fix: codex review — safe ack, dream stops worker, graceful errors

Worker:
- Only ack inbox items if Claude succeeds (prevents losing work on crash)
- Graceful error if colony not in PATH (no panic)
- Check colony inbox exit code before parsing
- Per-agent prompt path (/tmp/colony-{name}-prompt.md)

Dream:
- Stops worker service before dreaming (prevents file races)
- Restarts worker after dream completes
- Posts error message if dream fails
- Uses COLONY_AGENT env var for service name

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/colony-agent/src/dream.rs

@@ -13,12 +13,17 @@ pub fn run_dream() {
         return;
     }
 
-    // 2. Announce dream
+    // 2. Stop worker to prevent file races
+    let agent_name = std::env::var("COLONY_AGENT").unwrap_or_else(|_| "agent".into());
+    let worker_service = format!("agent-{}-worker", agent_name);
+    let _ = Command::new("systemctl").args(["stop", &worker_service]).status();
+
+    // 3. Announce dream
     let _ = Command::new("colony")
         .args(["post", "general", "💤 dreaming... back in a few minutes", "--type", "plan", "--quiet"])
         .status();
 
-    // 3. Invoke Claude for dream cycle
+    // 4. Invoke Claude for dream cycle
     eprintln!("dreaming... ({} lines of memory)", line_count);
 
     let prompt = format!(
@@ -31,7 +36,7 @@ pub fn run_dream() {
         line_count
     );
 
-    let status = Command::new("claude")
+    let dream_ok = match Command::new("claude")
         .args([
             "--dangerously-skip-permissions",
             "-p",
@@ -39,16 +44,24 @@ pub fn run_dream() {
             "--max-turns",
             "10",
         ])
-        .status();
+        .status()
+    {
+        Ok(s) if s.success() => { eprintln!("dream completed"); true }
+        Ok(s) => { eprintln!("dream exited with status: {}", s); false }
+        Err(e) => { eprintln!("failed to run claude for dream: {}", e); false }
+    };
 
-    match status {
-        Ok(s) if s.success() => eprintln!("dream completed"),
-        Ok(s) => eprintln!("dream exited with status: {}", s),
-        Err(e) => eprintln!("failed to run claude for dream: {}", e),
+    // 5. Restart worker
+    let _ = Command::new("systemctl").args(["start", &worker_service]).status();
+
+    // 6. Announce return
+    if dream_ok {
+        let _ = Command::new("colony")
+            .args(["post", "general", "👁 back from dreaming", "--type", "plan", "--quiet"])
+            .status();
+    } else {
+        let _ = Command::new("colony")
+            .args(["post", "general", "⚠ dream failed, back online", "--type", "error", "--quiet"])
+            .status();
     }
-
-    // 4. Announce return
-    let _ = Command::new("colony")
-        .args(["post", "general", "👁 back from dreaming", "--type", "plan", "--quiet"])
-        .status();
 }

crates/colony-agent/src/worker.rs

@@ -4,10 +4,18 @@ use std::time::Duration;
 /// Run one pulse cycle. Returns true if Claude was invoked (there was work).
 pub fn run_pulse(config_path: &str) -> bool {
     // 1. Check inbox
-    let inbox_output = Command::new("colony")
-        .args(["inbox", "--json"])
-        .output()
-        .expect("failed to run colony inbox");
+    let inbox_output = match Command::new("colony").args(["inbox", "--json"]).output() {
+        Ok(o) => o,
+        Err(e) => {
+            eprintln!("colony not found or failed: {}. is it in PATH?", e);
+            return false;
+        }
+    };
+
+    if !inbox_output.status.success() {
+        eprintln!("colony inbox failed (exit {}), skipping pulse", inbox_output.status);
+        return false;
+    }
 
     let inbox_str = String::from_utf8_lossy(&inbox_output.stdout);
     let inbox: Vec<serde_json::Value> = serde_json::from_str(&inbox_str).unwrap_or_default();
@@ -50,13 +58,17 @@ pub fn run_pulse(config_path: &str) -> bool {
     prompt.push_str("Use `colony read <channel>` to get more context if needed. ");
     prompt.push_str("Log what you did to memory/memory.md.\n");
 
-    // 5. Write prompt to temp file
-    let prompt_path = "/tmp/colony-agent-prompt.md";
-    std::fs::write(prompt_path, &prompt).expect("failed to write prompt file");
+    // 5. Write prompt to temp file (per-agent path to avoid conflicts)
+    let agent_name = std::env::var("COLONY_AGENT").unwrap_or_else(|_| "agent".into());
+    let prompt_path = format!("/tmp/colony-{}-prompt.md", agent_name);
+    if let Err(e) = std::fs::write(&prompt_path, &prompt) {
+        eprintln!("failed to write prompt file: {}", e);
+        return false;
+    }
 
     // 6. Invoke Claude Code
     eprintln!("invoking claude with {} inbox items + heartbeat", inbox.len());
-    let claude_status = Command::new("claude")
+    let claude_ok = match Command::new("claude")
         .args([
             "--dangerously-skip-permissions",
             "-p",
@@ -64,22 +76,24 @@ pub fn run_pulse(config_path: &str) -> bool {
             "--max-turns",
             "20",
         ])
-        .status();
-
-    match claude_status {
+        .status()
+    {
         Ok(s) if s.success() => {
             eprintln!("claude completed successfully");
+            true
         }
         Ok(s) => {
             eprintln!("claude exited with status: {}", s);
+            false
         }
         Err(e) => {
             eprintln!("failed to run claude: {}", e);
+            false
         }
-    }
+    };
 
-    // 7. Ack inbox items
-    if !inbox.is_empty() {
+    // 7. Only ack inbox items if Claude succeeded
+    if claude_ok && !inbox.is_empty() {
         let ids: Vec<String> = inbox.iter()
             .filter_map(|i| i["id"].as_i64().map(|id| id.to_string()))
             .collect();
@@ -89,6 +103,8 @@ pub fn run_pulse(config_path: &str) -> bool {
                 .args(&ids)
                 .status();
         }
+    } else if !claude_ok && !inbox.is_empty() {
+        eprintln!("claude failed — NOT acking {} inbox items (will retry next pulse)", inbox.len());
     }
 
     true