Port import skill to MCP-first conventions

[chrisghill]

Re-lands the /massdriver:import workflow from add-import-skill, rewritten against main's MCP-first conventions, and rewrites the README for users rather than as a tool inventory.

Import skill, ported to MCP-first

The original branch predates the MCP server, so the mechanics were rebuilt rather than rebased:

• Control-plane operations move to MCP tools — mass component add → add_component, mass instance version --release-channel → update_instance with latest+dev, mass instance deploy --plan → create_deployment with action: PLAN + get_deployment_logs, mass environment default → set_environment_default. Adds a get_viewer auth check and the OCI repo/grant step that add_component now requires.
• Drops reading ~/.config/massdriver/config.yaml for state backend credentials, which main now forbids outright. TF_HTTP_USERNAME/TF_HTTP_PASSWORD come from already-exported env vars or from the user.
• Corrects the backend guidance — TF_HTTP_* only applies when the http backend is selected, so the procedure writes a throwaway backend_import.tf rather than assuming no backend block works.
• Takes the state URL from get_instance's statePaths[].stateUrl instead of hand-building the /state/<id>/<step> URL.
• New safety step — tofu import runs through Bash, so the safety hook does not guard it; the workflow now requires explicit confirmation before writing to a production instance's state. Also adds the local provider-credential step tofu import actually needs, state recovery via tofu state rm / orphan_instance, and the resource type instructions field for Path C.
• Agent frontmatter matches current convention (skills: massdriver, no tools or model keys); drops v2 wording and the massdriver/-prefix rule.

Version 4.2.0 → 4.3.0.

README rewrite

The README is user-facing — no agent reads it — so naming individual MCP tools throughout was noise.

• Command "What it does" lists describe outcomes instead of annotating each step with the tool that performs it.
• "How It Works" was an inventory of the MCP tool surface; it now covers what the plugin does for you — blueprint-once semantics, log streaming, safe dry runs, development release isolation, day 2, environment waves.
• Tool names are kept in Safety Guardrails, where they're load-bearing: those are the exact strings users see in permission prompts, so knowing approve_deployment is hard-blocked and get_*/list_* are auto-approved explains behavior they'll actually observe.
• The MCP-first paragraph read as a prerequisite warning; it's now a lede about what the plugin does, with the MCP server as the reason Docker is required rather than the headline.

Terminology: "citizen engineer" → "citizen developer"

"Citizen developer" is the established industry term for the audience the architect agent targets. Straight swap across the architect agent, its command, and the README; no behavior or guidance changes.

🤖 Generated with Claude Code