commandkit/docs/roadmap/implementation.md

# PromptKit request validation implementation plan

Status: Stages 1–7 complete; implementation is ongoing.

## Authority, scope, and baseline

Follow the stages below in order. Each stage is sized for one Luna implementation
prompt and identifies its changes, evidence, and completion condition. The
[accepted feature roadmap](promptkit-validation.md) owns the target state and
user choices. The [architecture policy](../policy/architecture.md), accepted
[ADRs](../adr/), and existing public contracts remain authoritative.

The planning baseline is commit `1119bda6b720171c708f3e6d03c5c8690b65ea81`.
The implementation now selects released PromptKit v0.15.0; the accepted roadmap
adds `promptkit requirements validate PROMPT_ID`.

The deliverable is the complete accepted feature, including the independently
optional callback, suite-returning helper, nil-builder path, file bindings,
YAML confirmation, maintained consumer, and canonical documentation. Existing
discovery, catalogs, definitions, configuration operations, and application
commands keep their current contracts.

## Working instructions

- Start with [development guidance](../development.md) and applicable policies.
  Prefer the repository knowledge graph for code discovery; verify snippets
  against current source when graph evidence is incomplete or stale.
- Implement in `integrations/ckprompt`, using public core, ckconfig, ConfigKit,
  and PromptKit APIs. Core and ckconfig need no new public API or PromptKit
  dependency for this feature.
- Use `Engine.ValidateRequest` directly. Discovery completeness is not a
  validation gate. Avoid parallel native validation rules, identity enforcement,
  request remerging, automatic execution, or prepared-run retention.
- Preserve caller ownership of callback objects, input, and state. Omitted
  factories supply zero values, including nil pointers. No arbitrary deep
  copies, automatic pointer repair, or concurrent-dispatch guarantee is added.
- Keep machinery proportionate. Private file and helper names below identify
  a concrete implementation route; equivalent simple organization within the
  owning package is acceptable. Public names and signatures are settled.
- Apply `gofmt` to changed Go files. Add focused tests alongside the stage that
  owns each behavior; reuse existing helpers where useful without duplicating
  PromptKit's suite. Current tests are evidence, not a source of new policy.
- Preserve the ignored local workspace and any existing replacements. Do not
  commit workspace files or sibling-directory replacements. The selected
  upstream PromptKit must be the released v0.15.0 module.
- Commit each completed nonempty stage under the
  [agent commit policy](../policy/agents.md#git-commit-policy). Record concise
  completion markers and actual validation evidence beneath the corresponding
  stage here. Never mark a future stage complete merely because its tests pass
  incidentally. Do not push, tag, publish, or rewrite historical release notes
  as part of this sprint.

The validation commands in this plan are for the implementation sprint. This
planning pass creates documentation only.

## Stage 1: Adopt the released PromptKit dependency and notices

**Primary files:** `integrations/ckprompt/go.mod`, its `go.sum`, and
`integrations/ckprompt/docs/oss/attribution.md`.

1. Upgrade only the selected PromptKit requirement to v0.15.0. Keep Go 1.26.0,
   ConfigKit v0.7.1, and the existing CommandKit prerequisite version selections.
   Review any actual transitive graph changes rather than upgrading unrelated
   dependencies.
2. Resolve module metadata with the real released PromptKit. If unpublished
   CommandKit prerequisite versions prevent workspace-off operations, use
   temporary copies of ckprompt's module and sum files as described under
   Stage 9. Add replacements only for core and ckconfig in those copies, run
   the targeted upgrade/tidy there, then remove those temporary replacements
   before transferring the resulting metadata to the owning checked-in files.
   Preserve unrelated existing checksum entries unless normal tidy removes them.
3. Update the **integration's** PromptKit attribution entry from v0.14.0 to
   v0.15.0 and pin its source/notice links to
   `635674c066d0b432187f96e72b9c03d085443ac5`. The root attribution document is
   not the owner of this entry. Review released copyright/license material;
   update other entries or texts only for actual changes in selected material.
4. Confirm the released dependency provides `Engine.ValidateRequest` and that
   no committed manifest includes a sibling PromptKit replacement.

**Validation:** Run ckprompt's existing suite with the upgraded dependency;
inspect its selected module version and metadata diff. Use the module practices
in Stage 9 if required. Prose and attribution checks follow
[OSS](../policy/oss.md) and [documentation](../policy/documentation.md) policy.

**Complete when:** ckprompt selects released PromptKit v0.15.0, its current
suite passes, notices identify that release, and module boundaries are unchanged.

**Completed:** 2026-10-12. Updated ckprompt to released PromptKit v0.15.0 and
refreshed its integration attribution at commit
`635674c066d0b432187f96e72b9c03d085443ac5`. `GOTOOLCHAIN=local go test
-count=1 ./...` passed from `integrations/ckprompt`; `go list -m` selected v0.15.0.
The released source contains `Engine.ValidateRequest`; PromptKit's LICENSE and
COPYRIGHT are unchanged from v0.14.0. The dependency graph, other versions, and
checked-in replacement metadata are unchanged.

## Stage 2: Add the public capability and bundled validation adapter

**Primary files:** `integrations/ckprompt/adapters.go`, a new owning-package
`validation.go`, and focused adapter tests.

1. Add these exact public contracts, with GoDoc:

   ```go
   type ValidationRequest struct {
       Selection ckconfig.Selection
       Request   promptkit.RunRequest
   }

   // Additional field in Suite[I, S]:
   ValidateRequest func(
       context.Context, ValidationRequest, I, S,
   ) (commandkit.Acquired[struct{}], error)

   func WithRequestValidation[I, S any](
       suite Suite[I, S],
       load func(context.Context, ckconfig.Selection, I, S) (*promptkit.Engine, error),
       build func(context.Context, ValidationRequest, I, S) (promptkit.RunRequest, error),
   ) Suite[I, S]
   ```

2. Return the supplied suite with its engine-backed `ValidateRequest` callback
   attached. Preserve other callbacks, factories, selectors, checks, and
   attached issues. With a nil loader, retain a wiring error on the returned
   suite and return without constructing a usable new validation callback.
   When adding an issue, use an independent issues slice so the operation does
   not overwrite a source suite's retained metadata through slice capacity.
   Do not deep-copy caller-owned callback objects or selector bindings here.
3. Leave `FromEngine`'s existing five capabilities unchanged. A nil builder is
   valid and means forwarding the seed request unchanged.
4. The attached callback checks cancellation before building, uses the builder
   if present, checks cancellation again after successful building, acquires
   the engine through `loadPromptKitEngine`, and invokes
   `engine.ValidateRequest(ctx, concreteRequest)`. Pass the same `I`, `S`, and
   invocation context to builder and loader. Use the envelope's original
   configuration selection for loading.
5. A builder failure returns its supported error identity and prevents engine
   loading. A loader failure returns the acquired configuration diagnostics and
   error. A native failure returns that error with the acquired diagnostics.
   Check operation errors before substituting a later context error, preserving
   PromptKit's native cancellation category. Reuse `promptKitContextError` and
   the existing engine-loading diagnostic adaptation where applicable.
6. Treat `Acquired[struct{}]` as diagnostics plus an empty value, not a Boolean
   result or nil-result protocol. Builder output is the authoritative concrete
   request: forward it without reapplying seed fields, merging maps, or checking
   its prompt identity independently. Successful builder calls have no separate
   diagnostic-return field in this selected API.

**Validation:** Add direct adapter tests for composition preserving existing
suite members, nil-builder success, builder failure before loading, and retained
nil-loader issues. Confirm installation reports retained issues even after all
capabilities are cleared. Run the focused tests and ckprompt suite.

**Complete when:** the public API and adapter work without command dispatch,
with explicit capability opt-in and the selected wiring/error behavior.

**Completed:** 2026-10-12. Added `ValidationRequest`, the optional suite callback,
and `WithRequestValidation`. The callback forwards a nil-builder seed through
PromptKit validation, preserves original selection and invocation objects, and
retains configuration diagnostics. Adapter tests cover suite composition,
successful nil-builder validation, builder failure before loading, independent
retention of nil-loader issues, and installation reporting after capabilities
are cleared. Focused adapter tests and
`GOTOOLCHAIN=local go test -count=1 ./...` passed from
`integrations/ckprompt`.

## Stage 3: Verify native delegation, request construction, and cancellation

**Primary files:** a focused `integrations/ckprompt/validation_adapters_test.go`
or equivalent owning adapter tests.

1. Use a real offline PromptKit engine with synthetic prompts/profiles and a
   counting or fail-on-call model client. Demonstrate successful concrete
   validation and a missing required binding, asserting zero model calls.
2. Compare representative adapter failures with native `ValidateRequest` using
   the same concrete request, or assert the documented native category with
   `errors.Is`. Cover an invalid selection and missing/unloadable artifact
   without reproducing every native preparation rule.
3. Demonstrate that a builder receives the seed, context, selection, typed input,
   and shared state, then can add inline bindings and request settings. Use an
   observable native effect, such as supplying a needed variable/artifact or
   invalid appended message, to detect discarded builder output. The adapter
   must not silently restore seed fields afterward.
4. Cover a nil engine result and a loader's real `*configkit.LoadError`, including
   its retained diagnostic records and error identity. Reuse the existing
   fixture and diagnostic helpers rather than a new abstraction framework.
5. Cover cancellation before building, after successful building, and during
   or after loading. Verify later work is skipped and already acquired
   diagnostics survive. For cancellation returned during native artifact
   loading, assert PromptKit's applicable operation category; do not require
   a general `errors.Is(err, context.Canceled)` relationship for native errors.

**Validation:** Run these focused adapter tests and the ckprompt suite. Prefer
behavioral assertions; exact ordering assertions belong only to the explicitly
selected builder-before-loader lifecycle.

**Complete when:** the adapter's native authority, builder forwarding, failure
handling, and cancellation boundaries have meaningful public-boundary evidence.

**Completed:** 2026-10-12. Added real-engine adapter coverage for native success,
missing required input, invalid profile, and an unloadable file reference. The
builder test supplies a required inline artifact and verifies context, selection,
typed input, and state reach both callbacks; validation succeeds without a model
call. Tests cover nil engines, retained `*configkit.LoadError` diagnostics and
identity, cancellation before and after building, cancellation during and after
loading, and PromptKit's native artifact-load category when an artifact reader
returns cancellation. Focused adapter tests and
`GOTOOLCHAIN=local go test -count=1 ./...` passed from
`integrations/ckprompt`.

## Stage 4: Implement the validate command and requirements group composition

**Primary files:** `integrations/ckprompt/command.go`, a new private
`validation_command.go` in the same package, and basic command tests.

1. Include `Suite.ValidateRequest` in `Command`'s capability/absence check.
   Assemble one `requirements` group from the independently supplied show and
   validate declarations. Continue attaching retained suite issues and using
   the existing root-only family declaration. Update help summaries to describe
   inspection and validation without changing catalog/definition routing.
2. Introduce a private per-invocation validation input holding `ValidationRequest`,
   the format, captured selector values, and prepared application input. Build
   a normal typed `CommandSpec` named `validate`, with one required prompt ID.
   Use `CaptureFlags` and `prepareSuiteInput` for application selectors/checks.
3. Bind the following fields; defaults and repetition follow the accepted
   roadmap and current requirements bindings:

   | CLI input | Destination |
   | --- | --- |
   | Prompt ID operand | `Request.PromptID` |
   | `--config`, `--overlay` | Presence-aware `Selection.Config`, `Selection.Overlay` |
   | `--version` | `Request.PromptVersion` |
   | `--profile` | `Request.ProfileID` |
   | `--session-id` | `Request.SessionID` |
   | Repeatable `--var KEY=VALUE` | `Request.Vars` |
   | Repeatable `--input NAME=PATH` | `Request.Inputs[NAME] = promptkit.File(PATH)` |
   | `--format` | Private format, default `yaml` |
   | Suite selectors | Captured bindings applied to initialized `I` |

4. Parse file assignments at the first `=` and preserve the name and path
   exactly. Parse every occurrence, then let the last duplicate name win.
   Missing separators are usage errors; empty names/paths go to native
   validation. Do not read files, normalize paths, expand them, or rebase them
   while binding. Reuse variable-assignment mechanics where this is simple,
   keeping their existing behavior unchanged.
5. Reject unsupported formats before application input preparation. Use the
   existing YAML-only format check where appropriate. Missing operands,
   malformed bindings, and unknown flags retain existing usage handling.
6. In the handler, check cancellation, initialize state using the existing
   zero-value convention, check cancellation again, and invoke the captured
   validation callback. On error, return `definitionFailure` with diagnostics.
   After successful validation, check cancellation before encoding/delivery.
   A successful empty acquisition value is valid; there is no inspection value
   to require or dereference.
7. Encode a private YAML DTO with explicit `prompt_id` and `valid` fields. Use
   the command's requested ID and `valid: true`, with no resolved metadata or
   request dump. Reuse `encodePromptYAMLBytes` and `writePromptReport`, returning
   diagnostics for core delivery. Do not change established report/diagnostic
   delivery order, short-write handling, or failure bookkeeping.

**Validation:** Add a basic assembled command test with a capturing custom
callback and semantic YAML decoding. Verify the correct seed reaches the
callback, and both show and validate can be installed in one requirements group.
Run focused tests and the ckprompt suite.

**Complete when:** the new command invokes the selected provider and produces
the agreed YAML confirmation using existing core mechanics.

**Completed:** 2026-10-12. Added `requirements validate` with typed selector
preparation, presence-aware configuration selection, PromptKit request flags,
repeatable variable and file bindings, and YAML confirmation output. `Command`
now composes supplied show and validate declarations in one requirements group
and retains issue reporting for absent capabilities. The assembled command test
verifies callback values, first-`=` file paths, semantic YAML output, and both
operations in help. Focused command tests and
`GOTOOLCHAIN=local go test -count=1 ./...` passed from
`integrations/ckprompt`.

## Stage 5: Cover command grammar, capability availability, and input lifetimes

**Primary files:** a focused `integrations/ckprompt/validation_command_test.go`
and existing requirements command tests where an extension is simpler.

1. Exercise validation-only, discovery-only, both, and absent suites. Check
   installed help/availability and the existing namespace collision/root-only
   behavior for the new validation-only declaration, reusing core coverage
   rather than retesting the registry in full.
2. Verify construction, installation, help, and version leave input/state
   factories, builders, and loaders uninvoked. Capture a command, then mutate
   source callbacks/selector metadata to confirm the established eager-capture
   behavior includes the new callback. Callback-owned referenced objects stay
   caller-owned.
3. Capture standard request fields and typed selectors in a custom provider.
   Cover explicit empty config/overlay presence, empty scalar native semantics,
   singleton repetition, and repeated variable/file assignments. Include paths
   containing `=`, empty names/values, whitespace/commas, duplicate names, and
   a malformed earlier occurrence followed by a valid one. Inspect native file
   references without requiring nonexistent paths to pass actual validation.
4. Cover missing/extra operands, unsupported format, unknown flags, malformed
   `--var`/`--input`, and required suite selectors. Confirm applicable syntax
   failures stop before application preparation/acquisition. Input-check
   failures stop state creation and provider work using current classifications.
5. Run multiple invocations on one App and verify parsed Vars/Inputs, selection,
   and selector presence do not leak into later calls. Verify one reached input
   preparation/check and state creation per selected validation invocation,
   and omitted factories supply zero values including nil pointers. Do not
   promise isolation for aliases created by application factories.

**Validation:** Run the focused command tests and ckprompt suite. Decode each
YAML output into a fresh value; avoid output-prefix or key-order assertions.

**Complete when:** the new capability obeys the selected standard grammar,
availability, capture, and invocation-lifetime contracts.

**Completed:** 2026-10-12. Added validation-only availability, namespace
collision and root-only placement coverage; construction/install/help/version
callback timing; eager callback and selector capture; exact standard request,
selector, empty-presence, repeated-assignment, first-`=` file, and duplicate
last-wins behavior; repeated-invocation isolation; usage preflight; input-check
failure ordering; and nil-pointer zero-value factories. A missing required
captured selector is a usage outcome before `Suite.CheckInput`, state creation,
or provider work. Under the existing `CaptureFlags`/`prepareSuiteInput` flow,
`Suite.NewInput` runs before `FlagValues.ApplyTo` reports that missing selector.
Focused validation command tests and
`GOTOOLCHAIN=local go test -count=1 ./...` passed from
`integrations/ckprompt`.

## Stage 6: Cover native command outcomes and report delivery

**Primary files:** validation command tests and small real-engine fixtures.

1. Exercise `WithRequestValidation` through `Command` and `App.Execute`, with
   real native success, missing required binding, and one selection failure.
   Include a useful nil-builder request built entirely from standard flags.
   Assert no model call and preserve native error identity; compare categories,
   not exact human-readable error wording.
2. Cover custom-provider diagnostics on success and error. Builder, loader,
   and native failures are operational outcomes/status 1. CLI usage failures
   are status 2. Retain warning diagnostics without turning warning severity
   alone into failure. Reuse existing input-check handling.
3. Verify cancellation before acquisition and after a successful callback
   prevents later work/report output. Preserve callback errors and native error
   categories rather than overwriting them with a general context error.
4. Reuse existing writer doubles to cover report errors, short writes, and
   diagnostic writer errors. Preserve `ReportIncomplete`, `DeliveryError`,
   diagnostic-delivery state, and supported cause identity as established by
   the shared delivery contract. Do not add rollback or blocking-writer
   interruption guarantees.
5. Decode the success report and assert only the agreed confirmation fields
   are emitted, without values/artifact contents/rendered messages. Keep the
   existing partial-discovery-success tests as regression coverage; do not
   duplicate their whole matrix for validation.

**Validation:** Run focused tests and the complete ckprompt suite, including
its maintained composed consumer tests.

**Complete when:** assembled validation follows native semantics and existing
outcome/report delivery contracts under representative successes and failures.

**Completed:** 2026-10-12. Exercised the composed command with real PromptKit
success, missing required input, and invalid profile selection; success emits
only `prompt_id` and `valid`, and the fail-on-call model client remains unused.
Command tests cover builder, ConfigKit loader, custom provider, and native
failures; warning diagnostics on success and error; cancellation before and
after acquisition; callback-error precedence; report writer errors, short
writes, and diagnostic writer errors with the established outcome fields.
Focused outcome tests and `GOTOOLCHAIN=local go test -count=1 ./...` passed from
`integrations/ckprompt`.

## Stage 7: Extend the maintained composed consumer

**Primary files:** `integrations/ckprompt/examples/composed/main.go`, its tests,
README, and synthetic fixtures under its existing `fixtures/` directory.

1. Compose `WithRequestValidation` into the existing `FromEngine` suite, keeping
   its factories, campaign selector, catalogs, definitions, discovery, and
   separately installed config/session commands.
2. Add a typed builder that starts from the seed and uses the existing
   application configuration loader and invocation cache. Load the selected
   campaign through that shared state, then default a missing `topic` variable
   to a synthetic campaign-specific topic and missing `mode` to `""`. Use map
   membership, preserving supplied empty values. Keep seed identity, scalar
   context, and operator artifact bindings. The engine loader can then reuse
   that application load through the same cache. This is example behavior,
   not a CommandKit cache/defaults policy.
3. Keep the existing synthetic prompt template and required `topic` artifact.
   Add secret-free `fixtures/topic.txt` and `fixtures/notes.txt`. Do not supply
   the required topic artifact implicitly in the builder: its absence is the
   useful validation failure. The existing profile uses an inert endpoint and
   no real credentials; validation remains offline.
4. Add README commands, from the composed directory:

   ```sh
   go run . promptkit requirements validate briefing --campaign spring --input topic=fixtures/topic.txt
   go run . promptkit requirements validate briefing --campaign spring --input topic=fixtures/topic.txt --var mode=full --input notes=fixtures/notes.txt --format yaml
   go run . promptkit requirements validate briefing --campaign spring
   ```

   The first two succeed; the last intentionally fails for the missing required
   topic binding. Preserve the existing show examples, explaining their distinct
   discovery result. Describe success as passing preparation at that invocation.
5. Add focused composed-consumer evidence for successful file-backed validation,
   missing-topic failure, and the builder preserving explicitly empty variables
   while supplying defaults only for absent keys. Test builder behavior directly
   where needed to observe precedence without inventing a library enforcement
   rule. Preserve current campaign and discovery tests.

**Validation:** Run composed tests and the documented successful/failing
commands. Record the expected classifications and semantic reports rather than
making the intentionally failing example a failed completion check.

**Complete when:** the maintained consumer demonstrates the accepted feature,
typed selection, shared state, and application-owned defaults without execution.

**Completed:** 2026-10-12. Composed `WithRequestValidation` with the existing
engine loader, campaign selector, input/state factories, and invocation cache.
The typed builder loads the selected campaign through that cache and defaults
only absent `topic` and `mode` variables, preserving explicit empty values and
operator bindings. Added offline topic and notes fixtures and extended the
consumer README with both successful file-backed commands and the intentional
missing-topic failure. Composed tests passed with
`GOTOOLCHAIN=local go test -count=1 .`; both successful README commands emitted
the expected two-field YAML confirmation, and the missing-topic command exited
with status 1 for the native required-artifact error.

## Stage 8: Update canonical documentation for implemented behavior

**Primary files:** GoDoc, `docs/cli.md`, `docs/consumers/guide.md`, README,
`docs/development.md`, and `docs/internal/overview.md` as affected.

1. Complete GoDoc for `ValidationRequest`, `Suite.ValidateRequest`,
   `WithRequestValidation`, and command factory behavior. Update the existing
   `InspectRequirements` comment so omission refers to `requirements show`,
   rather than implying the entire requirements group is absent. State optionality,
   nil-builder forwarding, retained wiring issues, builder-before-loader order,
   callback ownership, and native error/diagnostic treatment.
2. Extend the CLI requirements section with separate show/validate descriptions,
   exact flags and repetition rules, file-binding semantics, YAML confirmation,
   and status meanings. Scope existing discovery-only statements appropriately:
   concrete validation loads resources and renders through native preparation;
   it returns encountered errors and no snapshot/model output.
3. Update consumer guidance from five independently optional callbacks to six,
   distinguishing `FromEngine`'s five supplied callbacks from explicit validation
   composition. Include minimal full-suite and validation-only/nil-builder
   wiring, and link to the maintained consumer for the complete builder example.
   Recommend shared validation/execution request construction without prescribing
   downstream architecture.
4. Update module summaries and implemented internal inventory to include concrete
   validation. Audit statements about no rendering/resource loading/credential
   resolution so metadata-inspection descriptions do not accidentally cover the
   new capability. Keep native semantics linked to adopted upstream contracts.
5. Leave accepted ADRs, policy boundaries, historical releases, and unrelated
   deferred proposals intact. No new ADR or policy revision is required by this
   accepted feature. Current API/CLI documentation becomes the canonical owner
   when implementation lands.

**Validation:** Review examples against implemented signatures and commands;
check local links/anchors and Markdown whitespace. Reuse Stage 7's unchanged
example evidence. Check prose for accidental stronger guarantees or consumer
mandates.

**Complete when:** current documentation accurately describes the implemented
capability under its assigned owners, with no stale five-callback or
metadata-only claims covering request validation.

## Stage 9: Validate the complete end state and finish

1. Format changed Go files and inspect the final diff for scope, generated junk,
   local replacements, and unrelated changes. Keep `deferred.md` intact unless
   a statement actually becomes stale through this feature.
2. Run the complete workspace suites from all three module directories:

   ```sh
   GOTOOLCHAIN=local go test -count=1 ./...
   (cd integrations/ckconfig && GOTOOLCHAIN=local go test -count=1 ./...)
   (cd integrations/ckprompt && GOTOOLCHAIN=local go test -count=1 ./...)
   ```

   Root `./...` does not traverse nested modules. These ordinary suites do not
   authorize a release publication exercise or live model service tests.
3. Separately verify each module with `GOWORK=off`, `GOTOOLCHAIN=local`, and
   `-mod=readonly`. For unpublished CommandKit prerequisites, create temporary
   absolute-path copies of each affected module's `go.mod` and `go.sum`, keeping
   paired names such as `ckprompt.mod` and `ckprompt.sum`. Run from that module's
   real directory with `-modfile=/absolute/temp/path/ckprompt.mod`.

   - Core: use its actual manifest; no optional-library replacements.
   - ckconfig: replace only its required core version with the current root.
   - ckprompt: replace only its required core and ckconfig versions with their
     current local directories.
   - Use version-specific replacements for the versions in the manifests.
     PromptKit and ConfigKit remain real released dependencies. Temporary
     replacements are validation scaffolding, never committed metadata.
   - Prepare any needed cache/checksums in the temporary copies, then record
     their hashes before running readonly tests. Verify those files are unchanged
     afterward. A workspace-only pass does not establish isolated resolution.

   In each module, use `go test -mod=readonly -count=1 ./...` with the applicable
   `-modfile` and environment above. Inspect `go list -m all` and
   `go list -deps -test ./...` with the same isolated manifest: core has no optional
   integration dependency; ckconfig has no PromptKit/ckprompt dependency;
   ckprompt selects released PromptKit v0.15.0. Do not change checked-in
   prerequisite versions merely to make isolated checks easier. Remove temporary
   module files after retaining the relevant validation evidence.
4. Review the Stage 7 commands and canonical docs together. Check that the
   nil-builder path and typed builder are both covered, show still permits
   partial discovery, errors preserve native categories, and only the agreed
   YAML confirmation is emitted on validation success.
5. Run relevant prose whitespace/link checks, including staged documentation.
   Confirm the integration attribution matches the selected released dependency.
   No new coverage threshold, benchmark, platform matrix, release gate, or
   threat scope is introduced by this plan.
6. Mark completed stages with actual evidence, update the feature roadmap's
   implementation status, and record any genuine remaining limitation accurately.
   Commit outstanding sprint changes using a concise plain-English message,
   inspect status afterward, and report the final commit(s) and validation.
   Retain planning documents until their cleanup is requested.

**Complete when:** every accepted roadmap capability is implemented, all
applicable evidence above passes, public documentation and module metadata are
aligned, and the workspace is clean. No unresolved product or API decisions
were identified in preparing this plan.