Add per-agent scoping to skill add --mcp

[xsh310]

What

Adds an --agents option to ucode skill add --mcp so a skill schema can be added to a chosen subset of configured coding agents instead of all of them. This mirrors ucode mcp add --agents, closing the gap between how the skills MCP path and the general MCP path scope work per agent.

How

• add_skills_command(locations, agents=None) forwards agents to setup_mcp_clients, which scopes the returned client set. The per-client skill-location map is updated only for the targeted clients; the rest keep their scope untouched.
• In cli.py, skill add gains the --agents option, bootstraps any named agent that is not configured yet via _configure_agents_for_mcp, and guards that --agents is only valid with --mcp (downloaded skills use shared directory families and cannot be agent scoped).
• Without --agents, behavior is unchanged: every configured agent is updated.

Tests

• test_mcp.py: per-agent add touches only the named client's scope and re-registers only that client; a global add reaches every configured client while preserving per-client divergence; adding a schema an agent already has is a no-op.
• test_cli.py: --agents is parsed, bootstrapped, and forwarded for --mcp; an empty --agents list is rejected; --agents is rejected in download mode.

uv run pytest tests/test_mcp.py tests/test_cli.py tests/test_lint.py is green.

Manual verification (installed build)

Installed the stack tip with uv tool install --reinstall . (ucode 0.1.0+91.g5c0dc1e) and exercised the command in an isolated sandbox: a temporary HOME, stub claude/codex binaries that record every registration ucode issues, and DATABRICKS_BEARER set so auth is satisfied offline. State was seeded with both agents configured and no skills.

Scenario	Command	Result
Global add (all agents)	ucode skill add --location shared.skills --mcp	Both agents scoped to shared.skills; each re-registered once
Per-agent add	add_skills_command(["claude.only"], agents={"claude"})	claude=[shared.skills, claude.only], codex=[shared.skills]; only claude re-registered; summary showed divergent per-client URLs
Guard: no --location	ucode skill add	ERROR --location is required for ucode skill add.
Guard: --agents without --mcp	ucode skill add --location a.b --agents claude	ERROR --agents is only supported when using --mcp
Guard: empty --agents	ucode skill add --location a.b --mcp --agents ,	ERROR No agents provided for --agents ...

The --agents add path first runs an interactive agent bootstrap, so the per-agent case drove add_skills_command(..., agents=...) directly against the installed package with only that prompt stubbed; state I/O and per-client agent registration ran for real. CLI parsing and bootstrap are covered by the mocked unit tests.

Stacking

This is the first of a stacked series that adds per-agent scoping to the skills MCP connection, built on top of xshen/per-client-skill-scopes (per-client skill scope map) and skill-managed-download-only (managed skills delivered by download only). Reviewing the diff against its base branch shows just this change. It rebuilds behavior originally designed by Arthur Jenoudet on the current per-client-map state model.

This pull request and its description were written by Isaac.