Session Reconciliation Protocol Design

Issue: #99 - Agent should clean up orphaned PTY sessions on reconnect Status: Draft Author: Architecture Designer Date: 2026-02-02

Behavior update (2026-07, #634): the agent no longer kills tracked

workloads on transport loss. `cleanup_sessions()` (SIGTERM-all on every

stream loss or connect failure) was removed; `running_commands` and the

child processes are preserved across reconnects, the output channel is

persistent (output produced while disconnected buffers and flushes after

reconnect), and the re-registration `SessionReport` re-reports the live

sessions. The server's `import_reported_sessions` + `reconcile_sessions`

is now the sole kill authority for reconnect scenarios. Disk persistence

of session state remains future work.

Problem Statement

When the management server restarts, it loses track of active PTY sessions on agents. The agents continue running PTY processes (typically tmux sessions) that:

1. Send output for command IDs the server no longer recognizes 2. Cause terminal corruption when users reconnect to a "new" session 3. May leak resources (orphaned tmux/PTY processes) over time

Current Behavior

Agent                             Server (restarted)
  |                                    |
  |-------- stdout for cmd-abc ------->| "unknown command: cmd-abc" (warning)
  |                                    |
  |<------ new shell: cmd-xyz ---------|
  |                                    |
  |  (tmux "main" still running        |
  |   from old cmd-abc, conflicting    |
  |   with new cmd-xyz for "main")     |

Desired Behavior

Agent                             Server (restarted)
  |                                    |
  |-------- Registration ------------->|
  |<------- RegistrationAck -----------|
  |                                    |
  |<------- SessionQuery --------------|
  |                                    |
  |-------- SessionReport ------------>| (reports cmd-abc active)
  |                                    |
  |<------- SessionReconcile ----------| (kill cmd-abc, unknown to server)
  |                                    |
  |<------ new shell: cmd-xyz ---------|
  |                                    |
  |  (clean slate, no conflicts)       |

Design Decision

Recommendation: Hybrid Approach (Server-Initiated with Agent-Reported State)

After analyzing the options, I recommend a hybrid approach that combines the strengths of server-initiated and agent-initiated reconciliation:

Aspect	Server-Initiated	Agent-Initiated	Hybrid (Recommended)
Complexity	Low	High	Medium
Race conditions	Few	Many	Few
Recovery from partial failures	Limited	Better	Best
Preserves sessions on brief network issues	No (kill-all)	Yes	Yes
Server is authoritative	Yes	No	Yes

Why Hybrid?

1. Server is authoritative - Only the server knows which sessions are valid 2. Agent has ground truth - Only the agent knows what's actually running 3. Graceful reconciliation - Preserves sessions that the server still recognizes 4. Handles all edge cases - Both reconnect and server restart scenarios

Protocol Design

New gRPC Messages

Add to `proto/agent.proto`:

// =============================================================================
// Session Reconciliation Messages
// =============================================================================

// Server queries agent for active sessions
message SessionQuery {
  // If true, agent should report all active sessions
  // If false, agent should only report sessions matching session_ids
  bool report_all = 1;

  // Optional: specific session IDs to query (if report_all is false)
  repeated string session_ids = 2;
}

// Agent reports its active sessions to server
message SessionReport {
  string agent_id = 1;
  repeated ActiveSession sessions = 2;
  int64 timestamp_ms = 3;
}

// Description of an active session on the agent
message ActiveSession {
  string command_id = 1;        // UUID assigned by server
  string session_name = 2;      // e.g., "main", "claude"
  SessionType session_type = 3; // Interactive, Headless, Background
  string command = 4;           // Original command string
  int64 started_at_ms = 5;      // When session was created
  int32 pid = 6;                // Process ID (for debugging)
  bool is_pty = 7;              // Whether this is a PTY session
  string session_id = 8;        // Stable identity assigned by the server
}

// Session types matching dispatcher.rs
enum SessionType {
  SESSION_TYPE_UNKNOWN = 0;
  SESSION_TYPE_INTERACTIVE = 1;
  SESSION_TYPE_HEADLESS = 2;
  SESSION_TYPE_BACKGROUND = 3;
}

// Server instructs agent to reconcile sessions
message SessionReconcile {
  // Sessions to keep (server recognizes these)
  repeated string keep_session_ids = 1;

  // Sessions to terminate (server doesn't recognize)
  repeated string kill_session_ids = 2;

  // Kill all sessions not in keep_session_ids (nuclear option)
  bool kill_unrecognized = 3;

  // Grace period before SIGKILL (seconds)
  int32 grace_period_seconds = 4;
}

// Agent confirms reconciliation complete
message SessionReconcileAck {
  string agent_id = 1;
  repeated string killed_session_ids = 2;
  repeated string kept_session_ids = 3;
  repeated string failed_to_kill = 4;  // Sessions that couldn't be terminated
  int64 timestamp_ms = 5;
}

Updated Message Envelopes

Add to `AgentMessage.payload` oneof:

message AgentMessage {
  oneof payload {
    // ... existing fields ...
    SessionReport session_report = 8;
    SessionReconcileAck session_reconcile_ack = 9;
  }
}

Add to `ManagementMessage.payload` oneof:

message ManagementMessage {
  oneof payload {
    // ... existing fields ...
    SessionQuery session_query = 8;
    SessionReconcile session_reconcile = 9;
  }
}

Sequence Diagrams

Normal Agent Reconnection (Server Running)

Agent                                 Server
  |                                      |
  |--- Connect (stream open) ----------->|
  |                                      |
  |--- Registration --------------------->|
  |                                      |
  |<-- RegistrationAck -------------------|
  |                                      |
  |<-- SessionQuery (report_all=true) ----|
  |                                      |
  |--- SessionReport (sessions: [        |
  |      {cmd-abc, "main", Interactive}, |
  |      {cmd-def, "claude", Headless}   |
  |    ]) ------------------------------>|
  |                                      |
  |    [Server checks dispatcher:        |
  |     - cmd-abc: NOT in active_sessions|
  |     - cmd-def: in active_sessions]   |
  |                                      |
  |<-- SessionReconcile (               |
  |      kill: [cmd-abc],               |
  |      keep: [cmd-def]                |
  |    ) --------------------------------|
  |                                      |
  |    [Agent kills cmd-abc PTY]         |
  |                                      |
  |--- SessionReconcileAck (            |
  |      killed: [cmd-abc],             |
  |      kept: [cmd-def]                |
  |    ) ------------------------------->|
  |                                      |
  |    [Reconciliation complete]         |

Server Restart (Agent Already Connected)

Agent                                 Server (new instance)
  |                                      |
  |    [gRPC stream broken]              |
  |                                      |
  |--- Connect (reconnect) -------------->|
  |                                      |
  |--- Registration --------------------->|
  |                                      |
  |<-- RegistrationAck -------------------|
  |                                      |
  |<-- SessionQuery (report_all=true) ----|
  |                                      |
  |--- SessionReport (sessions: [        |
  |      {cmd-old1, "main", Interactive},|
  |      {cmd-old2, "debug", Interactive}|
  |    ]) ------------------------------>|
  |                                      |
  |    [Server has empty dispatcher:     |
  |     NO sessions recognized]          |
  |                                      |
  |<-- SessionReconcile (               |
  |      kill_unrecognized=true,        |
  |      keep: []                       |
  |    ) --------------------------------|
  |                                      |
  |    [Agent kills ALL sessions]        |
  |                                      |
  |--- SessionReconcileAck (            |
  |      killed: [cmd-old1, cmd-old2]   |
  |    ) ------------------------------->|
  |                                      |
  |    [Clean slate achieved]            |

Brief Network Interruption (Graceful Recovery)

Agent                                 Server
  |                                      |
  |    [Network blip, 5 second outage]   |
  |                                      |
  |--- Connect (reconnect) -------------->|
  |                                      |
  |--- Registration --------------------->|
  |                                      |
  |<-- RegistrationAck -------------------|
  |                                      |
  |<-- SessionQuery (report_all=true) ----|
  |                                      |
  |--- SessionReport (sessions: [        |
  |      {cmd-xyz, "main", Interactive}  |
  |    ]) ------------------------------>|
  |                                      |
  |    [Server checks dispatcher:        |
  |     - cmd-xyz: STILL in pending      |
  |       (not yet cleaned up)]          |
  |                                      |
  |<-- SessionReconcile (               |
  |      keep: [cmd-xyz]                |
  |    ) --------------------------------|
  |                                      |
  |--- SessionReconcileAck (            |
  |      kept: [cmd-xyz]                |
  |    ) ------------------------------->|
  |                                      |
  |    [Session preserved!]              |

Implementation Details

Server-Side Changes

1. Dispatcher Enhancement (`management/src/dispatch/dispatcher.rs`)

impl CommandDispatcher {
    /// Get all known command IDs for an agent
    pub fn get_known_command_ids(&self, agent_id: &str) -> Vec<String> {
        let mut ids = Vec::new();

        // From pending commands
        for entry in self.pending.read().iter() {
            if entry.1.agent_id == agent_id {
                ids.push(entry.0.clone());
            }
        }

        // From active sessions
        if let Some(sessions) = self.active_sessions.read().get(agent_id) {
            for info in sessions.values() {
                if !ids.contains(&info.command_id) {
                    ids.push(info.command_id.clone());
                }
            }
        }

        ids
    }

    /// Reconcile agent sessions after reconnect
    pub async fn reconcile_sessions(
        &self,
        agent_id: &str,
        reported_sessions: &[ActiveSession],
    ) -> SessionReconcile {
        let known_ids = self.get_known_command_ids(agent_id);

        let mut keep = Vec::new();
        let mut kill = Vec::new();

        for session in reported_sessions {
            if known_ids.contains(&session.command_id) {
                keep.push(session.command_id.clone());
            } else {
                kill.push(session.command_id.clone());
            }
        }

        SessionReconcile {
            keep_session_ids: keep,
            kill_session_ids: kill,
            kill_unrecognized: false,  // Explicit list preferred
            grace_period_seconds: 5,
        }
    }

    /// Handle reconciliation acknowledgment
    pub fn handle_reconcile_ack(&self, agent_id: &str, ack: SessionReconcileAck) {
        // Remove killed sessions from our tracking
        for killed_id in &ack.killed_session_ids {
            self.pending.write().remove(killed_id);
        }

        // Update active_sessions tracking
        if let Some(mut sessions) = self.active_sessions.write().get_mut(agent_id) {
            sessions.retain(|_, info| !ack.killed_session_ids.contains(&info.command_id));
        }

        info!(
            agent_id = %agent_id,
            killed = ?ack.killed_session_ids,
            kept = ?ack.kept_session_ids,
            failed = ?ack.failed_to_kill,
            "Session reconciliation complete"
        );
    }
}

2. gRPC Handler Enhancement (`management/src/grpc.rs`)

async fn handle_agent_message(
    registry: &Arc<AgentRegistry>,
    dispatcher: &Arc<CommandDispatcher>,
    output_agg: &Arc<OutputAggregator>,
    agent_id: &str,
    msg: AgentMessage,
    tx: mpsc::Sender<ManagementMessage>,
) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
    match msg.payload {
        Some(agent_message::Payload::Registration(reg)) => {
            // ... existing registration handling ...

            // Trigger session reconciliation after registration
            let query = SessionQuery {
                report_all: true,
                session_ids: vec![],
            };
            tx.send(ManagementMessage {
                payload: Some(management_message::Payload::SessionQuery(query)),
            }).await?;
        }

        Some(agent_message::Payload::SessionReport(report)) => {
            info!(
                agent_id = %agent_id,
                session_count = report.sessions.len(),
                "Received session report"
            );

            // Generate reconciliation instruction
            let reconcile = dispatcher.reconcile_sessions(agent_id, &report.sessions).await;

            tx.send(ManagementMessage {
                payload: Some(management_message::Payload::SessionReconcile(reconcile)),
            }).await?;
        }

        Some(agent_message::Payload::SessionReconcileAck(ack)) => {
            dispatcher.handle_reconcile_ack(agent_id, ack);
        }

        // ... existing handlers ...
    }
}

Agent-Side Changes

1. Session Tracking (`agent-rs/src/main.rs`)

impl AgentClient {
    /// Build session report from running_commands
    async fn build_session_report(&self) -> SessionReport {
        let running = self.running_commands.lock().await;

        let sessions: Vec<ActiveSession> = running
            .iter()
            .map(|(cmd_id, cmd)| ActiveSession {
                command_id: cmd_id.clone(),
                session_name: cmd.session_name.clone().unwrap_or_default(),
                session_type: cmd.session_type.into(),
                command: cmd.command.clone(),
                started_at_ms: cmd.started_at.elapsed().as_millis() as i64,
                pid: cmd.pid.map(|p| p.as_raw()).unwrap_or(0),
                is_pty: cmd.pty_control_tx.is_some(),
                session_id: cmd.session_id.clone().unwrap_or_default(),
            })
            .collect();

        SessionReport {
            agent_id: self.config.agent_id.clone(),
            sessions,
            timestamp_ms: chrono_timestamp_ms(),
        }
    }

    /// Kill sessions as instructed by server
    async fn kill_sessions(&self, session_ids: &[String], grace_seconds: i32) {
        for cmd_id in session_ids {
            if let Some(cmd) = self.running_commands.lock().await.get(cmd_id) {
                // Send SIGTERM first
                if let Some(pid) = cmd.pid {
                    let _ = nix::sys::signal::kill(pid, Signal::SIGTERM);
                }
            }
        }

        // Wait for grace period
        tokio::time::sleep(Duration::from_secs(grace_seconds as u64)).await;

        // SIGKILL any survivors
        for cmd_id in session_ids {
            if let Some(cmd) = self.running_commands.lock().await.remove(cmd_id) {
                if let Some(pid) = cmd.pid {
                    let _ = nix::sys::signal::kill(pid, Signal::SIGKILL);
                }
            }
        }
    }
}

2. Handle Reconciliation Messages

async fn handle_inbound(&self, msg: ManagementMessage, output_tx: mpsc::Sender<AgentMessage>) {
    match msg.payload {
        // ... existing handlers ...

        Some(Payload::SessionQuery(query)) => {
            info!("Received session query (report_all={})", query.report_all);

            let report = self.build_session_report().await;
            let _ = output_tx.send(AgentMessage {
                payload: Some(agent_message::Payload::SessionReport(report)),
            }).await;
        }

        Some(Payload::SessionReconcile(reconcile)) => {
            info!(
                "Received session reconcile: keep={}, kill={}",
                reconcile.keep_session_ids.len(),
                reconcile.kill_session_ids.len()
            );

            let to_kill = if reconcile.kill_unrecognized {
                // Kill everything not in keep list
                let running = self.running_commands.lock().await;
                running.keys()
                    .filter(|id| !reconcile.keep_session_ids.contains(id))
                    .cloned()
                    .collect()
            } else {
                reconcile.kill_session_ids.clone()
            };

            self.kill_sessions(&to_kill, reconcile.grace_period_seconds).await;

            // Build acknowledgment
            let kept: Vec<String> = self.running_commands.lock().await.keys().cloned().collect();
            let ack = SessionReconcileAck {
                agent_id: self.config.agent_id.clone(),
                killed_session_ids: to_kill,
                kept_session_ids: kept,
                failed_to_kill: vec![],  // TODO: track actual failures
                timestamp_ms: chrono_timestamp_ms(),
            };

            let _ = output_tx.send(AgentMessage {
                payload: Some(agent_message::Payload::SessionReconcileAck(ack)),
            }).await;
        }
    }
}

Edge Cases

1. Race: Command Arrives During Reconciliation

Scenario: Server sends new command while agent is processing SessionReconcile.

Mitigation:

Agent processes messages sequentially on the inbound stream
New commands queued after reconciliation completes
Server waits for ReconcileAck before dispatching new sessions

2. Agent Crashes Mid-Reconciliation

Scenario: Agent dies after receiving SessionReconcile but before sending Ack.

Mitigation:

Server times out waiting for Ack (30s default)
Server re-queries on next agent reconnect
Sessions may be orphaned until next reconnect, but not duplicated

3. Server Crashes Mid-Reconciliation

Scenario: Server dies after sending SessionQuery but before receiving Report.

Mitigation:

Agent's SessionReport is lost to the void
On new server startup, agent reconnects and process repeats
No harm done; agent state unchanged

4. Partial Kill Failure

Scenario: Some sessions fail to terminate (zombie processes, etc.)

Mitigation:

Agent reports `failed_to_kill` list in Ack
Server logs warning and may retry with SIGKILL
Operator can manually investigate via SSH

5. tmux Session vs PTY Process Mismatch

Scenario: tmux session exists but tracked PTY was killed.

Mitigation:

Agent tracks both PTY process PID and tmux session name
Kill tmux session by name (`tmux kill-session -t <name>`) as backup
Report includes both for debugging

6. Multiple Agents with Same Session Names

Scenario: agent-01 and agent-02 both have "main" session.

Mitigation:

Sessions tracked per-agent in dispatcher (`agent_id -> session_name -> info`)
Reconciliation scoped to single agent
Native host instances receive separate private `TMUX_TMPDIR` namespaces, so

same-user agents cannot discover or terminate one another's tmux sessions

No cross-agent interference

Metrics and Observability

Prometheus Metrics

# Session reconciliation counts
agentic_session_reconciliations_total{agent_id, result="success|partial|failed"}

# Sessions killed during reconciliation
agentic_sessions_killed_total{agent_id, reason="orphaned|server_restart|manual"}

# Sessions preserved during reconciliation
agentic_sessions_preserved_total{agent_id}

# Reconciliation latency
agentic_reconciliation_duration_seconds{agent_id}

Structured Logging

{
  "timestamp": "2026-02-02T10:00:00Z",
  "level": "info",
  "event": "session_reconciliation_complete",
  "agent_id": "agent-01",
  "sessions_killed": 2,
  "sessions_preserved": 1,
  "sessions_failed": 0,
  "duration_ms": 1234
}

Testing Strategy

Unit Tests

1. `test_session_report_building` - Verify agent builds correct report from running_commands 2. `test_reconcile_decision_logic` - Server correctly identifies orphaned vs valid sessions 3. `test_kill_signal_sequence` - SIGTERM followed by SIGKILL after grace period

Integration Tests

1. `test_reconnection_preserves_valid_sessions` - Session survives brief disconnect 2. `test_server_restart_cleans_orphans` - All sessions killed after server restart 3. `test_partial_kill_failure_handling` - Graceful handling of unkillable processes

E2E Tests

1. `test_e2e_server_restart_recovery` - Full server restart with active terminal session 2. `test_e2e_network_partition_recovery` - Simulated network outage with session preservation

Rollout Plan

Phase 1: Protocol Implementation (Non-Breaking)

1. Add new protobuf messages (backward compatible) 2. Implement server-side reconciliation logic 3. Server sends SessionQuery but handles missing response gracefully

Phase 2: Agent Implementation

1. Agent responds to SessionQuery with SessionReport 2. Agent handles SessionReconcile instructions 3. Feature flag: `ENABLE_SESSION_RECONCILIATION=true`

Phase 3: Full Enablement

1. Enable by default 2. Monitor metrics for reconciliation success rate 3. Tune grace periods based on observed behavior

Appendix A: Alternative Designs Considered

Option 1: Server-Initiated Kill-All

Simple approach where server sends "kill all sessions" on agent reconnect.

Pros: Simple, no new protocol needed Cons: Destroys valid sessions on brief network issues

Option 2: Agent-Initiated Cleanup

Agent detects server restart and cleans up autonomously.

Pros: Agent is self-healing Cons: Agent can't know which sessions server still expects; race conditions

Option 3: Session Persistence to Disk

Server persists session state to disk, recovers on restart.

Pros: Sessions survive server restart Cons: Complex persistence layer; state drift between disk and agent

Issue #99: Agent should clean up orphaned PTY sessions on reconnect (this design)
Issue #XX: Session persistence across server restarts (future enhancement)
Issue #XX: Multi-agent session handoff (future enhancement)

Session Reconciliation Protocol Design

Session Reconciliation Protocol Design

Problem Statement

Current Behavior

Desired Behavior

Design Decision

Recommendation: Hybrid Approach (Server-Initiated with Agent-Reported State)

Why Hybrid?

Protocol Design

New gRPC Messages

Updated Message Envelopes

Sequence Diagrams

Normal Agent Reconnection (Server Running)

Server Restart (Agent Already Connected)

Brief Network Interruption (Graceful Recovery)

Implementation Details

Server-Side Changes

1. Dispatcher Enhancement (`management/src/dispatch/dispatcher.rs`)

2. gRPC Handler Enhancement (`management/src/grpc.rs`)

Agent-Side Changes

1. Session Tracking (`agent-rs/src/main.rs`)

2. Handle Reconciliation Messages

Edge Cases

1. Race: Command Arrives During Reconciliation

2. Agent Crashes Mid-Reconciliation

3. Server Crashes Mid-Reconciliation

4. Partial Kill Failure

5. tmux Session vs PTY Process Mismatch

6. Multiple Agents with Same Session Names

Metrics and Observability

Prometheus Metrics

Structured Logging

Testing Strategy

Unit Tests

Integration Tests

E2E Tests

Rollout Plan

Phase 1: Protocol Implementation (Non-Breaking)

Phase 2: Agent Implementation

Phase 3: Full Enablement

Appendix A: Alternative Designs Considered

Option 1: Server-Initiated Kill-All

Option 2: Agent-Initiated Cleanup

Option 3: Session Persistence to Disk

Appendix B: Related Issues