When you tell an agent to handle a complex backend migration, hitting a hard cutoff 20 steps in ruins the workflow. We bumped the ceiling and smoothed out the integration edges.
What's new
- Bluesky syndication: Release announcements now fan out to Bluesky with rich text tags and external link cards.
-
Claude auto-trust: Project linking now directly mutates your
~/.claude.jsonto pre-accept workspace trust, skipping the interactive dialog entirely. -
Serena auto-install: The Serena MCP binary now self-installs during project setup and is actively probed by the
doctorcommand.
What's fixed
- Premature cutoffs: We raised the max agent turns from 20 up to 40 for backend, frontend, and mobile implementation agents to prevent mid-task failures.
-
Monorepo state leaks: Running commands from inside a sub-package no longer leaves stray
.agents/directories behind. State is correctly anchored to the project root. -
Fragile JSON parsing: Agent publishing steps now use
jsonrepairwith a 3x retry loop, preventing trailing commas or unclosed braces from aborting a release. - OpenCode dispatch: Subagents now properly inherit the user's default model instead of a hardcoded slug, and routing correctly falls back to native task tools.
What's better
-
Smarter MCP resolution: Serena now uses
--project-from-cwdto gracefully handle execution from nested subdirectories. -
Installation UX: The install flow detects native OpenCode tasks and gracefully defaults to a
mixedpreset when a strict single-vendor lock is unnecessary. - Context isolation: The Antigravity MCP transport now correctly stamps its own context instead of bleeding the Claude Code defaults.
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 who orchestrate more than they prompt. Next up: expanding our deepsec vulnerability scanner capabilities.
Top comments (1)
The 20-to-40 turn bump is one of those unglamorous changes that probably removes a lot of real workflow pain, especially for backend migrations where the last 20% is where cleanup and verification live. I also like the monorepo root anchoring fix for stray
.agents/directories and thejsonrepairretry loop around publishing; both are small signs the tool is being exercised in messy repos, not just demos. The one boundary I'd treat very carefully is auto-mutating~/.claude.jsonfor workspace trust. That kind of convenience is valuable, but it needs to be obvious, auditable, and easy to undo because trust state becomes part of the.