promptkit/docs/roadmap/configkit-assessment.md

# ConfigKit Adoption Assessment

Status: assessment with agreed design direction; upstream API details remain
to be settled before the feature roadmap and implementation plan. No runtime
changes are part of this assessment.

## Goal And Recommendation

Use ConfigKit to standardize PromptKit's configuration semantics and simplify
its loading machinery, improving ConfigKit where a reusable capability is
justified. Preserve substantive consumer capabilities, not historical loading
sequences, file syntax, or Go signatures. Document intentional breaking changes
and their migration paths.

Recommend adoption through an eager, immutable definition-loading boundary.
ConfigKit should own YAML interpretation, source-tier combination, strict typed
decoding, presence, and source diagnostics. PromptKit should own definition identity, named profile
relationships, LLM-specific rules, and execution. Do not retain the lazy loading
architecture merely to minimize migration, or introduce a second generic
configuration framework inside PromptKit.

The expected benefits are consistent configuration semantics for applications
using both libraries and removal of repeated discovery, partial metadata parsing,
lazy error arbitration, and repeated profile-chain resolution. Some of those
savings come from the loading redesign rather than ConfigKit itself. ConfigKit
adds ongoing value by owning the generic decoding behavior and its tests.

## Prerequisite: Proposed Upstream ConfigKit Work

Design, implement, and release these two ConfigKit capabilities before the
dependent PromptKit refactor. The remaining recommendations assume this upstream
work; finalize their integration details against its agreed public contracts.
These are selected architectural prerequisites, not a claim that adoption is
technically impossible using today's reader-based API.

1. **Native `fs.FS` sources and source-relative context.** Support embedded and
   caller-supplied filesystems as first-class configuration sources. Define source
   identity, relative path resolution, ownership, and how enabled imports and
   overlays operate within that filesystem. This is broadly useful for libraries
   bundling default configuration. Passing an opened file to today's `Reader`
   supplies bytes and a label, but does not make filesystem-relative loading
   native to ConfigKit; that workaround is not the target architecture.
2. **Source-backed underlays with origins and programmatic input.** Adopt
   ConfigKit's working design: strictly compose contributions within each tier,
   then recursively overlay explicitly ordered tiers. Support embedded, host-file,
   byte/reader, and labeled programmatic sources without temporary files or YAML
   serialization round trips. Prefer the proposed string-keyed programmatic
   mappings: omitted keys convey absence; present values, including nil/null,
   convey explicit input. Do not require typed-value ingestion as a prerequisite.

Adopt ConfigKit's recursive combination model rather than preserving PromptKit's
whole-profile shadowing or atomic provider-map replacement across source tiers.
No hidden replacement pass should run after ConfigKit loading. Named inheritance
is a separate optional upstream request, not part of these prerequisites.

The upstream design should show how the two capabilities work together, including
source origins and efficient ownership/copying. Resolve these API questions before
turning the assessment into a decision-complete PromptKit implementation plan.
This assessment does not authorize code changes in either repository.

## Evidence And Scope

Reviewed local baselines on 2026-09-29:

- PromptKit: `12d6d80024e80c552d122a076c22429bf6f2f6c2` (v0.11.2).
- ConfigKit: `67c03690181c5af8dde6be325abac241dae6732f` (v0.4.2).

Design update: reviewed ConfigKit's working
`docs/roadmap/sources-and-layering.md` with checkout HEAD
`36172ed20e832b371506d6aceb02dc6c8f50f71b`. The reviewed roadmap included
uncommitted changes, so that commit alone does not identify the reviewed text.
It remains a proposed target, not implemented behavior. The recommendations here
now follow its underlay semantics and support its two open-decision recommendations:
labeled programmatic mappings and direct source-context queries. Verify the
final upstream contracts before preparing dependent implementation details.

PromptKit evidence: [consumer workflows](../consumers/pkg-promptkit.md),
[file formats](../formats.md), [source implementation](../internal/sources.md),
[runner](../internal/runner.md), and the implementation owners identified below.
ConfigKit evidence: its reviewed `docs/config.md`, `docs/roadmap/deferred.md`,
`lifecycle.go`, `source.go`, and ADR 0001 in the sibling checkout. That checkout
is review context, not a proposed build dependency or required directory layout.

ConfigKit currently exposes `Load[T]` over file, byte, reader, and defaults-only
sources. It supports strict decoding, presence/origin records, defaults,
recorded resolution, semantic validation, explicit imports, one selected
filesystem overlay, and secret handling. Its parser and generic conversion
machinery are private. It has no public arbitrary layer-composition, generic
inheritance, or `fs.FS` source API. Its deferred scope explicitly excludes several
of these generalizations; an upstream enhancement requires a deliberate decision.

This is a source-and-contract assessment, not a performance benchmark or full
correctness audit. It does not claim measured allocation, latency, or line-count
savings. Current code and canonical contracts remain authoritative until an
approved implementation changes them.

## 1. Capability Inventory

| Preserve the ability to… | Replaceable mechanics and representation |
| --- | --- |
| Define named, versioned prompts, inspect their declared inputs, and select a version explicitly | Directory scans on each operation, YAML metadata prepasses, filename layout, constructor signatures |
| Use inline and file-backed message templates, variables, input helpers, session templates, roles, and cache control | Timing of definition/content reads and private repository interfaces |
| Configure named profiles, alias/inherit another profile, and override selected settings | Zero-as-absence fields, per-call chain traversal, raw YAML decoding types |
| Supply embedded application defaults, operator definitions, maintained catalogs, and programmatic profiles | Repository-wrapper stacks and exact source-option API |
| Register custom/local backends, use built-ins or endpoint-only profiles, and control backend capacity | Configuration DTO shapes and registry assembly order |
| Prepare, inspect, execute, or freeze an execution for later single-claim execution | Source reloading during ordinary preparation; exact preparation error ordering |
| Override per-run settings, credentials, reasoning, session, and output contract; append rendered messages | Public value signatures if presence or clearer configuration separation requires revision |
| Supply inline/file artifacts and inject artifact readers, model clients, and HTTP clients | Assembly APIs; runtime collaborators need not be YAML-decodable |
| Validate output, use provider-native structured output, bounded repair, transport retries, cancellation, and observation | Internal wiring, not the substantive execution behaviors |
| Receive results, usage, identities, hashes, validation outcomes, and inspectable failures | Configuration error types and when definition failures appear |

Preserve independent profile and prompt inspection as useful operations. They
need not retain the old guarantee that engine construction has never examined
other definitions. Preserve custom backend identity and capacity behavior,
optional upstream credentials, and the distinction between failed generated
content and operational errors. This project does not authorize silently
dropping execution features to simplify the configuration layer.

## 2. Responsibility Mapping

| Concern | Recommended owner and disposition |
| --- | --- |
| YAML grammar, scalar interpretation, one-document enforcement, unknown fields, typed ranges, null legality | Delegate to ConfigKit; remove corresponding PromptKit parser logic and redundant parser tests |
| Field presence and source coordinates | Delegate to ConfigKit; retain only the compact records needed for resolution and diagnostics |
| File/directory/embedded source selection and definition ID indexing | Retain a small PromptKit collection loader; ConfigKit loads documents, not a prompt catalog |
| Configuration filesystem sources, relative imports, and generic layer origins | Delegate to the prerequisite ConfigKit APIs; do not maintain parallel reader-based loading machinery |
| Definition discovery and relative message-content references | Retain PromptKit's domain resource adapters, using ConfigKit source context and standard `io/fs` facilities where sufficient |
| Profile-source priority and combination | PromptKit selects tiers; ConfigKit owns strict within-tier composition and recursive cross-tier overlay; eliminate repository fallback and whole-definition shadowing |
| Named profile inheritance and backend selection | PromptKit owns domain relationships unless the separate inheritance proposal is adopted; integrate resolution with ConfigKit's recorded lifecycle |
| LLM setting bounds, role/session normalization, output rules, reserved provider fields | Retain shared PromptKit domain owners; ConfigKit validation callbacks may invoke them |
| Definition defaults | Use ConfigKit defaults where they do not erase inheritance presence; apply effective execution defaults after inheritance |
| Programmatic configuration | Prefer labeled mapping input with explicit presence; evaluate whether typed PromptKit convenience constructors justify conversion code, without YAML round trips or zero-value inference |
| Runtime clients, observers, artifact readers, request data | Retain normal Go injection and request APIs; do not load them through ConfigKit |
| Credentials | Retain optional environment-name and direct request-key semantics; do not replace these with ConfigKit's required typed secret resolution |
| Catalog manifest, ownership, reserved names, secret-safety checks | Retain the catalog adapter; reuse the common ConfigKit profile decoder |
| JSON Schema resources and generated-output validation | Retain JSON Schema implementation; these are not application configuration documents |
| Application imports, CLI overrides, deployment overlays, secrets-directory bootstrap | Leave with consuming applications unless a concrete PromptKit need is accepted |
| Generic source/decoding diagnostics | Reuse ConfigKit; add only PromptKit identity/context and domain relationship diagnostics |

## 3. Recommended Target Loading Model

### Definition Snapshot

Create one immutable definition snapshot during engine construction. Discover
the explicitly supplied sources, assemble their contributions through ConfigKit,
decode the effective configuration, normalize domain values, and resolve relationships
before publishing the engine. No provider call or credential availability check
belongs in this phase. Publish no partially initialized engine on failure.

Prefer keyed collections, such as `profiles.<name>`, assembled before one typed
loading lifecycle instead of independently decoding each profile and recombining
the results. Multiple files can contribute to those collections; this does not
require a monolithic file. Reconsider the previous one-typed-document-per-definition
recommendation. Settle prompt ID/version representation and catalog format migration
in the feature roadmap. Support OS directories, individual files, embedded/custom
`fs.FS` sources, and programmatic input through the new source model; source selection
and definition discovery remain PromptKit concerns, not a second composition engine.

Perform domain-specific resolution through ConfigKit's recorded lifecycle before
semantic validation and result publication. Reuse its presence and provenance;
do not add a generic merge/defaulting pass or reconstruct origins after loading.
Runtime client assembly and resource interpretation remain outside typed decoding.

Load and freeze referenced message text with definitions. Validate static prompt
structure and template syntax at loading, using the same template function setup
as execution; render only when request variables and artifacts exist. Preserve
source-relative content resolution and containment without expanding the threat
model. A snapshot is a set of captured reads, not an atomic filesystem transaction.

Resolve each effective content reference using its contributing source's context,
not the highest tier's directory. Support ConfigKit's proposed direct context
query rather than a consumer-maintained filesystem registry. Composite origins
must not be assigned an arbitrary directory; ambiguous or absent context needs
an actionable resource error or explicit consumer-supplied context. ConfigKit
does not snapshot external text: PromptKit opens it explicitly and freezes it
before engine publication. Caller-owned filesystem lifetimes remain caller-owned;
close resource handles acquired by PromptKit without retaining ConfigKit load handles.

Keep request artifacts and schema compilation at preparation. A request can
replace the output contract, so an unused default schema should not need eager
compilation. Prepared execution continues to freeze its effective resources.
The snapshot covers definitions/message text, not environment values, request
files, every schema file, or execution state.

### Composition And Validation

Use explicit source categories in the existing useful priority order:
programmatic profiles, operator source, application fallback, PromptKit-owned
profiles, then external catalogs. Express these as explicit ordered tiers, not
source discovery order. Within a tier, mappings compose recursively and overlapping
non-mapping contributions conflict, even when equal. Disjoint contributions to
the same profile may therefore combine; do not preserve a blanket duplicate-ID
rejection. Across tiers, mappings overlay recursively, retaining omitted lower
children. Effective prompt identities/versions must remain addressable, with an
omitted version ambiguous when multiple versions exist; finalize their keyed
representation without reintroducing whole-definition shadowing.

All sources undergo syntax, acquisition, and composition checks; higher tiers
cannot hide those failures. Decode and validate effective values after assembly,
not every intermediate tier against the final typed schema. A replaced lower
value may consequently have a type that would not fit the final schema. An unused
invalid effective definition still fails construction. Catalog ownership,
reserved-name, and secret-safety checks remain distinct source-policy requirements,
not a reason to impose intermediate typed validation on ordinary inputs.

Preserve profile templates that become usable through descendants or runtime
overrides: eager loading must not require every consumer profile to be an
independently runnable endpoint/model pair. Validate supplied fields and graph
integrity eagerly (including missing bases and cycles), and check final execution
requirements after runtime selection/overrides. Prompt default-profile references
can be checked against the assembled collection; a prompt with no default remains
valid for callers that supply a profile explicitly.

ConfigKit's ordering is typed defaults, source-underlay tiers, strictly composed
main input, selected overlay, then CLI overrides. PromptKit's API must map its
source priorities explicitly onto supported roles; do not mislabel programmatic
input as CLI flags. Underlay-only loading avoids inventing a main document.
Imports remain opt-in, and the named-overlay convenience need not be exposed by
PromptKit merely because ConfigKit supports it.

A prompt's `version` is a domain identifier, not automatically a ConfigKit
application-schema version. Do not configure `VersionRule` for prompt versions.
If a separate configuration schema version is later needed, respect ConfigKit's
designated main-input ownership and pre-underlay check timing; do not assume
last-wins version selection across catalogs.

### Source Presence And Separate Named Inheritance

Adopt ConfigKit source-layer semantics: absence retains lower input; explicit
zero, false, and empty strings replace it; mappings combine recursively; empty
mappings retain lower children; sequences replace as units; null explicitly
replaces, subject to final schema legality. No reset/delete operator is assumed.
Preserve presence through domain resolution, backend/framework
default application, request overrides, and outbound request construction.
Otherwise accepting `temperature: 0` during decoding would accomplish nothing.

Use ConfigKit metadata for both file and programmatic input. A final Go value
alone does not distinguish absent from explicit zero. Avoid eagerly defaulting
profile fields in ways that erase that distinction before relationship resolution.

Provider `extra_params` maps combine recursively across source tiers too: `{}`
does not clear lower keys. Retire the proposed atomic source-map replacement and
null-as-unspecified conventions; do not emulate them after loading. An operator
overriding only `model` also retains a lower `base_profile` declaration and other
omitted settings. Diagnose and document these consequences explicitly.

Named inheritance still selects relationships between effective entries; source
layering does not implement it. The optional upstream request remains pending.
Finalize named-inheritance null/map semantics separately, favoring consistency
with source layering rather than automatically retaining atomic maps or null-as-
inherit. Any domain exception needs explicit justification. This also does not
silently change the existing request-level override contract, which is a separate
runtime boundary to review in the feature roadmap.

Resolve shared ancestors once per snapshot. Use recorded domain resolution to
keep resulting values and origins consistent. Confirm the upstream lifecycle can
express required relationship provenance; do not claim generic layer history
already explains named inheritance or invent a parallel audit framework.

### Reloading, Errors, And Application Integration

Later definition edits take effect through a newly constructed engine. Do not add
watchers, mutable registry reloads, or a second lazy compatibility mode. Applications
choose how to swap engines and let in-flight work finish. Old prepared handles
remain associated with their original engine. Simultaneously active engines have
separate backend capacity pools; migration guidance must not imply a global limit
across a reload overlap.

Return configuration failures from construction with file/path/location and
useful definition identity where available. Recommend preserving ConfigKit's
public `LoadError` through wrapping for `errors.As`, together with a PromptKit
configuration category; do not copy its entire diagnostic API into the facade.
Domain-only relationship failures remain PromptKit errors with relevant source
locations. Start with deterministic fail-fast loading across files and ConfigKit's
existing multi-diagnostic behavior within a load, rather than inventing a batch
diagnostic subsystem. Do not expose raw credentials or add implicit logging.

Runtime failures continue to distinguish invalid requests, missing selections,
rendering/artifact/schema errors, provider/transport errors, capacity rejection,
and content-validation results. Cancellation remains effective at
PromptKit-controlled boundaries; ConfigKit's current synchronous `Load` has no
context argument, so do not promise interruption inside it or create goroutines
to simulate that guarantee.

Applications using ConfigKit can present the same underlying loading diagnostics
for their application settings and PromptKit files. They can also translate
application-owned typed settings into PromptKit profiles/backends. Do not require
applications to encode Go clients into YAML or adopt ConfigKit merely to call
PromptKit's runtime API. No full effective-configuration/explanation API is proposed
without an identified consumer requirement.

## 4. Simplification Accounting

| Current implementation | Expected disposition |
| --- | --- |
| `internal/profile/filesystem_repository.go`: metadata scans, selected strict decode, stream-end checks, malformed-file selection arbitration | Replace with source acquisition, ConfigKit assembly/decoding, and indexed effective values; remove the metadata prepass and per-lookup filesystem path |
| `internal/promptdef/filesystem_repository.go`: metadata selection and repeated decoding/content opening | Replace with eager prompt loads and an ID/version index; keep domain normalization and contained content reading |
| Profile filesystem, `fs.FS`, eager, memory, and overlay repository variants | Collapse loading into source adapters plus immutable collections; preserve only interfaces that still serve a real boundary |
| `internal/profile/resolving_repository.go` | Replace per-operation traversal and zero-sentinel merging with once-per-snapshot presence-aware typed resolution; graph logic remains |
| `internal/engine` raw repository-stack assembly | Replace with ConfigKit source tiers and one effective typed collection; engine capacity/client assembly remains |
| `internal/catalog` profile metadata extraction | Obtain explicit field information from ConfigKit; retain manifests, ownership, namespace, connection-metadata and secret checks |
| `internal/filecatalog` | Retain/reduce discovery and containment helpers actually needed by the eager loader; do not claim the entire package disappears |
| `internal/domain` source YAML tags and runtime/source coupling | Separate configuration DTOs from execution values where presence requires it; reuse domain validation rather than duplicating it |
| Direct `gopkg.in/yaml.v3` use for definition loading | Remove once all such consumers migrate; ConfigKit brings its own private YAML dependency |

New integration work is real: collection DTOs, ConfigKit calls and error wrapping,
source-context access, collection indexing, programmatic mapping ergonomics, and
conversion to runtime definitions. Catalog documents may need new keyed shapes;
do not hide that migration behind a permanent second loader.
Some defensive copies remain necessary at public/injected-client boundaries;
ConfigKit's detached load result is not permission to expose shared engine maps.
`internal/jsonvalue` still serves programmatic values, request parameters, and
prepared schema ownership, so adoption does not eliminate it.

Maintenance savings depend on actually deleting the old lazy paths and their
special-case tests. Move generic YAML/scalar behavior coverage to ConfigKit's
ownership; PromptKit retains representative adapter tests plus domain, precedence,
snapshot, inheritance, embedded-source, error, and execution integration tests.
Do not duplicate ConfigKit's entire parser suite or discard execution regression
coverage because construction fixtures need migration.

Runtime benefit: ordinary runs no longer rediscover definition directories,
redecode definitions, reread template text, or traverse inheritance chains.
Cost: construction loads even unused definitions and retains their text and
resolved values. ConfigKit also creates metadata and projections; keep only needed
records after loading. These are reasoned tradeoffs, not measured performance
claims. Adoption is not justified by the decoder replacement alone.

### Algorithmic Efficiency

Efficiency is an explicit design criterion for both the upstream work and this
shared-library refactor. Prefer optimal or near-optimal algorithms relative to
input and produced-data size, not a blanket O(n) claim with an undefined n.

- Acquire/parse each accepted source occurrence once, assemble input, and run
  one typed loading lifecycle rather than decoding intermediate tiers. Reuse
  captured content instead of reparsing it for metadata, selection, and validation.
- Build identity indexes in one pass, with expected constant-time map lookup.
  Do not scan whole collections for every lookup or precedence decision.
- Process inheritance graph structure in O(V + E), where V is the number of
  profiles and E the inheritance edges. Resolve shared ancestors once. Account
  separately for the size of materialized settings and retained origins; copying
  large inherited maps to many independent results necessarily costs more than V.
- Avoid whole-collection copying per layer, repeated serialization, and repeated
  prefix/history traversal. Analyze layer processing against total input nodes
  and produced values/provenance, including intentionally retained history.
- Keep configuration discovery and decoding out of ordinary execution. Copy only
  where ownership requires it; do not trade caller isolation for speed.

These are design targets, not promised complexity for every operation. Sorting
for deterministic discovery may require O(F log F) for F files, and processing
message text costs work proportional to its bytes. Construction time, retained
memory, and frequent-call costs all matter. Prefer straightforward indexes and
clear immutable ownership over speculative global caches or intricate optimizers.
The implementation plan should include focused benchmarks for consequential
repeated-work risks, such as shared-ancestor profiles and many layers, rather than
machine-dependent timing gates or benchmarks for every helper.

## 5. Additional Upstream Opportunities

Native filesystem sources and public layered loading with origins are the
prerequisites described near the beginning, not optional follow-ups. The following
remain optional and do not block adoption unless the upstream design establishes
a concrete dependency:

| Candidate | Recommendation and tradeoff |
| --- | --- |
| Generic named configuration inheritance | Requested separately in ConfigKit's `docs/roadmap/feature-request-promptkit-inheritance.md`, primarily for shared semantics across consumers. ConfigKit decides whether to adopt it; PromptKit can retain a small domain resolver if declined. Presence-aware source layering belongs to the prerequisite work, not this optional proposal. |
| Context-aware loading | Optional for reusable cancellable construction if large-source consumers need it. Promise cooperative cancellation only; no detached reads or instant interruption of arbitrary readers. |

Do not propose a public raw AST, permissive identity-only decoder, or dynamic
subtree conversion solely to preserve PromptKit's current scanning design.
Do not upstream backend/model selection, credential timing for inference,
profile catalog reservations, or prompt-template functions.

ConfigKit and PromptKit both declare LGPL-3.0-or-later. Adoption must still update
dependency notices under the OSS policy and inspect the exact selected release
and its transitive notices. Their current Go minimums do not require a PromptKit
toolchain increase. Use a released module dependency, not a sibling-directory
replacement. This assessment proposes upstream work, but makes no upstream edits
or licensing change.

## 6. Decisions And Migration

The agreed direction now incorporates ConfigKit's working roadmap. This supersedes
the earlier recommendations for independently typed documents, whole-definition
shadowing, intermediate typed validation, and atomic maps/null-as-unspecified at
the source boundary.

1. **Loading lifetime and failure scope.** Keep eager construction and frozen
   definitions/content. Syntax/acquisition/composition failures remain unmaskable;
   typed and domain validation apply to effective values. A separate reusable
   public snapshot lifecycle remains possible but is not currently needed.
2. **Source organization.** Prefer keyed collections composed across files and
   tiers before typed loading. Settle exact profile and prompt/version shapes,
   source-role mapping, and catalog migration in the feature roadmap. Independent
   typed loads followed by custom combination are not the selected direction.
3. **Presence and relationships.** Accept recursive source-tier overlay, retained
   children for empty mappings, explicit null replacement, and effective metadata.
   Named inheritance remains separate pending the upstream decision; settle its
   domain semantics before implementation rather than assuming either automatic
   equivalence or an exception to ConfigKit's rules.
4. **Diagnostic boundary.** Keep wrapping ConfigKit's public loading errors rather
   than mirroring them. Retain actionable domain relationship and resource errors
   without duplicating ConfigKit's generic diagnostic/provenance system.

Support the upstream recommendations for labeled mapping-based programmatic input
and direct filesystem resource-context queries. If real consumer examples show
mapping construction to be repetitive, consider small typed PromptKit conveniences
with explicit presence rather than requiring an upstream typed-ingestion API.
Final upstream signatures and ownership details remain open dependencies of the
feature roadmap, not implemented guarantees. Resolve the named-inheritance
direction there as well; a PromptKit-owned resolver remains viable if ConfigKit
declines or defers that optional request.

Expected consumer migration:

- Update definitions for ConfigKit's strict scalar rules, rejected YAML
  anchors/aliases/tags, input limits, and null/type checks. Quote prompt version
  strings rather than introducing a special numeric-version conversion merely
  for historical YAML coercion.
- Move standalone definitions into the agreed keyed contribution format. Review
  same-name contributions: disjoint fields within a tier can combine, and higher
  tiers retain omitted lower fields, including `base_profile` references.
- Review zero/false/empty/null values explicitly. Provider maps merge across
  tiers, `{}` does not clear them, and null replaces rather than means omission.
  Use labeled programmatic mappings or approved presence-aware conveniences.
  Document the separately selected inheritance semantics when settled.
- Handle effective-definition errors at construction, fix syntax/composition
  failures even in overridden inputs, and deliberately rebuild/swap engines
  after definition changes.
- Keep runtime credential lookup optional, avoid baking resolved secrets into
  definition snapshots, and preserve request-scoped key overrides.
- Update maintained catalog data for collection shapes and any incompatible syntax;
  coordinate catalog releases before the PromptKit dependency update. Preserve
  maintained profile names and catalog ownership rules, not every old scalar
  spelling. Do not validate live provider/model availability.
- Update error inspection, constructor/source options, and programmatic profile
  literals as required. Keep representative preparation, execution, inspection,
  custom backend, fallback, and embedded workflows in the migration guidance.

The subsequent feature roadmap should carry forward these choices, resolve the
upstream API details, and specify the target end state. The staged implementation
plan must sequence the prerequisite ConfigKit releases before dependent PromptKit
work, then cover deletions/replacements, contract tests, canonical documentation,
examples, attribution, and minor-release migration notes. Neither document should
become a reason to preserve obsolete compatibility machinery.