Target Platforms Reference

Last Updated: January 2026

This document tracks the current state of Calvin's target platforms and their output formats.

Platform Overview

Platform	Type	Current Version	Model	Status
Claude Code	CLI + VS Code Extension	2.0.70	Claude Sonnet 4.5	Stable
Cursor	IDE (VS Code fork)	2.2	Multi-model	Stable
VS Code + Copilot	IDE + Extension	1.95+	GPT-4 / Claude	Stable
Antigravity	IDE (VS Code fork)	1.11.17 Preview	Gemini 3 Pro	Beta
Codex	CLI	0.72.0	GPT-5-codex	Experimental
OpenCode	CLI	0.x	Multi-model	Experimental

Claude Code

Vendor: Anthropic
Latest Version: 2.0.70 (December 16, 2025)
Model: Claude Sonnet 4.5

Recent Changes (v2.0.70)

Fixed terminal flickering issues
3x memory optimization for large conversations
MCP tool permission wildcards support

Output Locations

Asset Type	Output Path	Scope
Commands	`.claude/commands/<id>.md`	Project
Commands	`~/.claude/commands/<id>.md`	User
Settings	`.claude/settings.json`	Project
Settings Local	`.claude/settings.local.json`	Project (gitignored)
Agents	`.claude/agents/<id>.md`	Project
Agents	`~/.claude/agents/<id>.md`	User
Skills	`.claude/skills/<id>/SKILL.md`	Project
Skills	`~/.claude/skills/<id>/SKILL.md`	User
Memory	`CLAUDE.md`	Project

Format: Commands

<!-- Generated by Calvin. Source: .promptpack/1-planning/1-plan-init-project.md -->

# Project Initialization

Arguments: $ARGUMENTS

[... prompt content ...]

Format: Settings

{
  "permissions": {
    "deny": [".env", ".env.*", "*.pem", "*.key", "id_rsa", ".git/"]
  }
}

Format: Agents

Claude Code agents use YAML frontmatter with specific fields for subagent configuration:

---
name: code-reviewer
description: Use proactively to review code changes for security, performance, and maintainability.
tools: Read, Grep, Glob, Bash
model: sonnet
permissionMode: acceptEdits
skills: engineering-standards
---

You are a senior code reviewer. Analyze changes and provide:
- Security concerns (with severity)
- Performance implications
- Suggested improvements with minimal diff

Do NOT make changes directly. Output actionable review comments.

<!-- Generated by Calvin. Source: .promptpack/agents/reviewer.md. DO NOT EDIT. -->

Field	Required	Description
`name`	Yes	Agent identifier for routing
`description`	Yes	Used by Claude for auto-delegation decisions
`tools`	No	Comma-separated tool whitelist (inherits all if omitted)
`model`	No	`sonnet`, `opus`, `haiku`, or `inherit`
`permissionMode`	No	How to handle permission prompts
`skills`	No	Comma-separated skills to preload

Cursor

Vendor: Cursor Inc.
Latest Version: 2.2 (December 10, 2025)
Model: Multi-model (Claude, GPT-4, etc.)

Recent Changes (v2.2)

Debug Mode: AI-assisted bug reproduction and fixing
Plan Mode: Enhanced "think before code" capability
Agent Review: Dedicated code review agent in sidebar

Output Locations

Asset Type	Output Path	Scope
Rules	`.cursor/rules/<id>/RULE.md`	Project
Commands	`.cursor/commands/<id>.md`	Project
Agents (fallback)	`.cursor/commands/<id>.md`	Project
MCP Config (manual; validated by `calvin check`)	`.cursor/mcp.json`	Project
Skills	`.claude/skills/<id>/SKILL.md`	Project
Skills	`~/.claude/skills/<id>/SKILL.md`	User

Note: Cursor skill support uses Claude Code's skill paths (.claude/skills/). Note: Agents are compiled to .cursor/commands/ only when Claude Code target is NOT enabled. When both are enabled, agents go to .claude/agents/ only. Note: Calvin does not currently generate .cursor/mcp.json; it only validates it (allowlist + JSON schema) when present.

Format: Rules

---
description: Code Style Guidelines
globs: ["**/*.ts", "**/*.tsx"]
alwaysApply: true
---

# Code Style Guidelines

[... policy content ...]

Format: Commands

<!-- Generated by Calvin -->

# Generate Tests

$ARGUMENTS

[... action content ...]

VS Code + GitHub Copilot

Vendor: Microsoft / GitHub
Latest Version: VS Code 1.95+
Model: GPT-4, Claude (configurable)

Output Locations

Asset Type	Output Path	Scope
Instructions (merged)	`.github/copilot-instructions.md`	Project
Instructions (split)	`.github/instructions/<id>.instructions.md`	Project
Agents	`.github/instructions/<id>.instructions.md`	Project
Agents Summary	`AGENTS.md`	Project

Note: Agents are compiled as instruction files and also listed in the AGENTS.md summary.

Format: Instructions

---
applyTo: "**/*.py"
---

# Python Code Style

[... policy content ...]

Notes

Instructions only affect Chat, not inline completions
chat.useAgentsMdFile setting must be enabled for AGENTS.md
Skills are not supported on VS Code + Copilot.

Antigravity

Vendor: Google DeepMind
Latest Version: 1.11.17 Preview (December 8, 2025)
Model: Gemini 3 Pro

About

Antigravity is Google's agent-first IDE, launched November 18, 2025. It's not just code completion—it's a "task control center" where AI agents autonomously:

Write and modify code
Run terminal commands
Operate browsers for testing
Collaborate asynchronously

Output Locations

Asset Type	Output Path	Scope
Rules	`.agent/rules/<id>.md`	Project
Workflows	`.agent/workflows/<id>.md`	Project
Agents	`.agent/workflows/<id>.md`	Project
Rules (global)	`~/.gemini/antigravity/global_rules/`	User
Workflows (global)	`~/.gemini/antigravity/global_workflows/`	User
Browser Allowlist	`~/.gemini/antigravity/browserAllowlist.txt`	User

Note: Antigravity does not support SKILL.md skills. Note: Agents are compiled as workflows; agent-specific fields (tools, model, permissionMode, skills) are ignored.

Format: Workflows

---
description: Generate Tests
---

# Generate Unit Tests

[... workflow content ...]

Security Modes

Mode	Description
Off	No auto-execution
Auto (Balanced)	Prompt for approval
Turbo	Auto-execute (dangerous)

Calvin defaults to recommending Auto mode.

Codex

Vendor: OpenAI
Latest Version: 0.72.0 (December 13, 2025)
Model: gpt-5-codex / gpt-5.1-codex-max

About

Codex is OpenAI's CLI programming agent, installed via:

npm install -g @openai/codex

Output Locations

Asset Type	Output Path	Scope
Prompts	`.codex/prompts/<id>.md`	Project
Prompts	`~/.codex/prompts/<id>.md`	User
Agents	`.codex/prompts/<id>.md`	Project
Agents	`~/.codex/prompts/<id>.md`	User
Skills	`.codex/skills/<id>/SKILL.md`	Project
Skills	`~/.codex/skills/<id>/SKILL.md`	User

Note: Agents are compiled as prompts with $ARGUMENTS placeholder; agent-specific fields are ignored.

Format: Prompts

<!-- Generated by Calvin -->

# Generate Tests

Usage: /prompts:generate-tests <target_file>

Arguments: $ARGUMENTS
- $1: Target file or function

[... prompt content ...]

Parameter Placeholders

Placeholder	Description
`$ARGUMENTS`	All arguments as string
`$1` - `$9`	Positional arguments
`$KEY`	Named argument (KEY=value)

OpenCode

Vendor: SST
Latest Version: 0.x
Model: Multi-model (provider/model IDs)

About

OpenCode is an open-source terminal AI coding agent. Calvin supports OpenCode by compiling:

Agents to .opencode/agent/
Actions to .opencode/command/ (slash commands)
Policies to AGENTS.md (project) and ~/.config/opencode/AGENTS.md (user)
Skills to .opencode/skill/ (or reusing .claude/skills/ when Claude Code is also enabled)

Output Locations

Asset Type	Output Path	Scope
Agents	`.opencode/agent/<id>.md`	Project
Agents	`~/.config/opencode/agent/<id>.md`	User
Commands	`.opencode/command/<id>.md`	Project
Commands	`~/.config/opencode/command/<id>.md`	User
Policies	`AGENTS.md`	Project
Policies	`~/.config/opencode/AGENTS.md`	User
Skills	`.opencode/skill/<id>/SKILL.md`	Project
Skills	`~/.config/opencode/skill/<id>/SKILL.md`	User

Format: Agents

---
description: Review code changes for security, performance, and maintainability.
mode: subagent
model: anthropic/claude-sonnet-4-5
temperature: 0.3
tools:
  read: true
  bash: true
permission:
  edit: allow
---

[... agent prompt ...]

<!-- Generated by Calvin. Source: .promptpack/agents/reviewer.md. DO NOT EDIT. -->

Format: Commands

---
description: Run tests for a module
agent: build
subtask: false
---

Run tests for $ARGUMENTS

<!-- Generated by Calvin. Source: .promptpack/actions/test.md. DO NOT EDIT. -->

Adapter Compatibility Matrix

Feature	Claude Code	Cursor	VS Code	Antigravity	Codex	OpenCode
Project-scope commands	✅	✅	⚠️	✅	❌	✅
User-scope commands	✅	⚠️	❌	✅	✅	✅
Policy/Rules	✅	✅	✅	✅	❌	✅
Settings/Deny lists	✅	⚠️	❌	⚠️	❌	❌
MCP allowlist validation	❌	✅	❌	❌	❌	❌
Agents/Subagents	✅	⚠️¹	⚠️²	⚠️³	⚠️³	✅
Skills	✅	✅	❌	❌	✅	✅

Legend: ✅ Full support | ⚠️ Partial/experimental | ❌ Not supported

Notes:

Cursor agents: Falls back to .cursor/commands/ when Claude Code is not enabled. When both are enabled, agents go to .claude/agents/ only.
VS Code agents: Compiled as instructions (.github/instructions/<id>.instructions.md) and listed in AGENTS.md.
Antigravity/Codex agents: Compiled as workflows/prompts; agent-specific fields (tools, model, etc.) are ignored.
OpenCode skills: When Claude Code is enabled, skills are written to .claude/skills/ only (OpenCode reads Claude Code skills).

Version Tracking

Calvin exposes Calvin + adapter versions via calvin version:

$ calvin version

╭─────────────────────╮
│ Calvin v0.6.0       │
│ Source Format: 1.0  │
│                     │
│ Adapters:           │
│   - ClaudeCode   v1 │
│   - Cursor       v1 │
│   - VSCode       v1 │
│   - Antigravity  v1 │
│   - Codex        v1 │
│   - OpenCode     v1 │
╰─────────────────────╯

FilesExpand file tree

target-platforms.md

Latest commit

History

target-platforms.md

File metadata and controls

Target Platforms Reference

Platform Overview

Claude Code

Recent Changes (v2.0.70)

Output Locations

Format: Commands

Format: Settings

Format: Agents

Cursor

Recent Changes (v2.2)

Output Locations

Format: Rules

Format: Commands

VS Code + GitHub Copilot

Output Locations

Format: Instructions

Notes

Antigravity

About

Output Locations

Format: Workflows

Security Modes

Codex

About

Output Locations

Format: Prompts

Parameter Placeholders

OpenCode

About

Output Locations

Format: Agents

Format: Commands

Adapter Compatibility Matrix

Version Tracking