fix: reduce AI em-dashes in paper, use natural punctuation

Replaced em-dashes (---) with parentheses, semicolons, colons, or periods:
- Abstract: 'systems---APIs' → 'systems (APIs'
- Abstract: 'sessions---OS' → 'sessions, including OS'
- Intro: 'connection---this server' → 'connection (this server)'
- Souls: 'soul---a persistent' → 'soul, a persistent'
- Validation: 'pipeline---from' → 'pipeline, from'
- Discussion: 'changes---only' → 'changes; it only requires'
- Conclusion: 'contributions---' → 'contributions ('
- Limitations: 'sharing---each' → 'sharing; each'
- Crate descriptions: '---' → ').'

5 remaining dashes are in TikZ node labels and table cells (standard).

Author: omoshola-o

paper/paper-i-universal-connectivity/agenticconnect-paper.pdf

@@ -76,7 +76,7 @@
 % ABSTRACT
 % ============================================================================
 \begin{abstract}
-Autonomous AI agents require access to external systems---APIs, databases, servers, email, cloud infrastructure---yet existing connectivity tools force fragmented integration: one library per protocol, each with separate authentication, error handling, and retry logic. No tool learns from past interactions, and knowledge gained from one connection is unavailable to others. We present AgenticConnect, a universal external interface engine that unifies 18~protocol families behind a single MCP (Model Context Protocol)~\cite{mcp2024} server exposing 123~tools. The system introduces three components: (1)~\emph{Connection Souls}, persistent profiles that accumulate knowledge about remote systems across sessions---OS fingerprints, performance baselines ($P_{50}$, $P_{95}$, $P_{99}$ latency), and error pattern histories capped at 100 records per connection; (2)~an \emph{Intelligent Retry Fabric} with a five-class failure taxonomy (transient, permanent, rate-limit, authentication, network) and per-endpoint circuit breakers~\cite{nygard2018release} that open after $N$ consecutive failures (default $N{=}5$); and (3)~an encrypted \emph{Credential Vault} using AES-256-GCM with PBKDF2-HMAC-SHA256 key derivation (100{,}000~iterations) supporting eight authentication methods including OAuth~2.0~\cite{rfc6749} with automatic token lifecycle management. We implement AgenticConnect in 6{,}203~lines of Rust across four crates with eight engine modules. The system achieves sub-microsecond failure classification latency, processes 10{,}000 failure records without memory growth (history bounded at 500 entries), and encrypts 100{,}KB payloads with zero data corruption. All 164~tests pass including 20~paper-claim validation tests. The system requires no cloud services, stores all data in a single SQLite~\cite{sqlite2024} database, and speaks standard MCP~\cite{mcp2024} over stdio.
+Autonomous AI agents require access to external systems (APIs, databases, servers, email, cloud infrastructure), yet existing connectivity tools force fragmented integration: one library per protocol, each with separate authentication, error handling, and retry logic. No tool learns from past interactions, and knowledge gained from one connection is unavailable to others. We present AgenticConnect, a universal external interface engine that unifies 18~protocol families behind a single MCP (Model Context Protocol)~\cite{mcp2024} server exposing 123~tools. The system introduces three components: (1)~\emph{Connection Souls}, persistent profiles that accumulate knowledge about remote systems across sessions, including OS fingerprints, performance baselines ($P_{50}$, $P_{95}$, $P_{99}$ latency), and error pattern histories capped at 100 records per connection; (2)~an \emph{Intelligent Retry Fabric} with a five-class failure taxonomy (transient, permanent, rate-limit, authentication, network) and per-endpoint circuit breakers~\cite{nygard2018release} that open after $N$ consecutive failures (default $N{=}5$); and (3)~an encrypted \emph{Credential Vault} using AES-256-GCM with PBKDF2-HMAC-SHA256 key derivation (100{,}000~iterations) supporting eight authentication methods including OAuth~2.0~\cite{rfc6749} with automatic token lifecycle management. We implement AgenticConnect in 6{,}203~lines of Rust across four crates with eight engine modules. The system achieves sub-microsecond failure classification latency, processes 10{,}000 failure records without memory growth (history bounded at 500 entries), and encrypts 100{,}KB payloads with zero data corruption. All 164~tests pass including 20~paper-claim validation tests. The system requires no cloud services, stores all data in a single SQLite~\cite{sqlite2024} database, and speaks standard MCP~\cite{mcp2024} over stdio.
 \end{abstract}
 
 % ============================================================================
@@ -87,7 +87,7 @@ \section{Introduction}
 
 AI agents built on large language models~\cite{vaswani2017attention} increasingly operate as autonomous actors: managing infrastructure, interacting with APIs, querying databases, and communicating with external services. Yet the connectivity layer between an agent's reasoning capabilities and the external world remains fundamentally fragmented.
 
-Consider an agent tasked with deploying a service update. It must SSH into a server to check prerequisites, make HTTP calls to a container registry, query a PostgreSQL database for configuration, and send a webhook notification on completion. Today, this requires four separate libraries (\texttt{russh} for SSH, \texttt{reqwest}~\cite{reqwest2024} for HTTP, \texttt{sqlx} for PostgreSQL, \texttt{ring}~\cite{ring2024} for HMAC signatures), four authentication configurations, four error handling strategies, and four retry policies. Knowledge gained from one connection---``this server times out during peak hours''---is not available to any other connection. Every session starts from zero.
+Consider an agent tasked with deploying a service update. It must SSH into a server to check prerequisites, make HTTP calls to a container registry, query a PostgreSQL database for configuration, and send a webhook notification on completion. Today, this requires four separate libraries (\texttt{russh} for SSH, \texttt{reqwest}~\cite{reqwest2024} for HTTP, \texttt{sqlx} for PostgreSQL, \texttt{ring}~\cite{ring2024} for HMAC signatures), four authentication configurations, four error handling strategies, and four retry policies. Knowledge gained from one connection (``this server times out during peak hours'') is not available to any other connection. Every session starts from zero.
 
 Current approaches to agent connectivity fall into three categories, each with limitations that become critical at scale:
 
@@ -164,7 +164,7 @@ \section{Introduction}
 
 \begin{enumerate}[leftmargin=*, topsep=3pt, itemsep=1pt]
 \item A \textbf{protocol-agnostic connection engine} that auto-detects and handles 18~protocol families through URL scheme parsing, port mapping, and TCP banner analysis (Section~\ref{sec:protocols}).
-\item \textbf{Connection Souls}: persistent profiles that accumulate OS fingerprints, latency baselines, and error histories. Each soul enables predictive connectivity---the agent knows a server's typical behavior before connecting (Section~\ref{sec:souls}).
+\item \textbf{Connection Souls}: persistent profiles that accumulate OS fingerprints, latency baselines, and error histories. Each soul enables predictive connectivity: the agent knows a server's typical behavior before connecting (Section~\ref{sec:souls}).
 \item An \textbf{Intelligent Retry Fabric} with a formal five-class failure taxonomy and per-endpoint circuit breakers~\cite{nygard2018release} that open after configurable consecutive failures (Section~\ref{sec:retry}).
 \end{enumerate}
 
@@ -224,10 +224,10 @@ \subsection{Crate Structure}
 The system comprises four Rust crates:
 
 \begin{itemize}[nosep, leftmargin=*]
-  \item \textbf{\texttt{agentic-connect}} (2{,}252 lines) --- Core library containing type definitions and eight engine modules: ConnectionStore, RetryEngine, CredentialVault, DbConnection, HttpClient, ProtocolDetect, TlsInspect, and WebhookEngine.
-  \item \textbf{\texttt{agentic-connect-mcp}} (2{,}537 lines) --- MCP server with JSON-RPC 2.0 over stdio, tool registry dispatching 123~tools, and session management.
-  \item \textbf{\texttt{agentic-connect-cli}} (41 lines) --- CLI binary (\texttt{acnx}) for terminal-based connection management.
-  \item \textbf{\texttt{agentic-connect-ffi}} (25 lines) --- C-compatible FFI bindings (\texttt{cdylib} + \texttt{staticlib}).
+  \item \textbf{\texttt{agentic-connect}} (2{,}252 lines). Core library containing type definitions and eight engine modules: ConnectionStore, RetryEngine, CredentialVault, DbConnection, HttpClient, ProtocolDetect, TlsInspect, and WebhookEngine.
+  \item \textbf{\texttt{agentic-connect-mcp}} (2{,}537 lines). MCP server with JSON-RPC 2.0 over stdio, tool registry dispatching 123~tools, and session management.
+  \item \textbf{\texttt{agentic-connect-cli}} (41 lines). CLI binary (\texttt{acnx}) for terminal-based connection management.
+  \item \textbf{\texttt{agentic-connect-ffi}} (25 lines). C-compatible FFI bindings (\texttt{cdylib} + \texttt{staticlib}).
 \end{itemize}
 
 % Figure 2: Architecture diagram
@@ -323,7 +323,7 @@ \subsection{Protocol Detection}
 \subsection{Connection Souls}
 \label{sec:souls}
 
-Every connection in AgenticConnect accumulates a \emph{soul}---a persistent profile stored in SQLite that grows richer with each interaction. The soul contains four components:
+Every connection in AgenticConnect accumulates a \emph{soul}, a persistent profile stored in SQLite that grows richer with each interaction. The soul contains four components:
 
 \begin{itemize}[nosep, leftmargin=*]
   \item \textbf{System fingerprint}: OS, server software and version, supported protocols, timezone, detected services. Populated on first connection through banner analysis and response headers.
@@ -741,7 +741,7 @@ \subsection{Comparison with Existing Approaches}
 \section{Empirical Validation}
 \label{sec:validation}
 
-Beyond the micro-benchmarks of Section~\ref{sec:evaluation}, we conducted four phases of end-to-end validation testing the complete MCP pipeline---from JSON-RPC request through tool dispatch and engine execution to response serialization.
+Beyond the micro-benchmarks of Section~\ref{sec:evaluation}, we conducted four phases of end-to-end validation testing the complete MCP pipeline, from JSON-RPC request through tool dispatch and engine execution to response serialization.
 
 \subsection{Phase 1: Type Foundation}
 
@@ -819,21 +819,21 @@ \subsection{Paper Claim Validation}
 \section{Discussion}
 \label{sec:discussion}
 
-\textbf{Protocol-level vs.\ service-level.} Integration platforms~\cite{zapier2024} require a custom integration for each service. AgenticConnect operates at the protocol level: any HTTP API works immediately because the engine understands HTTP, OAuth~2.0~\cite{rfc6749}, rate limiting, and retry natively. Adding support for a new API requires zero code changes---only configuring a connection. This distinction becomes critical at scale: an agent managing 50 different APIs does not need 50 integrations.
+\textbf{Protocol-level vs.\ service-level.} Integration platforms~\cite{zapier2024} require a custom integration for each service. AgenticConnect operates at the protocol level: any HTTP API works immediately because the engine understands HTTP, OAuth~2.0~\cite{rfc6749}, rate limiting, and retry natively. Adding support for a new API requires zero code changes; it only requires configuring a connection. This distinction becomes critical at scale: an agent managing 50 different APIs does not need 50 integrations.
 
 \textbf{Connection Souls as accumulated intelligence.} Traditional clients treat every connection as stateless and independent. Connection Souls transform connectivity from a transactional operation into a learning process. Over time, the agent's understanding of its environment grows: server performance baselines inform timeout settings, error patterns predict failures before they occur, and capability maps enable context-aware command execution. This mirrors how human operators build intuition about the systems they manage.
 
 \textbf{The circuit breaker as a safety mechanism.} Without circuit breakers, a failing API can cause cascade failures across dependent systems. An agent making 100 requests to a failing endpoint wastes time and quota, and the retries themselves can worsen the overload. Circuit breakers~\cite{nygard2018release} provide automatic isolation: the failing endpoint is quarantined, other endpoints continue operating, and the circuit re-opens automatically when the endpoint recovers.
 
-\textbf{Limitations.} (1)~Protocol detection relies on port-based heuristics when TCP banners are unavailable; a full TLS handshake analyzer would improve accuracy for encrypted protocols. (2)~Database support is SQLite-only~\cite{sqlite2024} in the core build; PostgreSQL and MySQL require the \texttt{sqlx} feature flag. (3)~Browser automation tools require headless Chrome, detected gracefully at runtime with informative error messages. (4)~OAuth~2.0 token refresh requires the HTTP feature flag. (5)~Connection Souls do not yet support cross-agent sharing---each agent instance maintains its own soul database.
+\textbf{Limitations.} (1)~Protocol detection relies on port-based heuristics when TCP banners are unavailable; a full TLS handshake analyzer would improve accuracy for encrypted protocols. (2)~Database support is SQLite-only~\cite{sqlite2024} in the core build; PostgreSQL and MySQL require the \texttt{sqlx} feature flag. (3)~Browser automation tools require headless Chrome, detected gracefully at runtime with informative error messages. (4)~OAuth~2.0 token refresh requires the HTTP feature flag. (5)~Connection Souls do not yet support cross-agent sharing; each agent instance maintains its own soul database.
 
 % ============================================================================
 % 7. CONCLUSION
 % ============================================================================
 \section{Conclusion}
 \label{sec:conclusion}
 
-We presented AgenticConnect, a universal external interface engine for AI agents that unifies 18~protocol families behind 123~MCP tools. The three core contributions---Connection Souls (persistent system knowledge), Intelligent Retry Fabric (five-class failure taxonomy with circuit breakers), and the encrypted Credential Vault (AES-256-GCM, 8 auth methods)---transform connectivity from a stateless, per-protocol problem into a learning, unified system.
+We presented AgenticConnect, a universal external interface engine for AI agents that unifies 18~protocol families behind 123~MCP tools. The three core contributions (Connection Souls (persistent system knowledge), Intelligent Retry Fabric (five-class failure taxonomy with circuit breakers), and the encrypted Credential Vault (AES-256-GCM, 8 auth methods) transform connectivity from a stateless, per-protocol problem into a learning, unified system.
 
 The system achieves sub-microsecond latency for failure classification and circuit breaker checks, processes 10{,}000 failure records with bounded memory (500-entry cap), and encrypts 100{,}KB payloads without corruption. All 164+ tests pass, including 20 paper-claim validation tests that verify every quantitative claim in this paper.