rebalance

Human output. Read ${CLAUDE_PLUGIN_ROOT}/skills/build-item/references/plain-language-output.md once per top-level run; composed skills reuse it. Apply it to every message.

Re-slice a live DevStride plan's NOT-STARTED leaves to a new delivery profile in place. Never touch shipped or in-flight work; carry every existing spec into its successor; delete nothing.

The profile itself — what grain and specDepth mean for each of prototype / standard / enterprise, the resolution order, and the root marker this skill rewrites — is defined ONCE, in ${CLAUDE_PLUGIN_ROOT}/skills/plan/references/delivery-profiles.md. Read that file first and apply it by citation; this skill never restates its table.

Arguments — a grouping item to re-slice (a whole plan root, or a single release unit — this org's Epic — to scope the re-slice to that unit; e.g. I20100) and the target profile as a bare word anywhere in the arguments (prototype, standard or enterprise — the contract's argument form), optionally with --dry-run; this skill REQUIRES both, so missing either → ask, never guess: $ARGUMENTS

IMPORTANT — the DevStride MCP targets PRODUCTION (api.devstride.com). Every create_item, add_relationship, bulk_update_items, update_item and archive_item call this skill makes is a real, immediate, user-visible change to the live plan — there is no draft/sandbox mode. Nothing is written before the explicit sign-off in step 2, and a dry run writes nothing at all.

IMPORTANT — this is a Gantt-grooming skill: it changes DevStride data only, never the repo. Building the re-sliced stories is /devstride:build-item's job, afterwards.

IMPORTANT — the same Workflow-drafting split plan enforces applies here. Interactive judgment — which leaves merge, where a coarse leaf splits, every "does this grouping still deliver a vertical slice?" call, and the sign-off — happens in the main conversation. A Workflow is used only for bulk drafting of successor specs once the shape is agreed, one agent per release unit, and its agents DRAFT ONLY: they never call an MCP write tool. Workflow output is a proposal to review, not a commit.

0. Safety gates — all of them, before a single read is trusted

• Parse $ARGUMENTS. Resolve the item number with get_item(view:"full"). It must be a grouping item (a container level — this org: Module/Capability/Epic), never a leaf; a leaf argument → stop and ask for its release unit or plan root. The target profile is the bare word prototype, standard or enterprise wherever it appears in the arguments (the contract's resolution-order item 1); no such word, or any other word in its place → ask. If either is missing, ask — a wrong root re-slices somebody else's plan, and a guessed profile is the wrong grain applied at production speed. Note --dry-run if present.
• Resolve this org's work-type roles at runtime with get_work_type_hierarchy (and get_workspace_context for lanes and priorities): which types are executable leaves, which level is the release unit (the level /devstride:build-item branches and releases at), and which are plain containers above it. Bind every role word in this file to what the call returns — never assume "Story"/"Epic" spelling or hierarchy depth. When the release-unit level is ambiguous from structure alone, use the consuming repo's .claude/ds-config.json hierarchyRoles.releaseUnit as the tie-breaker if the repo is known, otherwise ask. Say "grouping item" or the org's real type name in anything user-facing — never "container".
• Refuse to run while a build loop is active on this plan — two writers on one plan is the collision. Two checks, either a hard stop: an In Progress leaf under the root with a live branch (get_item_branches, or git ls-remote --heads origin "*/I<number>-*" when the repo is known); or the handoff project memory naming this root (or an ancestor/descendant) as mid-iteration. Report what tripped and stop — never wait it out, never move an In Progress item to clear the gate. An In Progress leaf with NO branch is not an active loop but stays untouchable (step 1).
• Confirm the organization-wide dependency auto-scheduler is OFF — apply the canonical Enable Link Mode rule in ${CLAUDE_PLUGIN_ROOT}/skills/rationalize-gantt/references/auto-scheduler-off.md: its READ-ONLY check now (get_organization_metadata, have the user disable the setting if on, never change it automatically) and staticMode omitted from every write this skill makes. Its probe-date verification is a WRITE (update_item on a live item and a restore), so it does not run here: it runs at the top of step 3, after sign-off, and never on a dry run. Step 3 rewrites edges and step 3f rewrites dates; with propagation on, the backend would overwrite both behind you.
• Re-run the loop check immediately before step 3 writes anything — the proposal in step 2 can take a while, and a loop started during it is exactly the collision the gate exists to prevent.