DEV Community

gracefullight
gracefullight

Posted on

oh-my-agent: cross-vendor refactor guards and isolated eval

#ai #agents #programming #productivity

oh-my-agent 12.4.0 just shipped with a cross-vendor refactor guard and isolated harness evaluation. When an agent edits too many lines in a single file without refactoring, the runtime now blocks the stop event and dispatches a dedicated refactor engineer to clean up the mess before returning control.

What's new

  • Cross-vendor refactor guard: A new refactor_guard hook tracks line budgets across edit tools. If an agent exceeds 500 lines in a single file, the hook blocks termination and dispatches refactor-engineer. This works across all 9 supported vendors.
  • Isolated harness evaluation: The oma harness command now compares scoped .agents overlays in paired temporary workspaces, using deterministic artifact checks and hash-bound record and replay.
  • Agent Plugins support: A standard git clone of the repository now acts as an installable Agent Plugins 1.0.0 package, exposing plugin.json, mcp.json, and the skills tree directly at the repository root.
  • Translation profiles: oma-translator now uses per-language profiles (en, ko, ja, zh) to isolate register systems and typography rules from the shared pattern taxonomy.

Controlling an agent's context requires precision. The new refactor guard ensures that large changes are forced into a refactoring phase before the session ends, keeping files maintainable.

What's fixed

  • SCM decisions like commit splits and merges are no longer recorded as durable facts in the agent memory store. They are already in git history, and skipping them prevents repository metadata from crowding out real cross-session decisions.
  • oma link now renders the union of vendor hooks and subagents for shared files like AGENTS.md, preventing the file from flip-flopping when different vendors trigger the merge.
  • User-added MCP servers in .cursor/mcp.json are preserved during updates. The CLI now merges its managed servers rather than overwriting the entire file.
  • OpenCode authentication checks now correctly derive the provider from the row's model specification, resolving false offline reports when users authenticated with alternative providers.

What's better

  • Replaced unreachable context budget claims in the documentation with measured, reproducible token costs. The two-layer skill design avoids roughly 75 percent of skill-loading overhead on complex tasks.
  • Scoped Single Source of Truth (SSOT) protection strictly to definition files. Agents are now explicitly instructed that run outputs in .agents/results/ are expected, preventing them from deleting session handoff state to restore protection.
  • Korean translations across all 22 pages are synced to parity with the English source, normalizing instruction registers and fixing legacy structural drift.

Installation 

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/first-fluke/oh-my-agent/main/cli/install.sh | bash
Enter fullscreen mode Exit fullscreen mode 
# Windows (PowerShell)
irm https://raw.githubusercontent.com/first-fluke/oh-my-agent/main/cli/install.ps1 | iex
Enter fullscreen mode Exit fullscreen mode

oh-my-agent is built for teams who orchestrate more than they prompt. Next up: multi-turn rollback primitives for stalled reasoning loops.

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

Top comments (0)