Files
commandkit/docs/roadmap/invalid-configuration-outcomes.md

238 lines
13 KiB
Markdown

# Invalid configuration outcomes
Status — proposed implementation roadmap. The user has accepted the targeted
integration behavior. The decisions at the end identify remaining policy and
API choices for the implementation plan.
## Purpose
Make standard ckconfig and ckprompt commands return usage status `2` for
identifiable invalid configuration input, while returning operational status `1`
for cancellation and other operational failures. Preserve actionable upstream
diagnostics and supported typed error identity through the integration boundary.
CommandKit owns this CLI classification. ConfigKit and PromptKit own loading,
validation, diagnostic meaning, and their public error contracts. Application
commands own their domain outcomes. These responsibilities follow the
[architecture policy](../policy/architecture.md#ownership-boundaries).
## Current baseline
The reviewed baseline is CommandKit commit
`1e92a87706e44684ee34395245e5b633474677e0`, following the publication of
ckconfig v0.7.0 and ckprompt v0.7.0. Both integration modules select core
v0.6.0 and ConfigKit v1.2.0. ckprompt selects ckconfig v0.7.0 and PromptKit
v0.21.0. Their manifests own the dependency selections.
The consumer report reproduced malformed operator YAML in Audita, including
its automatic `prompts.yml` companion and an explicitly selected application
configuration. Delegated inspection returned `yaml.invalid` with status `1`,
while Audita's processing command returned the same diagnostic with status `2`.
The report's reproduction is retained in Audita's
[upstream handoff](https://forge.maximumdirect.net/eric/audita/src/commit/a355f05a273d6ac9d577eab8d30caa60cb74eee8/docs/roadmap/implementation.md).
The review confirmed the behavior through Audita's existing regression tests and
a temporary public-API reproduction against the aligned CommandKit workspace.
The latter covered catalog listing and all five configuration operations in both
the application and PromptKit scopes, including wrapped ConfigKit loading errors
and already-canceled contexts.
- [FromConfigKit](../../integrations/ckconfig/configkit_adapter.go) and
[the PromptKit adapters](../../integrations/ckprompt/adapters.go) preserve the
ConfigKit loading error and adapt its diagnostic evidence.
- [configFailure](../../integrations/ckconfig/operations.go) and
[definitionFailure](../../integrations/ckprompt/command.go) construct operational
outcomes for the affected failures. Configuration comparison and validation
also reach the former helper.
- ConfigKit's public `LoadError` represents several kinds of loading failure,
including malformed YAML, acquisition errors, and schema or callback problems.
Its public diagnostic codes provide the classification evidence. Its error
contract intentionally exposes safe diagnostics and preserves its own typed
identity across CommandKit adaptation.
- `config sources` can report discovered source evidence after failed loading,
and `config validate` can report failed validation. Cancellation observed
before acquisition leaves stdout empty in the reviewed command paths.
The current [CLI contract](../cli.md#outcome-status) assigns loading failures
status `1`, with more specific statements in the configuration comparison,
requirements, validation, and smoke sections. This roadmap proposes a deliberate
change to that published status behavior.
## Targeted end state
| Evidence available to the integration | Selected outcome |
| --- | --- |
| ConfigKit error diagnostics that identify invalid configuration input under the selected mapping | Usage, status `2` |
| Caller cancellation or deadline observed at an existing operation checkpoint, or present in the returned error chain | Operational, status `1` |
| I/O, schema or wiring defects, callback execution failures, and other recognized operational failures | Operational, status `1` |
| An unrecognized failure, including a loading error with insufficient classification evidence | Operational, status `1` |
| Successful acquisition with warnings | Follow the command's successful-result behavior |
Classification uses upstream structured evidence. Diagnostic text, filenames,
and source labels serve presentation and provenance. The exact input-error code
mapping is proposed under Open Decisions.
The selected outcome retains the original returned cause, including ordinary
wrapping and joined errors, so supported `errors.Is` and `errors.As` queries
continue to work. Existing explanations, source context, diagnostic visibility,
and operation-specific reporting lifecycles apply to the newly classified
failures. Shared rendering and delivery failures follow the
[core output-failure rules](../cli.md#output-failures), including their effect on
usage outcomes.
Apply the classification to configuration acquisition failures throughout the
standard suites, including failures returned by application callbacks when they
carry the supported ConfigKit evidence. CLI syntax, resource lookup, request
execution, and application-specific error semantics retain their owning
contracts. The work concerns interpreting returned failure evidence at the CLI
boundary.
## Work needed
### 1. Define the configuration-failure classification
- Inventory ConfigKit v1.2.0's public diagnostic codes and their documented
meaning. Record the selected input-error cases explicitly using upstream
constants. Review all code families, including composition, imports,
inheritance, versions, overrides, secrets, and configured limits.
- Use error-severity diagnostics as failure evidence. Warning records can
accompany a failure or a successful result.
- Handle wrapped `LoadError` values and standard joined error branches. A failure
selects usage only when its failing evidence is classified as invalid input.
An operational or unclassified failing branch makes the combined result
operational. Preserve the original cause tree.
- Account for cancellation and deadlines before choosing usage. When an existing
checkpoint observes cancellation alongside an acquisition failure, preserve
both causes using the established error-joining approach.
- Keep classification within integration ownership, using the sharing approach
selected below. Existing core outcome constructors and reporting provide the
required mechanics.
### 2. Apply it throughout ckconfig
Wire the classification into acquisition failure handling for `config show`,
`config sources`, `config validate`, `config explain`, and `config diff`.
Follow each operation's current preparation, aggregation, and reporting flow.
In particular, source and validation reports can accompany a failed load.
Detailed validation filters control displayed findings, while the complete
underlying failure supplies classification evidence. Comparison must classify a
failure on either loading side and retain the side-specific diagnostic context.
Include native ConfigKit scopes, PromptKit configuration scopes, and custom
scope callbacks returning the supported evidence. Scope selection and input
preflight continue through their existing contracts.
### 3. Apply it throughout ckprompt
Cover engine-acquisition failures in catalogs, named prompt/profile/backend
inspection, requirements discovery, concrete request validation, prepared-request
preview, individual smoke operations, and aggregate smoke testing.
Acquisition evidence may arrive through loaders or application request builders.
Use the same classification policy wherever it reaches the standard command's
failure boundary. For joined aggregate smoke errors, an operational or
unclassified failure takes precedence over input failures. Follow the existing
aggregate execution and report lifecycle when selecting the final outcome.
### 4. Verify observable integration behavior
Extend focused public-command tests where they provide regression protection.
Use upstream errors as fixtures to establish CommandKit's classification and
preservation behavior, following the
[integration testing policy](../policy/testing.md#upstream-integration-testing-ownership).
Cover these distinct risks with a proportionate set of cases.
- Malformed main and companion configuration produce status `2` with
`yaml.invalid` and the typed ConfigKit cause available through wrapping.
- Representative decoding and selected semantic-validation errors receive the
intended classification. Operational and unrecognized loading diagnostics
receive status `1`, including mixed failing evidence and warnings alongside
errors.
- The five configuration operations and affected PromptKit command families use
the policy, including comparison-side failures and aggregate smoke outcomes.
- Cancellation and deadlines receive status `1` and preserve their supported
identity. Already-canceled invocations retain the reviewed empty-output
behavior. A failed acquisition combined with cancellation preserves both causes.
- Source and validation reports retain their operation-specific failure evidence.
Detailed display filters do not determine failure classification.
- Output failures continue to follow the shared reporting contract when the
command has selected usage status `2`.
Run the complete deterministic suites independently for each changed integration
and with the aligned local workspace as described in the
[developer guide](../development.md#documentation-validation). Exercise the
maintained examples where their existing assertions cover affected outcomes.
The Audita reproduction supplies supporting consumer evidence. CommandKit-owned
tests provide the maintained regression coverage.
### 5. Document and prepare the integration releases
Update the CLI status table and affected operation sections with the implemented
classification. Update any affected exported GoDoc and consumer guidance under
their canonical owners. Explain the classification boundary and fallback through
the chosen API's documentation.
Prepare ckconfig v0.8.0 followed by ckprompt v0.8.0, with ckprompt selecting the
published ckconfig version. The status-contract change warrants minor releases
under the [compatibility policy](../policy/release.md#versioning-and-compatibility).
The release notes should explain that identifiable invalid configuration now
selects status `2`, alongside the operational classification and preserved error
evidence. Follow the [release procedure](../release.md) when publication is
authorized. Implementation completion and publication are tracked separately.
## Completion evidence
Implementation is complete when the accepted classification is exercised through
the public integration commands, the focused regression cases and selected
integration suites pass, and the canonical documentation describes the
implemented behavior. Review the final diff for the agreed scope and retire this
roadmap under the normal documentation lifecycle after recording any unfinished
release work.
## Open Decisions
### Exact diagnostic-code mapping
**Recommendation.** Use an explicit mapping based on ConfigKit v1.2.0's public
code semantics. Include YAML-content and decoding errors, explicit
`version.error`, `resolve.error`, and `validate.error` findings, and other codes
whose documented meaning clearly identifies invalid configuration content.
Classify schema/options defects, acquisition failures, hook execution failures,
invalid callback paths or records, and validation mutation as operational.
Keep ambiguous limit, secret, and source-selection cases operational until their
individual meaning supports input classification. Review and record those cases
when preparing the implementation plan.
This uses the evidence upstream already supplies and distinguishes reported
configuration rejection from callback machinery failing. Source evidence
alone may be insufficient to establish whether application-supplied data came
from an operator, so the rule follows the meaning recorded in upstream
diagnostics.
**Alternative.** Add an application-supplied classification hook for cases that
depend on source ownership or application policy, alongside the built-in mapping.
Applications can then identify input failures using knowledge their loaders
possess. This supports genuinely contextual cases, but expands the public
integration contracts and creates additional configuration and precedence rules.
The current reproduction is addressable through existing typed diagnostics, so
the additional hook has limited demonstrated value in this scope.
### Shared classification interface
**Recommendation.** Place one narrowly documented classification function in
ckconfig and have ckprompt call it through its existing dependency. Keep command
explanation construction and reporting in their owning handlers. Select its
exact signature during implementation planning based on the needed cause and
cancellation inputs. A shared implementation reduces drift across the two
standard families and fits ckconfig's ownership of ConfigKit adaptation. The
public function adds an API maintenance obligation.
**Alternative.** Maintain small private classifiers in both integration modules,
with matching behavioral cases. This keeps the helper private and supports
package-local evolution, but duplicates the diagnostic mapping and joined-error
rules. Future upstream code changes would require coordinated edits in both
modules. Prefer this approach if the shared signature would expose more policy
or machinery than the demonstrated integration need warrants.