AgentBoard v0.10.1

Site and MCP tools whose input schemas use pattern, format, $ref, or uniqueItems now appear and run. AgentBoard publishes third-party schemas to the model exactly as the author wrote them and no longer validates model arguments against them locally. The tool owns its constraints and reports violations itself, which is how native Chrome and MCP clients already behave; MCP-idiomatic tools return those rejections as results the model reads directly.

• No tool is withheld for the constraint vocabulary its schema uses. A single pattern or uniqueItems: true used to remove the whole tool silently.
• Large schemas work. A per-call validation budget refused realistic calls to tools with a few hundred schema nodes as "arguments do not match the declared schema"; the budget left with the validator.
• The checks that protect the whole request stay: an object root, bounded size, and keyword values of the type providers accept. One malformed schema still cannot take its sibling tools down with it.
• read_page validates its own arguments against the bounds it advertises and names the field it rejects.

Reliability

• Pages that rewrite an identical URL with history.replaceState after every interaction no longer cause the tool catalog to be rebuilt each time. Those rebuilds interrupted running agents and used up the sidebar's auto-continue budget within a few tool calls.
• When the extension refuses a WebMCP tool call, the sidebar and the model see the actual reason instead of a generic "WebMCP tool execution failed".
• User scripts load on pages whose Content Security Policy omits blob:; the page's nonce now reaches user-provided scripts.
• Extension pages no longer emit module preload hints, which cost a duplicate fetch on chrome-extension:// pages and filled the console with unused-preload warnings.

Full changelog: v0.10.0...v0.10.1