Nx Multi Repo Migrate

Nx Multi-Repo Migrate

Migrate a set of repos to one target nx version, then open linked draft PRs. Think of it like a pharmacist filling the same prescription for several patients: same drug (target version), but each patient (repo) has different allergies (package manager quirks) — get those wrong and the dose silently fails.

Input

• Target version — e.g. 23.0.0-beta.25. Verify it exists: npm view nx@<version> version.
• Repos — an explicit list, or the repos already in a Polygraph session. When none is given, the default set is nx, ocean, nx-labs, nx-examples, nx-console (all in the nrwl org).

Procedure

1. Set up the session

Use the polygraph skill to discover repos, select the org, and start (or join) the session. It owns auth and session lifecycle — don't reimplement any of that here.

2. Delegate the migration to a child agent per repo

This is the Polygraph way: each repo's work runs in its own child agent (spawn_agent), not in the parent. Delegate to every repo in the session — in parallel — and poll with show_agent until each is terminal. Hand each child the migration instruction below (substitute the target version).

Migrate this repository to nx <VERSION>.

1. Branch from the current default branch, not the clone's checkout. Fetch first so you don't inherit a stale clone or an in-place working-dir branch, then create the branch from origin/<base> (master or main): git fetch origin <base> && git checkout -B migrate-nx-<VERSION> origin/<base>.
2. Detect the package manager from the lockfile (package-lock.json=npm, yarn.lock=Yarn Berry, pnpm-lock.yaml=pnpm, bun.lock/bun.lockb=bun).
3. Install first, so node_modules is at the repo's current (pre-migrate) nx version. nx migrate reads the "from" version from node_modules, not package.json — if node_modules is already at the target, it finds zero migrations and silently skips them. Verify with node -p "require('./node_modules/nx/package.json').version".
4. Run nx migrate <VERSION> (updates package.json, writes migrations.json).
5. Install again — mutable. Do NOT set CI=true (it makes Yarn Berry immutable / pnpm frozen, so the install and migrations fail silently). pnpm needs --config.confirm-modules-purge=false; Yarn Berry needs YARN_ENABLE_IMMUTABLE_INSTALLS=false.
6. Commit the version bump first (before running migrations, so it stays isolated from the migration edits): stage package.json + the lockfile — NOT migrations.json — and commit chore(repo): migrate to nx <VERSION> (never mention AI/Claude).
7. Run migrations with the orchestrated loop when the target has it (nx ≥ 23.2.0-beta.8). First git status --porcelain and delete any untracked junk (agent-litter dotfiles like .bashrc/.zshrc/.claude/* from children using the repo dir as $HOME) — the loop's checkpoint commit does git add -A and will sweep them into a commit. Then run NX_MIGRATE_ORCHESTRATOR=true <pm> nx migrate --run-migrations --create-commits --commit-prefix="chore(repo): [nx migration] " (--create-commits is required: a custom --commit-prefix errors without it — commits only default on when no custom prefix is passed; keep the mutable-install env vars from step 5 on this and every loop command — the loop runs installs itself). This turns migrate into a durable loop that hands you one step at a time:
   • Each <nx_migrate_step> block carries a command (nx migrate --run-migration=<id> --run-id=<id>) and a next reconcile command (nx migrate --run-id=<id>). Run the command verbatim, then run next to record the outcome and get the following step. Repeat until a block with action="complete".
   • Prompt-based (AI) migrations: the worker prints an <nx_migrate_prompt> block. Apply the prompt to the workspace yourself (same "passing baseline" rules as step 8), then write the handoff file at the path the dispense names — JSON { "status": "success", "summary": "<what you did>" } ("status": "failed" + summary if you couldn't; success + "outcome": "skipped" if N/A) — then run next. No tools/ai-migrations/ pile is left behind in this mode.
   • Failed/died steps: the dispense lists the valid --step-action options (retry, retry-clean, skip; adopt for a died worker whose changes actually landed). Prefer retry-clean when offered; skip only for a genuinely inapplicable step.
   • Commits are created by nx, one per migration, and default ON in this mode — do not also hand-commit. The message goes to git via stdin, so the classic (-in---commit-prefix shell crash (see step 8) does not apply here. If nx refuses because .nx/migrate-runs isn't gitignored, add that entry to .gitignore and re-run.
   • Crash/timeout recovery: run state lives in .nx/migrate-runs/<run-id>. Re-running nx migrate --run-migrations (same env) resumes the same run instead of starting over — use this instead of re-migrating after a killed child.
   • If no <nx_migrate_step> block appears (target older than 23.2.0-beta.8, or the agent-detection gate — CLAUDECODE in the env — didn't trip and the classic loop ran), fall back to step 8.
8. Classic fallback — do NOT use --create-commits. nx shells its --commit-prefix="chore(repo): [nx migration] " through /bin/sh unescaped, and the ( crashes it (Syntax error: "(" unexpected), which silently drops migrations. Instead run one nx migrate --run-migrations pass (apply the whole list, not a subset), then commit each migration's edits by hand, e.g. chore(repo): [nx migration] <name> (git commit -m handles the parens fine). Then apply the AI migrations yourself — you are the agent nx defers them to. --run-migrations applies the deterministic codemods (importantly remove-removed-typescript-eslint-extension-rules, which strips typescript-eslint v8-removed rules like @typescript-eslint/no-extra-semi; leaving one in a flat config crashes ESLint's loader → nx "Failed to process project graph" → red CI) AND writes prompt-only migrations to tools/ai-migrations/**/*.md, printing "Next steps for the AI agent driving this run: apply the deferred prompts." That is addressed to you (the child) — read each prompt and make the described changes; do NOT leave them for a human. Honor each prompt's "passing baseline": keep lint/typecheck passing, never disable a rule the user explicitly configured, and disable a newly preset-enabled rule with a short comment rather than editing source to satisfy it. (nx auto-skipping its nested agentic flow inside an agent is the review skipping — NOT permission to skip the migrations.)
9. Verify before declaring done: nx run-many -t lint --skip-nx-cache must resolve the project graph and pass (the removed-rule crash only shows at graph-processing time), plus typecheck/build affected projects where feasible. Fix migration-introduced breaks; surface genuine framework-major incompatibilities (Angular/React/TS majors) for a human rather than hacking around them.
10. Delete tools/ai-migrations/ (classic path only — the orchestrated loop doesn't create it) and migrations.json; leave .nx/migrate-runs/ alone (gitignored scratch, kept for post-mortems). If migrations changed deps, re-install and commit the lockfile update.
11. Report: old→new version, packages bumped, which mode ran (orchestrated vs classic), deterministic migrations run (+ commits), each AI prompt and how you applied it (or why N/A, per handoff), any --step-action resolutions used, final lint/typecheck/build status, and any unresolved failures — type/name collisions, framework-major breaks. Leave true blockers for a human; do not invent workarounds.

Completing a partial / already-at-target run. If node_modules is already at the target, nx migrate <VERSION> finds zero migrations. To (re)apply migrations that a prior run skipped — the deterministic remove-removed-* codemod or the AI prompts — regenerate the full list with an explicit --from: nx migrate <VERSION> --from=nx@<original-version>. Migrations detect already-applied state and no-op, so this safely re-runs only what's missing, then finish with steps 7–11 above.

Package-manager cheat sheet:

Migrations can rewrite source: a multi-beta jump (e.g. beta.23→beta.25) pulls migrations from every intervening version, so it may rewrite real code (e.g. CreateNodesContextV2→CreateNodesContext). The child should review the non-dep diff before committing. A single-beta jump on an already-current repo often legitimately has none.

3. Push + open a PR per repo, as each child finishes

Don't barrier on the slowest repo. The moment a child reports success, push_branch that repo (branch migrate-nx-<VERSION>) and create_pr for that repo alone — so its CI starts immediately and one slow repo (e.g. one stuck fighting the sandbox) doesn't gate the others:

for each repo, as its child reaches terminal success (not in a barrier):
  push_branch(repo) → create_pr([repo])

The PRs stay linked because they all join the same Polygraph session — the link is the session, not the single batched call. Commit-message scope repo passes nx's commitlint. Print the Polygraph session URL once all are open.

Verify once: a single batched create_pr writes every PR body with its sibling cross-references at creation time; with incremental creation, confirm Polygraph back-fills the earlier PRs' bodies with links to the later ones (vs. each PR only linking to the session). If it doesn't back-fill and you need the in-body cross-links, fall back to one batched create_pr after all children finish.

Verification checklist (per repo, before opening PRs)

• package.json nx + @nx/* at the exact target version (not silently downgraded to latest by an age gate)
• Migrations ran (not skipped because node_modules was already at target), including the deterministic remove-removed-* codemods
• AI-migration prompts applied by the child (not just written): orchestrated runs have a success handoff per prompt step; classic runs have tools/ai-migrations/ applied then deleted. migrations.json deleted either way
• nx run-many -t lint --skip-nx-cache resolves the project graph and passes; typecheck/build checked where feasible
• Version-bump commit (chore(repo): migrate to nx <VERSION>) plus one chore(repo): [nx migration] … commit per applied migration/prompt on migrate-nx-<VERSION> (nx-authored in orchestrated mode, hand-made in classic — never both)
• Orchestrated runs ended with the complete block — no step left failed/died/awaiting-prompt-outcome, and no unresolved commit-debt / failed-install warnings in it
• Child reports verified against actual repo state (git log, git status, git show --stat on loop commits) — delegates can report steps as done that never ran, and a checkpoint commit can contain swept-in junk; a checkpoint whose only contents are junk is safe to drop (git reset --hard to the commit before it) when the run's migrations made no file changes
• Any collision / compile / framework-major errors surfaced in the child's report for a human to resolve

Gotchas from real runs

These each cost real time on a live 5-repo run. Plan for them up front.

Fresh betas/canaries are hidden by release-age gates → silent downgrade to latest. A <24h-old target is filtered out by supply-chain age gates in up to three places on an nx-dev box: ~/.npmrc min-release-age=1 (npm/bun), ~/.config/pnpm/rc minimum-release-age=1440 (pnpm), and a ~/.yarnrc.yml registry pointed at a local age-gating proxy (http://localhost:7190) that is often down (→ ECONNREFUSED). When the target is filtered, nx migrate does not error — it silently resolves the whole @nx/* group to the newest visible version (e.g. latest 23.0.1 instead of 23.1.0-beta.5), so the repo "migrates" to the wrong version. Bypass per-command (do NOT edit global config): npm_config_min_release_age=0 npm_config_minimum_release_age=0 (npm/pnpm/bun), plus for Yarn Berry YARN_NPM_REGISTRY_SERVER=https://registry.npmjs.org/ YARN_NPM_MINIMAL_AGE_GATE=0. pnpm's nx-migrate temp-dir pnpm add also needs PNPM_CONFIG_STRICT_DEP_BUILDS=false (else ERR_PNPM_IGNORED_BUILDS aborts it). Always verify each repo landed on the exact target version, not latest. (Note: pnpm ignores the npm-style min-release-age key but honors its own minimum-release-age; that's why a pnpm repo may resolve the beta while a yarn/npm sibling silently downgrades.)

pnpm dies under the Bash sandbox; bun/yarn don't. As of Claude Code 2.1.172 the Bash tool sandboxes by default. pnpm's content-addressed store + clonefile() reflink + node_modules purge trip macOS rules — com.apple.provenance xattr removal, creating .vscode/.idea dirs in the virtual store — plus outbound TLS, so pnpm install fails with ERR_PNPM_EPERM / reflink / Operation not permitted, while bun and yarn install cleanly. Polygraph children carry their own sandbox (~/.polygraph/config.json → agentOptions.claude.sandbox), separate from ~/.claude/settings.json → sandbox.enabled; either one only reaches already-spawned processes after a restart. If a pnpm child stops on a sandbox/EPERM error, do not let it invent workarounds (xattr stripping, TLS shims, store redirection). Instead, disable the sandbox + restart, or migrate that repo from the unsandboxed parent: the initiator repo is in-place, and clones live at ~/.polygraph/sessions/<id>/repos/<org>/<repo> — run the same install→migrate→install steps there with the sandbox off, then push.

The base can move after you start. Step 1 (branch from origin/<base>) handles the initial state, but the default branch can still advance mid-run — e.g. a separate version-bump PR merges underneath you, as happened when ocean's main jumped beta.23→beta.25 below an open migrate PR and turned it conflicting. Detect it with the behind-count (git rev-list --count migrate-nx-<V>..origin/<base>) and watch for open bump PRs; when the base moves, redo the branch onto the fresh base — only the repos whose base actually advanced need it. Redoing onto a newer base can also shrink the diff: a beta.25→rc.0 redo is dep-only, whereas the old beta.23→rc.0 ran 16 migrations and rewrote source.

The initiator repo runs in-place in your working dir, so migrating it switches branches and churns node_modules. Restore it afterward — or run its migration in a throwaway worktree off the real base (git worktree add -B migrate-nx-<V> /tmp/wt origin/<base>) so the working copy is never touched. But a fresh full install in the worktree duplicates the huge node_modules and can ERR_PNPM_ENOSPC (inode/disk pressure on top of the other clones' installs). Avoid it: run the nx migrate planning step in the main checkout (reuse its already-installed node_modules so migrate can bump the whole @nx/* group — without node_modules it only bumps nx itself), copy package.json+migrations.json onto the worktree branch, restore the main checkout; when there are no migrations to run, just pnpm install --lockfile-only in the worktree instead of a full install. Clean up the worktree with git worktree remove after pushing (the branch ref persists).

A concrete source collision. The CreateNodesContextV2→CreateNodesContext rename migration collided with a vendored local interface CreateNodesContext extends CreateNodesContextV2, producing a self-referential extends CreateNodesContext (TS2310). Surface it for a human; the minimal fix is aliasing the import: import { CreateNodesContext as NxCreateNodesContext } from '@nx/devkit'. (That rewrite is a beta.24 migration — starting from beta.25 skips it entirely.)

Push/auth pitfalls. (1) The SSH agent can drop mid-run (communication with agent failed) — SSH git push then fails; retry, or have the user re-ssh-add. (2) A read-only GH_TOKEN env var can shadow a write-capable keychain login: every write (push, pr edit, pr merge --auto) returns Resource not accessible by personal access token. Prefix gh writes with env -u GH_TOKEN to fall back to keychain auth. (3) Polygraph push_branch does an internal pull --rebase, so it cannot force-update a rebased branch — use a direct git push --force (SSH/HTTPS) for those. (4) Polygraph create_pr intermittently 401s (Bad credentials) on nrwl/nx specifically while succeeding on sibling nrwl repos in the same batch — just retry the failed repo; it usually goes through on the 2nd–3rd attempt. (5) The personal GH_TOKEN can push to nrwl/nx but is denied (403) on some other nrwl repos (e.g. nrwl/nx-examples) and cannot create PRs on nrwl/nx — so for those, use Polygraph push_branch/create_pr (backend auth), and since push_branch is fast-forward-only, prefer adding a new commit over amending when you need to update an already-pushed branch. nrwl/nx PR creation may still need the pushed-branch + pre-filled compare-URL fallback if create_pr keeps failing.