusting it.

Case studies and supported hosts

The project’s public case studies cover guided refactors in real open-source code:

• DuckDB: +17 / -17 lines, with 2,104 assertions passing across 8 test files; distinct continuation and recovery paths were preserved.
• FastAPI: +16 / -33 lines, with 49 tests passing and no public API change.
• LangChain: The sync/async boundary was preserved, with 4 tests passing.

Boffin describes these as reproducible case studies, not a controlled A/B benchmark.

Installation requires Node.js 18 or newer. Cursor users can run npx boffinit cursor from a project. Claude Code uses its plugin marketplace commands, while Codex users run codex plugin marketplace add MicSm/boffin followed by codex plugin add boffin@boffin.

Codex does not automatically trust plugin hooks. Users must run /hooks once to review and trust them; until then, the plugin’s skills work but automatic per-session activation remains disabled.

For OpenCode, users run npx boffinit opencode and open the project in OpenCode. Always-on guidance is delivered through opencode.json and .boffin/AGENTS.md; /boffin, /boffin-review, and the corresponding skills provide on-demand access.

Boffin offers lite, full, and max profiles that adjust cleanup ambition rather than correctness. The safety floor remains unchanged across profiles, including trust-boundary validation, data-loss prevention, security, and accessibility requirements. The project explicitly says it does not sandbox commands or restrict filesystem and network access, and it does not replace repository tests or human code review.

Boffin is available under the MIT License. Uninstall commands for Cursor and OpenCode remove only each host’s managed files, leaving shared .boffin/packs and .boffin/VERSION in place if another host remains installed.