feat(ai): name agents after their model, not hardcoded "oracle"

Agents now join under their model tag (e.g. "qwen2.5:3b" — model name +
parameter size) or profile label, derived at /ai start and tracked on
App.agent_name, so the roster shows what's actually answering. Updates the
broker re-grant + /ai stop revoke paths, the Python --name default
(falls back to provider.model), and the demo/bootstrap/README/docs refs.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

README.MD

@@ -275,7 +275,8 @@ directly:
 
 ```bash
 .venv/bin/python -m cmd_chat.agent <server_ip> <port> \
-    --name oracle --password <room-pw> --provider ollama --model qwen2.5:3b
+    --password <room-pw> --provider ollama --model qwen2.5:3b
+# joins as its model tag ("qwen2.5:3b") unless you override with --name
 # cloud (opt-in): --provider anthropic --model claude-opus-4-6   (needs ANTHROPIC_API_KEY)
 ```
 

bootstrap-ai.sh

@@ -129,6 +129,6 @@ fi
 echo
 echo "AI ready. start a local agent against a running room with:"
 echo "    .venv/bin/python -m cmd_chat.agent <host> <port> \\"
-echo "        --name oracle --password <room-pw> --provider ollama --model $MODEL --no-tls"
+echo "        --password <room-pw> --provider ollama --model $MODEL --no-tls"
 echo
 echo "tip: pick a different model with  HH_AI_MODEL=llama3 ./bootstrap-ai.sh"

cmd_chat/agent/__main__.py

@@ -2,8 +2,8 @@
 
 Examples
 --------
-    # local Ollama (default, recommended)
+    # local Ollama (default, recommended) — joins as "qwen2.5:3b"
-    python -m cmd_chat.agent 127.0.0.1 3000 --name oracle \
+    python -m cmd_chat.agent 127.0.0.1 3000 \
         --password hunter2 --model qwen2.5:3b --no-tls
 
     # cloud, opt-in
@@ -102,7 +102,8 @@ def main() -> None:
     )
     ap.add_argument("server", nargs="?", help="room host (omit with --list-models/--check)")
     ap.add_argument("port", type=int, nargs="?", help="room port")
-    ap.add_argument("--name", default="oracle", help="agent's room display name")
+    ap.add_argument("--name", default=None,
+                    help="agent's room display name (default: the model tag, e.g. qwen2.5:3b)")
     ap.add_argument("--password", default=None, help="room password")
     ap.add_argument("--provider", default="ollama",
                     help="ollama | anthropic | openai | module:Class")
@@ -184,8 +185,11 @@ def main() -> None:
     if code_provider is not None:
         print(f"sandbox/code path → {code_provider.name}/{code_provider.model}", file=sys.stderr)
 
+    # Default the room handle to the model tag (model name + parameter size,
+    # e.g. "qwen2.5:3b") so the roster shows what's actually answering.
+    name = args.name or provider.model
     bridge = AgentBridge(
-        args.server, args.port, name=args.name, provider=provider,
+        args.server, args.port, name=name, provider=provider,
         password=args.password, insecure=args.insecure, no_tls=args.no_tls,
         system_prompt=args.system, context_window=args.context_window,
         token_budget=args.token_budget, embedder=embedder, rag_top_k=args.rag_top_k,

docs/demo-save-load-poc.md

@@ -34,10 +34,10 @@ the chat provider is Ollama and a `qwen2.5-coder` is present (it is — pulled).
 
 1. **Title card** — "Ephemeral by default. Persistent on demand."
 2. **Summon** — alice: `/sbx launch docker` → "summoned" sandbox bubble.
-3. **Spawn the coder** — alice: `/ai start` → `oracle online — ollama/qwen2.5:3b`
+3. **Spawn the coder** — alice: `/ai start` → `qwen2.5:3b (ai) online — ollama/qwen2.5:3b`
    (the coder model rides along for `!task`).
 4. **Build, by the fast model** — alice:
-   `/ai oracle !write /root/fib.py that prints the first 10 Fibonacci numbers, then run it`
+   `/ai qwen2.5:3b !write /root/fib.py that prints the first 10 Fibonacci numbers, then run it`
    → agent drives the shared shell; `fib.py` is written and executed; the
    sandbox pane shows the Fibonacci output.
 5. **Freeze it** — alice: `/sbx save buildbox` →
@@ -74,7 +74,7 @@ render.
 
 - TUI doesn't bind Ctrl-U (it inserts a literal `u`); clear input with `BSpace`.
   Send text with `send-keys -l "<text>"` then a separate `Enter`; don't race renders.
-- Agent name is hardcoded `oracle`; only one `/ai start` per room.
+- Agent joins under its model tag (e.g. `qwen2.5:3b`); only one `/ai start` per client.
 - Keep `!task` phrasing single-line; the agent's drive output lands in the sandbox
   pane, not chat.
 - `/sbx load` refuses if a sandbox is already running — stop first.

hh/demo-save-load.sh

@@ -2,8 +2,8 @@
 # demo-save-load.sh — PoC harness for the "persistent sandbox" video beat.
 #
 # Flow (see docs/demo-save-load-poc.md):
-#   session A: /sbx launch docker → /ai start → /grant oracle →
+#   session A: /sbx launch docker → /ai start → /grant qwen2.5:3b →
-#              /ai oracle !build fib.py & run it  →  /sbx save buildbox  → quit
+#              /ai qwen2.5:3b !build fib.py & run it  →  /sbx save buildbox  → quit
 #   prove:     container purged on quit, but hh-snap:buildbox image survives
 #   session B: fresh client → /sbx load buildbox → the model's code is intact
 #
@@ -131,17 +131,17 @@ wait_for "$A_SESS" 'summoned|sandbox|ready|online' 60 >/dev/null
 snap_evid "$A_SESS" 02-sandbox
 
 # ---- 4. spawn the coder agent + grant drive --------------------------------
-step "spawn oracle (qwen2.5:3b chat, qwen2.5-coder:1.5b for !task)"
+step "spawn agent (qwen2.5:3b chat, qwen2.5-coder:1.5b for !task)"
 say "$A_SESS" "/ai start"
-wait_for "$A_SESS" 'oracle|online|ollama' 45 && ok "oracle announced" \
+wait_for "$A_SESS" 'online|ollama|qwen' 45 && ok "agent announced" \
-  || note "no 'online' line yet - agent log: ${TMPDIR:-/tmp}/hh-agent-oracle.log"
+  || note "no 'online' line yet - agent log: ${TMPDIR:-/tmp}/hh-agent-qwen2.5:3b.log"
-say "$A_SESS" "/grant oracle"
+say "$A_SESS" "/grant qwen2.5:3b"
 sleep 1
 snap_evid "$A_SESS" 03-agent
 
 # ---- 5. fast model builds code in the sandbox ------------------------------
 step "fast qwen builds /root/fib.py in the sandbox"
-say "$A_SESS" "/ai oracle !create /root/fib.py that prints the first 10 fibonacci numbers space-separated on one line, then run it with python3"
+say "$A_SESS" "/ai qwen2.5:3b !create /root/fib.py that prints the first 10 fibonacci numbers space-separated on one line, then run it with python3"
 # Give the CPU coder model room to think, then poll for the file.
 WT=150 wait_cmd docker exec "$CTR" test -s /root/fib.py
 NEED='0 1 1 2 3 5 8 13 21 34'

hh/film-save-load.sh

@@ -147,9 +147,9 @@ wait_for 'summoned|sandbox|ready|online' 60 >/dev/null
 sleep 1.5
 
 # ---- 5. spawn fast qwen agent (auto-grant drive) ---------------------------
-step "spawn oracle (auto-grant sandbox drive)"
+step "spawn agent (auto-grant sandbox drive)"
 say "/ai start $CODER allow"
-wait_for 'oracle|online|ollama|qwen' 45 && ok "oracle online" || note "no online line yet"
+wait_for 'online|ollama|qwen' 45 && ok "agent online" || note "no online line yet"
 sleep 1.5
 
 # ---- 6. the fast model builds code in the sandbox --------------------------
@@ -157,7 +157,7 @@ sleep 1.5
 # through the PTY. Validate-by-running; retry once; abort before save if it
 # still fails (no silent fallback in a film).
 step "fast qwen writes /root/fib.py and runs it"
-TASK="/ai oracle !create /root/fib.py with exactly two lines and nothing else: line 1 is  nums = [0, 1, 1, 2, 3, 5, 8, 13, 21, 34]  and line 2 is  print(*nums)  then run it with: python3 /root/fib.py"
+TASK="/ai $CODER !create /root/fib.py with exactly two lines and nothing else: line 1 is  nums = [0, 1, 1, 2, 3, 5, 8, 13, 21, 34]  and line 2 is  print(*nums)  then run it with: python3 /root/fib.py"
 BUILT=0
 for attempt in 1 2; do
   note "build attempt $attempt"

hh/src/app.rs

@@ -161,9 +161,6 @@ pub struct PendingVm {
     pub needs_install: bool,
 }
 
-/// Display handle the summoned Python agent joins under (see `spawn_agent`).
-const AGENT_NAME: &str = "oracle";
-
 pub struct App {
     pub me: String,
     pub lines: Vec<ChatLine>,
@@ -212,6 +209,11 @@ pub struct App {
     /// (`/ai start <name> allow`). Re-applied in the broker Ready handler, since
     /// launching a sandbox resets the ACL back to just the owner.
     pub agent_sbx_allow: bool,
+    /// Display name the summoned agent joined under — derived from its model or
+    /// profile at `/ai start` (e.g. "qwen2.5:3b", model name + parameter size).
+    /// Tracked so the broker Ready handler can re-grant its drive and `/ai stop`
+    /// can revoke the right ACL entry.
+    pub agent_name: Option<String>,
 }
 
 impl App {
@@ -244,6 +246,7 @@ impl App {
             ai_stream: std::collections::HashMap::new(),
             spin: 0,
             agent_sbx_allow: false,
+            agent_name: None,
         }
     }
 
@@ -1085,7 +1088,9 @@ pub async fn run(params: net::ConnParams, mut session: Session, mut theme: Theme
                         app.sudoers.insert(app.me.clone()); // owner = superuser
                         if app.agent_sbx_allow {
                             // Re-apply a `/ai start … allow` grant the launch reset.
-                            app.drivers.insert(AGENT_NAME.to_string());
+                            if let Some(n) = &app.agent_name {
+                                app.drivers.insert(n.clone());
+                            }
                         }
                         send_frame(&out_tx, &session.room, json!({
                             "_sbx":"status","state":"ready","backend": backend.label(), "rows": rows, "cols": cols
@@ -1722,7 +1727,10 @@ fn handle_command(
             app.sys("⛧ dismissed the AI agent");
             // Drop any sandbox drive the agent held so a dead handle can't act.
             app.agent_sbx_allow = false;
-            let revoked = app.drivers.remove(AGENT_NAME) | app.sudoers.remove(AGENT_NAME);
+            let revoked = app
+                .agent_name
+                .take()
+                .is_some_and(|n| app.drivers.remove(&n) | app.sudoers.remove(&n));
             if revoked && app.sandbox.is_some() {
                 broadcast_acl(out_tx, room, app);
             }
@@ -1749,12 +1757,6 @@ fn handle_command(
                 Some(head) if head.is_empty() || head.ends_with(' ') => (head.trim(), true),
                 _ => (raw, false),
             };
-            // Tolerate a leading agent-name token (`/ai start oracle allow`):
-            // there's only one agent, so treat it as addressing, not a profile.
-            let raw = match raw.strip_prefix(AGENT_NAME) {
-                Some(rest) if rest.is_empty() || rest.starts_with(' ') => rest.trim(),
-                _ => raw,
-            };
             // A bare name (no ':' tag, no '/' path) is a models.toml profile;
             // anything else is treated as a literal Ollama model tag.
             let (profile, model): (Option<&str>, &str) = if raw.is_empty() {
@@ -1764,10 +1766,14 @@ fn handle_command(
             } else {
                 (Some(raw), raw)
             };
-            let name = AGENT_NAME;
+            // Name the agent after what it runs, for clarity in the roster: a
+            // profile keeps its label; a direct Ollama model uses its tag
+            // (e.g. "qwen2.5:3b" — model name + parameter size).
+            let name = profile.unwrap_or(model);
             match spawn_agent(params, &app.password, name, profile, model) {
                 Ok(child) => {
                     *agent = Some(child);
+                    app.agent_name = Some(name.to_string());
                     app.agent_sbx_allow = grant_sbx;
                     let desc = match profile {
                         Some(p) => format!("profile {p}"),

hh/src/ui.rs

@@ -567,7 +567,8 @@ fn draw_roster(f: &mut Frame, area: ratatui::layout::Rect, app: &App, theme: &Th
     f.render_widget(roster, area);
 }
 
-/// Animated "⠋ oracle is thinking…" title shown while AI agents generate a reply.
+/// Animated "⠋ <agent> is thinking…" title shown while AI agents generate a
+/// reply. The name(s) come from `app.ai_typing`, i.e. each agent's own handle.
 fn ai_thinking_title(app: &App) -> String {
     const FRAMES: [&str; 10] = ["⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏"];
     let glyph = FRAMES[(app.spin / 2) % FRAMES.len()];