@@ -1,6 +1,7 @@
 ---
 name: releasing
 description: Use when cutting or publishing a release of the @metaobjectsdev npm/TypeScript packages — e.g. "do a release", "release 0.x.y", bumping the published npm version, promoting merged work to npm latest, or refreshing version references after a release.
+disable-model-invocation: true
 ---
 
 # Releasing the npm/TypeScript packages
@@ -40,6 +41,15 @@ Do not bump or publish anything until every check is green. Print each result.
 | **Target version is free** | `npm view @metaobjectsdev/cli@<version> version` (a 404 = free); `git tag -l v<version>` | the target version is NOT already published or tagged (npm versions are permanent — a taken version can never be reused) |
 | CHANGELOG ready | `sed -n '1,20p' CHANGELOG.md` | an entry exists or you will add one (see Phase 8) |
 | **`metamodelVersion` moved if the metamodel did** | `node scripts/check-metamodel-version.mjs` | exit 0. It diffs `expected-registry.json` against the last release tag and fails if the vocabulary changed without the version moving. Read its PROSE warning too — a rule can change with no machine-readable footprint (#210 changed only a `rules` string). Fix with `--set <version>`; detail with `--explain`. |
+| **You CAN publish** | `npm whoami` | prints your username (see the note under this table) |
+
+**Check that you CAN publish.** `scripts/release.mjs` runs `npm whoami` in preflight; do the same
+by hand if you are publishing manually, and use an **Automation / bypass-2FA** token. A "Publish"
+token passes `whoami`, shows `read-write`, passes `--dry-run`, and then `EOTP`s on every write,
+at which point `bun publish` falls back to a browser flow that cannot complete non-interactively
+(`--otp=` is ignored; a code must be piped on stdin). Verify BEFORE the version bump: the bump is
+committed and pushed well before the first publish, so a dead credential strands the cut with
+other registries already shipped. See `docs/RELEASING.md` → Prerequisites.
 
 If the code-review check fails (no PR, or unmerged/dirty tree): **stop** — releases
 ship reviewed, merged `main`, never a working tree.
@@ -162,12 +172,9 @@ Add the `CHANGELOG.md` entry (Keep-a-Changelog format) if not already committed.
 
 `npm deprecate <pkg>@<version>-rc.1 "superseded; use <version>"` for each package.
 
-Retire the RC's `next` tag by **repointing, not deleting**: `npm dist-tag rm` 403s for every
-token we hold (local and CI both — npm's bypass-2FA package-access rule, not a missing OTP).
-The Actions → **npm dist-tag** dispatch that repointed the whole lockstep set from the
-publish token **cannot run — GitHub Actions is disabled on this repository (2026-09-16),
-and the replacement path is being decided**. Repoint per package from your machine instead:
-`npm dist-tag add <pkg>@<version> next` (`add` is not restricted). See `docs/RELEASING.md` §4.
+Retire the RC's `next` tag by **repointing, not deleting**: `npm dist-tag rm` is refused
+for every token we hold. Repoint per package with `npm dist-tag add <pkg>@<version> next`,
+or dispatch the npm dist-tag workflow for the whole set. See `docs/RELEASING.md` §4.
 
 ## Phase 10 — Propagate versions (docs + websites)
 
@@ -211,24 +218,14 @@ version FAILS by design — skip what's done, never force.
 - About to `bun publish` to `latest` without human confirmation → stop.
 - About to push the site repo without confirming the live deploy → stop.
 
-## OIDC Trusted Publishing — now DATED, not optional
-npm removes **direct publishing from 2FA-bypass tokens around January 2027**
-([changelog](https://github.blog/changelog/2026-07-31-restricting-npm-bypass-2fa-granular-access-tokens/));
-since 2026-07-31 such tokens already cannot change package access — including
-**deleting a dist-tag**. Measured 2026-09-10: the local token and the CI `NPM_TOKEN`
-both 403 on `DELETE …/dist-tags/next` and `--otp` changes neither, so this is a grant
-boundary and no OTP session was going to clear it. Phase 9 repoints with `dist-tag add`
-instead, which is not restricted. The replacement is **OIDC Trusted Publishing** via
-`publish-npm.yml` (which exists — inoperative while GitHub Actions is disabled on this
-repository, 2026-09-16), the same mechanism `publish-csharp.yml` already
-uses for NuGet; the remaining work is a trusted-publisher registration per package
-on npmjs.com. Caveat: OIDC can't drive `npm dist-tag` yet, so RC→promote still needs
-a token.
-
-Separately, and still genuinely optional: **Changesets** (per-PR intent files → a
-reviewable "Version Packages" PR → automated bump + CHANGELOG). Adopting it would
-replace Phases 1/3/7's manual steps; the preflight gate, smoke test, and
-stop-and-confirm in this skill stay.
+## npm credentials and Trusted Publishing
+The npm token type, the read-only auth preflight, and the dated plan to move publishing to
+OIDC Trusted Publishing are in `docs/RELEASING.md` → Prerequisites. Read them before the
+first publish of a cut.
+
+Separately, and optional: **Changesets** (per-PR intent files → a reviewable "Version
+Packages" PR → automated bump + CHANGELOG). Adopting it would replace Phases 1/3/7's manual
+steps; the preflight gate, smoke test, and stop-and-confirm in this skill stay.
 
 ## Verify with the committed script, not ad-hoc shell
 `node scripts/release-verify.mjs [--preflight] [<version>] [--smoke]` is the one place
@@ -238,13 +235,3 @@ instead of improvising checks per release: the 0.24.0 cut was verified by a doze
 throwaway scripts, and the improvised auth check was a **write** against a published
 package that could not be undone. **Never probe credentials by mutating a published
 package.** It is read-only; extend it there when a release needs a check it lacks.
-
-## Phase 0 addendum — check that you CAN publish
-`scripts/release.mjs` now runs `npm whoami` in preflight. Do the same by hand if you
-are publishing manually, and use an **Automation / bypass-2FA** token: a "Publish"
-token passes `whoami`, shows `read-write`, passes `--dry-run`, and then `EOTP`s on
-every write, at which point `bun publish` falls back to a browser flow that cannot
-complete non-interactively (`--otp=` is ignored; a code must be piped on stdin).
-Verify BEFORE the version bump — the bump is committed and pushed well before the
-first publish, so a dead credential strands the cut with other registries already
-shipped. See `docs/RELEASING.md` → Prerequisites.
@@ -16,7 +16,7 @@ The first four ship per-language today across the five ports (TS / C# / Java / P
 2. **Runtime metadata** — load metadata at runtime, drive behavior dynamically (CRUD, validation, relationships, dynamic admin UIs, LLM tool registration). On Kysely (TS), a DB-API 2 driver via ObjectManager (Python), modernized JDBC + Spring-tx via OMDB (Java), Exposed (Kotlin). **C# runtime metadata is on the roadmap** — its EF Core output is generated code, not a metadata-driven runtime; the runtime `MetaObjects` package depends only on YamlDotNet and ships no data access.
 3. **Drift detection** — `meta verify` catches divergence between code and metadata (covers entity codegen, prompt templates, output parsers, schema). Quality-of-life on top of codegen + runtime.
 4. **Prompt construction** — a prompt is code, not a string scattered across services. Declare a prompt's payload as a typed projection (payload bloat becomes a diff), keep its text external and provider-resolved, and render it deterministically: snapshot-testable, cache-stable (no whitespace change silently breaking exact-prefix prompt-cache hits), and drift-checked at build time so a renamed field can't degrade a prompt. Conformance-gated, so the guarantee holds in every language port. **Render + payload-VO codegen + `verify` + parser-on-receipt for a *responding* `template.prompt` — one carrying `@responseRef` (FR-006) — + the output-format prompt fragment & tolerant `extract` parser (FR-010) ship in all five ports today** (since 0.24.0 the whole inbound tier keys off `@responseRef`; a `template.output` is outbound-only and emits no parser — ADR-0052) — the library-side building blocks of the pillar are complete. The one remaining library-side piece is MCP exposure of declared prompts/tools (see `spec/roadmap.md`); the application-level consolidation (eval harness, end-to-end declared-prompt orchestration) and consumer adoption are exercised in adopter projects, not in this library repo. Designed in `docs/superpowers/specs/2026-05-22-fr-004-cross-language-prompt-construction-design.md`.
-5. **Requirements and testing** — declare what the software is supposed to *do* in the same model as the entities, so a capability claim is checkable instead of prose. The other four pillars keep the code honest about the *model*; this one asks whether the thing you said the software does is actually built — an absence no test can fail on, because a test exercises code that exists. `requirement.functional` (existence: `meta verify` WARNS when nothing implements it) and `requirement.architectural` (universality: `meta verify` FAILS a live policy applied to nothing, and does not check that each claimed node complies) are registered vocabulary in all five ports, with the loader enforcing the closed `@status` enum. `@implementedBy` is **resolved, not trusted** — it names a real member of the real model, so a claim whose implementation was renamed or deleted fails the build instead of going quietly stale. `meta verify` reports the ledger on every run (unresolved links, entities no claim covers, gaps recorded versus gaps nobody has ruled on) plus an authoring lint whose findings can never fail a build; `meta docs` renders it for humans and for agents. **The port split, stated exactly: the vocabulary loads and validates in all five ports; the requirement gate runs in every port's `verify` (ADR-0057 — `docs/CONFORMANCE.md` "Split coverage"); every port ships an ejectable `requirement-tests` generator — which scaffolds a test per claim; the seven authoring-lint advisories are TypeScript-only.** A green run proves referential integrity, never that a status is true or an implementation correct (`docs/features/requirements.md`, "What a green run does not prove"). A project that declares no `requirement.*` nodes sees no change at all. Note the standing carve-out: `agent-context/skills/metaobjects-fit-assessment/SKILL.md` deliberately does NOT treat `requirement.*` as an assessment axis — see the ruling in that file before "finishing the job" there.
+5. **Requirements and testing** — declare what the software is supposed to *do* in the same model as the entities, so a capability claim is checkable instead of prose. The other four pillars keep the code honest about the *model*; this one asks whether the thing you said the software does is actually built — an absence no test can fail on, because a test exercises code that exists. `requirement.functional` (existence: `meta verify` WARNS when nothing implements it) and `requirement.architectural` (universality: `meta verify` FAILS a live policy applied to nothing, and does not check that each claimed node complies) are registered vocabulary in all five ports, with the loader enforcing the closed `@status` enum. `@implementedBy` is **resolved, not trusted** — it names a real member of the real model, so a claim whose implementation was renamed or deleted fails the build instead of going quietly stale. `meta verify` reports the ledger on every run (unresolved links, entities no claim covers, gaps recorded versus gaps nobody has ruled on) plus an authoring lint whose findings can never fail a build; `meta docs` renders it for humans and for agents. **The port split, stated exactly: the vocabulary loads and validates in all five ports; the requirement gate runs in every port's `verify` (ADR-0057 — `docs/CONFORMANCE.md` "Split coverage"); every port ships an ejectable `requirement-tests` generator — which scaffolds a test per claim; the seven authoring-lint advisories are TypeScript-only.** A green run proves referential integrity, never that a status is true or an implementation correct (`docs/features/requirements.md`, "What a green run does not prove"). A project that declares no `requirement.*` nodes sees no change at all. Note the standing carve-out: `agent-context/skills/metaobjects-fit-assessment/SKILL.md` deliberately does NOT treat `requirement.*` as an assessment axis — the ruling is in `docs/superpowers/specs/2026-07-12-metaobjects-fit-assessment-maintainer-notes.md`; read it before "finishing the job" there.
 
 6. **Libraries** — reusable declared design, shipped as metadata and opted into by name (`"libraries": ["iam"]` in `.metaobjects/config.json`). Not a sixth verb: a library is the **reuse unit that composes the other five** — entities, requirements, the generators its design implies, its runtime packages — as one named, opt-in, drift-gated artifact. What makes it a pillar rather than a folder of YAML is the fifth: without requirements a library is a schema snippet; with them it is design an adopter's build is held to, which is the same test the requirements pillar passes (*does it change what an agent can be checked against?*). **A library is LAYERED and its core layer is INERT** — the core declares no `source.rdb`, and a sourceless object generates nothing and migrates to nothing (#248), so `["iam"]` adds zero tables and zero generated code while making the design present and resolvable; `["iam", "iam/db"]` is the separate opt-in that proposes the schema. **Copy is the expected mode** (`meta eject <library>` — ADR-0034's ruling applied to metadata), and an ejected copy still named in `libraries` is refused at load (`ERR_LIBRARY_PACKAGE_COLLISION`) rather than merging asymmetrically. Object coverage activates on ADOPTER-authored requirements only, so a library cannot volunteer a project for a gate it did not ask for. `iam` (preview) and `ai` (stable) ship today; rows appear in `meta gen --list` beside the generators. See [docs/features/libraries.md](docs/features/libraries.md) and `docs/superpowers/specs/2026-09-13-fr-043-feature-and-nfr-packages-design.md`.
 
 
@@ -827,6 +827,21 @@ until you regenerate.
 
   One rule is relaxed: an authored `validator.length @min: 0` on a `@required` string now
   admits the empty string, as the generated schema already did.
+- **The shipped agent skills fit the skill-authoring limits.** `metaobjects-audit` (601 lines to
+  263), `metaobjects-authoring` (804 to 479) and `metaobjects-fit-assessment` (666 to 481) moved
+  sections into `references/` files, each with a row and an "open it when" condition in the
+  skill's reference table, which now sits at the top of `metaobjects-audit`,
+  `metaobjects-prompts` and `metaobjects-runtime-ui` as well. The text moved and none was cut; `meta init`
+  and `meta init --refresh-docs` write the new files into `.claude/skills/`. Every reference
+  file over 100 lines opens with a `## Contents` list, and `metaobjects-codegen` links its five
+  TypeScript topic files and every language file from `SKILL.md`. A test in the SDK holds all
+  three rules.
+- **The hosted fit-assessment prompt is built, not copied.** The capability sheet and the JSON
+  contract left `SKILL.md`, so `node scripts/build-assess-prompt.mjs --out <file>` writes the
+  single-file form served at `metaobjects.dev/assess.md`, inlining both. The output is the old
+  prompt, apart from one dropped phrase ("in this file"). Its maintainer notes moved to
+  `docs/superpowers/specs/2026-07-12-metaobjects-fit-assessment-maintainer-notes.md`, and the
+  skill's `scaffold: false` marker is now `metadata: { scaffold: false }` (no code reads it).
 
 ### Deprecated
 
 
@@ -97,9 +97,10 @@ If your agent can't fetch URLs (or you want it to follow the prompt verbatim), s
 the file into your workspace instead — `curl -fsSL https://metaobjects.dev/assess.md
 -o metaobjects-assess.md` (don't commit it) — and say: *"Read
 `metaobjects-assess.md` and run the assessment it contains against this
-repository."* The prompt is one Markdown file
-([source](agent-context/skills/metaobjects-fit-assessment/SKILL.md)); read it first
-if you like — you should never point your agent at a prompt you haven't vetted.
+repository."* The file is built from the
+[skill](agent-context/skills/metaobjects-fit-assessment/SKILL.md) and its
+`references/` by [`scripts/build-assess-prompt.mjs`](scripts/build-assess-prompt.mjs);
+read it first if you like — you should never point your agent at a prompt you haven't vetted.
 
 The catch, stated plainly: it runs in **your** agent on **your** tokens. The quick
 pass takes roughly 5–30 minutes of agent time, depending on the agent and the
 
@@ -18,7 +18,7 @@ The **scaffolded** set is exactly the six skills in the SDK's `SKILL_NAMES`
 additionally holds **`metaobjects-fit-assessment`**, a *pre-adoption* tool that is
 deliberately **not** in `SKILL_NAMES` and so is never scaffolded (a not-yet-adopted
 target has no `.claude/skills/metaobjects-*` to receive it); it carries
-`scaffold: false` in its front-matter as the human marker. Its design + retro-test
+`scaffold: false` under `metadata:` in its front-matter as the human marker (no code reads it). Its design + retro-test
 validation live in `docs/superpowers/specs/2026-07-12-metaobjects-fit-assessment-design.md`.
 
 Design: `docs/superpowers/specs/2026-06-02-downstream-agent-context-design.md`.