Deploy Code Review

Deploy code review

Create a nested project session that performs the review. The main session is the coordinator and must not perform the delegated review itself.

Fixed reviewer settings

• model: gpt-5.6-sol
• reasoning effort: medium
• context tier: long_context (1.1M context)
• mode: autopilot
• execution location: local
• workspace type: branch
• coordination with the main session: enabled
• idle notification: once
• default agent: omit agent

Normally omit base_branch so the nested session uses the current local checkout and current HEAD. Do not create a worktree unless the user explicitly requests one.

Build the review handoff

Before creating the nested session, collect the review instructions already available in the main conversation. The nested session automatically receives applicable repository Copilot instruction files. Determine whether the AI IDE where this skill is installed exposes a code-review skill. If it does, the nested reviewer must invoke that skill. If it does not, the nested reviewer must run the IDE's /review command instead. Do not repeat stable repository guidance or generic review rules already supplied by those sources.

The kickoff must remain standalone and use this compact structure:

Scope:
Intended behavior:
Known risks and exclusions:
Output and permissions:

Include:

• repository name and absolute path only when the nested workspace cannot identify them reliably;
• current branch plus the exact comparison base, pull request, commit range, or working-tree scope;
• task-specific intended behavior and conversation-only decisions needed to distinguish defects from deliberate behavior;
• known risks, exclusions, and validation limits that materially affect the review;
• whether findings go to the main session, staged inline comments, GitHub, or a combination; and
• any user-authorized exception to the default read-only behavior.

Do not separately repeat the same requirement as context, priority, risk, and definition of done. When available, the invoked code-review skill governs high-confidence findings, severity ordering, exact changed-line references, and implementation-ready guidance. Otherwise, rely on the IDE's /review command for its standard review behavior.

Aim for roughly 250-500 words. Exceed that range only for genuinely broad multi-subsystem reviews or substantial conversation-only requirements. Long context is capacity, not a prompt-length target.

Preserve the user's additional context and instructions accurately. Do not replace specific requirements with a generic summary. If the comparison target or review scope is genuinely absent and cannot be inferred from the current session, ask one focused question before creating the reviewer.

Create the reviewer

Use create_session with:

name: Code review - <branch-or-pr>
execution_location: local
workspace_type: branch
coordinate_with_creator: true
notify_on_idle: once
kickoff.model: gpt-5.6-sol
kickoff.reasoning_effort: medium
kickoff.context_tier: long_context
kickoff.mode: autopilot

When the code-review skill is available, the first sentence of the kickoff must be:

Invoke the code-review skill immediately as your first action and follow it as the source of truth for the entire review.

When no code-review skill is available, the first sentence of the kickoff must instead be:

Run the /review command immediately as your first action and use its review workflow as the source of truth for the entire review.

Do not invoke code-review or run /review in the main session as a substitute for requiring the nested reviewer to use the available review workflow.

Completion

Wait for the nested session's completion message. An idle notification alone is not proof that the review completed. If the reviewer reports a blocker caused by missing context, send the missing information to the same nested session rather than creating a replacement.

Return the review result to the user without weakening severity, dropping file or line references, or omitting suggested changes. State the nested session name or ID only when useful for traceability.