Skip to content

[pstack] Shipping: independent cross-repo PRs and clean-install verification #1259

Open

Description

@gook-lab

Version: pstack 0.15.36, run in Claude Code via docs/harness/claude-code.md.

What happened

I used poteto-mode to land six PRs across four small repos. One repo held a new shared package, and three repos consumed it through a git-subdirectory dependency (github:<org>/<repo>#<tag>&path:packages/<pkg>). The PRs depended on each other across repos, but none of them formed a branch stack.

playbooks/shipping.md covers only a stack in one repo. Its walk from the lowest unmerged PR, its patch-id re-check and its retarget-on-merge step have nothing to act on when the dependency runs between repos. I worked out the order by hand: package first, then a tag, then the consumers, then rebases.

One consumer PR passed every local check, including an independent verifier subagent. It failed on its first CI run, with 270 test files reporting Cannot find package. The cause was in the lockfile. pnpm had written the git resolution without path:, so a cold store installed the whole repo root in place of the subpackage. Every local run used a warm pnpm store, which already had the right tarball, and that hid the error. The verifier never ran in a clean environment, so it never saw the failure.

Proposal

  1. Shipping, step 1. When a PR changes a manifest or lockfile, the verifier installs from a fresh clone with an empty package store, for example pnpm install --frozen-lockfile --store-dir <empty>. A verdict from a warm cache does not count for that PR.
  2. Shipping, new short section on cross-repo order. For PRs that depend on each other across repos, land them in dependency order: producer, then its version or tag, then consumers. A consumer's verdict counts only if it ran against the producer's landed tag, not a local link. Read the first CI result on each consumer before landing the next one.
  3. Claude Code adapter. "One verifier per PR, each a Cursor cloud agent" has no direct equivalent. Spell out the substitute: a fresh isolation: "worktree" subagent that did not write the code. Note that a worktree shares the host's package store, so for manifest changes it is not a clean environment unless the brief passes an empty --store-dir.

Why it matters

The playbook's rule "green CI is not a verdict" is right. In this run, though, CI was the only check that ran in a clean environment. A verifier that reuses the author's cache misses exactly the class of failure that only shows up after merge.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions