Host support: JetBrains Junie — extension marketplace, skills, guidelines, MCP and gated hook profiles

[ScriptedAlchemy]

Scope / research status

P2 feature expansion. Research date 2026-09-06; Agent Bundle baseline 5fc8de37c9695b816d55bdcf18a6bcb1dfee69f5. Read current JetBrains/Junie documentation and first-party extension source; no native installation or authenticated execution was run.

Junie is no longer accurately described as only a manually configured AGENTS.md reader. Its CLI supports extension marketplaces, packaged skills/MCP/agents/commands/guidelines/hooks. IDE integrations and ACP still require separate capability checks. This is one Junie-family issue, not a request for a custom IntelliJ Platform plugin.

Primary specifications and examples

• Extension management, marketplace formats, install scope and cache
• Official extension structure
• Minimal first-party extension manifest
• Skill format, discovery, disabled state and custom locations
• Guidelines precedence and AGENTS.md
• MCP settings and ACP differences
• Hook JSON, native results, packaging, client support and Early Access gate
• IDE plugin versus integrated AI Chat

Pin host version/channel and parser revision. Do not combine an old IDE limitation, the new CLI guide, and a beta hook API into one unsupported universal profile.

Native extension and marketplace

Native marketplaces use .junie-extension/marketplace.json; Claude-compatible marketplace manifests are also accepted. Extensions are downloaded into a user cache and enabled by project/user reference files. Installation is available in-session; removing a reference need not delete the shared cache.

The first-party repository gives this package structure:

composite-root/
  extension.json
  skills/review/SKILL.md
  guidelines/project.md
  mcp/.mcp.json
  hooks/hooks.json             # only for a supported installed hook profile
  hooks/before.junie.mjs

{"name":"example-plugin","description":"Project review capabilities"}

The marketplace can point at this extension directory; do not force a second physical copy of the shared compiled application. Reuse the existing Claude projection where its exact supported subset works, or emit the native extension marker and documented component paths. Verify lookup precedence when both are present.

Documented interactive management:

/extensions
/extensions marketplace add https://github.com/example/plugin
/extensions install example-plugin

The /plugin and /plugins aliases are native UI aliases, not proof of equivalent standalone shell subcommands. The UI supports local directories, Git repositories and direct marketplace.json URLs. Project references live in .junie/extensions.json; user references and cache live under ~/.junie/extensions/, with an explicit override available. Use native identifiers and preserve cached data used by other projects.

Skills and persistent guidance

Skills are found under project/user .junie/skills, trusted .agents/skills, installed extensions and configured extra locations. Project skills can override user names, and disabled state persists. Preserve explicit names and descriptions instead of depending on fallback description extraction. Native /skills is a management UI in TUI but a different listing/control surface in ACP.

Guidelines precedence is materially important: project .junie/AGENTS.md wins; otherwise root AGENTS.md is combined with Junie playbook/rules; legacy guidelines are fallback. Do not create a new highest-precedence file that hides existing authored guidance. Global and project content may both apply. Test required and forbidden cross-host discovery before allowing the composite combination.

MCP configuration

Independent project MCP configuration is .junie/mcp/mcp.json, with user equivalent under ~/.junie. Installed extensions can carry their own server definitions. Avoid registering the same canonical server twice.

{"mcpServers":{"example":{"command":"node","args":["/absolute/installed-plugin/mcp/example.mjs"],"env":{}}}}

The example is an operator setting, not a published absolute build path. Use a proven extension-root token/binding for packaged content and native secrets/OAuth. /mcp can configure and authorize in TUI, but ACP only lists local/client-provided servers; do not expect an interactive installation wizard to work headlessly.

Hook capability boundary

The hook guide is Early Access. Installed extension hooks use hooks/hooks.json, and ${JUNIE_EXTENSION_ROOT}/${CLAUDE_PLUGIN_ROOT} resolve to the installed extension. Default repository-local hook config is ignored unless explicitly selected; never enable it globally as a workaround.

Map exact event semantics. A synchronous PreToolUse hook can adjust input or deny; asynchronous hooks cannot block. PermissionRequest is not observational: an apparently harmless successful callback can authorize an action. Do not register a no-op observer on that event. Session and Stop effects differ, and TUI/batch/ACP firing coverage is not uniform. Pin and test these differences; shared field names do not establish full Claude parity.

Acceptance and implementation

• Pin CLI/IDE/ACP profiles and the Early Access hook requirement; unsupported combinations yield clear diagnostics.
• Native and Claude-compatible marketplace/extension discovery is deterministic with one canonical application identity.
• Static skills/guidelines install with no synthetic MCP/Flight/state process.
• Test default/custom skill roots, duplicate names, disabled state and guideline precedence without overwriting user files.
• Execute one generated MCP operation from an installed relocated source-free package, with correct schema/result/cancellation and native authorization.
• Hook before-tool denial prevents a sentinel action; async/observer/permission-request semantics never widen authority accidentally.
• Verify exact root-token expansion, quoting, timeout and unsupported-client behavior.
• Project/user enablement and shared cache survive update/removal without deleting another installation's content or state.
• Add Composite conformance: test cross-host auto-discovery and precedence, not only file collisions #651 composite and Small-plugin acceptance: skills and plain hooks without an unnecessary MCP/RSC runtime or empty Workbench #652 small-plugin acceptance, plus a real successful native run distinct from listed resources or emitted requests.
• Document setup in English/Chinese with versioned source provenance.

Use existing adapters, content lowering, canonical manifests/executable records and receipts. No IDE extension authoring, authentication provisioning, Brave-mode enablement, new agent orchestration, automatic cache purge or blanket removal of AB4106.

[ScriptedAlchemy]

Landed in #727, squash-merged as b5cf7575e0 on main.

Junie reads the artifact Agent Bundle already emits, so this needed no adapter, projection, or compiler change — the support is one pinned record in packages/agent-bundle/src/adapters/capabilities/portable-1.0.0.json, which feeds the generated en/zh hosts reference, the generated INSTALL.md, and the discovery tests. Read against junie.jetbrains.com/docs on 2026-09-06 (the agent-skills page is dated 01 September 2026; no CLI version is published on the page).

Tier: skills. Registration points Junie at the emitted tree:

junie --skill-location <plugin directory>/skills

skill-locations in config.json makes the same folder persistent; .junie/skills/<skill>/, ~/.junie/skills/<skill>/, and .agents/skills/ in a trusted project are the default roots. /skills lists what loaded.

Surface	State	Evidence
skills	partial	agent-skills requires a folder per skill containing SKILL.md with YAML frontmatter whose name is required and description optional — satisfied by every skill this projection emits — and reads extra folders given with --skill-location. Partial, not full: the emitted root is not a default location, so it loads only once registered, and "if a project-level and a user-level skills have the same name, the user-level skill will be skipped".
manifest	unsupported	Junie's documented units are agent skills, guidelines, and MCP configuration; no page defines a plugin manifest, so the emitted root plugin.json is read by nothing there.
mcp	unsupported	MCP servers are read from .junie/mcp/mcp.json or ~/.junie/mcp/mcp.json (junie-cli-mcp-configuration). --mcp-location adds "additional folders where MCP servers should be found", but no page states the file name it looks for inside them, and the documented server fields are command/args/env and url/headers with no type discriminator, so the emitted plugin-root mcp.json is not a proven Junie MCP source.
placeholders	unsupported	No ${PLUGIN_ROOT} or ${PLUGIN_DATA} expansion is documented for Junie MCP configuration or skills, so Agent Plugins 1.0.0 §9.1 reserved behavior is not provided.
hooks	unsupported	Agent Plugins 1.0.0 defines no hooks, so this projection emits no hooks document for Junie to read.

Deferred: the MCP row, on one missing fact. If JetBrains documents the file name --mcp-location looks for inside an added folder, and whether the spec's type discriminator is accepted, the emitted mcp.json may become a Junie MCP source and the row moves without any code change — the record is the only thing that would need editing. Nothing else here is open.