Graduate @bounded-systems/verbspec to a shared, standalone library

[bdelanghe]

Goal

Turn @bounded-systems/verbspec from a prx-internal workspace package into a shared layer across the @bounded-systems libraries (prx, keeperd, …): author a verb once as a typed VerbSpec, project it to every surface (CLI, MCP, OpenAPI, OpenRPC, JSON-RPC/NDJSON, Anthropic tool), no drift.

Done so far

• Extract the spec-driven CLI core into @bounded-systems/verbspec (PR Extract verbspec into @bounded-systems/verbspec package #594).
• Add OpenRPC + JSON-RPC/NDJSON projections so daemon-shaped libraries are a first-class surface (PR Add OpenRPC + JSON-RPC/NDJSON projections to verbspec #595).

Remaining

• Decouple from Zod via Standard Schema. Today the package hard-depends on Zod v4 (z.toJSONSchema, input.parse/safeParse). Adopting the ~standard interface lets consumers bring Valibot/ArkType. Note: z.toJSONSchema is Zod-specific, so a validator-agnostic JSON-Schema path is needed (per-lib converter, or require schemas that can emit JSON Schema).
• Graduate to its own repo and publish under @bounded-systems/verbspec. PolyForm-Noncommercial ⇒ a private registry (GitHub Packages or similar), not public npm.
• Repoint prx at the published package (drop the workspace paths mapping).
• Wire keeperd as the first external consumer: model sign / verify / status / commit / push / getPublicKey as a VerbSpec registry, serve the NDJSON protocol via dispatchNdjson, and derive lib/keeper.ts client types from the verb output schemas.

Open design questions

• Source of truth: Zod vs an IDL. Keep Zod as the authoring source and emit OpenRPC/OpenAPI (current direction — schema is the runtime validator), or move to a declarative IDL (TypeSpec / OpenRPC) and generate Zod/TS? Recommendation: stay Zod-sourced; reach for TypeSpec only if authoring across many services starts to dominate.
• OpenRPC params style. Currently by-name Content Descriptors from the input's top-level properties; revisit if any verb needs by-position params.

Tracking the shared-layer effort discussed in session.

https://claude.ai/code/session_01XvpnMDCnzCvn4MjDL15vi1