Imported from Pipelex/pipelex (.claude/skills/add-migration/SKILL.md). Install upstream with npx skills add Pipelex/pipelex --skill add-migration. Copyright stays with the author.

Add a Migration Entry

A configuration surface — pipelex.toml and its tiers, telemetry.toml, pipelex_service.toml, the inference backend definitions in inference/backends/ — ships a checked-in ledger at pipelex/migration/ledgers/<surface-id>.toml recording, as data, every shape change it has ever undergone. A user's existing file is repaired by replaying that ledger over it, so a schema change with no entry is a schema change that breaks every machine in the field.

docs/migration-ledger.md is the normative contract. Read it before writing an entry — this skill is the procedure, that document is the law, and where the two disagree the contract wins.

The one rule

Derive the entry from the gate's refusal, never from memory. The coverage check recomputes the surface's fingerprint and names every path, every enumerated spelling and every value domain the change moved. That output is the fingerprint diff, already rendered. Composing an entry by hand from your own reading of the diff is how a rename gets misspelled and every user's file is migrated onto a key the schema rejects, with the tool reporting success.

Step 1: See what the gate says

make check-migration-schemas (alias cmig) is deliberately not in make agent-check — it is a golden check, and a fail-regenerate-fail cycle in the loop agents run constantly is how a gate goes permanently green while catching nothing. So run both gates explicitly:

make cl      # check-ledger: is the ledger legal, and is replaying it harmless?
make cmig    # check-migration-schemas: is every schema change accounted for?

Each failure names the surface and a kind, and every message says what to do. The table below is not the whole list of kinds — it is the map of which ones mean an entry is owed, and what it owes:

A gate that only asks for a regeneration is not asking for an entry. An additive change — a new optional key, a widened type, a relaxed bound — costs a make umig and nothing else. Stop at step 5 in that case.

Step 2: Decide the safety

safety governs whether the applier may act, and it is not a matter of taste:

• safe — the operations mechanically complete the repair. The file is rewritten after one confirmation. Everything structural is safe: renames, moves, deletions, and a remap_value whose old spelling is genuinely no longer legal.
• unsafe — no operation in the vocabulary can repair the file, so the entry is reported and never applied. This is the form for a tightened numeric bound, a completeness rule a validator expresses, anything where the tool would have to choose a value on the user's behalf.

An unsafe entry may legitimately carry no operations at all — and then declared_narrowed_paths is mandatory, because the engine questions a document before reporting an entry, and an entry with neither operations nor a declaration answers "nothing to say" for every file there will ever be. The declaration is spelled as the fingerprint at the entry's own version records the paths, * segments included, and make cl refuses a path that version does not have.

guidance is independent of safety: any entry may carry it, and it is the explanation a person reads, never the mechanism. Anything expressible as operations must be operations.

Step 3: Write the entry

Two edits to pipelex/migration/ledgers/<surface-id>.toml, in one change:

1. In [surface], raise current_schema_version to N.
2. Append a [[migration]] block. Entries are contiguous and named for their version — the id must be <surface-id>@<N> and to_schema_version must be N, or the file is refused when it is parsed.

[[migration]]
id                = "example-config@2"
to_schema_version = 2
introduced_in     = "0.46.0"
breaking          = true
safety            = "safe"
title             = "Short imperative summary of the shape change"
description       = "One or two sentences a release note can quote."
guidance          = """
What a user should understand or decide. Never the mechanism.
"""

[[migration.ops]]
kind       = "rename_table_key"
table_path = ["reporting"]
key        = "output_config"
new_key    = "output"

introduced_in is orientation only — nothing branches on it. A breaking entry ships in a minor bump per the house convention, so write the next minor above the current pyproject.toml version. /release step 3b confirms it against the version actually being cut.

The operation vocabulary is structural only — rename_table_key, move_key, delete_key, delete_table, remap_value. There is deliberately no operation that writes a value into a user's file; one would be refused when the ledger is parsed. docs/migration-ledger.md → "The operation vocabulary" has each kind's fields, the move_key placement rule, and what the * wildcard segment means (and where it is legal, which is only at an open node — and as a remap_value key, which is the one shape that can repair a member renamed beneath an open mapping).

Operations target the narrowest path that expresses the change. A ledger entry is permanent data that outlives the current field census of its parent table, so deleting a parent because it happens to hold one field today is wrong — unless the parent itself is what retires, in which case its name becomes reserved too.

Step 4: When the reserved-path rule fires

make cl refuses a name an earlier version retired:

schema version 4 has 'reporting.output' again, but schema version 2 retired it — a retired path stays retired, because reusing the name makes removed material legal again. Pick another name

There is no escape hatch and there should not be one. Propose a replacement to the user, and propose a good one: prefer a name that reads naturally for what the field now means over a decorated variant of the retired one (output_target, not output_v2 or output_). The same rule applies to enumerated spellings — a value a remap_value moved away from can never come back.

Renaming the field means changing the model, so the fingerprint moves again: go back to step 1 rather than patching the entry.

Step 5: Regenerate the goldens

make umig

This writes fingerprint@N.json and defaults@N.toml for every surface, at the version its ledger now declares. Older versions are never rewritten — they are the frozen history the chain is made of — and the new version's files are new files, so nothing is erased.

Read the diff. It is the only signal this command produces: line by line, which paths the change added, removed or renamed. Running umig without reading its diff throws that away.

If umig refuses, it is the guard doing its job: you are about to overwrite a head golden that records material the models no longer have, which would erase exactly what the coverage gate exists to catch. Read what it names. Almost always the answer is that step 3 has not happened yet — bump the version and write the entry, and the refusal disappears because there is no stored golden at N to erase. make umigf is not the general escape: it is for one situation only, a golden that predates a change to the fingerprint format over a schema version that has not been released. Never hand-edit a golden.

Step 6: Prove it

make cl && make cmig

Both green means three separate claims hold: the entry is legal and replaying it over a healthy file changes nothing (cl), every path the change moved is accounted for, and — the transform goldens — the operations actually perform the migration from defaults@N-1 to defaults@N, and the migrated document is accepted by the current model.

Then the usual gates: make agent-check, and make agent-test before wrapping up.

Step 7: The changelog

The ledger and the changelog are deliberately separate artifacts saying the same thing to different readers, and /release step 3b is where they are checked against each other. For each entry with breaking = true, add a bullet to ## [Unreleased] beginning with the literal label:

- **Migration:** `telemetry-config@2` lifts the flat first-generation telemetry
  file into today's `[custom_posthog]` section, keeping every value the user
  chose. Four settings the current shape has no home for are dropped, and the
  entry says which and why. Run `pipelex migrate`.

House style: a bold label, then two to four complete sentences. Say what moved and what a user has to do; do not restate the operation list.

Special case: a change that predates the first fingerprint

An entry may carry pre_history = true when the change it describes happened before the surface had a fingerprint, so no diff accounts for it. Such an entry declares its own declared_removed_paths (an empty declaration is refused), none of those paths may appear in any fingerprint at or below its version, and it ships a hand-authored goldens/<surface-id>/before@N.toml so the transform check has a pair to verify. This is rare and it is not a way past the gate — reach for it only when there genuinely is no diff, and read docs/migration-ledger.md → "Pre-history entries" first.

Inserting a pre-history entry below entries that already exist

The change you are accounting for may be older than an entry already in the ledger — a key that moved before a later entry renamed the table it lived in. The new entry then has to run first, because its operations address the historical spelling, so it takes a version the ledger already uses and everything above it is renumbered. Ids, versions and current_schema_version are forced to agree, so this is one atomic edit, and the goldens move with it:

1. Insert the entry at @N and renumber every entry above it (@N → @N+1, and so on); bump current_schema_version to the new top.
2. git mv goldens/<surface-id>/fingerprint@N.json fingerprint@N+1.json, and the same for defaults@N.toml. Repeat upwards, highest first.
3. Copy fingerprint@N-1.json → fingerprint@N.json and defaults@N-1.toml → defaults@N.toml. A pre-history entry's fingerprint pair must show no diff at all, and a copy makes that true by construction rather than by luck.
4. Hand-edit one line per moved or copied fingerprint: the "schema_version" in its body. A golden whose body disagrees with its filename is refused when it is read, and make umig reads the stored head before it writes anything — so a fingerprint still saying N under the name @N+1 makes umig raise instead of regenerating. This is the one sanctioned exception to "never hand-edit a golden", and it is bounded to that single line.
5. Hand-author goldens/<surface-id>/before@N.toml for the new entry.
6. Run make umig. It must be a byte-pure no-op on everything the moves produced — verify with shasum -c over the directory, or by diffing each moved golden against its pre-move self in HEAD and finding only the version line. That no-op is the proof the renumber was pure; any other diff is a finding, not something to regenerate around.
7. Sweep the tree for the old id (git grep -n '<surface-id>@N' across pipelex/ tests/ docs/ CHANGELOG.md .claude/). Every hit is either the new entry or a stale quote of the renumbered one; there is no third kind. Where a test or a docstring named a golden by version but meant "the shape before the change", make it say so rather than bumping the number — the construct is what goes stale, not the value.
8. Both ids need a changelog mention: a renumber otherwise reads as a new breaking entry that nobody wrote a bullet for.

What this skill must never do

• Never hand-edit a fingerprint or defaults golden, with the single exception of the "schema_version" line during a renumber, above. They are generated; the generator is make umig and the diff is the review artifact.
• Never make a gate green by widening it. If a check refuses the change, either the change owes an entry or the check found a real defect. Editing the check is a separate decision with its own review.
• Never reuse a retired path or spelling to make the reserved-path rule quiet.
• Never invent a value. A remap_value mapping carries the new spelling of an old one — a translation the user's own file determines, which is why it can be safe. What no operation may do is write a value the file does not already fix: a default, or a replacement for a value a tightened bound now refuses. If the repair needs one of those, the entry is unsafe and a person makes the choice.