oafish

Output compression for AI coding agents. Same answer, fewer tokens.

Before

"Sure! I'd be happy to help you with that. The issue you're experiencing is likely caused by a missing null check in the authentication middleware. What's happening is that when the token is undefined, the code attempts to call .split() on it, which throws a TypeError. You'll want to add a guard before that line."

After (/oafish)

"Auth middleware: token undefined → .split() throws. Add null guard before line 42."

~75% fewer output tokens. Full accuracy.

Install

curl -fsSL https://oafish.bitclaw.com/install | bash

Detects which agents are installed and wires up each one automatically.

| Agent | Install method | Auto-activates | |---|---|---| | Claude Code | Plugin (hooks + skill) | Yes — every session | | Cursor | .cursor/rules/oafish.md | Yes | | Windsurf | .windsurf/rules/oafish.md | Yes | | Cline | .clinerules/oafish.md | Yes | | GitHub Copilot | .github/copilot-instructions.md | Yes | | opencode | ~/.config/opencode/AGENTS.md | Yes | | Gemini CLI | GEMINI.md | Yes | | Codex / Roo / Amp / Goose | npx skills add bitclaw/oafish | On invoke |

Or copy rules/oafish.md manually into any agent's rules directory.

Usage

/oafish           activate (full mode)
/oafish lite      professional tight — keeps full sentences
/oafish ultra     maximum compression — arrows, abbreviations
stop oafish       deactivate
normal mode      deactivate

Mode persists across turns. No drift. Deactivates only on explicit command.

Intensity levels

| Level | What changes | Example | |---|---|---| | lite | No filler/hedging. Full sentences. | "Your component re-renders because you create a new object reference each render." | | full | Drop articles, fragments OK. | "New object ref each render. Inline prop = re-render. Wrap in useMemo." | | ultra | Abbreviations, arrows for causality. | "Inline prop → new ref → re-render. useMemo." |

Auto-clarity: oafish suspends for security warnings, destructive operations, and anywhere compression risks misread. Resumes after.

How it works

For Claude Code, oafish installs three hooks:

• SessionStart (activate) — injects compression rules at session open, writes active mode to ~/.config/oafish/.active
• UserPromptSubmit (tracker) — handles /oafish commands, reinforces mode each turn
• PostToolUse (compress) — digests verbose Read/Bash/MCP outputs before they fill context

The compress hook is the high-leverage part: large file reads and bash output can consume thousands of context tokens per call. oafish intercepts them and injects a terse digest instead:

[oafish] Read: auth.ts — 312L | ts | fns: login, logout, validateToken, refreshSession
[oafish] Bash: npm test — 47 passing, 2 failing: auth.test.ts:88, user.test.ts:112

For file-based agents (Cursor, Windsurf, etc.), the rules file is written directly — no hooks, same compression behavior via prompt rules.

Benchmarks

Benchmarks measure token reduction against a live model.

export ANTHROPIC_API_KEY=sk-...   # or add to .env.local
uv run benchmarks/run.py

See CONTRIBUTING.md for prerequisites.

Build from source

Requires Bun.

bun install
bun run build       # compiles hooks/src/*.ts → hooks/dist/*.js

Hook source is TypeScript (hooks/src/). Compiled JS is committed to hooks/dist/ so installs require no build step on the user's machine.

License

MIT