fix(container): pin the User-Agent to the installed Chromium and assert it at build

pkill -f 'Xvfb :99' also matches the RUN shell's own command line, so the
cleanup could terminate the build shell.

containers/Dockerfile

@@ -58,7 +58,9 @@ RUN apt-get update && \
     libcap2-bin \
     gdb \
     libnss3-tools \
-    chromium fonts-liberation
+    chromium fonts-liberation fonts-noto-core fonts-noto-color-emoji \
+    xvfb x11-utils xdotool dbus-x11 \
+    imagemagick
 
 
 RUN setcap cap_net_raw,cap_net_admin,cap_net_bind_service+eip $(which nmap)
@@ -114,11 +116,38 @@ RUN npm install -g retire@latest && \
     ln -sf ast-grep /home/pentester/.npm-global/lib/node_modules/@ast-grep/cli/sg
 
 ENV AGENT_BROWSER_EXECUTABLE_PATH=/usr/bin/chromium
-ENV AGENT_BROWSER_USER_AGENT="Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/131.0.0.0 Safari/537.36"
-ENV AGENT_BROWSER_ARGS="--disable-blink-features=AutomationControlled,--no-first-run,--no-default-browser-check,--lang=en-US"
+# Must match the installed Chromium major; the build asserts it below.
+ENV AGENT_BROWSER_USER_AGENT="Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/150.0.0.0 Safari/537.36"
+# AGENT_BROWSER_ARGS is comma-separated, so no flag value may contain a comma
+# (window geometry comes from the virtual display instead).
+ENV AGENT_BROWSER_ARGS="--disable-blink-features=AutomationControlled,--no-first-run,--no-default-browser-check,--lang=en-US,--password-store=basic,--use-mock-keychain,--disable-dev-shm-usage"
 ENV AGENT_BROWSER_SCREENSHOT_DIR=/workspace/.agent-browser-screenshots
+ENV AGENT_BROWSER_IDLE_TIMEOUT_MS=180000
 RUN /home/pentester/.npm-global/bin/agent-browser doctor --offline --quick
 
+# `agent-browser --headed` exits 0 even when Chrome fails to launch, so assert
+# headed mode here. `--no-sandbox` is build-only: buildkit has no unprivileged
+# user namespaces. The stale X lock/socket must go or the runtime Xvfb refuses.
+RUN set -eu; \
+    chrome_major="$(chromium --version | grep -oE '[0-9]+' | head -1)"; \
+    printf '%s' "${AGENT_BROWSER_USER_AGENT}" | grep -q "Chrome/${chrome_major}\." || { \
+        echo "AGENT_BROWSER_USER_AGENT disagrees with installed Chromium ${chrome_major}"; exit 1; \
+    }; \
+    Xvfb :99 -screen 0 1280x800x24 -nolisten tcp >/tmp/xvfb-build.log 2>&1 & \
+    xvfb_pid=$!; \
+    for _ in $(seq 1 20); do xdpyinfo -display :99 >/dev/null 2>&1 && break; sleep 0.5; done; \
+    xdpyinfo -display :99 >/dev/null; \
+    DISPLAY=:99 agent-browser --headed --args "${AGENT_BROWSER_ARGS},--no-sandbox" \
+        open about:blank >/tmp/headed-check.log 2>&1 || true; \
+    if grep -qi 'Missing X server\|exited early' /tmp/headed-check.log; then \
+        echo "headed browser check failed:"; cat /tmp/headed-check.log; exit 1; \
+    fi; \
+    DISPLAY=:99 xwininfo -root -children | grep -qi chromium; \
+    agent-browser close >/dev/null 2>&1 || true; \
+    kill "${xvfb_pid}" 2>/dev/null || true; \
+    wait "${xvfb_pid}" 2>/dev/null || true; \
+    rm -rf /tmp/.X99-lock /tmp/.X11-unix/X99 /tmp/xvfb-build.log /tmp/headed-check.log
+
 RUN set -eux; \
     TS_PARSER_DIR="/home/pentester/.tree-sitter/parsers"; \
     mkdir -p "${TS_PARSER_DIR}"; \

containers/docker-entrypoint.sh

@@ -117,6 +117,46 @@ echo ". /etc/profile.d/proxy.sh" >> ~/.zshrc
 
 echo "✅ System-wide proxy configuration complete"
 
+# A virtual display so the agent can fall back to headed Chrome when a target
+# rejects headless; without it headed mode dies but still exits 0.
+DISPLAY_NUM="${STRIX_DISPLAY_NUM:-99}"
+DISPLAY_GEOMETRY="${STRIX_DISPLAY_GEOMETRY:-1280x800x24}"
+
+if ! xdpyinfo -display ":${DISPLAY_NUM}" >/dev/null 2>&1; then
+  # Nothing is answering, so a leftover lock/socket is stale; Xvfb refuses to
+  # start with one present.
+  rm -f "/tmp/.X${DISPLAY_NUM}-lock" "/tmp/.X11-unix/X${DISPLAY_NUM}" 2>/dev/null || true
+  Xvfb ":${DISPLAY_NUM}" -screen 0 "${DISPLAY_GEOMETRY}" -nolisten tcp \
+    > /tmp/xvfb.log 2>&1 &
+  for _ in $(seq 1 20); do
+    xdpyinfo -display ":${DISPLAY_NUM}" >/dev/null 2>&1 && break
+    sleep 0.5
+  done
+fi
+
+if xdpyinfo -display ":${DISPLAY_NUM}" >/dev/null 2>&1; then
+  echo "✅ Virtual display :${DISPLAY_NUM} ready (${DISPLAY_GEOMETRY})"
+else
+  echo "⚠️  Xvfb failed to start; headed browsing is unavailable. Xvfb log:"
+  cat /tmp/xvfb.log 2>/dev/null || echo "(no log available)"
+fi
+
+# Best-effort session bus: without it headed Chrome spews dbus errors that read
+# like fatal failures in tool output.
+if [ ! -S /run/dbus/system_bus_socket ]; then
+  sudo mkdir -p /run/dbus
+  sudo dbus-daemon --system --fork > /tmp/dbus.log 2>&1 || true
+fi
+
+cat << EOF | sudo tee /etc/profile.d/browser.sh
+export DISPLAY=:${DISPLAY_NUM}
+EOF
+
+echo ". /etc/profile.d/browser.sh" >> ~/.bashrc
+echo ". /etc/profile.d/browser.sh" >> ~/.zshrc
+
+. /etc/profile.d/browser.sh
+
 echo "Adding CA to browser trust store..."
 sudo -u pentester mkdir -p /home/pentester/.pki/nssdb
 sudo -u pentester certutil -N -d sql:/home/pentester/.pki/nssdb --empty-password

strix/agents/prompts/system_prompt.jinja

@@ -263,7 +263,13 @@ Remember: A single well-validated high-impact vulnerability is worth more than d
 <multi_agent_system>
 AGENT ISOLATION & SANDBOXING:
 - All agents run in the same shared Docker container for efficiency
-- Each agent has its own: browser sessions, terminal sessions
+- Each agent has its own terminal sessions
+- Browsers are NOT per-agent by default: `agent-browser` with no `--session` is one
+  shared browser, so a concurrent agent's navigation invalidates your page and refs.
+  Pass `--session <your-agent-name>` for any browser work of your own — then it is
+  yours alone. Each session is a full Chromium (~340 MB) on this shared box, so keep
+  one, not several, and `agent-browser --session <name> close` when you're done with
+  the target; an idle browser is reclaimed automatically after 3 minutes
 - All agents share the same /workspace directory and proxy history
 - Agents can see each other's files and proxy traffic for better collaboration
 

strix/core/inputs.py

@@ -201,9 +201,10 @@ def make_model_settings(
     request_timeout: float | None = None,
     prompt_cache: bool = True,
     extra_headers: dict[str, str] | None = None,
+    has_tools: bool = True,
 ) -> ModelSettings:
     model_settings = ModelSettings(
-        parallel_tool_calls=False,
+        parallel_tool_calls=False if has_tools else None,
         retry=DEFAULT_MODEL_RETRY,
         include_usage=True,
         extra_args=request_timeout_extra_args(request_timeout),

strix/interface/main.py

@@ -224,6 +224,7 @@ async def warm_up_llm(show_model_warning: bool = True) -> None:
                 request_timeout=llm.timeout,
                 prompt_cache=False,
                 extra_headers=settings.dedupe.extra_headers,
+                has_tools=False,
             )
             if deduper_extra:
                 merged = {**(deduper_settings.extra_args or {}), **deduper_extra}

strix/interface/scan_setup.py

@@ -78,6 +78,7 @@ async def preflight_model_connection(
         request_timeout=resolved_settings.llm.timeout,
         prompt_cache=False,
         extra_headers=resolved_settings.llm.extra_headers,
+        has_tools=False,
     )
     await asyncio.wait_for(
         model.get_response(

strix/llm/compaction.py

@@ -294,6 +294,7 @@ async def _summarize(model: str, prompt: str, max_tokens: int) -> str | None:
         request_timeout=llm.timeout,
         prompt_cache=False,
         extra_headers=llm.extra_headers,
+        has_tools=False,
     ).resolve(ModelSettings(max_tokens=max_tokens))
     try:
         response = (

strix/report/dedupe.py

@@ -62,6 +62,7 @@ def _dedupe_model_settings(
         # must never receive the main endpoint's credentials. A dedicated model
         # gets its own DEDUPE_LLM_EXTRA_HEADERS instead.
         extra_headers=dedupe.extra_headers if dedupe.model else llm.extra_headers,
+        has_tools=False,
     )
     extra = _dedupe_extra_args(dedupe)
     if extra:

strix/skills/tooling/agent_browser.md

@@ -58,6 +58,26 @@ agent-browser screenshot
 The browser stays running across commands so these feel like a single
 session. Use `agent-browser close` (or `close --all`) when you're done.
 
+The default session is **shared with every other agent in the sandbox** — if
+another agent navigates it, your page and your refs are gone from under you. So
+claim your own by passing `--session <your-agent-name>` on **every** command:
+
+```bash
+agent-browser --session recon-3 open https://example.com
+agent-browser --session recon-3 snapshot -i
+agent-browser --session recon-3 close        # when done with the target
+```
+
+The examples in the rest of this skill omit `--session` to keep them readable;
+keep passing yours. Each session is a separate Chromium (~340 MB) on a shared
+box, so hold one rather than several, and close it when you're finished.
+
+A browser left idle for 3 minutes is reclaimed automatically to free memory for
+the other agents; the next command relaunches it, but the page, tabs, refs and
+cookies are gone. If you're authenticated and about to go do something else for a
+while, save the state first (see
+[Persist session across runs](#persist-session-across-runs)).
+
 ## Reading a page
 
 ```bash
@@ -169,6 +189,59 @@ After any page-changing action, pick one:
 Avoid bare `wait 2000` except when debugging — it makes scripts slow and
 flaky. Timeouts default to 25 seconds.
 
+## Bot protection (read before fighting a login)
+
+Many targets sit behind a bot check (Cloudflare Turnstile, hCaptcha, Datadome).
+You will see it in a snapshot as a challenge iframe plus inputs that go
+`[disabled]` when you submit:
+
+```
+- textbox "Email address" [disabled, ref=e18]: user@example.com
+- button "Continue with email" [disabled, ref=e12]
+- Iframe "Widget containing a Cloudflare security challenge" [ref=e15]
+  - checkbox "Verify you are human" [checked=false, ref=e22]
+```
+
+`checked=false` that never flips means the challenge is refusing you, not that
+the click missed. Headless Chrome is itself one of the strongest signals these
+systems key on, so switch to a real browser window instead of retrying:
+
+```bash
+agent-browser close --all              # a running daemon makes --headed a no-op
+export AGENT_BROWSER_HEADED=1          # applies to every later command
+agent-browser open https://target.tld/login
+agent-browser get url                  # confirm it actually launched
+```
+
+The sandbox provides a virtual display and `DISPLAY` is already exported, so this
+works with no setup. Three traps:
+
+- **Set the env var, don't just pass `--headed` to `open`.** The flag is
+  per-invocation: the next bare `agent-browser snapshot` tries to start a
+  *headless* daemon and dies with `Multiple targets are not supported in headless
+  mode`. Export `AGENT_BROWSER_HEADED=1` (or pass `--headed` to every command).
+- **`--headed` is ignored when a daemon is already running** — it prints
+  `⚠ --headed ignored: daemon already running`. Always `close --all` first.
+- **A failed launch still exits 0.** Read the output text: `✗ Chrome exited
+  early` or `Missing X server` means you are not headed, whatever the exit code
+  says. Confirm with `agent-browser get url` before concluding anything.
+
+Then behave like a person rather than a script:
+
+- Drive real input events — `click`, `hover`, `keyboard type` — never `eval` with
+  `element.value = ...`. Assigning `value` directly leaves React-controlled
+  inputs internally empty, so the form submits blank or stays disabled.
+- `focus` the field, then `keyboard type "text"` when `fill` appears to work but
+  the app doesn't react.
+- Click the challenge checkbox by its ref inside the iframe, then
+  `wait --text` / `wait --url` for the *result*; don't re-click while it verifies.
+- Save the session once you're through (`state save`, or `--session-name`) so a
+  browser restart doesn't send you back to the challenge.
+
+If the challenge still refuses after a couple of honest attempts, stop. Report
+that the target is gated and hand back a bounded result — burning your whole
+window on one login costs more coverage than the login was worth.
+
 ## Common workflows
 
 ### Log in
@@ -307,6 +380,16 @@ agent-browser --session b fill @e1 "bob@test.com"
 `AGENT_BROWSER_SESSION=myapp` sets the default session for the current
 shell.
 
+Use a session named after yourself for your own work — that's what keeps a
+concurrent agent from navigating the page out from under you. Every session is a
+separate Chromium though, so hold one at a time rather than a collection, and
+close each one when its flow is finished:
+
+```bash
+agent-browser --session a close
+agent-browser --session b close
+```
+
 ### Mock network requests
 
 ```bash
@@ -367,8 +450,11 @@ agent-browser dialog dismiss          # cancel
 
 ## Readiness & recovery
 
-The first `agent-browser open` in a session launches the headless-Chrome
-daemon; later commands reuse it. Distinguish the two failure modes and react
+The first `agent-browser open` in a session launches the Chrome daemon (headless
+unless you pass `--headed`); later commands reuse it. A daemon left idle for 3
+minutes shuts itself down to free memory for the other agents, so an `open` after
+a long gap is a fresh browser rather than a resumed one — expect to re-navigate,
+and re-`state load` if you were logged in. Distinguish the failure modes and react
 differently — do **not** blindly re-run the same failing command in a loop:
 
 - **Daemon / connection failure** (`Failed to connect`, `connection refused`,
@@ -457,7 +543,7 @@ and [references/authentication.md](references/authentication.md).
 ```bash
 --session <name>        # isolated browser session
 --json                  # JSON output (for machine parsing)
---headed                # show the window (default is headless)
+--headed                # real browser window (default is headless); see "Bot protection"
 --auto-connect          # connect to an already-running Chrome
 --cdp <port>            # connect to a specific CDP port
 --profile <name|path>   # use a Chrome profile (login state survives)

tests/test_inputs.py

@@ -299,6 +299,16 @@ def test_make_model_settings_forces_required_for_anyllm_routed_openai_model() ->
     assert settings.tool_choice == "required"
 
 
+def test_make_model_settings_disables_parallel_tool_calls_by_default() -> None:
+    assert make_model_settings("none", model_name="gpt-4o").parallel_tool_calls is False
+
+
+def test_make_model_settings_omits_parallel_tool_calls_without_tools() -> None:
+    settings = make_model_settings("none", model_name="gpt-4o", has_tools=False)
+
+    assert settings.parallel_tool_calls is None
+
+
 def test_make_model_settings_sets_request_timeout() -> None:
     settings = make_model_settings(
         "none",