When you tell an agent to scaffold a Next.js app, it picks the wrong version, ignores your lint config, and ships a save button with no storage backing it. The last four weeks of oh-my-agent were spent fixing exactly that, plus shipping a benchmark that actually measures it.

What's new

• Nine new skills: oma-deepsec (Vercel deepsec driver), oma-docs (doc reference drift detection), oma-observability (33 files routing MELT+P signals across L3/L4/mesh/L7), oma-academic-writer, oma-hwp (HWP/HWPX to Markdown via kordoc), oma-image (multi-vendor generation across codex, pollinations, gemini), oma-scholar (Knows sidecar paper records), oma-search (intent-based search with trust scoring), oma-skill-creator
• cursor promoted to a first-class vendor with cursor-only preset, composer-2 routing, and --yolo auto-approve
• New /docs and /deepsec workflows, both detected via keyword triggers in 11 languages
• oma model:check, model:probe, model:propose diff the local registry against OpenRouter and cursor agent --list-models, then scaffold a models.yaml patch you can paste in
• Auto-update CLI on outdated install, gated by auto_update_cli in oma-config.yaml
• Windows support via install.ps1, with junction and hardlink fallbacks when symlinks raise EPERM

What's fixed

• Hook keyword detector: NFKC normalization closes fullwidth IME bypasses (ｐａｒａｌｌｅｌ no longer slips past), hyphen-rejecting word boundary, two-tier CLI invocation guard so claude review this code still routes to workflows but claude exec --foo does not
• codex exec -i is variadic in clap and was swallowing the prompt as a second reference image. Argv now terminates with -- before the instruction
• Path traversal closed on --out and --reference paths; magic-byte MIME validation on reference images
• Windows path separators normalized across oma doctor, oma link, and the gitignore IO layer

What's better

• New 5-axis benchmark in benchmarks/: Functional (35), Spec (15), Visual (20), Engineering (20), Efficiency (10). oh-my-agent scored 80.6 first place, ahead of omc 74.1, superpowers 72.9, vanilla 70.7, ecc 70.2. Multi-judge averaging across 3 rounds, not single-shot
• Config consolidated to one oma-config.yaml with model_preset and inline agents map. Five built-in presets ship in the CLI; legacy agent_cli_mapping auto-migrates on install via migration 008
• 1492 tests passing, CI matrix now covers ubuntu/macos/windows with pre-push and pre-commit hooks
• oma docs i18n detects translation drift across 11 locales; oma docs lint flags em-dashes and wrong-language placeholders. Used internally to restructure ~1900 em-dashes across 173 files per CJK typography conventions
• Skills auto-register from .agents/skills/ frontmatter at build time, so adding a skill no longer requires touching a hardcoded SKILLS map

Installation

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/first-fluke/oh-my-agent/main/cli/install.sh | bash

# Windows (PowerShell)
irm https://raw.githubusercontent.com/first-fluke/oh-my-agent/main/cli/install.ps1 | iex

oh-my-agent is built for teams running parallel agents across multiple vendor CLIs and getting tired of each one drifting. Next up: deeper PR-gate integration for /deepsec and a reciprocal-skill pass on oma-observability.

https://github.com/first-fluke/oh-my-agent