ConfigKit/docs/roadmap/sequences.md

# Sequence Identity And Combination

Status: feature roadmap with OD-1 through OD-7 settled. This document captures
the targeted behavior for review before preparing a staged implementation plan.
It is not an implementation plan or an implemented contract.

## Motivation

Allow configuration authors to assemble collections from independently authored
records, declare their identity in configuration, and progressively extend a
baseline collection through higher layers. Authors should also be able to
replace the accumulated collection or replace individual records by identity.

The motivating example is a prompt collection whose defaults supply some
records and whose application or installation layers add or supersede records.
Across tiers, authors need three operations: whole-sequence replacement,
duplicate-rejecting append, and whole-record replacement by identity. The latter
retains unmatched lower-tier records and adds higher-tier records with new
identities. Matching records are replaced in full. Sibling inheritance through
`extends` uses the same sequence operations and effective-identity defaults.

## Current Baseline

The [import contract](../config.md#explicit-imports) owns current selector,
destination, placement, and ordering behavior. An import with
`placement: sequence-item` places each selected source mapping as one item at
its effective destination. Item-local helper imports can contribute fields
before placement.

The application currently registers sequence identity through
`Definition.SequenceCompositions`, using one literal string field at each
registered mapping path. Within a strict tier, distinct identities append and
duplicate identities fail. Repeated unregistered sequences conflict. Across
precedence tiers, a higher sequence replaces the lower sequence wholesale.
The [composition contract](../config.md#strict-document-composition) owns these
rules; [layered loading](../config.md#layered-loading) owns tier precedence.

Current [sibling inheritance](../config.md#sibling-inheritance) recursively
overlays an inheriting mapping onto its resolved parent and replaces supplied
sequences wholesale. The proposed inheritance extension is settled in OD-7.

Source-backed underlays and `Definition.Defaults` have different lifecycle
roles. Underlays participate in source assembly; typed defaults are applied
later. Extending one does not automatically imply extending the other.

## Requested Capabilities

The proposed `configkit.sequences` block introduces a `key` declaration that
accepts either one field or a list of fields for a compound identity.
Fields are treated as a set; identity components may be strings, numbers,
booleans, or explicitly supplied null values.
Applications retain an optional interface for registering a collection
and imposing its identity definition. Configuration declarations must conform
to an application-owned constraint when one exists. The exact Go interface
remains to be designed.

```yaml
configkit:
  sequences:
    - destination: [prompts]
      key: [namespace, id]
      mode: append
  imports:
    - dir: prompts
      pattern: '*.yaml'
      destination: [prompts]
      placement: sequence-item
```

Each selected file would contribute a record such as:

```yaml
namespace: support
id: welcome
text: Welcome
```

Here `mode: append` illustrates retaining a lower-layer collection and adding
distinct records from the higher layer. The `mode` field expresses the operation;
OD-5 defines its names and effective-identity default. Identity and combination
controls belong to the collection declaration; import shorthand is deferred.

Letters below denote record identities; `B*` is a higher-tier record with the same
identity as `B` and a different payload. The outcomes and ordering shown are
settled below.

| Lower-tier sequence | Higher-tier sequence | Operation | Proposed result |
| --- | --- | --- | --- |
| `[A, B]` | `[C]` | `append` | `[A, B, C]` |
| `[A, B]` | `[C]` | `replace` | `[C]` |
| `[A, B]` | `[B*]` | `append` | Duplicate-identity error |
| `[A, B]` | `[B*]` | `replace` | `[B*]` |
| `[A, B]` | `[B*]` | `upsert` | `[A, B*]` |
| `[A, B]` | `[B*, C]` | `upsert` | `[A, B*, C]` |

Whole-sequence replacement discards the lower sequence. A repeated identity in the
retained combined sequence is an error under the proposed append semantics,
including when the records have identical or disjoint payloads.

## Settled Design Choices

### Collection Declarations

**OD-2: settled.**

Use a root `configkit.sequences` block containing a sequence of collection
declarations. Each entry identifies a `destination` and may declare `key`,
`mode`, or both. Application sequences remain ordinary YAML data. This declaration
form serves imported items, inline records, and sequences supplied by ordinary
mapping imports, including inline sequences in selected overlays.

`destination` uses the existing sequence of literal mapping keys. Resolve it
relative to the declaring source's placement: a fragment mounted under
`catalog` can declare `[prompts]` locally. A sequence-item source can declare a
nested collection within its own item without referring to an eventual numeric
index. Preserve those source and item-local scope boundaries.

Identity and mode can be supplied independently. Omitted identity can use an
application registration; an explicit matching declaration is accepted, and a
conflicting declaration is invalid. Whole-sequence replacement does not itself
require identity matching, but application-owned constraints still apply.
Identity-policy lifetime across tiers is settled in OD-3 below.

Mode describes the collection's combination instruction at source-precedence
and inheritance boundaries, as detailed in OD-7. A lower tier's
`append` does not silently select append for a later overlay. Gather collection
declarations before composing their relevant scope so their effect does not
depend on which contribution is assembled first. Identity agreement is settled
in OD-3; operation agreement and omission among peers are settled in OD-4.

Permit sequence declarations in selected overlays and make their use independent
of filesystem import enablement. This deliberately changes the current blanket
prohibition on overlay `configkit` declarations. Enabling imports in overlays
is not part of this decision. Consume sequence controls before application
decoding, as with other structural directives.

For example, an overlay can replace a complete record by identity:

```yaml
configkit:
  sequences:
    - destination: [prompts]
      key: id
      mode: upsert

prompts:
  - id: foo
    text: Revised definition
```

The explicit `mode: upsert` line can be omitted here because `key: id` supplies
an effective identity. An overlay can instead select duplicate-rejecting append:

```yaml
configkit:
  sequences:
    - destination: [prompts]
      key: id
      mode: append

prompts:
  - id: baz
    text: Another prompt
```

Or it can explicitly select whole-sequence replacement:

```yaml
configkit:
  sequences:
    - destination: [prompts]
      mode: replace

prompts:
  - id: bar
    text: Bar definition
  - id: baz
    text: Baz definition
```

Explicitly selecting the default operation is valid. Removing the last example's
`configkit` block produces the same replacement behavior only when no effective
identity is available. For an identified collection, explicit `mode: replace`
overrides the inferred `upsert`; it is not redundant. OD-5 defines the selection
rule for `replace`, `append`, and `upsert`.

Start with this common collection declaration and defer `key` and `mode`
shorthand on imports. A single declaration form covers inline overlays and mixed
inline/imported collections without duplicate authoring surfaces. Local wrappers
around sequence values would require distinguishing application data from
control mappings; the existing reserved namespace keeps that distinction clear.

### Identity Definitions And Lifetime

**OD-3: settled.**

#### Field Sets

`key` accepts a string or a nonempty list of strings naming direct, literal
fields in each record. `key: id` and `key: [id]` are equivalent. A name such as
`metadata.id` denotes that literal field, not a nested lookup.

Treat the selected fields as a set: remove repeated names and disregard authored
field order when comparing definitions. For example, `[namespace, id]`,
`[id, namespace]`, and `[namespace, id, id]` define the same identity. An empty
list is invalid because it supplies no identity fields. These equivalence rules
also apply when comparing configuration declarations with application
registrations.

Use a consistent field order internally to construct compound identity tuples.
Preserve component types and tuple boundaries; concatenating unstructured values
must not cause otherwise distinct identities to collide.

#### Scalar Components And Equality

Each identified record must be a mapping, and every designated identity field
must be present. Components may be strings, numbers, booleans, or explicit null.
Missing fields and mapping or sequence components are invalid. A present null is
a valid component and is distinct from an absent field. Two single-field null
identities are equal; in a compound identity, all components must compare equal.

Strings compare exactly, with no trimming, case folding, or normalization.
Empty strings remain valid. Scalar categories remain distinct: the string
`"1"`, the number `1`, the boolean `true`, and null are different identities.

Numbers compare by mathematical value across integer, fractional, and exponent
spellings: `1`, `1.0`, and `1e0` are equal, as are `0` and `-0`. Use exact numeric
normalization for identity comparison without rounding through a floating-point
conversion. This comparison does not rewrite authored values or alter existing
typed decoding and projection semantics. The existing parser's rejection of
non-finite numeric forms remains unchanged.

#### Agreement And Introduction

Within a strict tier, identity declarations for the same collection must agree.
Equivalent field sets coalesce; different sets produce an error. Across tiers,
append and replacement by identity use one agreed definition for the records
being combined:

| Lower collection identity | Higher declaration | Result when combining records |
| --- | --- | --- |
| `id` | Omitted | Use `id` |
| `id` | `id`, `[id]`, or `[id, id]` | Accept the equivalent definition |
| `id` | `[namespace, id]` | Error: different identity definitions |
| None | `id` | Establish `id` and validate all participating records under it |

An application registration supplies identity when configuration omits it and
remains authoritative over explicit declarations. If neither the application nor
either participating collection supplies an identity, append and replacement by
identity fail; ConfigKit does not infer an `id` convention. Introducing identity
for a previously unkeyed lower sequence requires checking both lower and incoming
records, including uniqueness within each contribution.

#### Replacement And Policy Lifetime

YAML-owned identity follows the collection that survives an operation:

| Operation | Effective YAML-owned identity |
| --- | --- |
| Append | Retain or establish the agreed identity |
| Replacement by identity | Retain or establish the agreed identity |
| Whole-sequence replacement | Use the replacing tier's declaration and discard the lower collection's YAML-only identity |

Without an application constraint, whole-sequence replacement may introduce a
different identity, such as replacing an `id`-keyed collection with a `name`-keyed
collection. A whole-sequence replacement with no new identity declaration and no
application registration produces an unkeyed sequence. No separate identity
removal directive is needed.

Application registrations continue to constrain the replacement collection.
If the application requires `id`, declaring `name` is invalid under every mode.
The policy reset above applies when a sequence is actually replaced; effects of
absent contributions, null collections, and kind mismatches are settled in OD-6.
YAML-only identity declarations receive assembly-time enforcement, as settled
in OD-1 below.

### Combination Scope, Declarations, And Ordering

**OD-4: settled.**

#### One Operation Per Collection Per Tier

Strictly compose a tier's contributions first, then apply one selected operation
per collection against the accumulated lower sequence. A collection declaration
governs all contributions to that collection in its tier, including imported and
inline records; it does not govern only records in the declaring file.

Collection scope is the resolved destination within its composition scope.
Resolve relative destinations first and preserve separate item-local scopes for
collections inside different imported items. Apply the same combination
semantics to selected overlay contributions.

For example, with an agreed identity, higher-tier fragments contributing `[C]`
and `[D]` assemble `[C, D]`. Against lower `[A, B]`, append or replacement by
identity produces `[A, B, C, D]`; whole-sequence replacement produces `[C, D]`.
The tier operation does not sequentially replace peer source contributions.

#### Operation Agreement And Omission

Coalesce agreeing explicit modes for the same collection within a tier and
reject conflicting explicit modes. Omission expresses no preference; apply the
default only after reconciling declarations.

| Modes declared within one tier | Effective operation |
| --- | --- |
| `append` and `append` | `append` |
| `append` and omitted | `append` |
| `replace` and omitted | `replace` |
| `append` and `replace` | Error |
| All omitted | `upsert` with effective identity; otherwise `replace` |

Explicitly selecting the default is a real instruction and can conflict with
another explicit selection. Different modes in different tiers are valid; a
lower tier's mode does not carry forward. Neither a root declaration nor a later
declaration wins over conflicting peers. Diagnostics identify the collection
destination and conflicting declaration locations. OD-5 defines mode names and
the default-selection rule.

#### Strict Duplicate Handling

Duplicate identities among peer contributions or within one contributed list
remain strict-composition errors under all three modes. For example, peer
fragments contributing `[B]` and `[B*]` conflict even under replacement by
identity: neither has higher precedence. Lower `[B]` and higher `[B*]` can
instead produce `[B*]` at a precedence boundary. Whole-sequence replacement
does not permit duplicate identities within an identified higher-tier sequence.

#### Ordering

Retain existing dependency-ready source scheduling within each tier and authored
item order within each list. Gathering collection declarations does not change
assembly order, and identity values do not sort records.

Across tiers:

- Append retains lower records in order, followed by higher records in assembled
  order.
- Whole-sequence replacement retains only the higher sequence, in its assembled
  order.
- Replacement by identity replaces matched records in their existing lower
  positions and appends new identities in higher-tier order.

For example, lower `[A, B, C]` and higher `[C*, A*, D]` produce
`[A*, B, C*, D]` under replacement by identity. Updating a record's contents
preserves its position; whole-sequence replacement gives the higher tier control
of the complete sequence order.

OD-5 settles default selection and operation naming. OD-6 settles absent
contributions, empty sequences, null, and kind mismatches.

### Mode Names And Effective-Identity Defaults

**OD-5: settled.**

Use the mode names `replace`, `append`, and `upsert`, with one spelling per
operation:

| Mode | Behavior at source-precedence and inheritance boundaries |
| --- | --- |
| `replace` | Replace the complete lower sequence with the assembled higher sequence |
| `append` | Retain lower records and append higher records; duplicate identities fail |
| `upsert` | Replace matching records in full, retain unmatched lower records, and append new identities |

`upsert` preserves the existing positions of matching lower records and adds new
identities in higher-tier order, as settled in OD-4. It does not recursively merge
fields within matching records.

Select the operation at each collection's precedence boundary in this order:

1. Reconcile explicit modes within the current tier under OD-4; conflicting
   explicit modes fail.
2. If an explicit mode is supplied, use it.
3. Otherwise, select `upsert` when an effective identity is available.
4. Otherwise, select `replace`.

Effective identity may come from an application registration, a current-tier
declaration, or the lower collection, subject to OD-3's agreement rules.
Selection depends on identity availability, not on whether the current tier
explicitly repeats a `key` declaration. A plain overlay sequence can therefore
default to `upsert` when identity was established elsewhere. A lower tier's mode
does not carry forward.

Determine the inferred mode using the available identity before applying the
operation. Explicit `replace` still follows OD-3: discard the lower collection's
YAML-only identity with the lower sequence, use the replacing tier's identity
declaration if supplied, and preserve application-owned constraints. Selecting
`append` or `upsert` explicitly still requires an effective identity.

This default governs combination between tiers and during sibling inheritance
under OD-7. Within a tier, identified
sequences append distinct records and reject duplicate identities regardless of
mode. Unidentified sequences retain their existing strict-composition conflict
behavior; this decision does not introduce unkeyed concatenation.

### Whole-Record Replacement By Identity

The `upsert` operation replaces each matching lower record with the
complete higher record, retains unmatched lower records, and adds new identities.
For example, given these lower and higher records:

```yaml
# Lower record
id: B
text: Original
temperature: 0.5
```

```yaml
# Higher record
id: B
text: Revised
```

The resulting record at this precedence boundary contains `id` and `text`.
The lower record's `temperature` is discarded. Subsequent defaulting or other
lifecycle operations retain their separately defined semantics. Recursive
field-by-field overlay of matching records is outside this selected operation.

### Selected Overlay Control

Selected overlay authors must be able to choose any of the three operations
for a sequence supplied by the overlay, including records authored inline.
The operation applies against the existing effective sequence at that path.
For an existing sequence `[foo, bar]`:

| Overlay contribution | Selected operation | Result |
| --- | --- | --- |
| `[foo*]` | Replacement by identity | `[foo*, bar]` |
| `[baz]` | Append | `[foo, bar, baz]` |
| `[bar, baz]` | Whole-sequence replacement | `[bar, baz]` |

Here `foo*` has the same identity as `foo` and supplies its complete replacement
record. Append retains its duplicate-identity error semantics. The retained
records, replacement behavior, and ordering shown are settled, including the
OD-4 rule that matched records retain their lower positions.

The settled OD-2 collection declaration supplies the overlay authoring surface.
OD-5 supplies operation names and effective-identity default selection. The current
[selected-overlay contract](../config.md#selected-overlays) remains the baseline
until the new controls are implemented.

### Optional Application-Owned Identity Constraints

Retain an interface through which applications may register a collection and
impose its unique-key definition. If the application requires `key: id`, a
configuration declaration of `key: [id, name]` is invalid: the compound key
would permit repeated `id` values with different names. The application identity
constraint applies whichever cross-tier operation configuration selects.

This is an authoritative constraint, distinct from an application default that
only fills an omission. Configuration can define identity for collections
without an application-owned constraint. Omitted configuration identity uses a
registered definition when present, and an explicit matching declaration is
accepted, as settled in OD-2. OD-1 below distinguishes the final-result guarantee
for application-owned constraints from assembly-time enforcement for YAML-only
declarations.

### Final-Result Guarantee For Application Registrations

**OD-1: settled, including the YAML-only follow-up below.**

> On successful loading, the final effective collections covered by application
> registrations satisfy their declared identity constraints.

Retain assembly-time identity checks and validate the final effective
collections before publication. This covers the loaded state after defaults,
inheritance, overrides, and application resolution have taken effect. Those
operations retain their separately defined combination semantics. The guarantee
does not require every intermediate callback state to satisfy the constraint
and does not extend to caller mutations after loading returns.

Keep the promise limited to the declared identity constraints. It does not by
itself require registered paths to exist or be non-null. Identity validation checks
actual sequences; the application schema and validation determine whether absence,
null, or another shape is acceptable, as settled in OD-6. Final validation examines
effective typed values, while assembly-time checks retain their role in
validating authored contributions. A final scalar value alone does not prove
that the field was explicitly supplied in a source.

The existing canonical final-value representation and shared publication path
provide the implementation foundation. Diagnostics should identify the affected
collection and item paths, using available origin information without exposing
identity values. The guarantee requires neither a new mutation-history system
nor exact causal reconstruction for records supplied by defaults or callbacks.

### Assembly-Time Enforcement For YAML-Only Identities

Enforce identities declared only in YAML during structural composition,
precedence assembly, and identified collection operations during sibling
inheritance. They do not carry a final-result guarantee after defaulting, CLI
overrides, or application resolution. Those operations may produce a collection
that no longer satisfies a YAML-only identity, including duplicate identities.
OD-7 defines the combination and validation boundaries.

Applications requiring a final identity constraint register it through the
application-owned interface. This keeps YAML declarations aligned with the
structural operations they control and avoids extending their policy lifetime
through later replacements, callbacks, and nested-record changes solely for
final validation.

### Absence, Empty Sequences, Null, And Kind Mismatches

**OD-6: settled.**

At a collection's precedence boundary:

| Higher-tier contribution | `replace` | `append` / `upsert` |
| --- | --- | --- |
| Absent | Leave the lower value unchanged | Leave the lower value unchanged |
| Explicit `[]` | Clear the sequence | Contribute no records |
| Explicit null or another non-sequence value | Ordinary replacement, subject to application schema | Error: the operation requires a sequence |

A declaration alone does not clear a collection. An optional import selecting
nothing contributes nothing; it does not erase records supplied elsewhere.
An absent lower collection can be initialized by an incoming sequence under
any mode, subject to identity requirements. An explicitly null or other
non-sequence lower value is distinct from absence: append and upsert against it
fail, while replacement remains available. An empty incoming sequence does not
bypass these kind checks or the operation's identity requirements.

With effective identity available, an omitted mode selects upsert. Consequently,
`prompts: []` alone does not clear an identified collection. Clearing it requires
explicit `mode: replace`. Replacing an identified collection with null likewise
requires explicit replacement and an application schema that permits null.

An application identity registration does not itself require the collection path
to exist or be non-null, and does not independently reject non-sequence values.
Identity validation checks actual sequences; application schema and validation
own the permitted shape. Null as a collection is distinct from a present null
identity component inside a record, which is valid under OD-3.

### Lifecycle Coverage And Sibling Inheritance

**OD-7: settled.**

Apply the sequence operations during source composition and precedence assembly,
selected overlays, and sibling inheritance through `extends`. Preserve the
existing combination semantics of `Definition.Defaults`, CLI overrides, and
application resolution callbacks.

#### Inheritance Operations And Controls

Complete source-tier assembly before resolving inheritance, retaining the
existing dependency-ready inheritance algorithm and scope rules. At each
inheritance combination, the resolved parent's collection is the lower value
and the inheriting mapping's local collection is the higher value. Use the same
operation selection as source precedence: an explicit mode wins; otherwise use
`upsert` with effective identity and `replace` without it. There is no separate
default-append rule for inheritance.

Use the inheriting destination collection's controls. Its selected mode governs
both source-precedence and inheritance combinations at that destination, without
separate per-phase mode settings. A mode governing the parent's source collection
does not automatically become an instruction at the inheriting destination merely
because records are inherited. Retain the relevant controls until inheritance
uses them, then keep them out of application decoding. Identity agreement,
application constraints, and replacement lifetime follow OD-3.

For example:

```yaml
configkit:
  sequences:
    - destination: [parent, items]
      key: id

child:
  items:
    - id: foo
      value: A
    - id: bar
      value: B
parent:
  extends: child
  items:
    - id: baz
      value: C
```

Despite the mapping names, `parent` inherits from `child` in this example.
The effective identity at `[parent, items]` selects `upsert`, yielding records
`[foo, bar, baz]`. If the local contribution instead supplies a revised `foo`,
the result is `[foo*, bar]`: replace the complete record in its existing position.
Adding `mode: append` would reject that repeated identity; `mode: replace` would
retain only the local sequence.

Mappings continue to combine recursively. Identified sequence records remain
whole-record replacements, not recursive field merges. Identity supplies the
matching key; ConfigKit does not infer identity from a field named `id` or from
record position. Append and upsert continue to require effective identity, and
this extension does not introduce unkeyed sequence concatenation.

Apply OD-4 ordering and OD-6 presence and kind rules at inheritance boundaries.
An omitted local sequence retains the inherited sequence. An explicit empty
sequence clears it under replace, while append and upsert contribute no records.

#### Other Lifecycle Stages

Source-backed underlays participate in the new combination model, including
application-supplied mapping sources. Typed `Definition.Defaults` remain the
initial decoded state: supplied sequences replace default sequences. Source
`mode: append` does not append to a typed default list.

CLI overrides retain explicit replacement for supplied sequences. Application
resolution retains its existing assignment and recorded-transformation semantics;
ConfigKit does not automatically combine a callback's assigned sequence with
its previous value. Application-registered identities still receive final
validation under OD-1, independently of how these stages produce their values.

#### Validation Timing

Allow item-local helper imports to complete, and consume applicable non-main
source-version headers, before checking the application payload's identity for
collection assembly. A helper may supply an identity field; a consumed header
cannot supply a component after removal. Validate each identified contribution
before combination and uniqueness in the resulting identified collection,
including inheritance operations. An upsert match across a precedence boundary
is a replacement, not a duplicate error; duplicates within a contribution remain
errors.

Later inheritance, typed defaults, or callbacks do not repair an identity missing
at an earlier assembly boundary. Final application-registered validation examines
the final effective typed collection, while YAML-only enforcement remains limited
to the structural operations described above.

## Architectural Direction

The settled collection declaration gives imported and inline records at the same
destination a shared identity policy. Declarations are gathered before composing
the relevant scope.

OD-3 settles identity agreement and lifetime; OD-4 settles tier-wide operation
selection, strict peer composition, and ordering at precedence boundaries.

This requires collection policy to survive import expansion through precedence
assembly and the applicable inheritance operations. Correctness includes final
values, projected item paths, source destinations, provenance, history,
sensitivity, and passive resource contexts.
Appending changes final indexes; replacing changes which contributions survive.
The implementation must account for both through the existing loading lifecycle.

## Scope And Decision Ownership

This is narrower than the deferred
[identified-record sequences proposal](future.md#1-identified-record-sequences),
which also considers recursive merging of matching records and record-specific
inheritance. Those capabilities remain separate candidates. Compound keys,
duplicate-rejecting append, and whole-record replacement by identity do not
require adopting them.

The proposal would extend identity declaration ownership and revise the
cross-tier replacement decision in
[ADR 0014](../adr/0014-sequence-placement-and-composition.md), as well as sequence
replacement during inheritance in
[ADR 0009](../adr/0009-sibling-inheritance.md). Record
the architectural change in a superseding decision and reconcile affected
architecture policy before implementation. The current ADR and contracts remain
authoritative until that decision is made. This draft neither changes them nor
selects migration machinery.

A staged implementation plan should follow review of this settled roadmap.
It should cover ordinary and layered loading, item-local scopes, source-version
processing, identity validation, precedence, inheritance, and inspection together.
Public API design and test sequencing belong to that later plan.