Package Cascade

Package Cascade

Propagate a published package change up the dependency graph: re-pin and republish every package (ComponentModule) that references it, recursively, stopping at applications (which are reported as stale, never auto-published). This skill is thin, interactive orchestration over the dxs source cascade commands.

REQUIRED BACKGROUND

• ../datex-studio-shared/branch-setup.md — branch & connection selection
• references/cascade-workflow.md — the phase-by-phase runbook with exact dxs commands
• references/plan-schema.md — the cascade plan JSON shape
• references/interaction-patterns.md — tree rendering, pick/prune prompts, progress narration
• references/troubleshooting.md — publish failures, locks, conflicts, resume

When to use / not use

Use when the origin package is already published and you want its consumers updated. The user makes the change, commits, and publishes the origin themselves; this skill takes over from the published version.

Not for: authoring the package change itself (use the component-creator skills), single-branch caller analysis (use impact-analysis), or forward dependency inspection (dxs source deps).

Division of labour (do not cross this line)

The CLI owns all determinism: graph traversal, topological ordering, the transitive-reference rebuild, and conflict/cycle detection live in dxs source cascade plan and dxs source reference set. This skill owns interaction: present the plan, ask all-or-subset (including pruning the dependents of a deselected package from --select, walking the plan's own updates edges — the CLI runs exactly what you select and errors mid-run otherwise), drive execution, report. Never hand-edit AppConfig JSON, never compute the update order yourself, never reason about the transitive closure in prose — delegate to dxs. If you reach for jq on an appConfig, stop: that work belongs in dxs source reference set.

Workflow

Phase 0  Establish origin+org  — user names the package; resolve its uniqueIdentifier (dxs marketplace
         search "<name>" --type componentmodule, or marketplace list --type componentmodule to browse) and
         pass the RESOLVED uid to -p, never the display name the user typed (see safety gates); take the
         newest published version from the user; pick the target org (auto only when the user sees exactly
         one AND didn't name a different one — if they name an org you're not logged into, confirm + switch
         identity via dxs auth login, see references); CONFIRM (a wrong version poisons every pin, a wrong
         org syncs the wrong tenant).
Phase 1  Plan                  — dxs -O json source cascade plan -p <uid> -v <version> --org <orgId>
Phase 2  Present & select      — render the tree; ask ALL / subset / review-only (see
         references/interaction-patterns.md); a cycle or conflict = hard stop.
Phase 3  Execute bottom-up     — dxs source cascade run --plan plan.json [--select …];
         narrate each republish (n of N).
Phase 4  Report                — republished packages (+ new versions) and stale applications
         to update separately.

Follow references/cascade-workflow.md for the exact commands in each phase.

Sibling skills

• datex-studio-shared — branch & connection setup (see REQUIRED BACKGROUND above).
• release-notes-generator — draft the text you pass as --release-notes "<notes>" in Phase 3; remember it is applied verbatim to every republished node in that run, not per-node.
• commit-message-generator — cascade run always writes a fixed commit message (chore(deps): cascade update <uid>) and has no --commit-message flag, so this sibling only applies if you step outside cascade run and drive the per-node primitives (branch commit) by hand for a bespoke message — a current limitation, not the default path. That by-hand route needs -f to get past branch commit's confirmation prompt (dxs ≥ 0.4.19), and --dry-run to preview it; see references/cascade-workflow.md.

Safety gates

• Confirm the origin package + version + target org before planning (Phase 0). A wrong version poisons every pin; a wrong org syncs the wrong tenant.
• -p is the resolved uniqueIdentifier, never the display name. The uniqueIdentifier strips spaces/underscores from the name (display pkg_cascade_mid → uid pkgcascademid) and cascade plan matches it by exact equality — so a display name, even one that already looks id-shaped (e.g. pkg_cascade_leaf), matches no package and returns a silently empty plan that reads as "nothing to do." Pass the exact string marketplace search returned; sanity-check with dxs source referenced-from <uid> --org <id>.
• Cascade is single-tenant — Phase 1 passes --org <id> so only the target org's packages are planned. Consumers in other tenants are excluded (and could not be republished anyway; publishing is gated to the owning org).
• The target org may be one you're not logged into. organization list shows only the active identity's tenants, and cascade plan traverses cross-tenant read-only, so a plan can succeed against an org you cannot publish to — the failure surfaces only at run (404 on mainApplicationId). If the user names an org that isn't visible, do not silently fall back to a visible org and do not plan-and-fail: confirm the target, confirm the switch, then dxs auth login --tenant-id <tenantId> into that tenant before running. On dxs ≥ 0.5.0 that login returns immediately if an identity for the tenant is already cached, otherwise opens a browser the user completes — it only emits a device code when headless or when you pass --use-device-code / --no-browser. See references/cascade-workflow.md Phase 0 step 3a and references/interaction-patterns.md → "Target org you're not logged into".
• Never assume a branch/repo ID — the plan carries each node's Main branch id; surface it, don't guess.
• Cycle or version conflict = hard stop. cascade plan sets cycleDetected; reference set aborts on conflict. Report and do not force.
• Explicit go/no-go after presenting the plan — nothing mutates before the user picks (Phase 2).
• Applications are never auto-published — they appear only in staleApplications.
• Leave failures inspectable — on a failed node, keep its feature branch, report it, and offer resume via --select of the remaining nodes.
• Publish access — if the user lacks package-publish rights the first publish fails; surface it early.