oh-my-opencode-lite

Seven specialized agents, one orchestrator — delegate any task to the right specialist and ship faster.

oh-my-opencode-lite · Delegate-first orchestration · Thoth-mem persistence · Bundled SDD pipeline

Delegate-first OpenCode plugin with a seven-agent roster, root-session thoth_mem persistence, disk-backed delegation records, bundled requirements-interview, and a full SDD workflow.

oh-my-opencode-lite keeps the orchestrator lean, pushes discovery into specialists, persists important context, and ships the planning skills needed to move from ambiguous requests to verified implementation.

📦 Installation

Quick start

bunx oh-my-opencode-lite@latest install
opencode auth login
opencode

Then ask OpenCode to verify the roster:

ping all agents

Non-interactive install

bunx oh-my-opencode-lite@latest install --no-tui --tmux=no --skills=yes

Reset an existing generated config

bunx oh-my-opencode-lite@latest install --reset

When skills are enabled, the installer adds the recommended external skills and copies the bundled requirements-interview, cartography, plan-reviewer, executing-plans, and SDD skills into your OpenCode skills directory.

For LLM agents

Use the installer directly:

bunx oh-my-opencode-lite@latest install --no-tui --tmux=no --skills=yes

Or hand another coding agent this README:

Install and configure oh-my-opencode-lite by following:
https://raw.githubusercontent.com/EremesNG/oh-my-opencode-lite/refs/heads/master/README.md

JSON Schema

The package ships oh-my-opencode-lite.schema.json for editor autocomplete and validation:

{
  "$schema": "https://unpkg.com/oh-my-opencode-lite@latest/oh-my-opencode-lite.schema.json"
}

See docs/installation.md and docs/provider-configurations.md for the full setup flow.

🏛️ Seven-Agent Roster

The delegate-first philosophy is simple: the orchestrator coordinates while specialists execute. Read-only discovery work goes to async specialists for context isolation. Advisory and write-capable work stays sync so review, undo safety, and verification remain straightforward.

For blocking user decisions, the orchestrator uses a question tool — agents do not ask those questions in plain text. Independent work can be launched together when parallel dispatch is safe, and failed delegations are retried once before being reported back to the user.

🔑 Primary Agent

Orchestrator Root coordinator and sole primary agent. Role: The root coordinator. Handles delegation, sequencing, memory ownership, and SDD progress tracking. Does not read or modify source files directly. Mode: primary, non-mutating Dispatch: sync coordinator Recommended: anthropic/claude-opus-4-6 · openai/gpt-5.4 · kimi-for-coding/k2p5 Personality: Autonomous deep coordinator — multi-agent reasoning, works through delegation

🛠️ Specialist Subagents

Explorer Speed runner — fast parallel grep, codebase search. Role: Local codebase discovery and navigation. Fast parallel search, file reading, symbol lookup. Mode: read-only Dispatch: async via background_task Recommended: Grok Code Fast · openai/gpt-5.4-nano · anthropic/claude-haiku-4-5	Librarian All-rounder — large context + decent speed for research. Role: External docs and API research. Fetches documentation, finds public examples, validates version-specific behavior. Mode: read-only Dispatch: async via background_task Recommended: openai/gpt-5.4 · anthropic/claude-sonnet-4-6 · google/gemini-3.1-pro-preview	Oracle Deep reasoner — maximum strategic thinking capability. Role: Strategic advisor for debugging, architecture review, code review, and SDD plan review. Mode: read-only Dispatch: sync via task Recommended: openai/gpt-5.4 · anthropic/claude-opus-4-6 · opencode-go/glm-5
Designer Visual/multimodal — UI/UX reasoning, frontend engineering. Role: UI/UX implementation with visual verification. Owns approach, execution, and browser-based quality checks. Mode: write-capable Dispatch: sync via task Recommended: google/gemini-3.1-pro-preview · opencode-go/glm-5 · kimi-for-coding/k2p5	Quick Speed runner — well-defined tasks, fast turnaround. Role: Fast implementation for well-defined, bounded tasks. Optimized for speed over thoroughness. Mode: write-capable Dispatch: sync via task Recommended: openai/gpt-5.4-mini · anthropic/claude-haiku-4-5 · google/gemini-3-flash-preview	Deep Deep specialist — maximum coding capability for complex tasks. Role: Thorough implementation and verification. Handles correctness-critical, multi-file, edge-case-heavy work. Mode: write-capable Dispatch: sync via task Recommended: openai/gpt-5.4 · anthropic/claude-opus-4-6 · google/gemini-3.1-pro-preview

🧩 What oh-my-opencode-lite Adds

• Delegate-first orchestration with context isolation across specialists
• thoth_mem persistence for root-session memory workflows
• Disk-persisted delegation results that survive compaction and in-memory loss
• Bundled SDD pipeline: sdd-propose, sdd-spec, sdd-design, sdd-tasks, sdd-apply, sdd-verify, sdd-archive
• Requirements-interview skill for clarifying ambiguous work before implementation
• plan-reviewer for oracle review loops on task plans
• executing-plans for task-state ownership and progress tracking
• cartography for repository mapping and codemap generation
• Tmux integration for real-time agent monitoring
• Configurable presets, fallback chains, prompt overriding, and artifact-store modes

Delegation records persist under:

~/.local/share/opencode/delegations/<project-id>/<root-session-id>/<task-id>.md

SDD Pipeline

The bundled SDD workflow follows this path:

propose -> [spec || design] -> tasks -> apply -> verify -> archive

For medium work, the requirements interview can route into an accelerated path that starts at propose -> tasks. For complex work, the full path is used.

Artifacts can be persisted in four modes:

Mode	Writes to	Cost	Use when
thoth-mem	Memory only	Low	Fast iteration without repo planning files
openspec	openspec/ files only	Medium	Reviewable planning artifacts in the repo
hybrid	Both	High	Maximum durability; default
none	Neither	Lowest	Ephemeral iterations, no persistence

After sdd-tasks, the orchestrator can run an oracle review loop with plan-reviewer:

• Generate tasks.md
• Dispatch oracle with plan-reviewer
• If result is [REJECT], fix only the blocking issues
• Repeat until [OKAY]
• Continue into execution

During execution, executing-plans owns task-state tracking. Progress has two mandatory layers: todowrite for the visual task list, plus a persistent SDD artifact via tasks.md checkboxes and/or thoth_mem.

• - [ ] pending
• - [~] in progress
• - [x] completed
• - [-] skipped with reason

Requirements Interview

The bundled requirements-interview skill is the front door for ambiguous or substantial work. It is step-0 in the orchestrator prompt and runs before implementation when the request is open-ended, spans multiple parts of the system, or needs scope calibration.

Its workflow is six phases:

• Context gathering
• Interview
• Scope assessment
• Approach proposal
• User approval
• Handoff

After clarification, the skill routes the work into direct implementation, accelerated SDD, or full SDD based on complexity assessment.

🧩 Skills & MCP Servers

Bundled skills

Skill	Category	Purpose
requirements-interview	Clarification	Clarify intent, assess scope, and choose direct work vs accelerated or full SDD
cartography	Discovery	Generate and update hierarchical repository codemaps
plan-reviewer	Review	Validate tasks.md for real execution blockers and return [OKAY] or [REJECT]
sdd-init	SDD	Bootstrap OpenSpec structure and SDD context for a project
sdd-propose	SDD	Create or update proposal.md
sdd-spec	SDD	Write OpenSpec delta specs with RFC 2119 requirements and scenarios
sdd-design	SDD	Produce design.md with technical decisions and file changes
sdd-tasks	SDD	Generate phased tasks.md checklists
sdd-apply	SDD	Execute assigned SDD tasks and report structured results
executing-plans	Execution	Run task lists with durable progress tracking and verification checkpoints
sdd-verify	Verification	Create compliance-oriented verification reports
sdd-archive	Archive	Merge verified deltas into main specs and archive the change

Recommended external skills

Skill	Status	Typical use
simplify	Installed by --skills=yes	Keep solutions lean and reduce unnecessary complexity
agent-browser	Installed by --skills=yes	Browser automation for designer visual checks
test-driven-development	Optional companion	Useful before implementing bug fixes or features with deep
systematic-debugging	Optional companion	Useful for oracle and deep when failures need disciplined diagnosis

Built-in MCP servers

MCP	Purpose	Auth / runtime
websearch	Exa-backed web search	Optional EXA_API_KEY via env
context7	Official library and framework docs	Optional CONTEXT7_API_KEY via env
grep_app	Public GitHub code search	No auth required
thoth_mem	Local persistent memory and artifact storage	Local command, default npx -y thoth-mem@latest

🧠 Thoth-Mem is a persistent memory MCP server purpose-built for cross-session context. The orchestrator uses it to save architectural decisions, bug-fix learnings, SDD artifacts, and session summaries so the next session picks up where the last one left off — even after context-window compaction. It is included by default and runs locally via npx.

For targeted retrieval, Thoth-Mem uses a 3-layer recall protocol:

• mem_search — scan the compact index of IDs and titles
• mem_timeline — inspect chronological context around candidates
• mem_get_observation — read full content for selected records

After task completion, an automatic save nudge reminds the orchestrator to persist important observations. Session start also degrades gracefully if thoth_mem is unavailable, so memory errors do not block the main plugin flow.

Skill and MCP access in this project is prompt-driven. The generated plugin config focuses on model presets and runtime options rather than per-agent permission matrices.

📚 Documentation

• docs/installation.md
• docs/provider-configurations.md
• docs/tmux-integration.md
• docs/quick-reference.md
• docs/cartography.md
• docs/sdd-pipeline.md
• docs/skills-and-mcps.md
• AGENTS.md

Development

The project targets @opencode-ai/plugin and @opencode-ai/sdk v1.3.3.

Command	Purpose
bun run build	Build TypeScript into dist/
bun run typecheck	Run TypeScript type checking without emit
bun test	Run the Bun test suite
bun run lint	Run Biome linter
bun run format	Run Biome formatter
bun run check	Run Biome check with auto-fix
bun run check:ci	Run Biome check without writes
bun run dev	Build and launch the plugin in local dev mode

📄 License

MIT