v0.31.0

[0.31.0] - 2026-06-02

Changed

• Docs marked "Request sort" now drain automatically via a minion, instead of waiting for an agent that never came. The sidebar's sort marker always persisted correctly and list_pending_sorts always returned it — but nothing consumed the queue: 3 docs sat marked for 4–6 days while enrichment (identical surfacing scaffold) drained 217/223 docs over the same window. The difference was the directive, not the scaffold. Sort's footer/session-instruction carried advisory prose ("handle it inline, discuss with the user — no minion, sorting is a judgment call"); enrichment's carried an executable Agent(...) dispatch to an autonomous minion. An agent mid-task fires a fire-and-forget background dispatch; it does not derail the user's task for a judgment conversation. Sort now rides the same rail: sortFooter() and buildSortInstructions() emit a paste-ready dispatch for the new openwriter-sort-minion, which self-discovers via list_pending_sorts, picks a destination from workspace/container purpose hints, files the doc (move_item), retires the request (mark_sorted), and reports what moved. The "judgment call → must gate on a human" premise was self-defeating — propose_sort already moved the judgment to an async accept/reject step, and a sort-marked doc has no user-expected location to violate (the mark is the user delegating placement). Safety is reversibility + transparency (every move is one move_item to undo and is reported), not a pre-move gate — the same safety class as enrichment. SKILL.md firm rule 6 flipped from "no minion / inline" to "dispatch reflexively." → adr/sort-minion-drain.md
• Author's Voice model picker updated — Fast+ now maps to gemini-3.5-flash, and the standalone Gemini Pro option was dropped; the picker reads the selected tier live per request.
• Plugins tab: the model selector is now front-and-center. select config fields render at the top level of a plugin's settings (API keys/secrets stay in the collapsed "Settings" menu), and changing a config value now shows a trustworthy inline ✓ / ✗ reflecting the real save result — previously the save was fire-and-forget with no feedback.
• showToast is now OpenWriter's canonical toast primitive — consolidated and restyled with the context-menu design tokens.

Added

• autoSortDisabled workspace opt-out (mirrors enrichmentDisabled). Set via update_workspace_context; its docs drop from list_pending_sorts so the minion never auto-files them, falling back to the manual sidebar propose_sort → accept/reject flow. get_workspace_structure surfaces auto-sort: disabled in its header. propose_sort is retained unchanged for users who want to approve each move.
• Transform "ify" commands now refuse to run on an over-large document (5,000-word cap), with a toast — mirroring the Author's Voice selection guard. Stops a huge doc from triggering an unbounded-cost/time transform; client-side guard in the sidebar (transform-guard.ts) plus a server-side check in the publish plugin.
• Selection-based actions refuse over-large selections with a toast (client-side size guard, mirroring the AV API) — the same unbounded-cost protection for in-editor enhances.

Fixed

• Author's Voice now reads the selected model tier live per request, so a model-picker change takes effect immediately instead of being pinned at load.
• Plugin select fields no longer render a duplicate blank option.
• A plugin that fails to load is no longer persisted as disabled — so a transient load failure can't silently flip a plugin off.