Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/lnardev/opencode-config-agent/llms.txt

Use this file to discover all available pages before exploring further.

Skills are modular rule sets — each one a SKILL.md file — that the OpenCode Config Agent detects and injects automatically into every sub-agent prompt based on what code is being touched and what action is being performed. Unlike static prompts you invoke manually, skills are context-driven: the orchestrator matches them against the current task and inserts their compact rules before any work begins. You never need to remember to “load” a skill — the registry takes care of that on your behalf.

​
How Skills Work

Each skill lives in a SKILL.md file and is indexed by the skill-registry. When the orchestrator starts a session or fires its first delegation, it reads the registry once, caches the compact rules in memory, and then injects matching rules into every sub-agent prompt as a ## Project Standards (auto-resolved) block. Sub-agents never read SKILL.md files directly — they receive pre-digested, actionable rules already embedded in their launch prompt. This build-once / read-cheaply design keeps token costs low (50–150 tokens per skill per delegation) while ensuring every sub-agent works to the same standards regardless of how many agents are running in parallel.

​
Skill Resolver Protocol

Before launching any sub-agent that reads, writes, or reviews code, the orchestrator runs a four-step injection process:
1

Obtain the Skill Registry

The orchestrator checks its in-session cache first. On a cache miss it queries engram (mem_search with query: "skill-registry") and falls back to reading .atl/skill-registry.md from the project root. If no registry exists, it warns the user and proceeds with generic review only.
2

Match Skills by Context

Skills are matched on two dimensions simultaneously:
  • Code context — file extensions and paths of the target (.go → go-testing; creating a PR → branch-pr)
  • Task context — the action being performed (“adversarial review” → judgment-day; “create new skill” → skill-creator)
The Trigger field in each skill’s frontmatter is the source of truth for matching.
3

Inject Compact Rules

The orchestrator copies the matching compact rules blocks from the registry’s Compact Rules section into the sub-agent prompt before any task-specific instructions, under the header ## Project Standards (auto-resolved).
4

Include Project Conventions

If the registry has a Project Conventions section and the sub-agent will work on project code, the convention file paths and their notes are appended to the prompt — so the sub-agent can read project-specific rules without extra discovery hops.

​
Auto-Load Triggers

AGENTS.md defines exactly two hard auto-load entries — the agent is instructed to load these immediately upon detecting the matching context, before writing any code:
ContextSkill
Go tests, Bubbletea TUI testinggo-testing
Creating new AI skillsskill-creator
The remaining skills are context-triggered by the agent as needed and are documented in the README. Multiple skills can match and inject simultaneously.
ContextSkill
SDD workflow stepssdd-*
Creating pull requestsbranch-pr
Creating GitHub issuesissue-creation
Adversarial code reviewjudgment-day

​
Available Skills

Branch & PR

Issue-first PR creation workflow with strict branch naming, conventional commits, and automated GitHub Actions checks.

Go Testing

Table-driven tests, Bubbletea TUI testing with teatest, golden file comparisons, and mock patterns for Go projects.

Judgment Day

Parallel adversarial review — two independent judge sub-agents review the same target simultaneously, then findings are synthesized and fixed.

Skill Creator

Guided workflow for creating new SKILL.md files with correct frontmatter, compact rules, and trigger conditions.

Skill Registry

Scans all skill directories and convention files, generates compact rules, and writes .atl/skill-registry.md for the orchestrator.

Issue Creation

GitHub issue workflow — enforces template usage, the status:needs-review → status:approved approval chain, and the issue-first PR requirement.

​
Skill Directory Structure

Global skills are installed once and available across all projects. Project-level skills override global ones when both define the same skill name. 
# Global (user-level) skills
~/.config/opencode/skills/{skill-name}/SKILL.md

# Project-level skills (override global)
{project}/.agent/skills/{skill-name}/SKILL.md
The skill-registry scans all known skill directories — including ~/.claude/skills/, ~/.gemini/skills/, ~/.cursor/skills/, and ~/.copilot/skills/ — so the same skill library works across Claude Code, Gemini CLI, Cursor, and Copilot.
SDD-specific skills (sdd-*) and the _shared directory are internal to the Spec-Driven Development workflow and are intentionally excluded from the skill registry scan. They power the /sdd-* commands but are not covered as standalone skills here.

Build docs developers (and LLMs) love

Get started for freeTalk to us