v0.12.0-beta.14

Delta since previous beta (v0.12.0-beta.13) — 1 new changeset.

Minor Changes

• Agent writes now get advisory mermaid feedback, delivered on a new unified warnings channel. The write/edit MCP tools (and the underlying POST /api/agent-write-md / POST /api/agent-patch routes) validate every ```mermaid fence of the post-write document with the same mermaid version the editor renders with, and report parse failures as warnings entries (kind: "mermaid-parse-error" — fence locator + mermaid's line-numbered message) plus a ⚠ line in the tool response text — so the authoring agent can fix a broken diagram in the same session instead of the reader discovering the error chrome. Strictly advisory: writes land byte-faithfully regardless.

  The warnings array is the one advisory channel going forward: it also carries the existing write-integrity kinds (content-divergence, disk-edit-reconciled) on every mutating write surface (write, write-md, patch, frontmatter-patch, rollback / restore_version), so co-occurring advisories no longer mask one another. The single-valued warning field and the MCP structuredContent.document.contentDivergence nesting are deprecated — warning keeps emitting its highest-precedence integrity entry in parallel for one deprecation window; new consumers should read warnings.

  The palette tool's Mermaid entry now teaches the grammar sharp edges (raw ;/# terminate sequence-family message text — use commas or #59;/#35;; quote flowchart labels containing punctuation), and the tool-description footer points at it.