refactor!: rebuild skill on shared agents core - #29

Summary

skill becomes the skills edition of agents: a thin command layer over agents-tools/core, as specified in core's Documentation/Proposals/AgentInstructions.md (Products, Commands).

• Deletes Sources/Core and its tests; depends on core's AgentsCore and AgentsSkills (core master during development; release preparation pins a minor-bounded version range).
• skill add|list|remove|update|init behave as agents skills …; skill install|update|doctor behave as the top-level agents commands restricted to skills. Options, help, output, JSON schemas, remedies and exit codes match agents. Instruction selections, locked instruction records and target files stay untouched.
• Breaking: state moves to the shared agents.json / agents.resolved (project .agent/, user ~/.agents/); 0.2 state files are not read. The 0.2 watch ledger and review commands are removed; watch returns as a subscription in a follow-up PR before 0.3.0.
• CI runs swift-format lint --strict and swift build -c release on every PR. Documentation/Reference/Release.md documents releasing skill.
• README, CLI reference, architecture inventory, AGENTS.md, CONTRIBUTING and the bundled skills/skill-cli companion describe the new surface.

Issues

Resolved here: Resolves #19, Resolves #20, Resolves #21, Resolves #22, Resolves #23, Resolves #24, Resolves #25, Resolves #26.

Resolved in core and adopted by this change: Resolves #4, Resolves #7, Resolves #8, Resolves #14, Resolves #17 (core PRs agents-tools/core#4 and agents-tools/core#7).

Superseded by the proposal, to be closed with a link after merge: #9 (preview may fetch into mirrors; state stays untouched), #18 (watch is a subscription, not an install gate).

Still open, for a follow-up core change: #15 (core rejects a conflicting owner but names only the existing one; the issue asks for both) and #16 (needs a targeted core regression test for unqualified adds after other requirement kinds).

Validation

• swift build, swift test (agent configuration variables unset) → 28 tests pass; swift build -c release → pass; strict lint → 0 findings; git diff --check → clean.
• Subprocess tests run with an allowlisted environment; a full run with sentinel CODEX_HOME, CLAUDE_CONFIG_DIR and XDG_* directories left every sentinel unchanged.
• Parity: the same add → list → install after deletion → update → update --apply → remove flow with skill and with agents skills produced byte-identical agents.json, agents.resolved and installed trees at every step, in project and user scope.