Workspaces

A workspace is a standalone git repository that acts as the documentation, decision, and planning hub for work spanning multiple member repositories. The hub owns everything about the work; member repos own only code. Agent context is explicit (a manifest declares it), layered (loaded in a fixed order), and auditable (generated views drift-checked by script).

This skill composes existing skills rather than replacing them: grill-with-docs scopes, domain-modeling owns CONTEXT.md and ADRs, stacked-prs binds branches to the workspace by naming, worktrunk materializes interactive worktrees, and skills-skill standards govern the seeded project-local skills.

Routing

Match the request to a flow and follow its reference document exactly:

Before any flow, read references/workspace-layout.md for the hub anatomy and references/conventions.md for the binding rules, then read the hub's workspace memory (.claude/memory/MEMORY.md and the entries it indexes that touch your task) — memory is the always-read working-knowledge layer, and starting work without it repeats settled mistakes. When creating or editing the seeded skills, follow the skills-skill skill.

Scripts

All workspace mechanics go through one CLI (the consistent binding point; every workspace's Justfile delegates here):

bun $SKILL_DIR/scripts/workspace.ts <command> [--workspace <hub-dir>]

--workspace defaults to walking up from the current directory to the nearest workspace.yaml. Member state is read via git -C from the hub; no command enters a worktree.

--workspace may itself be a linked worktree of the hub, which is how hub work is normally done. Member paths then anchor on the hub's main working tree, so ../app still names the sibling checkout and not .claude/worktrees/app. Hub-internal paths (context layers, JOURNAL.md, the generated files) stay on the worktree, because those are the files it exists to edit.

Core rules (full detail in conventions.md)

1. One creator per branch+worktree pair. wt owns interactive worktrees, stacked-prs owns metadata only. Never run stacked-prs create for a child branch inside a wt worktree.
2. Hub owns planning. Plan files live in the hub, and member repos never commit workflow artifacts.
3. Binding is naming. Stacks <stack-prefix><task-slug> (no dots) and branches <branch-prefix><task-slug> use the same task slug.
4. Manifest is truth. Edit workspace.yaml, run just sync; never hand-edit CLAUDE.md. docs/index.md is the context allow-list.
5. Deviations are journaled before or with the change, cross-linked to phase, ADR, and PR. ADRs are superseded, never edited.
6. init is gated. Always confirm with the user before creating the hub repo, touching member git config, or registering sources.
7. Memory before work, memory as you think. The hub's .claude/memory/ (committed) is read before any flow's first action and written the moment an idea, decision-in-motion, surprise, or open question surfaces — during design especially, where the job is capturing and converging concepts, not generating variant floods. A settled entry graduates to an ADR, journal entry, or finding; the memory then points at that record.

Templates

references/templates/ holds the scaffold for every hub file (workspace.yaml, JOURNAL.md, constitution.md, index.md, phases.md, tasks.md, Justfile, memory-index.md → .claude/memory/MEMORY.md) and the three seed skills (skill-context.md, skill-domain.md, skill-conventions.md). Markdown and YAML templates use {{PLACEHOLDER}} markers; the Justfile template alone uses __PLACEHOLDER__ markers because just owns {{ }} interpolation there (and markdown formatters rewrite __x__ as bold).