Skip to content
Explanation

What ships, and what would stop it

The order is deliberate: the ledger and the diff come first, because a standard nobody consumes is a wiki page. Consumption and badges ship last, not first.

v0.1 — CLI, ledger, read API

scan, diff, attest and verify with tests, a ledger-as-repo with PR validation, the capability-diff GitHub Action, and the OpenAPI read spec.

In progress

Launch — publish real diffs

Run the tool against high-star public skills and publish the capability manifests and version diffs for anyone to argue with.

Next

v0.2 — eval harness

Baseline-versus-treated runs across pinned models, a judge distinct from the task model, caching by content hash, and a budget that hard-stops.

Planned

v0.3 — native consumption

Badges, a five-line fetch library for harnesses and registries, and a first conversation with a vendor about native support.

Conditional

Where v0.1 actually stands ​

The command surface, the Zod schemas, the generated OpenAPI contract and these docs are real. Inside them, the parts that matter most are still placeholders:

  • the scanner reads markdown only and records placeholder values on a match;
  • nothing is signed — attest writes a local record, verify always succeeds;
  • content_hash is not a digest yet, so records are not content-addressed;
  • ledger/, actions/capability-diff/ and evals/ are not in the repository.

Only diff plus its exit code is solid enough to gate a pull request on today. AGENTS.md in the repository root tracks the same list against the source files.

Launch ​

Run the tool against high-star public skills and publish the capability manifests and version diffs. One human-confirmed undeclared change is worth more than any adoption count.

v0.2 — eval harness ​

Baseline-vs-treated runs across pinned models, judge model distinct from task model, caching by content hash, --budget-usd hard-stop, signed delta records with per-model pass-rate pages.

v0.3 — native consumption ​

Badges, a five-line fetch library for harnesses and registries, and a first conversation with a vendor about native support. A standard nobody consumes is a wiki page — so this ships last, not first.

Kill criteria ​

Decided in advance, not rationalised later. We stop if Tessl or the harness vendors ship open model-bump regression for third-party skills, if ten teams running 20+ skills report no drift pain, or if nobody consumes the records within ~60 days of launch. The honest retrospective gets published either way.

Proof-first skill management for agent-first workflows