harness-command

name	harness-command
description	Help create and manage command entities in the agent-harness workspace. Use when the user wants to add a custom slash command, understands what command entities are, asks how commands differ from skills, or needs to know which providers support commands.
argument-hint	[command-id]
disable-model-invocation	true

harness-command

You are helping the user create and manage command entities in an agent-harness workspace.

What is a command entity?

A command entity defines a custom slash command — a reusable, invocable prompt that users trigger explicitly by typing /command-name in a supported AI coding tool. Commands are imperative one-shot actions: deploy a service, fix a GitHub issue, scaffold a component.

Source lives at: .harness/src/commands/<command-id>.md

How commands differ from skills

Aspect	skill	command
Source location	.harness/src/skills/<id>/SKILL.md (directory)	.harness/src/commands/<id>.md (single file)
Supporting files	Yes — additional files in the skill directory	No
Intended use	Reusable knowledge, conventions, recurring reference	Explicit one-shot actions with side effects
Auto-invocation by AI	Yes (model can load when relevant)	No — explicit invocation only
Provider support	Claude, Codex, Copilot	Claude, Copilot only (Codex: not supported)

Use a skill for knowledge that Claude should apply automatically (API conventions, code style). Use a command for workflows you always trigger manually (/deploy, /fix-issue, /release).

Source file format

.harness/src/commands/<command-id>.md is a Markdown file with YAML frontmatter:

---
description: "Short description shown in the slash command picker"
argument-hint: "[issue-number]"
---

# command-id

Task instructions go here. Use $ARGUMENTS to receive arguments from the invocation.

Example: Fix GitHub issue $ARGUMENTS following the project's coding standards.

Steps:
1. Read the issue description
2. Implement the fix
3. Write tests
4. Create a commit

Frontmatter fields:

Field	Required	Description
description	Yes	Non-empty string. Shown in the slash command picker. Validation fails if missing.
argument-hint	No	Hint shown at invocation (e.g. [issue-number] or [filename] [format]). Claude and Copilot only.

Body placeholders:

• $ARGUMENTS — all text after the command name at invocation
• $ARGUMENTS[N] / $N — individual positional argument (Claude Code only)

Provider output mapping

Claude Code

• Output path: .claude/commands/<command-id>.md
• Invocation: /<command-id> in the Claude Code chat input
• Format: Markdown with YAML frontmatter

Generated frontmatter includes description and, if set, argument-hint. The body is written as-is.

Example output at .claude/commands/fix-issue.md:

---
description: "Fix a GitHub issue"
argument-hint: "[issue-number]"
---

Fix GitHub issue $ARGUMENTS following the project's coding standards.

Claude Code also accepts .claude/commands/ files directly (legacy commands); harness manages these files for you.

Reference: https://code.claude.com/docs/en/slash-commands

GitHub Copilot

• Output path: .github/prompts/<command-id>.prompt.md
• Invocation: /<command-id> in Copilot Chat (VS Code, Visual Studio, JetBrains)
• Format: Markdown with YAML frontmatter; always includes agent: agent

Generated frontmatter always sets agent: agent (runs in agent mode). The description from your source is included. Body is written as-is.

Example output at .github/prompts/fix-issue.prompt.md:

---
agent: agent
description: "Fix a GitHub issue"
---

Fix GitHub issue $ARGUMENTS following the project's coding standards.

Note: Copilot prompt files are supported in VS Code, Visual Studio, and JetBrains IDEs. As of early 2026, Copilot CLI does not support custom slash commands from .github/prompts/.

Reference: https://code.visualstudio.com/docs/copilot/customization/prompt-files

OpenAI Codex CLI

Not supported. Codex has no renderCommand implementation. Codex deprecated its own custom prompts feature (~/.codex/prompts/) in favor of skills. Command entities are silently skipped when generating Codex artifacts.

If you need Codex-specific reusable prompts, use a skill entity instead.

Provider support matrix

Provider	Supported	Output path	Format
Claude Code	Yes	.claude/commands/<id>.md	Markdown + frontmatter
GitHub Copilot	Yes	.github/prompts/<id>.prompt.md	Markdown + frontmatter (agent: agent)
OpenAI Codex	No	—	—

Harness CLI commands

# Scaffold a new command entity and register it in manifest.json
npx harness add command <command-id>

# Preview what artifacts will be generated (dry run)
npx harness plan

# Write provider artifacts to disk
npx harness apply

# Remove a command entity and its source file
npx harness remove command <command-id>

# Remove but keep the source file
npx harness remove command <command-id> --no-delete-source

Complete worked example

1. Scaffold the entity:

npx harness add command fix-issue

2. Edit .harness/src/commands/fix-issue.md:

---
description: "Fix a GitHub issue by number"
argument-hint: "[issue-number]"
---

Fix GitHub issue $ARGUMENTS following the project's coding standards.

1. Read the issue description with `gh issue view $ARGUMENTS`
2. Understand the requirements
3. Implement the fix in the appropriate files
4. Write or update tests
5. Commit with a message referencing the issue

3. Apply:

npx harness apply

This writes:

• .claude/commands/fix-issue.md (if Claude provider is enabled)
• .github/prompts/fix-issue.prompt.md (if Copilot provider is enabled)

4. Invoke in Claude Code:

/fix-issue 1234

Override sidecars

Per-provider overrides are supported via YAML sidecar files:

.harness/src/commands/<id>.overrides.<provider>.yaml

version: 1
enabled: false        # disable this command for a specific provider
targetPath: "custom/path/command.md"   # override output path

If the user provides a command-id as $ARGUMENTS

Scaffold the command now:

npx harness add command $ARGUMENTS

Then open .harness/src/commands/$ARGUMENTS.md and fill in:

1. A clear description in the frontmatter
2. An argument-hint if the command takes arguments
3. Step-by-step instructions in the body
4. $ARGUMENTS placeholder where input should be inserted

Run npx harness apply when done.