Bump MCP

bump-mcp — Pin an MCP server to a new version in the nexus catalogue

roster/mcps.yaml is the closed catalogue of the five nexus MCPs (atlas-aci, junction, crystalium, tonberry, atomos). Under the default integrity.enforcement: strict posture it is what eidolons verify checks an installed image against — so a stale pin is not cosmetic. It means every consumer installs an old image and cannot reach newer versions with eidolons mcp use <name>@<ver>.

add-eidolon handles Eidolons (roster/index.yaml). This skill handles MCP servers (roster/mcps.yaml). They are different files with different shapes.

Invariants you must respect

1. Never copy a digest from a release note. Resolve every digest from the registry yourself. A wrong digest breaks eidolons verify for every consumer under strict.
2. Git tags are v-prefixed; image tags usually are not. git tag v2.0.0 but ghcr.io/rynaro/atlas-aci:2.0.0. imagetools inspect …:v2.0.0 will silently return nothing. Check both.
3. Backfill releases the catalogue skipped. If releases: jumps from 0.2.3 to the new version, every intermediate release is unreachable via mcp use. Record them, with digests resolved from the registry.
4. catalogue_version is for the file's shape, not its contents. Adding releases does not bump it. Adding a new field does.
5. Bump updated_at.
6. Do not sweep unrelated working-tree changes into the commit. This repo's tree frequently carries uncommitted .mcp.json / .gitignore edits. git add the exact paths.

Phase 0 — Confirm the upstream release is real

gh release view v<VER> --repo <owner>/<repo> --json tagName,isDraft,publishedAt

For kind: oci-image MCPs, confirm the image exists and is a multi-arch index:

docker buildx imagetools inspect ghcr.io/rynaro/<name>:<VER> --raw \
  | python3 -c 'import sys,json;d=json.load(sys.stdin);print(d["mediaType"], len(d.get("manifests",[])), "platforms")'

If the release workflow signs with cosign and attests SBOM/provenance, say so in the CHANGELOG — it is why the digest can be trusted.

Phase A — Capture and verify every digest

Not just the new one. Every version you will record.

for t in 0.3.0 0.3.1 0.4.0 2.0.0; do
  printf '%-8s %s\n' "$t" \
    "$(docker buildx imagetools inspect ghcr.io/rynaro/<name>:$t --format '{{.Manifest.Digest}}')"
done

Release dates: gh release view v<VER> --json publishedAt. Some tags have an image but no GitHub Release object — fall back to the git tag date (git log -1 --format=%aI <tag>) and say which you used.

Phase B — Edit roster/mcps.yaml

Under the MCP's versions: block:

• latest: → the new version
• pins.stable: → the new version
• releases: → add every version with digest: and released_at:
• top-of-file updated_at: → today

If you are backfilling, leave a comment saying why the catalogue skipped them. A future reader will want to know it was an oversight, not a policy.

Then re-verify what you wrote against the registry, reading the digests back out of the file rather than out of your own scrollback:

python3 - <<'PY' > /tmp/dg.txt
import yaml
d = yaml.safe_load(open('roster/mcps.yaml'))
e = [m for m in d['mcps'] if m['name'] == '<name>'][0]
for v, r in sorted(e['versions']['releases'].items()):
    print(v, r['digest'])
PY
while read -r v dg; do
  real=$(docker buildx imagetools inspect "ghcr.io/rynaro/<name>:$v" --format '{{.Manifest.Digest}}' 2>/dev/null)
  [ "$real" = "$dg" ] && echo "$v OK" || echo "$v MISMATCH recorded=$dg real=$real"
done < /tmp/dg.txt

Phase C — The test fixture is coupled to pins.stable

This is the trap. cli/tests/mcp_images.bats hardcodes the stable digest:

ATLAS_ACI_PINNED="sha256:…"

and S8 asserts "image present at the stable digest ⇒ DRIFT=no". If you bump the catalogue and forget the fixture, S8 keeps passing — vacuously. It asserts no-drift against a digest nothing pins any more. The hardcoded constant is a proxy for "the catalogue's stable digest".

Bump the fixture, and make sure a guard derives the value instead of trusting it. S8-guard does this today for atlas-aci (pure awk, bash 3.2 safe). If you add a fixture for another MCP, add its guard too.

Prove the guard has teeth before you trust it:

# revert the fixture to the OLD digest -> S8-guard and S8 must BOTH fail
bats cli/tests/mcp_images.bats | grep '^not ok'
# restore -> both pass

Note CRYSTALIUM_PINNED is deliberately a non-matching digest — S9 exercises the DRIFT=yes path. Do not "fix" it.

Phase D — CHANGELOG

Add to ## [Unreleased] under ### Changed. The house style is a dense paragraph, not a bullet fragment: what moved, the digest, why it matters, what breaks for consumers.

Link to an immutable ref. If you link to an upstream artefact under .spectra/changes/<id>/, link to the tag, not blob/main — the ESL lifecycle archives that folder to .spectra/changes/archive/<date>-<id>/ and blob/main will 404:

https://github.com/<owner>/<repo>/blob/v2.0.0/.spectra/changes/<id>/RETRO.md   ✅
https://github.com/<owner>/<repo>/blob/main/.spectra/changes/<id>/RETRO.md     ❌ breaks on archive

State breaking changes plainly. For a schema-epoch bump: "the on-disk DB is derived data with no in-place migration — re-index once."

Phase E — ESL change

This repo runs ESL. Open a change before editing (trivial routes need no spec):

tb() { docker run --rm -i --user "$(id -u):$(id -g)" \
  -v "$PWD:$PWD:z" -w "$PWD" --cap-drop ALL --security-opt no-new-privileges \
  ghcr.io/rynaro/tonberry@sha256:<pinned> "$@"; }

tb propose --change_id mcp-<name>-<ver-dashed> --maker vivi --checker kupo \
   --has_code true --spec_ref spec.md --project_root "$PWD"
tb right_size --change_id … --files_touched 5 --rubric_score 3 --tradeoff_present false --project_root "$PWD"

has_code: true — you are editing cli/tests/*.bats. Write spec.md, thread acceptance_checks via compose_manifest --patch, then transition. A catalogue bump right-sizes to lite.

Phase F — Verification

make schema                       # jq + yq structural checks CI runs
EIDOLONS_NEXUS="$(pwd)" bash cli/eidolons mcp show <name>    # CLI reads the catalogue
make test                         # parallel; needs GNU `parallel`
bats cli/tests/                   # sequential fallback (~2m30s) when `parallel` is absent

Before blaming your change for a red test, check whether it is red on a clean tree. Stash and re-run the single file:

git stash push -q roster/mcps.yaml cli/tests/mcp_images.bats
bats cli/tests/<file>.bats | grep '^not ok'
git stash pop -q

At time of writing, cli/tests/doctor_deep.bats DD-7 fails on a clean tree. Say so in the PR rather than silently accepting a red suite.

Phase G — Branch, commit, PR

Branch fix/mcp-<name>-<version-dashed> (mirrors the fix/roster-<eidolon>-<version> convention). Stage exact paths. Never push to main directly. Open a PR; do not merge without the user's word.

Traps, learned the hard way

• imagetools inspect ghcr.io/x:v2.0.0 returns nothing while :2.0.0 works. The v is a git convention, not a registry one.
• A hardcoded digest in a test is a proxy. Derive it from the catalogue or a bump will make the test pass vacuously.
• blob/main links to ESL change folders rot the moment the change is archived.
• The catalogue can silently skip releases for months. Check releases: against gh release list before you assume you are bumping by one.
• The local wiring (.mcp.json, eidolons.mcp.lock) is separate from the catalogue. Those are consumer artefacts, often untracked or carrying unrelated edits. Update them with eidolons mcp upgrade <name>, in their own commit, or not at all.