[joelvarty]

Adds a live, end-to-end test of agility sync (PROD-3016, under PROD-3015). Each run leases throwaway instances, syncs a source to a target with the CLI under test, checks the result, and always deletes the instances. ci/README.md explains when it runs, how to trigger it, and what it checks.

What it checks today (first slice):

• sync exits 0 with no failed items;
• agility verify finds no differences;
• an immediate sync --preflight plans no changes.

When it runs:

Trigger	Tier	CLI under test
Pull requests from this repo's branches	quick	the PR's build
v* tags	beta	the published npm package
Nightly	nightly	main, plus a sweep for left-behind instances
By hand	your choice	via Run workflow or gh workflow run

Fork and Dependabot PRs skip it.

The gate: npm-publish.yml's promote job now refuses to move a version to latest unless that version's tag run passed.

Cleanup, in three layers:

1. A release job runs whatever the outcome and deletes every instance labelled with the run.
2. Short expiry times, so anything a dead runner leaves is deleted automatically.
3. The nightly sweep.

Configuration: every identifier lives in the cli-dev-live GitHub Environment as a secret. Nothing is in the repo, and the public Actions logs mask it. Artifacts hold only the JSON results.

Current result: the check fails on two CLI bugs this test found. PROD-3025: linked content isn't dependency-ordered within a sync, so items that reference other referenced items are skipped. PROD-3026: a page whose name already exists on the target isn't adopted. This check stays red until they're fixed, and merging this holds new versions off latest until then.

🤖 Generated with Claude Code

[joelvarty]

First real run on GitHub, after the setup: run 41066708783 (attempt 2). Sign-in through OIDC, leasing, the test and the release all worked; both leases were deleted.

The test fails for two CLI bugs, now ticketed:

• PROD-3025: linked content isn't dependency-ordered inside the linked tier (get-content-item-types.ts puts every referenced item in one tier, pushed as one batch). So in FeaturedPost → Post → Category, the posts are remapped before their categories exist and get skipped. This causes 5 of the 6 failures, the verify differences, and most of the re-sync plan.
• PROD-3026: a source page whose name already exists on the target, with no mapping, fails outright. Containers in the same situation get adopted with a mapping row.

This check stays red until those are fixed, which is the point of it.

[joelvarty]

Handing this over to Kevin. It's already merged into release/1.3.0-beta.0 (#254), where the first sync now passes. What still fails is verify and the no-op preflight reporting the conflicts and drift the test keeps on purpose. Next steps for the engine:

1. Separate expected differences (kept conflicts, drift, known issue PROD-2775) from new ones, so a clean run passes.
2. content.add-chain finds no linked item to start from on the blog template.
3. Add change types for locales, moving items, delete-and-recreate under the same name, and damaged mapping files.

How it runs is in ci/README.md.

[5PK]

Added a reference-cycle fixture (f266d0c) so the live test covers content that references itself and each other (PROD-3076). Blog Starter has no such content.

• ci/permutation/fixtures.mjs adds a PermCycleNode model and list to the source before the first sync: A ↔ B and a self-reference, published. The link field is required and has a companion field.
• After round 0, a new cycles check confirms each copy on the target points at its partner's target ID, in both the link and the companion field.

That check only passes with #253, which adds the CLI side (dependency-ordered linked content, PROD-3025; the same-run two-pass push for cycles, PROD-3076). #253 already contains this branch, so merge this first, then #253.

Until #253 lands, the quick tier here also fails round 0's cycles check, on top of the round 1–2 verify and no-op failures it already shows with main's CLI. A local run of this harness with #253's CLI passed all of round 0, including cycles. One of the round 1 failures looks like a timing gap in the conflict check. The pre-sync preflight reported 0 conflicts, but the sync itself then reported the conflict and kept the target's value. The check only credits the preflight's count, so it read this as "silently kept".

🤖 Generated with Claude Code

[5PK]

Approving: the live sync test harness, plus the reference-cycle fixture. #253 follows with the CLI fixes the fixture needs.