prisma-next-migration-review

Prisma Next — Migration Review (Deployment + Concurrency)

Edit your data contract. Prisma handles the rest.

This skill is about reviewing migrations, not authoring them. It covers the questions that come up at deploy time and when multiple developers are landing migrations concurrently.

The skill teaches the system's mental model — what a ref is, what a marker is, what the migration graph is — and shows how to ask the system for its state. It does not prescribe rigid step-by-step procedures: most "review" questions are answered by understanding the model and querying the right thing. Rigid procedures are reserved for the rare case where there's literally one safe path.

When to Use

• User asks "what migrations will run when I merge this?" or "what's about to run on deploy?".
• User hit a concurrent-migration conflict (main advanced while their branch was open).
• User wants to wire up a staging / production ref so CI can deploy against it.
• User wants to run a migration against an environment that isn't the local dev DB.
• User asks about CI integration for migrations.

When Not to Use

• User wants to author a migration → prisma-next-migrations.
• User wants to fix a hash-mismatch / drift in a single env → prisma-next-migrations (re-plan path) or prisma-next-debug (envelope-driven).
• User wants to edit the contract → prisma-next-contract.

Key Concepts — the navigation model

Every migration question is a navigation from an origin to a destination. Once you have this model, the rest of the skill is just "which command asks the system about which navigation."

Origin

The origin is the database's current contract hash. The database carries a row in PN's marker table that records "this database is at hash X". When the CLI runs online (a --db <url> is provided, or db.connection is set in prisma-next.config.ts), PN reads the marker and that hash is the origin. Offline (no DB connection), the origin is unknown — many commands degrade to listing the on-disk migrations and skip the per-edge applied/pending status.

A live DB is therefore the authoritative source of origin. The "recorded marker" in any other artifact (refs, local cache, your assumptions) is a working copy that can drift; the live DB never does.

Destination

The destination is the contract hash you want the database to be at. Two ways to name a destination:

• A --to <name> — a named pointer to a hash, stored under migrations/app/refs/<name>. Refs are named after environments by convention (staging, production) to communicate "this is where production is expected to be". The ref itself is just a hash + an optional set of required invariants; it has nothing to do with which database you connect to.
• The current contract head — implicit when no --to is passed. This is the hash of the current contract.json on disk.

--to staging does not mean "connect to the staging database." It means "navigate the database I connected to (via --db or config) toward whatever hash this ref points at." Database selection is orthogonal: pass --db $STAGING_DATABASE_URL to actually point at staging.

The migration graph

The on-disk migrations form a directed graph: nodes are contract hashes; edges are migrations. Each migration declares a from hash and a to hash. A migration applies only when the database's current marker matches its from hash; running it advances the marker to its to hash.

migration status queries the graph for the path from origin to destination and reports per-edge status:

• applied — on the path from EMPTY_CONTRACT_HASH to the marker (history).
• pending — on the path from the marker to the destination (what would run).
• unreachable — on the path from EMPTY_CONTRACT_HASH to the destination, but the marker is on a different branch and won't reach it without first re-routing.

Diagnostic codes

migration status emits structured diagnostics on the result envelope (diagnostics[].code) so the agent can branch on the code rather than parsing the prose summary. Each diagnostic also carries severity (warn or info), a human message, and hints — the same hints the CLI prints under the summary line.

Code	Severity	Meaning in the navigation model	Next move
MIGRATION.UP_TO_DATE	info	Marker = destination; no edges to walk.	Nothing to do.
MIGRATION.DATABASE_BEHIND	info	Marker is an ancestor of the destination; N pending edges in between.	migrate --to <name> --db $URL.
MIGRATION.MISSING_INVARIANTS	info	Marker reached destination structurally but missing required invariants the ref declares.	migrate --to <name> --db $URL to take a path that covers them.
MIGRATION.NO_MARKER	warn	Online, but the database has no marker row — never initialised.	migrate --db $URL (first apply writes the marker).
MIGRATION.MARKER_NOT_IN_HISTORY	warn	Online; marker hash is not a node in the graph. The database was changed outside the migration system.	Decide which side is truth: db sign (accept DB as truth), db update (push contract to DB), contract infer (re-derive contract from DB), or db verify (inspect first). Not the same as MIGRATION.MARKER_MISMATCH: MARKER_NOT_IN_HISTORY is emitted during the runner's graph walk when the live marker is off the path being traversed; MARKER_MISMATCH fires earlier, at the CLI pre-DDL gate, when the marker hash is not a graph node at all.
MIGRATION.DIVERGED	warn	Multiple valid leaves; the destination is ambiguous.	Pass --to <name>, or ref set <name> <hash> to create one.
CONTRACT.AHEAD	warn	Contract head is not in the graph — the contract was edited without re-planning.	migration plan to extend the graph.
CONTRACT.UNREADABLE	warn	contract.json couldn't be read.	contract emit to regenerate it.

Graph-tree output

migration status (and migration list) render the migration graph as a colored lane tree in the terminal. Two flags control the rendering:

• --legend — prints the key for the tree glyphs and lane colors before the tree.
• --ascii — replaces box-drawing glyphs with pipe-safe ASCII characters (useful in CI logs or environments that don't support Unicode).

Both flags are also available on migration list and migration graph. migration log supports --ascii only (it renders a flat chronological table, not a tree).

Plan- and apply-time diagnostics

These codes surface on migration plan, ref set, and migrate — not on migration status. See Migration System § Recovery affordances and ADR 218.

Code	When	Meaning	Next move
MIGRATION.HASH_NOT_IN_GRAPH	migration plan (non-empty graph) or ref set	Resolved hash is not a node in the on-disk migration graph — typical when the default db ref points past the graph tip after dev-only db update cycles.	migration plan --from <reachable-ref> (e.g. --from production); or realign the ref with ref set db <graph-node-hash>.
MIGRATION.SNAPSHOT_MISSING	migration plan	A named ref has no pointer file (<name>.json), and the hash being resolved isn't a node in the migration graph either.	ref set <name> <hash> to create the ref, db update --advance-ref <name> to advance it, or pass a hash that is a graph node.
MIGRATION.MARKER_MISMATCH	migrate (pre-DDL, before the runner)	Live DB marker hash is not a graph node — drift the offline planner cannot see.	migration plan --from <graph-tip> if the marker is canonical; ref set db <marker-hash> if the on-disk graph is canonical; investigate out-of-band applies.
MIGRATION.PATH_UNREACHABLE	migrate (path resolution)	No migration path from the current marker to the resolved target in the on-disk graph.	Read the improved fix payload — it names fromHash / targetHash and suggests migration plan --from <from> --to <target>; run migration list to inspect the graph.

A CI gate should read diagnostics from --json output and decide based on severity plus code; see Workflow — CI below for the structure.

Workflow — "What's about to run on deploy?"

The user asks: "I'm about to merge this PR. What migrations are going to run when I deploy to staging?"

This is the navigation question: origin = staging's live marker; destination = the ref staging (or the contract head if you haven't set one). Ask the system:

pnpm prisma-next migration status --to staging --db "$STAGING_DATABASE_URL"

The command:

1. Reads the staging DB's marker (the origin).
2. Resolves staging to a contract hash (the destination).
3. Renders the path between them as an ordered list of migrations, with per-edge applied / pending / unreachable status, and an explicit summary line of the form "N migration(s) behind ref 'staging'".
4. Prints a header that names the config, migrations directory, the active ref, and the database connection (masked) — so the framing is visible in the output.

If you omit --db, the command runs offline: it lists the migrations on disk but cannot tell you what's applied, because it has no origin. That's fine for "what's on this branch?"; it's not fine for "what's about to run on staging?" — for that you need staging's live marker.

If you omit --to, the destination defaults to the contract head — which answers "is this branch's contract reachable from the database, and how?", not "what runs on deploy". Pass the ref explicitly when the question is about a specific environment.

migration status summarises each pending migration's operations by class (additive, widening, data, destructive) and reports a destructive-op count when destructive operations are present. Surface that count to the user before they merge or deploy — destructive operations are the class that warrants manual review.

Workflow — "What state is each environment at?"

Just migration status --db $URL for each environment's DB. The marker (origin) comes back from the DB itself; the summary line tells you whether the environment is at the contract head, at a named ref, ahead of head, or on a divergent branch.

Concept — concurrent migrations on the same branch point

This used to be called diamond convergence in some PN docs; the situation is the same regardless of the label.

What's happening. Two topic branches each authored a migration off the same parent contract hash. The first branch merges to main; the destination ref (e.g. production) advances to that branch's to hash. Your branch's migration still has its from hash pointing at the old parent. The migration graph, after rebase, no longer has a clean path through your migration:

• Your migration's from is no longer an ancestor of the new head.
• Or your migration's from is reachable, but the path through your migration arrives at a hash that's not the union of both branches' changes.

Either way, the on-disk plan is stale.

Resolution. The on-disk plan is stale because its from hash is no longer the head of the graph; apply the cluster's standard edit → plan → apply loop to the post-rebase state and the planner produces a fresh migration whose from matches the new head.

The one thing the planner can't do for you is port custom data-transform logic from the abandoned migration.ts into the new one — schema deltas are derived from the contract, but any hand-written data operations are yours to carry across before applying. There is no separate "revalidate" step, no special "diamond apply" flow.

Workflow — set, list, get, delete refs

Refs are small artifacts. There's no per-environment lifecycle; you just point a name at a hash.

pnpm prisma-next ref set production <contract-hash>
pnpm prisma-next ref list
# `ref get` was removed — use `ref list` and filter by name
pnpm prisma-next ref list | grep production
pnpm prisma-next ref delete production

ref set writes a file at migrations/app/refs/<name> carrying the hash and any required invariants. Refs are commit-friendly artifacts — keep them in git; the team agrees on what production points at the same way they agree on what main is.

Workflow — apply a migration against an environment

pnpm prisma-next migrate --to production --db "$PRODUCTION_DATABASE_URL"

The destination is the ref's hash; the origin is the production DB's live marker. The command computes the path between them and applies each pending migration in order, advancing the marker.

--db is the environment selection knob. --to is the destination-hash knob. They're independent.

Concept — ref-mismatch on CI / deploy

CI reports: "the recorded ref production is at hash X; the live DB is at hash Y."

The mismatch is a fact about two pieces of state that disagree. The investigation is the same regardless of which piece is wrong:

• DB ahead of the ref. Someone applied a migration outside CI without updating the ref in git. Re-record the ref with prisma-next ref set <ref-name> <db-marker-hash> (commit + push); then audit how the out-of-band apply happened.
• DB behind the ref. A previous deploy was rolled back, or the DB was restored from an older backup. Either re-apply forward with prisma-next migrate --to <ref-name> --db $URL, or re-route the ref backward to match what's actually deployed with prisma-next ref set <ref-name> <db-marker-hash>. The choice is the user's — name both options.
• DB on a different branch. An out-of-band schema change (manual SQL, ad-hoc migration) wrote something the migration graph doesn't model. Run prisma-next db verify to inspect the drift, then either prisma-next contract infer to re-derive the contract from the database, or edit the contract and run prisma-next migration plan so the database is the eventual destination.

ref set to silently align the ref with whatever the DB happens to be at is almost never the right move. It papers over drift that you'll pay for later.

Workflow — CI: verify a branch can advance the target environment

The gate is migration status --to <env> --db $URL: it computes the path from the live marker to the ref and reports it, without mutating anything. There is no --dry-run flag on migrate; the inspect / gate step is migration status.

For a human-readable ordered preview of the migration path before applying, use migrate --show --db $URL. For applied history after a deploy, use migration log --db $URL (flat chronological table).

- name: Verify staging is reachable
  run: |
    pnpm prisma-next mi

...