fix: prevent premature subagent termination, add attack planning framework, and fix TUI lag

strix/agents/prompts/system_prompt.jinja

@@ -217,6 +217,15 @@ VALIDATION APPROACH:
 - Chain vulnerabilities when needed to demonstrate maximum impact
 - Focus on demonstrating real business impact
 
+ATTACK PLANNING METHODOLOGY:
+Before testing a vulnerability type, reason deeply about the attack — do not spray payloads blindly.
+
+1. **Load the skill first**: Call `load_skill` with the relevant vulnerability name (e.g., `["sql_injection"]`, `["xss"]`, `["ssrf"]`, `["business_logic"]`) to get the full methodology — attack surfaces, detection channels, bypass techniques, technology-specific primitives, and testing workflow. Do not plan attacks from memory alone.
+2. **Use `think` to plan**: Before acting, reason through: target technology and how it handles input/sessions/auth → which endpoints and parameters are user-controlled → which techniques are most likely to succeed given the technology → what defenses (WAF, validation, encoding) are in place and how to bypass them → what encoding tricks or syntax variations could work → whether findings can be chained for greater impact
+3. **Match techniques to technology**: Do not test irrelevant attack vectors. NoSQL injection is pointless against PostgreSQL. XXE is irrelevant if no XML parsing occurs. SSRF requires server-side URL fetching. Select techniques based on what the target actually does.
+4. **Plan bypass strategies**: When standard payloads are filtered, reason about alternatives — encoding variations (double URL, unicode, NFKC), syntax tricks (comment splitting, case folding), alternative injection points (headers, cookies, path params), and protocol-level bypasses.
+5. **Chain for impact**: Think about how low-severity findings combine into high-impact attacks. IDOR + business logic = financial fraud. SSRF + cloud metadata = credential theft. XSS + CSRF = account takeover. Stored XSS + admin viewing = privilege escalation.
+
 VULNERABILITY KNOWLEDGE BASE:
 You have access to comprehensive guides for each vulnerability type above. Use these references for:
 - Discovery techniques and automation
@@ -259,8 +268,10 @@ WHITE-BOX TESTING - PHASE 1 (CODE UNDERSTANDING):
 - ONLY AFTER full code comprehension → proceed to vulnerability testing
 
 PHASE 2 - SYSTEMATIC VULNERABILITY TESTING:
-- CREATE SPECIALIZED SUBAGENT for EACH vulnerability type × EACH component
-- Each agent focuses on ONE vulnerability type in ONE specific location
+- ANALYZE the target's technology stack and attack surface to determine which vulnerability types are relevant
+- LOAD the relevant vulnerability skill via `load_skill` and use `think` to plan the attack strategy before spawning agents
+- CREATE SPECIALIZED SUBAGENTS for vulnerability types that match the target's technology and attack surface — not blindly for every category
+- Each agent focuses on ONE vulnerability type in ONE specific location with a task description informed by your analysis
 - EVERY detected vulnerability MUST spawn its own validation subagent
 
 SIMPLE WORKFLOW RULES:

strix/interface/tui/app.py

@@ -730,6 +730,9 @@ def __init__(self, args: argparse.Namespace):
         ]
         self._dot_animation_timer: Any | None = None
 
+        self._event_buffer: list[tuple[str, Any]] = []
+        self._event_flush_timer: Any | None = None
+
         self._setup_cleanup_handlers()
 
     def _build_scan_config(self, args: argparse.Namespace) -> dict[str, Any]:
@@ -1292,7 +1295,7 @@ def _get_animated_verb_text(self, agent_id: str, verb: str) -> Text:  # noqa: AR
 
     def _start_dot_animation(self) -> None:
         if self._dot_animation_timer is None:
-            self._dot_animation_timer = self.set_interval(0.06, self._animate_dots)
+            self._dot_animation_timer = self.set_interval(0.15, self._animate_dots)
 
     def _stop_dot_animation(self) -> None:
         if self._dot_animation_timer is not None:
@@ -1335,9 +1338,13 @@ def _agent_vulnerability_count(self, agent_id: str) -> int:
             if vuln.get("agent_id") == agent_id
         )
 
+    MAX_RENDER_EVENTS = 200
+
     def _gather_agent_events(self, agent_id: str) -> list[dict[str, Any]]:
         events = self.live_view.events_for_agent(agent_id)
         events.sort(key=lambda e: (e["timestamp"], e["id"]))
+        if len(events) > self.MAX_RENDER_EVENTS:
+            events = events[-self.MAX_RENDER_EVENTS:]
         return events
 
     def watch_selected_agent_id(self, _agent_id: str | None) -> None:
@@ -1407,10 +1414,25 @@ def scan_target() -> None:
         self._scan_thread.start()
 
     def _capture_sdk_event(self, agent_id: str, event: Any) -> None:
+        self._event_buffer.append((agent_id, event))
+        if self._event_flush_timer is None:
+            self._event_flush_timer = self.set_timer(0.15, self._flush_event_buffer)
+
+    def _flush_event_buffer(self) -> None:
+        self._event_flush_timer = None
+        if not self._event_buffer:
+            return
+        batch = self._event_buffer[:]
+        self._event_buffer.clear()
         try:
-            self.call_from_thread(self._record_sdk_event, agent_id, event)
+            self.call_from_thread(self._record_sdk_events_batch, batch)
         except RuntimeError:
-            self._record_sdk_event(agent_id, event)
+            for agent_id, event in batch:
+                self._record_sdk_event(agent_id, event)
+
+    def _record_sdk_events_batch(self, batch: list[tuple[str, Any]]) -> None:
+        for agent_id, event in batch:
+            self.live_view.ingest_sdk_event(agent_id, event)
 
     def _record_sdk_event(self, agent_id: str, event: Any) -> None:
         self.live_view.ingest_sdk_event(agent_id, event)

strix/interface/tui/live_view.py

@@ -14,10 +14,14 @@
 from strix.interface.tui.history import load_session_history
 
 
+MAX_EVENTS_PER_AGENT = 500
+
+
 class TuiLiveView:
     def __init__(self) -> None:
         self.agents: dict[str, dict[str, Any]] = {}
         self.events: list[dict[str, Any]] = []
+        self._events_by_agent: dict[str, list[dict[str, Any]]] = {}
         self._next_event_id = 1
         self._open_assistant_event_by_agent: dict[str, dict[str, Any]] = {}
         self._tool_event_by_call_id: dict[str, dict[str, Any]] = {}
@@ -115,10 +119,10 @@ def ingest_sdk_event(self, agent_id: str, event: Any) -> None:
             self._record_tool_output(agent_id, item)
 
     def events_for_agent(self, agent_id: str) -> list[dict[str, Any]]:
-        return [event for event in self.events if event.get("agent_id") == agent_id]
+        return list(self._events_by_agent.get(agent_id, []))
 
     def has_events_for_agent(self, agent_id: str) -> bool:
-        return any(event.get("agent_id") == agent_id for event in self.events)
+        return bool(self._events_by_agent.get(agent_id))
 
     def _ingest_raw_response_event(self, agent_id: str, data: Any) -> None:
         data_type = getattr(data, "type", "")
@@ -277,6 +281,16 @@ def _append_event(
         }
         self._next_event_id += 1
         self.events.append(event)
+        agent_events = self._events_by_agent.setdefault(agent_id, [])
+        agent_events.append(event)
+        if len(agent_events) > MAX_EVENTS_PER_AGENT:
+            dropped = agent_events[:len(agent_events) - MAX_EVENTS_PER_AGENT]
+            for d in dropped:
+                try:
+                    self.events.remove(d)
+                except ValueError:
+                    pass
+            del agent_events[:len(agent_events) - MAX_EVENTS_PER_AGENT]
         return event
 
     @staticmethod

strix/skills/coordination/root_agent.md

@@ -25,6 +25,30 @@ Before spawning agents, analyze the target:
 3. **Determine approach** - blackbox, greybox, or whitebox assessment
 4. **Prioritize by risk** - critical assets and high-value targets first
 
+## Attack Planning
+
+Before spawning subagents for a vulnerability type, **think deeply about the attack strategy**. Do not blindly spawn agents for every vulnerability category — reason about what is actually relevant to the target.
+
+### Step 1: Load Vulnerability Methodology
+
+Before planning an attack against a specific vulnerability type, call `load_skill` with the relevant skill name (e.g., `["sql_injection"]`, `["xss"]`, `["business_logic"]`, `["ssrf"]`, `["xxe"]`). This gives you the full methodology — attack surfaces, detection channels, bypass techniques, DBMS-specific primitives, testing workflow, and validation steps. **Do not plan attacks from memory alone — load the skill first.**
+
+### Step 2: Reason About the Attack
+
+Use `think` to reason through the attack plan before spawning agents. Structure your thinking:
+
+- **Target technology**: What framework, language, database, server is in use? How does it handle input, auth, sessions?
+- **Attack surface mapping**: Which endpoints, parameters, headers, cookies are user-controlled? Where does data flow to (database, filesystem, OS, other services)?
+- **Technique selection**: Given the technology, which attack techniques are most likely to succeed? Which are irrelevant? (e.g., NoSQL injection is irrelevant for a PostgreSQL backend; XXE is irrelevant if no XML parsing occurs)
+- **Defense analysis**: What WAF, input validation, encoding, or framework protections are in place? How might they be bypassed?
+- **Bypass strategy**: If standard payloads are filtered, what encoding tricks, syntax variations, or alternative injection points could work?
+- **Chaining opportunities**: Can findings be combined for greater impact? (e.g., IDOR + business logic = financial fraud; SSRF + metadata = cloud compromise; XSS + CSRF = account takeover)
+- **Priority ordering**: Which attacks should be attempted first based on likelihood of success and potential impact?
+
+### Step 3: Spawn Targeted Agents
+
+Only after loading the skill and reasoning through the attack, spawn subagents with specific, informed task descriptions. The task description should reflect your analysis — tell the subagent what technology to target, what techniques to prioritize, what bypasses to try, and what the success criteria are. A well-planned task description produces far better results than a generic "test for SQLi."
+
 ## Agent Architecture
 
 Structure agents by function:
@@ -84,9 +108,17 @@ Complex findings warrant specialized subagents:
 
 ## Completion
 
-When all agents report completion:
+**WAIT for all agents to self-terminate.** Do NOT call `stop_agent` on active children to clear them before finishing — this orphans their work and produces an incomplete report.
+
+When you believe all work is done:
+
+1. Call `view_agent_graph` to check every agent's status
+2. If ANY agent is still `running` or `waiting`, call `wait_for_message` to block until their completion reports arrive — do NOT stop them
+3. Only proceed when ALL agents show `completed` or `crashed` (these are the only safe terminal states)
+4. `stopped` agents were forcibly cancelled — their results are lost. If agents were stopped, their work must be re-done by spawning replacement agents
+5. Collect and deduplicate findings across all completion reports
+6. Assess overall security posture
+7. Compile executive summary with prioritized recommendations
+8. Invoke `finish_scan` with the final report
 
-1. Collect and deduplicate findings across agents
-2. Assess overall security posture
-3. Compile executive summary with prioritized recommendations
-4. Invoke finish tool with final report
+**Never use `stop_agent` as a shortcut to bypass the active-agent check in `finish_scan`.** The check exists to prevent incomplete reports. If a child is taking too long, send it a message asking for status or telling it to wrap up via `agent_finish` — don't stop it.

strix/tools/agents_graph/tools.py

@@ -595,6 +595,15 @@ async def stop_agent(
     to wrap up) for soft-stop scenarios. Reach for ``stop_agent`` when
     a child has gone off-track and won't self-correct.
 
+    **NEVER use ``stop_agent`` to clear active children so you can call
+    ``finish_scan``.** Stopped agents lose their in-progress work and
+    their completion reports never arrive. Instead, call
+    ``wait_for_message`` to block until children self-terminate via
+    ``agent_finish``, or ``send_message_to_agent`` asking them to wrap
+    up. The only valid use of ``stop_agent`` is for agents that are
+    genuinely off-track, stuck, or producing errors — not for agents
+    that are productively working but haven't finished yet.
+
     Args:
         target_agent_id: The 8-char id from ``view_agent_graph`` /
             ``create_agent``. Cannot stop yourself.

strix/tools/finish/tool.py

@@ -101,12 +101,13 @@ async def finish_scan(
     1. **Call ``view_agent_graph`` first.** Inspect every entry in the
        summary. If ANY agent is in ``running`` / ``waiting`` state,
        you MUST NOT call ``finish_scan`` yet —
-       wrap them up first via ``send_message_to_agent`` (ask them to
-       finish), ``wait_for_message`` (block until their report
-       arrives), or ``stop_agent`` (graceful cancel). Only ``completed``
-       / ``crashed`` / ``stopped`` agents are safe to leave behind.
-       Calling ``finish_scan`` while children are alive orphans their
-       work and produces an incomplete report.
+       call ``wait_for_message`` to block until their completion reports
+       arrive, or ``send_message_to_agent`` asking them to call
+       ``agent_finish``. **NEVER use ``stop_agent`` to clear active
+       agents so you can finish — this orphans their work and produces
+       an incomplete report.** Only ``completed`` / ``crashed`` agents
+       are safe to leave behind. ``stopped`` agents were forcibly
+       cancelled and their results are lost.
     2. All vulnerabilities you found are filed via
        ``create_vulnerability_report`` (un-reported findings are not
        tracked and not credited).
@@ -162,7 +163,10 @@ async def finish_scan(
                 "scan_completed": False,
                 "error": (
                     "Cannot finish scan while child agents are still active. "
-                    "Wait for completion, send them finish instructions, or stop them first"
+                    "Call wait_for_message to block until their completion reports "
+                    "arrive, or send_message_to_agent asking them to call agent_finish. "
+                    "Do NOT use stop_agent to clear active agents — their work will "
+                    "be lost and the report will be incomplete."
                 ),
                 "active_agents": active_agents,
             },