Installation

bunx omoslim@latest install

Non-interactive (copy-paste ready):

bunx omoslim@latest install --no-tui --kimi=yes --openai=yes --antigravity=yes --chutes=yes --opencode-free=yes --opencode-free-model=auto --tmux=no --skills=yes

Then authenticate:

opencode auth login

Run ping all agents to verify everything works.

Models are fully customizable. Edit ~/.config/opencode/omoslim.json (or .jsonc) to assign any model to any agent.

For LLM agents — paste into Claude Code, Cursor, or any coding agent:

Install and configure by following the instructions here:
https://raw.githubusercontent.com/alvinunreal/omoslim/refs/heads/master/README.md

Detailed guides: Installation · Antigravity · Tmux

How It Works

The plugin wires a team of specialist agents onto your OpenCode session. The Orchestrator is the only agent you talk to directly — it breaks your request into tasks and delegates each one to the right specialist.

Delegation flow

You → Orchestrator → delegate_task(agent="librarian") → packet returned inline
                   → delegate_task(agent="explorer", wait=false) ─┐
                   → delegate_task(agent="fixer",    wait=false) ─┤
                                                                   └─ packet_context(id1,id2)

Each specialist runs in its own session with only the tools it needs. When it finishes, its output is compressed into a packet — a structured ≤2,500-char summary — before being returned to the Orchestrator. Raw outputs never enter the Orchestrator's context.

Packets

Every delegate returns a YAML packet:

tldr:
  - Key insight 1
evidence:
  - file:src/foo.ts:42-80
  - thread:abc123#context
  - https://docs.example.com/api
recommendation: Single clear action
next_actions:
  - Step 1
  - Step 2

• Evidence accepts file:, thread:, cmd:, or URL pointers — never raw content.
• tldr / recommendation must be original synthesis — no URLs, no code blocks.
• Size-enforced: packets exceeding 2,500 chars are caught and replaced with a fallback that points to the archived thread.

Fallback packets

When a delegate's output can't be packaged (too large, invalid schema, extraction error), the Orchestrator receives a degraded packet with [fallback: reason] in its Options section:

## Options
[fallback: size limit exceeded]

The Orchestrator can then:

• Quick peek (500 chars): resolve_pointer(pointer="thread:abc123#context")
• Full compression: delegate_task(agent="summarizer", prompt="Summarize thread:abc123")

resolve_pointer

A surgical inspection tool for the Orchestrator — dereferences thread:, file:, or cmd: pointers from packets. Quota of 3 per user request, reset automatically at the start of each new task.

Parallel execution

For independent tasks, launch all delegates with wait=false and retrieve results in one shot:

delegate_task(..., agent="explorer",  wait=false)  → pkt_aaa
delegate_task(..., agent="librarian", wait=false)  → pkt_bbb
packet_context(task_id="pkt_aaa,pkt_bbb", timeout=120000)
→ merged packet with deduplication + conflict detection

The Agents

| Agent | Role | Delegates To | MCPs | |-------|------|--------------|------| | Orchestrator | Coordinator — the only agent you talk to | All | websearch | | Explorer | Codebase search (glob, grep, AST) | — | — | | Librarian | Live docs, API references | — | websearch, context7, grep_app | | Oracle | High-stakes architectural decisions | — | — | | Designer | UI/UX implementation | explorer | — | | Fixer | Fast parallel implementation | explorer | — | | Summarizer | Compresses oversized delegate output | — | — |

Auto-routing: omit agent in delegate_task and the task-router classifies the prompt and picks the right specialist.

Orchestrator

Master delegator. Routes tasks, merges results, integrates packet output into your session. Talks to all specialists; none of them talk back to it directly — only packets do.

Explorer

Parallel search specialist. Glob, grep, AST queries across the codebase. Fastest when you need to discover what exists before planning.

Librarian

Authoritative source for current library docs. Fetches official docs and examples via Context7 and grep.app. Use for APIs that change frequently — React, Next.js, AI SDKs, ORMs.

Oracle

Strategic advisor for high-stakes decisions and persistent problems. Slow, expensive, high-quality. Reserve for major architectural choices and bugs that survived 2+ fix attempts.

Designer

UI/UX specialist. Visual direction, responsive layouts, design systems. Delegates discovery to Explorer when needed.

Fixer

Fast parallel execution for well-defined tasks. Spawning multiple Fixers simultaneously is the standard pattern for 3+ independent changes.

Summarizer

Compresses oversized delegate output. Called by the Orchestrator when a packet fallback contains [fallback: size limit exceeded] and a full re-summary is needed.

Documentation

• Quick Reference — Presets, Skills, MCPs, Tools, Configuration
• Installation Guide — Detailed installation and troubleshooting
• Cartography Skill — Repo mapping + codemap generation
• Antigravity Setup — Antigravity (Google) provider configuration
• Tmux Integration — Real-time agent monitoring

License

MIT