Repository development doc. It isolates the product-specific facts behind the portable parallel-review contract so a product change updates this file only, not the packaged policies.
- Portable contract:
../shared/policies/parallel-review.md - PR application + capability table:
../skills/github-pr-review/policies/parallel-review.md - User-facing usage guide:
features/parallel-review.md
This file is not packaged into either Skill archive.
The Skill detects one of: none, ordinary isolated sub-agents, an
experimental agent-team mechanism (usable only when its prerequisite is
already set), or native concurrent agents. Uncertain → none → sequential.
The Skill never mutates the user's shell, global settings, or runtime
configuration to obtain a capability.
Agent Teams (a parent coordinating parallel sub-agents) are experimental
and disabled by default. They are enabled via the environment / settings
value CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. The Skill detects whether
Agent Teams are already available; it does not set the variable. When Agent
Teams are unavailable it uses ordinary Task/sub-agent calls if the runtime
offers them, otherwise sequential review. A review is never failed because
Agent Teams are off.
Source: Claude Code docs — https://docs.anthropic.com/en/docs/claude-code/costs
Cursor supports subagents with isolated context windows that run in
parallel. Project subagents can be defined under .cursor/agents/; Cursor
also recognises .claude/agents/ and .codex/agents/ for compatibility and
applies its own precedence between them. The parent reviewer delegates
independent analysis dimensions concurrently when the capability is enabled;
otherwise sequential.
Source: Cursor docs — https://prod.cursor.com/docs/subagents
Codex supports multiple agents working concurrently, with isolated
worktrees available for parallel work. Codex realises the same logical
review plan via its native agent/task capability. Codex worktrees are an
execution-isolation concern and are not responsible for the GitHub PR
checkout lifecycle (see repository-checkout.md) — the two are separate.
Source: OpenAI — https://openai.com/index/introducing-the-codex-app/
No .cursor/agents/, .claude/agents/, or .codex/agents/ directory is
added in this phase. The worker contract is fully specified by the packaged
shared/policies/parallel-review.md; adding near-duplicate prompt files for
three runtimes is the duplication the contract is meant to avoid, and such
files would not travel with the packaged Skill to a consumer anyway. If a
future need is demonstrated, prefer a single set under .claude/agents/
(which Cursor also reads) of thin adapters that point at the shared policy,
not three copies of the review instructions.