MCP tools

kumihimo mcp <plan> serves one plan; every tool signature is therefore free of path arguments. Errors carry the same messages the CLI prints.

Tool	What it does
`get_plan`	The whole graph: manifest meta plus every node's identity, edges (`needs`/`in`/`links`), mentions (`agents`/`skills`/`trains`), and fields (bodies elided)
`get_node(node_id)`	One node in full: raw and effective fields, edges, mentions, body
`add_node(node_id, kind, title?, body?, fields?, needs?, in_?)`	Create a node; edge targets must exist
`update_node(node_id, kind?, title?, body?, priority?, set_fields?, unset_fields?)`	Change a node; file comments survive
`remove_node(node_id, force?)`	Delete; refuses while referenced (naming referrers) unless `force` strips the edges too
`link(src, needs?/in_?/to?/agents?/skills?/trains?, rel?)`	Draw exactly one edge from the six; cycle-closing `needs` is refused with the path, and a wrong-kind mention target (`agents` wants `agent`, `skills` wants `skill`, `trains` wants `agent` or `skill`) is refused naming the kind it expected
`unlink(src, needs?/in_?/to?/agents?/skills?/trains?)`	Remove exactly one edge, mentions included; absent edges error
`rename_node(old, new)`	Move to a new id; file bytes untouched, every referrer and the view layout fixed
`check()`	Every validation finding, errors first
`braid(strategy?, where?, from_?, until?, in_?, for_agent?, diagram?, dry?)`	Compile; returns `{text, order, warnings}`. `for_agent` is `--for`: one agent's work orders, opening with `Ground with:` when it has a `retrieval` field. A `for_agent` that doesn't exist or isn't kind `agent` raises, naming which
`ready(for_agent?)`	Nodes whose own status is `todo` and whose dependencies are all satisfied (no status, or `done`/`settled`/`answered`). `for_agent` narrows to nodes whose `agents:` key names that id — `skills:`/`trains:` are deliberately not consulted. Validated the same way as `braid`'s `for_agent`: raises rather than returning an empty list for a bad id
`crew()`	Every `agent`/`skill`/`reference` node, sorted by kind then id: effective fields, `trained` date verbatim, mention counts (per mention key, plus consult-links for references)

Mutations write to disk immediately — files are the only truth — so a running kumihimo edit canvas follows every MCP change live.

trained/cadence values from crew() are never compared to a clock — no surface in this library decides staleness; the caller reading crew()'s output does (PLAN2 §3.6).