agents: tool-enabled LLM calls + DELETE action support

Add call_for_def() that threads model and tools from agent definitions
through to claude CLI. Tool-enabled agents get --allowedTools instead
of --tools "" and a longer 15-minute timeout for multi-turn work.

Add ActionKind::Delete with parse/apply support so agents can delete
nodes (used by organize agent for deduplication).

Use call_for_def() in run_one_agent instead of hardcoded call_sonnet.

poc-memory/src/agents/consolidate.rs

@@ -233,6 +233,8 @@ pub fn apply_consolidation(store: &mut Store, do_apply: bool, report_key: Option
                     println!("  REFINE {}", key),
                 knowledge::ActionKind::Demote { key } =>
                     println!("  DEMOTE {}", key),
+                knowledge::ActionKind::Delete { key } =>
+                    println!("  DELETE {}", key),
             }
         }
         println!("\nTo apply: poc-memory apply-consolidation --apply");

poc-memory/src/agents/knowledge.rs

@@ -51,6 +51,9 @@ pub enum ActionKind {
     Demote {
         key: String,
     },
+    Delete {
+        key: String,
+    },
 }
 
 #[derive(Debug, Clone, Copy, Serialize, Deserialize)]
@@ -180,11 +183,28 @@ pub fn parse_demotes(text: &str) -> Vec<Action> {
         .collect()
 }
 
+pub fn parse_deletes(text: &str) -> Vec<Action> {
+    let re = Regex::new(r"(?m)^DELETE\s+(\S+)").unwrap();
+    re.captures_iter(text)
+        .map(|cap| Action {
+            kind: ActionKind::Delete {
+                key: cap[1].to_string(),
+            },
+            confidence: Confidence::High,
+            weight: 1.0,
+            depth: 0,
+            applied: None,
+            rejected_reason: None,
+        })
+        .collect()
+}
+
 pub fn parse_all_actions(text: &str) -> Vec<Action> {
     let mut actions = parse_write_nodes(text);
     actions.extend(parse_links(text));
     actions.extend(parse_refines(text));
     actions.extend(parse_demotes(text));
+    actions.extend(parse_deletes(text));
     actions
 }
 
@@ -243,7 +263,7 @@ fn agent_base_depth(agent: &str) -> Option<i32> {
 
 pub fn compute_action_depth(db: &DepthDb, action: &Action, agent: &str) -> i32 {
     match &action.kind {
-        ActionKind::Link { .. } | ActionKind::Demote { .. } => -1,
+        ActionKind::Link { .. } | ActionKind::Demote { .. } | ActionKind::Delete { .. } => -1,
         ActionKind::Refine { key, .. } => db.get(key),
         ActionKind::WriteNode { covers, .. } => {
             if !covers.is_empty() {
@@ -336,6 +356,9 @@ pub fn apply_action(
                 false
             }
         }
+        ActionKind::Delete { key } => {
+            store.delete_node(key).is_ok()
+        }
     }
 }
 
@@ -582,7 +605,7 @@ pub fn run_one_agent(
         .ok_or_else(|| format!("no .agent file for {}", agent_name))?;
     let agent_batch = super::defs::run_agent(store, &def, batch_size)?;
 
-    let output = llm::call_sonnet(llm_tag, &agent_batch.prompt)?;
+    let output = llm::call_for_def(&def, &agent_batch.prompt)?;
 
     // Store raw output for audit trail
     let ts = store::compact_timestamp();
@@ -920,6 +943,9 @@ fn run_cycle(
                 ActionKind::Demote { key } => {
                     eprintln!("      DEMOTE {}", key);
                 }
+                ActionKind::Delete { key } => {
+                    eprintln!("      DELETE {}", key);
+                }
             }
 
             if apply_action(&mut store, action, agent_name, &timestamp, depth) {

poc-memory/src/agents/llm.rs

@@ -40,6 +40,8 @@ fn log_usage(agent: &str, model: &str, prompt: &str, response: &str,
 
 /// Maximum time to wait for a claude subprocess before killing it.
 const SUBPROCESS_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(300); // 5 minutes
+/// Longer timeout for agents with tool access (multi-turn conversations).
+const TOOL_AGENT_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(900); // 15 minutes
 
 /// Call a model via claude CLI. Returns the response text.
 ///
@@ -47,6 +49,19 @@ const SUBPROCESS_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(3
 /// parent daemon exits — no more orphaned claude processes.
 /// Times out after 5 minutes to prevent blocking the daemon forever.
 fn call_model(agent: &str, model: &str, prompt: &str) -> Result<String, String> {
+    call_model_with_tools(agent, model, prompt, &[])
+}
+
+/// Call a model via claude CLI, optionally with allowed tools.
+///
+/// When `tools` is empty, all tools are disabled (`--tools ""`).
+/// When `tools` has entries, they're passed as `--allowedTools` patterns
+/// (e.g. `["Bash(poc-memory:*)"]`), letting the agent call those tools
+/// in Claude's native tool loop.
+fn call_model_with_tools(agent: &str, model: &str, prompt: &str,
+                         tools: &[String]) -> Result<String, String> {
+    let timeout = if tools.is_empty() { SUBPROCESS_TIMEOUT } else { TOOL_AGENT_TIMEOUT };
+
     // Write prompt to temp file (claude CLI needs file input for large prompts)
     let tmp = std::env::temp_dir().join(format!("poc-llm-{}-{:?}.txt",
         std::process::id(), std::thread::current().id()));
@@ -54,8 +69,17 @@ fn call_model(agent: &str, model: &str, prompt: &str) -> Result<String, String>
         .map_err(|e| format!("write temp prompt: {}", e))?;
 
     let mut cmd = Command::new("claude");
-    cmd.args(["-p", "--model", model, "--tools", "", "--no-session-persistence",
-              "--strict-mcp-config"])
+    if tools.is_empty() {
+        cmd.args(["-p", "--model", model, "--tools", "", "--no-session-persistence",
+                  "--strict-mcp-config"]);
+    } else {
+        cmd.args(["-p", "--model", model, "--no-session-persistence",
+                  "--strict-mcp-config", "--allowedTools"]);
+        for tool in tools {
+            cmd.arg(tool);
+        }
+    }
+    cmd
         .stdin(fs::File::open(&tmp).map_err(|e| format!("open temp: {}", e))?)
         .stdout(std::process::Stdio::piped())
         .stderr(std::process::Stdio::piped())
@@ -87,7 +111,7 @@ fn call_model(agent: &str, model: &str, prompt: &str) -> Result<String, String>
     let cancel_flag = cancel.clone();
     let watchdog = std::thread::spawn(move || {
         // Sleep in 1s increments so we can check the cancel flag
-        let deadline = std::time::Instant::now() + SUBPROCESS_TIMEOUT;
+        let deadline = std::time::Instant::now() + timeout;
         while std::time::Instant::now() < deadline {
             if cancel_flag.load(std::sync::atomic::Ordering::Relaxed) {
                 return;
@@ -119,7 +143,7 @@ fn call_model(agent: &str, model: &str, prompt: &str) -> Result<String, String>
     match result {
         Ok(output) => {
             let elapsed = start.elapsed().as_millis();
-            if elapsed > SUBPROCESS_TIMEOUT.as_millis() - 1000 {
+            if elapsed > timeout.as_millis() - 1000 {
                 log_usage(agent, model, prompt, "TIMEOUT", elapsed, false);
                 return Err(format!("claude timed out after {:.0}s", elapsed as f64 / 1000.0));
             }
@@ -148,6 +172,11 @@ pub(crate) fn call_haiku(agent: &str, prompt: &str) -> Result<String, String> {
     call_model(agent, "haiku", prompt)
 }
 
+/// Call a model using an agent definition's model and tool configuration.
+pub(crate) fn call_for_def(def: &super::defs::AgentDef, prompt: &str) -> Result<String, String> {
+    call_model_with_tools(&def.agent, &def.model, prompt, &def.tools)
+}
+
 /// Parse a JSON response, handling markdown fences.
 pub(crate) fn parse_json_response(response: &str) -> Result<serde_json::Value, String> {
     let cleaned = response.trim();