npm.io
0.2.0 • Published 2d ago

@intentsolutions/refiner

Licence
Apache-2.0
Version
0.2.0
Deps
2
Size
52 kB
Vulns
0
Weekly
0
SummaryDependencyVersions

@intentsolutions/refiner

Skill Refiner orchestrator + I/O adapters + CLI — wave 2 of the Skill Refiner buildout (Phase A). The pure foundation it builds on is @intentsolutions/refiner-core (wave 1).

Published as @intentsolutions/refiner@0.1.0 to npm under the @intentsolutions scope per CEO directive 2026-06-21. CLI identity remains j-rig refine (unchanged).

This is the I/O half of the value-oriented Refiner discipline whose pure half lives in @intentsolutions/refiner-core. It supplies the four pieces of plan 027 § 4 Phase A build-order steps 4–7:

Step Piece Module
4 Content-addressed store + append-only event log + single mutable best-pointer store.ts (RefinerStore)
5 score() — delegate to the existing j-rig eval via an injectable shell-out score.ts
6 propose() — tiered (haiku|sonnet, never opus) Anthropic-backed RefinerModel wired to a refiner-core RefinerStrategy propose.ts
7 The 5 j-rig refine <cmd> CLI commands cli.ts

Reuse, not reinvention

Per the plan's "Reuses existing infrastructure" mandate:

  • score() delegates to j-rig eval — it shells out to the already-shipped j-rig eval <skill-dir> --json command and maps that command's scoreCard (pass_rate) into a refiner-core ScoreRecord. No new scorer is built.
  • propose() wires refiner-core's RefinerStrategy (NaiveInContextStrategy / SkillOptStyleStrategy, both from wave 1) to a model-completion seam. It reuses the repo's SDK-free Anthropic-Messages-API convention (an injectable transport, no @anthropic-ai/sdk dependency), the same approach the eval command's RealAnthropicProvider uses.

Injectable boundaries (so tests need no live evaluator / SDK / disk)

Every side-effecting boundary is injected:

  • RefinerStore takes a FileSystem seam → tests use an in-memory fake.
  • score() takes an EvalRunner seam → tests assert it delegates + maps without spawning j-rig.
  • propose() / createRefinerModel() take a CompletionClient → tests mock the model (and assert tiered routing rejects opus) without ANTHROPIC_API_KEY.

CLI

j-rig refine bootstrap <skill-dir>   # synthesize a held-out eval set (offline)
j-rig refine score <skill-dir>       # delegate scoring to `j-rig eval` (needs a model key)
j-rig refine propose <skill-dir>     # propose a bounded edit (needs ANTHROPIC_API_KEY)
j-rig refine apply <skill-dir> --proposal <hash>   # apply a stored proposal → new version (offline)
j-rig refine status <skill-id>       # show the store + event log (offline)

Not in this wave (gated / later)

The SkillVersion kernel entity, the skill-refiner-pass/v1 predicate URI + signed evidence emission, the Claude Code plugin + 3-layer hooks, and the synchronized npm release ceremony are all gated to later waves.

License: Apache-2.0.

Keywords

skill-refinerj-rigintent-evalevalskillclaudecli