Targets Guide

DZ Harness Hub adapts canonical skills for several coding agents. Use dz parity for the capability status of your installed version; this guide explains the target layouts. Optional --enrich generates platform-specific metadata where supported.

Supported Targets

claude-code

Anthropic’s official CLI for Claude. The default target.

Property	Value
Platform	Claude Code (Anthropic)
Skills directory	`.claude/skills/`
Settings file	`.claude/settings.json`
MCP support	Yes (native)
SKILL.md	Native (no conversion)
Default	Yes

# Explicit target (same as default)
dz setup --target claude-code

codex

OpenAI’s Codex CLI agent.

Property	Value
Platform	Codex (OpenAI)
Skills directory	`.agents/skills/`
Settings file	`.codex/config.json`
MCP support	Via adapter
SKILL.md	Native (no conversion)

dz setup --target codex

opencode

Open-source AI coding agent.

Property	Value
Platform	OpenCode
Skills directory	`.opencode/skills/`
Settings file	`.opencode/config.json`
MCP support	Via adapter
SKILL.md	Native (no conversion)

dz setup --target opencode

hermes

Hermes AI coding agent.

Property	Value
Platform	Hermes
Skills directory	`.hermes/skills/`
Settings file	`.hermes/config.json`
MCP support	Via adapter
SKILL.md	Native (no conversion)

dz setup --target hermes

openclaude

Open-source Claude-compatible agent.

Property	Value
Platform	OpenClaude
Skills directory	`.openclaude/skills/`
Settings file	`.openclaude/config.json`
MCP support	Via adapter
SKILL.md	Native (no conversion)

dz setup --target openclaude

Native agentskills.io Support

The targets described above use agentskills.io-shaped SKILL.md files. This means:

No format conversion required for any target
A skill can be reused across supported targets; check the target guide for platform-specific behavior
Cross-platform equivalence is verified by the cross-adapter-equivalence test suite

Optional Enrichment (`--enrich`)

While all platforms natively read SKILL.md, the --enrich flag generates additional platform-specific metadata for richer integration (UI metadata, implicit invocation, agent config):

Platform	Enrichment File	What It Adds
claude-code	(none needed)	Native — no enrichment required
openclaude	(none needed)	Native — no enrichment required
codex	`agents/openai.yaml`	UI display name, category, implicit invocation triggers, risk level
opencode	`.opencode/agents/<id>.md`	Agent definition with tool permissions
hermes	`hermes-config.yaml`	Skill config with timeout and retry settings
cursor	`.cursor/rules`	Cursor — skills compiled as project rules
windsurf	`.windsurf/rules`	Windsurf — skills compiled as rules (glob/model-decision triggers)
gemini	`.github/instructions`	Gemini Code Assist — instruction files
copilot	`.github/copilot-skills`	GitHub Copilot — skill instruction files
agents-md	`AGENTS.md`	Single managed AGENTS.md file — the lowest-common-denominator target for any agent that reads it

# Initialize with enrichment for Codex
dz init --target codex --preset devops --enrich

# Enrichment for OpenCode
dz init --target opencode --preset meta --enrich

4-Axis Risk Scoring for Codex

When using --enrich with Codex, DZ Harness Hub computes a 4-axis risk score for each skill and writes it to agents/openai.yaml. The score is informational; Codex does not enforce it as an invocation gate.

The 4 axes:

Axis	Description	Example
base_tool	Intrinsic risk of the primary tool used	`rm` = high, `cat` = low
file_sensitivity	Sensitivity of files the skill reads/writes	Secrets = high, docs = low
blast_radius	How many files/systems the skill can affect	Monorepo-wide = high, single file = low
irreversibility	Whether the action can be undone	Delete = high, read = low

Risk levels derived from the 4-axis score:

Level	Score Range	Examples
critical	≥ 0.70	`rm -rf`, destructive database migrations
high	≥ 0.50	`terraform apply`, infrastructure changes
medium	≥ 0.30	`terraform plan`, `docker-compose up`
low	< 0.30	`design-thinking`, `document-checker`, read-only skills

(Default thresholds from risk-scoring.ts: medium 0.3, high 0.5, critical 0.7.)

Review the skill’s permissions and the target’s actual approval settings before running sensitive actions. The risk_level metadata alone does not trigger a Codex confirmation.

Multi-Target Setup

You can deploy skills to multiple targets simultaneously:

# Set up for Claude Code and Codex
dz init --target claude-code --preset devops

dz init --target codex --preset devops --enrich

Target-Specific Behavior

Aspect	claude-code	codex	opencode	hermes	openclaude
Skill format	SKILL.md	SKILL.md	SKILL.md	SKILL.md	SKILL.md
Skills root	`.claude/skills/`	`.agents/skills/`	`.opencode/skills/`	`.hermes/skills/`	`.openclaude/skills/`
MCP config	`settings.json`	Adapter	Adapter	Adapter	Adapter
Memory path	`.dz/` (shared)	`.dz/` (shared)	`.dz/` (shared)	`.dz/` (shared)	`.dz/` (shared)
Enrichment	Not needed	`agents/openai.yaml`	`.opencode/agents/`	`hermes-config.yaml`	Not needed

Checking Target Status

# See which targets are configured and verify health
dz doctor

# Verify the installed skill structure for a target
dz verify --target codex

Switching Targets

There is no persistent “default target” setting — pass --target to each command:

# Initialize a different target
dz init --target codex --preset devops

# Initialize Hermes
dz init --target hermes --preset devops