feat: model-accessible diagnostics search for skill and MCP loading issues #310

Goal

When a user asks "why aren't you activating skill X?" or "why aren't you using MCP server Y?", the agent can run a targeted search over startup and loading diagnostics and explain the cause.

Context

#309 makes invalid plugin components non-fatal: skipped skills, skills directories, mcp.json files, and MCP server entries, plus duplicate plugin MCP server names, are logged with tracing::warn! and the rest of the plugin still loads. In the shipped CLI no subscriber delivers these events (telemetry::init installs none by default, and exported_targets() only covers agentkit_loop and agentkit_mcp). The reason a component is missing is therefore invisible to both the user and the agent.

Proposal

A built-in tool the model can call to query diagnostics recorded during plugin, skill, and MCP loading and hot-reload.

• Record diagnostics per runtime generation: component kind (skill, MCP server, plugin), plugin alias, component name, path, reason, and whether the component was skipped or the generation was rejected.
• Support targeted queries by component name, plugin alias, or kind, so the agent fetches only the entries relevant to the question.
• Cover at least: agentkit_plugins diagnostics (SkillSkipped, SkillsLocationInvalid, McpDisabled, McpServerSkipped, InvalidExtensionsField), skipped plugin MCP servers (unsupported transport, invalid cwd, escaping command path, invalid header), duplicate-name precedence, and MCP server connection or initialization failures.
• Keep output bounded and free of secrets (no header values, tokens, or env values).

Acceptance

• Given a plugin skill with invalid frontmatter, asking the agent why the skill is unavailable leads it to call the tool and report the plugin, the file path, and the spec violation.
• Given two plugins that declare the same MCP server, the agent can say which plugin owns the server and which declaration was skipped.