Apache Magpie

Search guides, workflows, and reference pages.

↑ ↓ to move · Enter to openApache Magpie docs
Docs/adapters

JetBrains Junie

Capability: platform:harness

Junie is JetBrains’ own coding agent, available as a CLI and as a JetBrains IDE plugin. It is a different harness from Claude Code running inside a JetBrains IDE, which needs nothing extra (see marketplace.md). Tracking issue: #321.

Harness contract

Surface Junie contract Magpie integration
Skills Open Agent Skills format; reads <project>/.agents/skills/ (trusted projects), ~/.agents/skills/, <project>/.junie/skills/, ~/.junie/skills/ The universal registry row; no Junie-specific relay
Invoke a skill Automatically, as /<skill-name>, or as $<skill-name> Same skill files as every other harness
Instructions AGENTS.md in the project root, .junie/rules/*.md, ~/.junie/AGENTS.md AGENTS.md is read as is
MCP .junie/mcp/mcp.json (project) with the usual mcpServers shape Adopter-owned; Magpie writes none
Action guard PreToolUse hook in ~/.junie/config.json (Early Access) agent-guard --junie
Sandbox junie --sandbox (sandbox-runtime, macOS and Linux) Not managed by Magpie
Approvals /brave levels and ~/.junie/allowlist.json Not managed by Magpie

Installation

Install an adopted project with the universal target, which is the default. Junie loads .agents/skills/ only in a trusted project; trust it from Junie itself, Magpie never grants it.

Action guard

Junie’s hook payload is Claude-shaped (tool_name, tool_input.command) but its block semantics differ. Junie honours exit code 2 and a top-level decision, and ignores Claude’s hookSpecificOutput. The no-argument agent-guard would route the payload to the Claude adapter, exit 0, and let every command through. Hooks must therefore pass --junie.

Add to ~/.junie/config.json (or a file passed with --config-location):

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "python3 <framework>/tools/agent-guard/src/agent_guard/__init__.py --junie",
            "timeout": 30
          }
        ]
      }
    ]
  }
}

Replace <framework> with the absolute path of the checkout, or of .apache-magpie/ for a snapshot adopter. Junie hooks live in user configuration, so Magpie does not write the file; the adopter adds the entry.

The adapter blocks with exit 2, the denial reason on stderr, and a {"decision": "deny", "reason": ...} object on stdout. Malformed events, other tools, and allowed commands exit 0 silently. Other non-zero exits only warn in Junie and let the tool run, so the adapter never uses them.

Verify

  1. Confirm the hook is listed in ~/.junie/config.json and the command path exists.
  2. Run git status through Junie and confirm it runs.
  3. Run git commit --no-verify --dry-run and confirm Junie reports a block from agent-guard[no-verify].
  4. Run echo '{"tool_name":"Bash","tool_input":{"command":"git commit --no-verify"}}' | python3 <framework>/tools/agent-guard/src/agent_guard/__init__.py --junie; echo $? and expect exit 2.

A static check of the config does not replace the live denial probe.

Known limitations

  • Hooks are in Junie’s Early Access Program; the schema may change, so re-verify after upgrades.
  • The guard covers shell commands only. Native file tools, MCP servers, and hooks themselves are outside it, and outside junie --sandbox too.
  • The sandbox policy is the global ~/.junie/sandbox.json; a project-local one is ignored, and network is unrestricted unless allowedDomains is set.
  • The payload shape was taken from the Junie documentation, not a captured live event. Capture one and add it to tools/agent-guard/tests/test_junie.py when a build is available.
  • There is no spec-loop runner profile for Junie yet.

Upstream references

Suggest a change