isV1 only triggered on a fixed set of top-level keys, so a config with just
$schema and a v1-shaped mcp block (servers directly under mcp, with enabled)
was parsed as v2 and silently produced zero servers. Detect the v1 mcp shape
(no servers wrapper, entries with a type) so these files migrate correctly.
Add MCP.callTool plus a dynamic jsonSchema mode for Tool.make, and register
each MCP server's tools as canonical Location-scoped tools via an McpTool
producer that reconciles on tools-changed notifications. Content-first tool
results preserve image/audio blocks; remove the arbitrary pagination page cap;
log mcp connection outcomes.
Tool naming keeps v1 parity (server_tool) so existing deny rules apply.
Add tools() and onToolsChanged() to the MCPClient.Connection abstraction:
list is capability-gated, paginated with cursor-dedup, and tolerates
unresolvable outputSchemas via a per-page fallback. Tool listing is folded
into connect so a listing failure marks the server failed rather than leaving
it connected with a silently empty list. Cache MCP.Tool values per server,
refresh on tools/list_changed notifications, and publish McpEvent.ToolsChanged.
MCP.tools() aggregates the cached tools across all connected servers.
Capture each server's initialize instructions on the Connection at connect
time and expose them via MCP.instructions(). Add McpGuidance, a system-context
source modeled on SkillGuidance, that renders the <mcp_instructions> block and
hides servers whose contributed tools are all denied for the agent.
instructions() returns just { server, instructions }; the guidance layer
fetches tools() separately and does the agent-aware permission correlation,
mirroring how SkillGuidance composes raw data with agent permissions.
ShellSelect.args() ran zsh/bash with -l and explicitly sourced .zshrc/.bashrc,
loading user functions and aliases that can shadow builtins and return non-zero
exit codes, breaking && chains. Match the old tool behavior: plain non-login
non-interactive shell -c command with cwd passed via spawn options.
Connect configured MCP servers in per-server child scopes at layer boot,
tracking status and a startup deferred per server. Query methods wait for
the relevant server(s) to finish startup before serving.
Drop the in-memory connect/disconnect/add overrides; runtime enable/disable
will come later through config updates. List methods (tools/instructions/
prompts/resourceCatalog) are aggregate across all servers with per-item
provenance; prompt/readResource stay server-scoped for routing.
Add a Connection abstraction in client.ts so the SDK Client never leaks
into the rest of core.
