Host support: Jules — repository instruction projection and explicit cloud-environment compatibility

[ScriptedAlchemy]

Scope / research evidence

P2 feature expansion, initially content/environment compatibility rather than an invented native plugin runtime. Research date 2026-09-06; Agent Bundle baseline 5fc8de37c9695b816d55bdcf18a6bcb1dfee69f5. Read Google's official Jules documentation. No Jules session, account connection, cloud task or deployment was performed.

Ponytail demonstrates repository instructions consumed by Jules. The primary references below establish that path, optional task setup, remote session management, and selected managed MCP connections. They do not establish an arbitrary third-party plugin package loader, custom local MCP registration, or general user-authored lifecycle hook API. Keep unverified capabilities unavailable until a primary specification and actual acceptance test exist.

Authoritative references

• Repository checkout, task setup, root AGENTS.md discovery
• Environment setup, validation and snapshots
• Running tasks
• Official MCP launch announcement and supported-service configuration
• Jules Tools CLI: remote sessions, not a local plugin harness
• Official documentation/changelog index
• AGENTS.md convention

Record the research date and actual runtime/tool versions in acceptance; a mutable cloud service does not have a local version guarantee merely because the helper CLI is pinned.

Supported first projection

Jules clones the selected GitHub repository/branch into its cloud environment and reads root AGENTS.md. Reuse the existing authored instruction/content pipeline. Do not overwrite a repository's existing instructions or silently expand a host-scoped rule into global guidance.

Proposed consumer layout:

consumer-repository/
  AGENTS.md
  agent-instructions/review.md
  agent-instructions/resources/...
  scripts/agent-environment-setup.sh  # only when explicitly authored/requested

This is repository content, not a Jules-specific plugin manifest. A reference in AGENTS.md does not automatically mean every linked file has entered the model context; preserve the explicit instruction and test relevant resource use. Do not promise skill discovery or slash commands solely because SKILL.md files are present.

Operator setup

1. In Jules, authorize the intended repository and choose the branch containing the generated/owned instructions.
2. Preserve the user's plan-review and execution settings.
3. In that repository's Configuration → Initial Setup, optionally reference an explicitly authored setup script. Use Run and Snapshot to validate it.
4. Start a new task using the intended branch and inspect the resulting plan/changes.

A setup script is task environment preparation, not a pre-tool permission hook. The documented environment provides common language tools, but bundled examples of tool versions are not a permanent dependency guarantee. Validate the actual runtime and install only the task's declared dependencies. Snapshots may retain earlier setup output: acceptance must prove the current artifact/version is used, not an old cached build.

Keep credentials in native operator settings, never AGENTS.md, committed setup scripts or generated artifacts. An absolute path on the developer's laptop is not a valid path inside the Jules VM.

MCP and CLI boundaries

The February 2 official MCP announcement documents Settings → MCP connections for selected services (Linear, Stitch, Neon, Tinybird, Context7 and Supabase), with service credentials supplied by the user. It explicitly describes an initially restricted integration list. This is not evidence that a generated arbitrary mcpServers JSON file will register an Agent Bundle server. Recheck official expansion before enabling such a capability; do not equate managed-provider MCP with arbitrary stdio plugin support.

Jules Tools installs as @google/jules; commands such as jules login, jules remote list --session and the documented remote task commands manage cloud sessions. Embedding that CLI is not a way to turn Jules into a local Workbench agent with native plugin hooks. Remote task scheduling/API orchestration is outside this adapter's responsibility.

Implementation and acceptance

• Define a clearly named Jules content projection using existing portable instruction lowering; unsupported tool/hook/plugin-package declarations are diagnosed rather than silently ignored.
• Preserve existing authored AGENTS.md, scopes, links and project instructions; no automatic summarization or destructive replacement.
• A content-only build adds no synthetic MCP, Flight, SQLite or background process.
• Test composite coexistence with other AGENTS.md readers, including forbidden scope leakage and deterministic output.
• Test branch/commit and resource identity in the actual cloud checkout; no local absolute paths or missing ignored generated files.
• Where setup is requested, validate runtime requirements, cold setup and warm-snapshot behavior against the same intended package version.
• Distinguish documented managed MCP connections from any newly verified arbitrary-server capability; no invented manifest or install command.
• Removal/update affects only owned repository content or explicitly registered setup references, not account grants, other tasks, secrets or unrelated files.
• Record a successful real task using the intended instructions separately from static file presence; redact service/account data.
• English/Chinese setup docs state the exact supported tier and link native documentation.

Non-goals

No Jules account provisioning, GitHub authorization changes, automatic remote task creation, scheduler, new model agent loop, fabricated hook API, or deployment of an MCP server. Reuse #651/#652 conformance and existing manifest/content/receipt machinery; do not broaden executable support without a verified host contract.

[ScriptedAlchemy]

Owner decision (2026-09-06): Jules support is no longer required. No implementation or compatibility claim should be added.