setup

This skill is the only framework artefact an adopter project commits. Every other apache-steward skill (security, pr-management, issue) is a gitignored symlink into the gitignored snapshot at <snapshot-dir> that this skill manages.

The adoption model is snapshot + agentic overrides + drift- aware updates (not submodule, not marketplace, not vendored copy):

• The framework is downloaded into <snapshot-dir> and gitignored in the adopter repo. The snapshot is a build artefact, not source.
• Three install methods are supported (see docs/setup/install-recipes.md for verbatim copy-pasteable recipes):
  • svn-zip — released, signed zip from ASF distribution (recommended for production once releases ship).
  • git-tag — pinned to a specific git tag.
  • git-branch — tracks a branch tip (default: main, the WIP path).
• Two lock files record the framework version. The committed one declares what the project pins to; the local one records what each machine actually fetched. Drift between them is surfaced and remediated by /magpie-setup upgrade.
• Symlinks make the framework's skills callable as if they lived in the adopter repo. Each symlink is named magpie-<framework-skill> — every framework skill is installed under a magpie- prefix so it is namespaced and never collides with the adopter's own skills (e.g. the snapshot's skills/pr-management-triage/ becomes magpie-pr-management-triage, invoked as /magpie-pr-management-triage). .agents/skills/ is the one canonical home: its magpie-* entries link into <snapshot-dir>/skills/<framework-skill>/. Every other agent target (.claude/skills/, .github/skills/, …) gets a thin per-skill relay symlink that points back at the canonical entry (.claude/skills/magpie-<n> → ../../.agents/skills/magpie-<n>) — no matter what layout the adopting project previously used (see agents.md). The symlinks are gitignored because their targets disappear on a fresh clone before /magpie-setup runs.
• Adopter-specific modifications to framework workflows live as agent-readable instructions under .apache-magpie-overrides/<skill-name>.md (committed). They invalidate or change steps the framework's skill would otherwise run. See overrides.md for the contract and docs/setup/agentic-overrides.md for the design rationale.

Local self-adoption (the framework checkout only). The one repo that cannot be adopted via the snapshot mechanism is the Apache Magpie framework checkout itself — a remote snapshot of the framework into itself would be circular. Instead it self-adopts with method:local: each canonical magpie-<skill> in .agents/skills/ is a committed symlink into the in-repo ../../skills/<skill>/ source, and every other active agent target (agents.md) — .claude/skills/ (Claude Code), .github/skills/ (GitHub's skill loader), and any present holdout — gets a committed relay symlink (magpie-<skill> → ../../.agents/skills/magpie-<skill>) — with no snapshot, no remote fetch, and no copy. This makes the framework's own skills callable while developing the framework, and every contributor gets them active on a fresh clone with no setup step. adopt detects the framework checkout structurally and routes there automatically (see adopt.md → Local self-adoption).

The two lock files

The framework's lock-file model splits what the project pins to (committed) from what this machine actually fetched (local). This split is the foundation of drift detection and the multi-installer support.

<committed-lock> — .apache-magpie.lock

Committed at the adopter repo root. The project's pin. Edited only by /magpie-setup; do not modify by hand.

# .apache-magpie.lock — committed; the project's pin.

method: <git-branch | git-tag | svn-zip>
url:    <see per-method format below>

# For method=git-branch:
ref:    main

# For method=git-tag:
ref:    v1.0.0          # the tag name
commit: <SHA>           # the commit the tag pointed to when committed

# For method=svn-zip:
ref:    1.0.0           # the version number
sha512: <hash>          # the released zip's SHA-512 (for re-fetch verification)

The next adopter who runs /magpie-setup adopt reads this file and re-installs to the same version the project declared. This is the core of the "adopt once, all subsequent users get the same thing" promise.

<local-lock> — .apache-magpie.local.lock

Gitignored at the adopter repo root. The local snapshot's fingerprint. Records what this machine fetched and when.

# .apache-magpie.local.lock — gitignored; per-machine.

source_method:    <git-branch | git-tag | svn-zip>
source_url:       <URL the snapshot was actually fetched from>
source_ref:       <branch / tag / version actually fetched>
fetched_commit:   <commit SHA on disk now>
fetched_at:       <ISO-8601 timestamp>

The drift check on every framework-skill invocation compares this against <committed-lock> and surfaces any mismatch as a proposed /magpie-setup upgrade.

Detail files in this directory

Golden rules

Golden rule 1 — never modify the snapshot. The <snapshot-dir> is a build artefact, gitignored, and read- only from an adopter's perspective. Every modification an adopter wants must go into .apache-magpie-overrides/ (where it is committed and survives the next upgrade). The skill, and any other framework skill consulting overrides at run-time, never writes to <snapshot-dir>.

Golden rule 2 — <committed-lock> is the project's pin; <local-lock> is per-machine truth. They serve different purposes and live in different places:

• <committed-lock> declares what version the project uses. Edited by the adopter who runs /magpie-setup adopt first (or who later runs /magpie-setup upgrade and accepts the new pin). Bumping it is a deliberate project-level action; the bump shows up in the git diff of the PR that proposed it.
• <local-lock> records what this machine installed. Updated silently by /magpie-setup adopt and /magpie-setup upgrade. Per-developer, per-checkout, per-worktree.

Golden rule 3 — drift surfaces, drift gets remediated. Every framework skill (and /magpie-setup verify) checks <committed-lock> vs <local-lock> at the top of its run. On mismatch the skill surfaces the gap and proposes /magpie-setup upgrade. The user accepts or defers; if they accept, upgrade:

1. Deletes <snapshot-dir> outright.
2. Re-installs per the committed lock (the new version the project chose).
3. Refreshes the gitignored framework-skill symlinks — adds any new framework skills the user's family pick covers, removes any framework skills that were renamed away or removed.
4. Reconciles agentic overrides against the new framework structure (surfaces conflicts; never auto-rewrites).
5. Updates <local-lock> to the new fetch.

Golden rule 4 — .gitignore keeps the adopter repo clean. Three things gitignored in the adopter repo:

• <snapshot-dir> (the entire framework snapshot — gigabytes potentially).
• <local-lock> (per-machine state).
• The magpie-* symlinks setup adopt creates in every active target dir — the canonical ones in .agents/skills/ (they target the gitignored snapshot) and the relays in .claude/skills/ / .github/skills/ / holdouts (they target the canonical entries) — both would dangle in a fresh clone. The one exception un-ignored in each dir is magpie-setup.

Committed: this skill (setup, as the canonical .agents/skills/magpie-setup/ plus its relays), the <committed-lock>, the .apache-magpie-overrides/ directory, the .gitignore entries themselves, any project-doc updates the adopt sub-action makes.

Golden rule 5 — .agents/skills/ is canonical; everything else just relays into it. Regardless of how an adopting project previously organised its .claude/skills/ or .github/skills/, adopt always wires the framework the same way: the canonical magpie-* links live in .agents/skills/, and every other active target (.claude, .github, holdouts) gets per-skill relay symlinks pointing back at the canonical entries (.claude/skills/magpie-<n> → ../../.agents/skills/magpie-<n>). The adopter's own native (non-magpie-) skills in those dirs are left untouched. See agents.md.

Golden rule 6 — copy this skill, symlink the rest; all under the magpie- prefix. This skill (source skills/setup/) is the only framework skill that gets copied into an adopter repo — committed as the canonical .agents/skills/magpie-setup/, with committed relay symlinks to it from .claude/skills/magpie-setup and .github/skills/magpie-setup. All other framework skills are symlinked (canonical link into the gitignored snapshot, plus relays), each named magpie-<framework-skill> (e.g. magpie-security-issue-import → <snapshot-dir>/skills/security-issue-import/). The magpie- prefix namespaces every framework skill so it never collides with an adopter's own skills. Mixing copy and symlink — copying a security skill, for instance — creates a maintenance hazard: copies drift from the framework's source- of-truth, and the drift-detection mechanism (which assumes the framework version is the one in <snapshot-dir>) silently mis-applies.

Golden rule 7 — agentic overrides are read at run-time. Every framework skill that supports overrides starts its run by checking .apache-magpie-overrides/<this-skill>.md for adopter-specific instructions and applying them before executing the default behaviour. The override file is plain markdown the agent interprets — no templating engine, no patch tool. See docs/setup/agentic-overrides.md for the contract.

Golden rule 8 — two families are always installed; the rest are opt-in. Two skill families are wired up unconditionally on every adopt / upgrade / worktree-init run and the user is never asked about them:

• setup-* — every framework skill whose source name starts with setup- except the bootstrap setup itself (which is copied as magpie-setup per Rule 6, not symlinked). Concretely: setup-isolated-setup-install, setup-isolated-setup-update, setup-isolated-setup-verify, setup-override-upstream, setup-shared-config-sync, plus any new setup-* skill the framework grows in the future — each symlinked as magpie-setup-*.
• list-* — the discovery family; every framework skill whose source name starts with list-. Today this is list-skills only (symlinked as magpie-list-skills); the prefix lets the framework grow a discovery family without re-prompting every adopter.

These two families are not exposed in the skill-families: prompt and not stored as user-selectable in the lock files; every sub-action that wires symlinks always covers them in addition to the user's opt-in family picks (security, pr-management, issue). Dropping them is not a supported configuration — the secure-setup and discovery flows the framework ships depend on those skills being callable.

Golden rule 9 — reload setup in-flight after a self-update. When a sub-action changes or creates the content of the committed setup skill (in practice: adopt recovering an out-of-date bootstrap, or upgrade's overwrite-from-snapshot step), the agent re-reads the modified files of this skill before continuing the rest of the current run. Concretely: after the copy lands on disk, re-load SKILL.md and the sub-action file you are currently executing (and any helper file you have already opened, such as agents.md or overrides.md), then resume from the step after the overwrite. The reload runs as the first thing that happens after the overwrite, before any further reconciliation, symlink work, or doc updates. The reason: the snapshot's skill version may have renamed steps, added new sub-actions, or changed the symlink contract; finishing the run against the old in-memory copy of the skill would silently mis-apply the new framework version the project just pinned to.

Sub-actions

The skill dispatches by the first positional argument:

Main-checkout-only sub-actions (adopt, upgrade, unadopt) detect their context via git rev-parse --git-dir ≠ git rev-parse --git-common-dir and refuse to run in a worktree with a pointer back to the main checkout. The worktree counterpart of adopt is worktree-init; for upgrade, every worktree automatically sees the refreshed snapshot once the main runs upgrade, because each worktree's <snapshot-dir> is a symlink to the main's.

adopt and upgrade always chain into worktree-init on every linked worktree as their final pass. The chain is unconditional — even on a fresh adoption with no linked worktrees yet (the pass becomes a no-op), even on an upgrade where every worktree already looks wired (worktree-init is idempotent, repairs broken symlinks, and adds new always-on-family entries the upgrade introduced). The user does not need to remember to cd into each worktree and re-run anything; the main-checkout sub-action propagates state outward to the worktrees by itself. See adopt.md Step 12.2 and upgrade.md Step 6c.

If the snapshot is missing (no <snapshot-dir>/) and <committed-lock> exists, the skill treats any sub-action as the recover-snapshot path: re-install per the committed lock first, then continue.

Inputs

What this skill is NOT for

• Not for installing the secure agent setup (sandbox, hooks, pinned tools). That is setup-isolated-setup-install.
• Not for upgrading framework tools installed on the host (bubblewrap, socat, claude-code itself). That is setup-isolated-setup-update.
• Not for syncing the user's ~/.claude-config across machines. That is setup-shared-config-sync.
• Not for committing framework changes. Framework PRs go against apache/airflow-steward directly — the snapshot is read-only.

Failure modes