English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

GitHub Copilot Links Stacked Sessions with Stacked Pull Requests: The Next Default Agent Coding Workflow

Forum topic · 小凯 · 2026-07-31

Summary

On July 30, 2026, GitHub introduced Stacked Sessions and Stacked Pull Requests in the Copilot App, chaining both AI conversations and pull requests into ordered sequences that land on the main branch. Previously these were separate: developers manually merged context across sessions and manually maintained base branches for PR chains. Now a stacked session automatically produces a corresponding stacked PR. GitHub's developer relations lead Cassidy Williams demonstrated the workflow by modernizing a 2014 personal project built with React 15, Less, and react-bootstrap: planning with Claude Opus 4.8, review via GPT-5.5 Rubber Duck, branch migration, and a cleanly separated react-bootstrap replacement in its own stacked PR—avoiding the scope creep of one giant PR. The design addresses three pain points of AI coding: avoiding massive PRs, preserving session context across chained tasks, and auto-syncing PR base branches. Stacked sessions support mixed models and are available to all Copilot plans, though developers must explicitly prompt for separate PRs. The article argues this shifts Copilot from an editor plugin toward an AI-native PR collaboration platform, and that the AI coding competition may move from longer-running tasks toward workflows that chunk long tasks for human review.

On July 30, 2026, GitHub announced Stacked Sessions and Stacked Pull Requests in the Copilot App via its official blog. This is not just multi-session support: GitHub has introduced a chained abstraction at both the session and PR levels—one session after another, one PR after another, until the work lands on the main branch.

What It Actually Is

Two concepts to clarify:

  • Stacked Sessions: within the same repository, you open a chain of tasks. Each session builds on the results of the previous one, with Plan mode providing a plan and each session being an independently executable unit.
  • Stacked PRs: a chain of pull requests, where each PR targets the branch of the previous PR, forming an ordered chain that ultimately lands on main.
  • Previously these were separate: running a long chain of Copilot tasks in the IDE meant manually merging context, and building PR chains meant manually maintaining base branches. GitHub has now connected them—stacked sessions automatically produce corresponding stacked PRs.

    Cassidy Williams' Real-World Retrofit

    GitHub's head of developer relations, Cassidy Williams, wrote a first-hand blog post about using these features to modernize a personal project from 2014. The project originally used React 15, Less, and react-bootstrap. She had previously attempted manual modernization but gave up because it was "too big and messy to extract value from." This time she threw the repo into the Copilot App:

    1. Step 1: In Plan mode she gave Claude Opus 4.8 a prompt covering a full Less teardown, a possible Tailwind migration, plus accessibility and responsive cleanup. The model produced a plan, then GPT-5.5 did a Rubber Duck review, with a few back-and-forth rounds to finalize it. 2. Step 2: When actually running the work, the workflow stalled because she had accidentally branched from the old dev branch instead of main. She asked Copilot to switch branches and redo it—none of the planning was wasted, as the system migrated her styling decisions intact to the dev branch. 3. Step 3: Midway through, she discovered that react-bootstrap itself references deprecated findDOMNode and componentWillReceiveProps. In Plan mode she asked whether to strip out react-bootstrap or upgrade it, and the model recommended a full library replacement. 4. Step 4 (the key step): She did not want the react-bootstrap replacement mixed into the same big PR—she explicitly called out this anti-pattern as "scope creep." With a single prompt, she had Copilot:

  • submit the existing work as a PR;
  • open a new stacked session on top of that PR, dedicated to the react-bootstrap replacement, inheriting prior context;
  • automatically generate a corresponding stacked PR targeting the original PR's branch.
  • The screen ultimately showed three stacked nodes at once: the first failed PR, the successful PR, and a draft PR for the react-bootstrap replacement. She wrote in the blog: "THIS WAS SO COOL. Stacked sessions *and* stacked pull requests? Is this the future? YES."

    Why This Design Matters

    From a workflow perspective, stacked sessions + stacked PRs solve three things at once:

  • Avoiding giant PRs: developers resist reviewing single PRs of thousands of lines, and the AI coding era makes such PRs especially easy to generate. Stacking forces PRs to be sliced along natural boundaries.
  • Preserving session context: each session used to be an island, requiring re-explanations of "what I want to do." Stacked sessions inherit context automatically, so new sessions need no re-setup.
  • Automatic PR chain syncing: base-branch relationships no longer need manual maintenance—merging one PR automatically advances the next PR's base.
  • In the competitive landscape of AI coding tools in the second half of 2026, this moves Copilot from "editor plugin" toward "PR collaboration platform." Cursor, Codex CLI, and Claude Code are still circling the "single session" level, while GitHub has added an AI-native abstraction on top of the PR—a collaboration unit developers already know.

    A Few Details

  • Coverage: the blog post targets all Copilot plan users, but stacked PRs are an existing GitHub platform feature; stacked sessions in the Copilot App are the new addition.
  • Model support: Cassidy used Claude Opus 4.8 with GPT-5.5 Rubber Duck review, showing stacked sessions are not tied to a single model and can be mixed.
  • Entry point: in the Copilot App main interface, once Plan mode produces a plan, the system automatically creates session nodes.
  • Risk: developers must explicitly ask for "later work in a separate PR"—otherwise the model tends to cram all changes into one PR. There is no default behavior; prompt engineering is required.
  • Practical Assessment

    Cassidy's case is a small personal project, but the two anti-patterns she highlighted—"giant PRs" and "session context loss"—are pitfalls every AI coding team hits. Stacked sessions + stacked PRs are not the only solution (there are tools building PR chains with git worktree, such as Graphite and Spruce), but GitHub's advantage is building it into the official repository abstraction with no third-party integration needed.

    My judgment: this workflow is a significant paradigm shift for the "AI coding + long tasks" scenario. The next thing to watch is whether Cursor, Codex, and Claude Code follow suit—if everyone adds stacking abstractions within six months, it means the AI coding tool competition in 2026 H2 will shift from "which model can run longer tasks" to "how workflows break long tasks into human-reviewable pieces."

    References:

  • GitHub official blog: https://github.blog/ai-and-ml/github-copilot/stacked-sessions-and-pull-requests-in-the-github-copilot-app
  • Copilot App: https://github.com/features/ai/github-app
  • Stacked PR documentation: https://gh.io/stacks

Tags

#github-copilot#stacked-pull-requests#stacked-sessions#ai-coding#developer-workflow#pull-requests#copilot-app#agent-coding

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178503831