@harness-one/devkit
v1.0.0
Published
Developer-time toolkit for harness-one: evaluation runners, scorers, component registry, drift detection
Maintainers
Readme
@harness-one/devkit
Developer-time toolkit for harness-one: evaluation runners, starter scorers,
generator-evaluator loops, data-flywheel helpers, component registry, drift
detection, and taste-coding utilities.
Eval + evolve tooling ships from this sibling package rather than a core
subpath. The runtime architecture-checker stays in core under
harness-one/evolve-check; everything dev-tool-shaped lives here.
Install
pnpm add -D @harness-one/devkitharness-one is a peer dependency. Node 18+.
Eval
import {
createEvalRunner,
createBasicRelevanceScorer,
createBasicFaithfulnessScorer,
createBasicLengthScorer,
createCustomScorer,
runGeneratorEvaluator,
extractNewCases,
} from '@harness-one/devkit';createEvalRunnerruns datasets through scorer pipelines and produces reports.createBasicRelevanceScorer/createBasicFaithfulnessScorer/createBasicLengthScorerare baseline starter scorers, not production-optimal judges.createCustomScorerlets you plug in domain-specific or LLM-as-judge scoring.runGeneratorEvaluatorimplements the generate → evaluate → feedback → retry loop.extractNewCasesturns low-scoring results into new regression cases for a data flywheel.
Evolve
import {
createComponentRegistry,
createDriftDetector,
createTasteCodingRegistry,
} from '@harness-one/devkit';createComponentRegistrytracks versioned prompts, tool definitions, guardrails, and retirement conditions.createDriftDetectorcompares live signals against a stored baseline and reports drift severity.createTasteCodingRegistrystores postmortem-derived engineering rules and can export them as Markdown.
Architecture Rules
The runtime architecture rule engine lives on the core subpath:
import { createArchitectureChecker } from 'harness-one/evolve-check';That split is intentional: architecture rules can gate production code, while devkit stays purely developer-time.
Related
harness-oneis the runtime this package evaluates and evolves.@harness-one/clican scaffold starter projects that already import devkit.- Root migration notes live in MIGRATION.md.
