Skip to content

minimal-consumer-artifact.md

Latest commit

 

History

History
97 lines (64 loc) · 4.29 KB

minimal-consumer-artifact.md

File metadata and controls

97 lines (64 loc) · 4.29 KB
Raw
summary V0 contract for the smallest real external consumer artifact and how capability should expand from it.
read_when
minimal consumer artifact
start small
outside-in packaging
research-agent install path
what is the real v0

Minimal Consumer Artifact Contract

This file exists to stop drift between three different things that are easy to conflate:

  1. the packed omniweb-toolkit package consumer path
  2. the exported OpenClaw workspace bundles under agents/openclaw/
  3. the eventual stronger goal of a truly simple outside install + runtime path

Current V0 Verdict

The current smallest proven external consumer artifact is:

  • a packed omniweb-toolkit package tarball
  • installed into a clean temporary consumer workspace
  • imported by package name
  • exercised through a safe no-spend proof

That path is already maintained by:

  • bun run check:package-consumer

What V0 Proves Today

check:package-consumer proves that a fresh consumer can:

  • install the packed package into a clean workspace
  • import omniweb-toolkit, omniweb-toolkit/agent, and omniweb-toolkit/types by package name
  • render a deterministic plan-only dry-run prompt with explicit no-publish / no-spend rules
  • run one safe live read
  • receive write-readiness feedback that reports missing env instead of trying to spend DEM

This is enough to call the package path a real external consumer artifact.

What V0 Is Not

V0 is not the same thing as:

  • a copied exported OpenClaw bundle acting as a full standalone npm consumer package
  • a full live-write runtime path
  • clone-and-go wallet-backed OmniWeb execution
  • a published npm install story from the public registry

Those are later capability layers.

Relationship To The Exported OpenClaw Bundle

The exported agents/openclaw/research-agent/ bundle has a different honest role:

  • it is a lightweight OpenClaw workspace artifact
  • it supports no-install inspection, skill loading, explanation, and lightweight starter smoke
  • it should degrade safely instead of assuming heavy runtime deps

That bundle is truthful and useful, but it is not the current v0 package-consumer artifact.

Expansion Rule

Do not jump straight from v0 to "full dependency load" or "everything should work live".

Expand capability in this order:

  1. v0 package consumer — install, import, dry-run prompt, one safe read, missing-env readiness
  2. simpler research-agent-facing entrypoint on top of the package proof
  3. lightweight OpenClaw bundle parity where the same minimal behavior is easy to reach
  4. dry-run runtime enrichment with deferred optional deps
  5. live-read runtime proof
  6. wallet-backed live-write proof

Each layer should be proven before the next one becomes part of the public claim.

Error Handling Rule

When a new layer fails:

  • narrow the failure to the smallest truthful claim
  • cut a specific follow-up bead
  • fix the layer without pretending the whole stack is already solved

That is the intended development model for this package.

Current Implementation Step

The minimal research-agent path now has four proven layers:

  • importable subpath: omniweb-toolkit/research-agent-minimal
  • packaged example: examples/research-agent-minimal.mjs
  • maintained package proof: bun run check:research-agent-consumer
  • maintained deferred dry-run proof: bun run check:research-agent-dry-run
  • maintained explicit live-read proof: bun run check:research-agent-live-read
  • maintained explicit live-write gate proof: bun run check:research-agent-live-write-gate
  • current proof scope: clean tarball install, package-name import, no-spend dry-run behavior, one safe live read in the package-consumer path, honest runtime capability reporting for missing credentials/dependencies, successful forced deferred dry-run runtime from the source workspace when optional deps are present, explicit read-only starter proof without wallet-backed execution, explicit live-write failure when wallet/runtime prerequisites are absent, and starter-facing routing that keeps evidence-backed publish claims downstream of the maintained attestation gate rather than treating it as optional

That keeps the evolution path honest: prove each layer separately instead of pretending the copied OpenClaw bundle or the package surface already guarantees the full stack.