Deploy Developer

Deploy Developer

Create a nested session for development work by calling create_session with these exact kickoff settings:

• model: gpt-5.6-luna
• reasoning_effort: max
• context_tier: long_context

The long_context value selects the model's 1.1M context setting. Preserve any explicitly requested session name, mode, execution location, workspace type, base branch, notification behavior, or coordination setting. Otherwise:

• Create the session in the current project by omitting project_id.
• Use local execution by omitting execution_location.
• Use an isolated worktree by omitting workspace_type.
• Coordinate with the creator by leaving coordinate_with_creator enabled.
• Set notify_on_idle to once.
• Choose a short, descriptive sentence-case session name based on the task.
• Do not set base_branch unless the user explicitly requests one or the task clearly depends on an in-progress branch.

If the user asks only to create a developer session but does not provide the development task, ask for the task before creating the session.

Build a precise kickoff

The nested session automatically receives applicable repository Copilot instruction files. Do not repeat stable repository guidance already defined there, such as general coding conventions, architecture, standard validation, or formatting rules.

Make kickoff.prompt standalone but concise. Include only task-specific facts that the nested session cannot reliably recover from the repository:

Task:
Scope:
Required behavior:
Known findings/risks:
Validation:
Deliverable and permissions:

Apply these rules:

• Preserve exact user requirements and conversation-only decisions.
• Prefer exact findings, acceptance criteria, symbols, and file paths over background narrative.
• State branch or commit dependencies when the task relies on in-progress work.
• Mention validation limitations only when they affect execution.
• State side-effect permissions once: whether to commit, push, deploy, post, or report only.
• Consolidate overlapping context instead of repeating it under multiple headings.
• Do not repeat instructions supplied by another skill the developer is explicitly told to invoke.

Aim for roughly 150-350 words. Exceed that range only when multiple subsystems, exact findings, or substantial conversation-only decisions require it. Long context is capacity, not a prompt-length target.