Coloph Migrations

coloph-migrate is a PostgreSQL migration CLI built for coding agents. It pairs goal-driven project skills with numbered SQL migrations, checksum checks, schema reconstruction, and deployed-code compatibility tests. Agents use the CLI to inspect, change, and diagnose schemas instead of writing migration machinery. The CLI is the public interface; Python modules are internal.

Quick start

Add it to the repository's development dependencies (and commit the updated pyproject.toml and lockfile):

uv add --dev coloph-migrations

The lockfile records the exact package version. To run a fixed version without adding a dependency, replace VERSION in this command:

uvx --from 'coloph-migrations==VERSION' coloph-migrate --help

Initialize the migration files at the repository root:

uv run coloph-migrate init

This command creates these files if they do not exist:

coloph-migrations.toml
migrations/0001_init.sql
.env
skills/change-database-schema/SKILL.md
skills/repair-database-schema/SKILL.md

init installs the bundled skills in the host project's ./skills directory. Configure your agent to discover that directory or load these files through the project's agent instructions. File installation alone does not guarantee automatic discovery in every agent host. Existing identical skills are left alone. If a skill differs, init fails before writing files. Reconcile or move that file before retrying, including when updating skills from a new release. Existing application-specific skills, such as db-schema, remain separate.

The skills cover schema changes and failure diagnosis. CLI and configuration details belong in this README and command help, not in skill descriptions.

The generated configuration contains:

migrations_dir = "migrations"
schema_snapshot = "migrations/schema.sql"

Set DATABASE_URL in .env. Process environment variables override .env. The COLOPH_MIGRATIONS_DATABASE_URL variable remains available as a higher-priority override.

Edit the initial migration, inspect it, and apply it:

printf 'CREATE TABLE account (id bigint PRIMARY KEY);\n' > migrations/0001_init.sql
uv run coloph-migrate plan
uv run coloph-migrate apply
uv run coloph-migrate snapshot

Create each later migration with a normalized name. The command verifies the existing sequence and writes the next padded number:

uv run coloph-migrate new add_accounts
# creates migrations/0002_add_accounts.sql

Use a project-specific starter file with new add_accounts --template path/to/template.sql.

Use an ignored coloph-migrations.local.toml for local credentials and overrides. COLOPH_MIGRATIONS_DATABASE_URL keeps the URL out of files and process arguments.

Common workflows

# Show applied and pending migrations
uv run coloph-migrate list

# Create the next numbered migration
uv run coloph-migrate new add_accounts

# Execute the selected migration chain in disposable PostgreSQL before applying it
uv run coloph-migrate dry-run
# Require that every migration is applied and its checksum still matches
uv run coloph-migrate check

# Verify migration history, the committed snapshot, and the current target
uv run coloph-migrate verify

# Repair diagnostic: compare the target with its recorded migration prefix
uv run coloph-migrate validate --match-applied

# Regenerate schema.sql from a disposable reconstruction
uv run coloph-migrate snapshot --fresh

# Check a new migration number against main and deployed Git refs
uv run coloph-migrate check-chain

# Run deployed code against the new schema before deployment
uv run coloph-migrate check-backwards

--json is a global option. Put it before the command:

uv run coloph-migrate --json list

Use apply --up-to 012 to apply versions before 012 (the boundary is exclusive). apply --reconstruction enables only the disposable-database policies configured for reconstruction.

Hooks

Configure hooks when migrations need setup or follow-up work that is not part of the migration SQL itself:

before_each_migration_sql = "migrations/before_each.sql"
after_each_migration_sql = "migrations/after_each.sql"
post_max_attempts = 5
post_statement_timeout_seconds = 30
post_lock_timeout_seconds = 10
reconstruction_after_hook_versions = ["0186"]

The pre-migration hook runs inside each migration transaction. If it fails, the migration rolls back. The post-migration hook runs after that migration commits, in a separate transaction. A failed post hook remains marked incomplete; the next apply retries incomplete hooks before it runs a new migration or reports success. Hook SQL must therefore support safe retries. post_max_attempts controls retries during one run, and post_statement_timeout_seconds controls the hook statement timeout.

Reconstruction runs post hooks at configured checkpoint versions and once after the selected schema is rebuilt. If a post-hook file is missing, restore it or configure the correct after_each_migration_sql path, then run apply again. If the hook SQL changed, restore the version that safely completes the pending work (or make the replacement safely retryable) before applying new migrations.

What it prevents

Problem	Example	Guardrail
Edited history	An applied migration is changed, renamed, or removed	plan and check reject the invalid history.
Bad ordering	A branch adds 007_*.sql while main already has 007_*.sql	check-chain detects collisions across refs.
Failed SQL	A migration's second statement fails	The runner rolls back its transaction; earlier migrations remain committed. See transaction-control limitations below.
Schema drift	A snapshot or target differs from executable migrations	verify reconstructs and compares all three schemas.
Transaction escape	A migration commits with END or another transaction command	dry-run compares PostgreSQL transaction IDs around every migration.
Unsafe checksum repair	Someone wants to accept modified applied SQL	repair-checksums requires schema equivalence first.
Incompatible deployed code	New schema breaks a tested query in deployed code	Configured check-backwards tests exercise deployed code against the final rebuilt schema.

Result boundaries and recovery

plan reports pending files and rejects checksum mismatches; it does not run migration SQL. list reports applied, pending, orphan, renamed, and checksum_mismatch. check rejects every non-applied status. Status commands can create the migration tracking table; verify requires the table to exist and does not create it.

Run verify before deployment. It fails unless the complete migration history is current and its rebuilt canonical schema matches both schema_snapshot and the target database. It reports each mismatch as a unified diff, ignores only -- schema-doc: snapshot comments, and changes neither the snapshot nor target. Use validate --match-applied only to diagnose a target at its recorded prefix. Neither command compares data contents.

repair-checksums --dry-run previews updates after comparing the target schema with the full reconstructed chain. Schema equality does not prove equivalent data transformations. Do not replace this check with direct tracking-table updates or treat ordinary pending migrations as checksum repairs.

The runner owns each migration transaction. Do not include transaction control in migration SQL. Before target changes, apply runs the selected chain in a disposable database and rejects migrations that change their transaction.

The before hook runs inside the migration transaction. The after hook runs after that migration and its record commit, in a separate transaction. If the after hook fails on the target, the migration remains applied and its hook is marked incomplete. A later apply retries incomplete hooks before new work.

check-backwards can return skipped; that is not a passed compatibility test. It detects added SQL files between the deployed revision and committed HEAD, then tests deployed code against a reconstruction from the current files. Use a committed, clean candidate for deployment checks. Tests cover the final schema, not each intermediate prefix or the production data set.

If migrations run while old code serves traffic, first deploy code that stops using an object. Drop or rename that object in a later deployment, after the compatible code is actually deployed. A commit on main alone is insufficient.

Configuration

migrations_dir = "migrations"
schema_snapshot = "migrations/schema.sql"
database_url = "postgresql://postgres:postgres@localhost:5432/app"
main_ref = "main"
deployed_ref = "deployed"
deployed_fetch_remote = "origin" # optional; refresh deployed_ref before backwards check

# Optional. The before file runs in the migration transaction. During normal
# apply, the after file runs in a separate transaction after each migration is
# recorded and committed. Failed after hooks are retried before later applies
# continue. During reconstruction, the after file runs at any configured
# checkpoint versions and once after the selected schema is fully rebuilt.
before_each_migration_sql = "migrations/before_each.sql"
after_each_migration_sql = "migrations/after_each.sql"

# Disposable-reconstruction options.
fresh_statement_timeout_seconds = 90
reconstruction_after_hook_versions = ["0186"]

# Optional. Fresh databases use local Docker when this environment variable is
# absent or set to "local-docker". A PostgreSQL URL selects a shared cluster;
# non-loopback URLs must use sslmode=verify-full. Loopback URLs can use the
# caller's SSL mode so an authenticated local TCP proxy remains transparent.
test_cluster_url_env = "TEST_POSTGRES_CLUSTER_DSN"

Explicit CLI flags override configuration files.

The local TOML file overrides the base TOML file. Database URL resolution uses this order: CLI flag, process COLOPH_MIGRATIONS_DATABASE_URL, process DATABASE_URL, the same two names in .env, then the merged TOML value. Relative TOML paths and .env resolve from the configuration directory. These details identify which database a command will use; credentials and cluster provisioning remain the host project's responsibility.

Fresh databases use Docker by default. A configured remote cluster must permit temporary database creation and deletion. Schema dumps still use Docker with a PostgreSQL client image, even when the temporary database is remote.

Configure backwards tests to accept the disposable database URL, for example:

backwards_setup_command = ["uv", "sync"]
backwards_test_command = ["uv", "run", "pytest"]
backwards_test_globs = ["tests/test_database_*.py"]
backwards_database_url_env = "TEST_DATABASE_URL"

The deployed test harness must use that environment variable. Test setup and test selection belong to the consuming project.

Snapshot files can contain -- schema-doc: comments immediately before the SQL statement they describe. Regeneration restores these notes when the following statement line still matches. They do not become live database metadata.

Command reference

coloph-migrate init
coloph-migrate new NAME [--template PATH]
coloph-migrate apply
coloph-migrate dry-run
coloph-migrate list
coloph-migrate plan
coloph-migrate check
coloph-migrate snapshot
coloph-migrate verify
coloph-migrate validate
coloph-migrate repair-checksums
coloph-migrate check-chain
coloph-migrate check-backwards

Put global --json before the command, for example coloph-migrate --json list.

apply --reconstruction applies the selected migration prefix, runs the configured after hook at explicit checkpoint versions, and then runs it once against the rebuilt schema. This keeps historical reconstructions from repeatedly validating every intermediate schema while preserving known migration-chain dependencies. Ordinary production apply remains fail-loud and keeps per-migration after hooks. It first performs the same migration chain as dry-run, so a failed disposable run leaves the target database unchanged.

Coloph dependency workflow

When Coloph needs a coloph-migrations behavior change, edit this package directly in its local checkout, test it here, commit and push the package change, then update Coloph's pinned Git dependency and lockfile to that exact commit. Do not patch installed site-packages or work around dependency behavior inside Coloph.

The test suite deliberately exercises broken numbering, explicit transaction control, failed migration rollback, pre/post-hook transaction boundaries, checksum drift, schema drift, and safe-versus-unsafe checksum repair.

Public writing

Do not publish links or issue references to private repositories. This rule applies to source files, documentation, issues, pull requests, comments, and release notes. Explain each problem with a self-contained example, the actual result, the expected result, and the practical impact. Separate proposed features from observed defects. Do not present missing tests alone as a defect.

License

GPL-3.0-only. The Coloph name and logo are not licensed for use as trademarks.