When a local browser_navigate (or any browser command) fails fast because
Chromium isn't on disk, attempt a one-shot binary download via
`agent-browser install` and retry instead of only printing a hint.
Scope is narrow on purpose:
- binary only, never `--with-deps` (that shells apt/needs root, so missing
system libraries stay a user action)
- gated by `security.allow_lazy_installs` (same opt-out as every lazy install)
- skipped in Docker (Chromium ships in the image)
- attempted once per process
Follow-up to #54353, which made the cold-start failure legible; this closes
the "doesn't actually install the missing browser" gap for the common case.
Local browser_navigate cold-starts the agent-browser daemon and Chromium;
60s was too short on slow Linux hosts and timeouts discarded stderr,
leaving users with a generic failure. Use a 120s floor on first open,
inject --no-sandbox in Docker, include captured daemon output plus install
hints when commands time out, and show "Failed to open" in the desktop
tool chip when navigation returns success=false.
_get_platform_tools() applies agent.disabled_toolsets as a final
override AFTER reading platform_toolsets.<platform>, so a toolset
listed there stays permanently OFF no matter what the toggle write
path saves. Blank Slate installs pre-populate this list with ~27
toolsets, making most of the desktop Toolsets UI un-enableable
(issue #49995).
Fix: _save_platform_tools() now removes any toolset the user just
explicitly enabled FOR THIS PLATFORM from agent.disabled_toolsets.
Toolsets the user did not touch, or that remain disabled on other
platforms, are left alone -- disabled_toolsets keeps working as a
cross-platform suppression list for anything not actively re-enabled.
Disabling a toolset (unchecking it) does not touch disabled_toolsets
at all -- only enables reconcile it.
Verified end-to-end with the exact repro from the issue: Blank Slate
config (disabled_toolsets=['todo','memory','browser'], cli=['file',
'terminal']) -> enable 'todo' via the toggle -> _get_platform_tools()
now resolves 'todo' as enabled while 'memory'/'browser' (untouched)
remain disabled.
Added 4 regression tests. Full tools_config suite: 101 passed
(97 existing + 4 new), no regressions.
Fixes#49995
main (cb982ad99) wired windows_hide_flags() into the auxiliary git/gh/wmic/
bash/powershell/taskkill legs but left two it didn't reach, plus the Electron
backend-launch leg it explicitly deferred. Cover them the same way:
- apps/desktop/electron/main.cjs: getNoConsoleVenvPython resolves the BASE
pythonw.exe instead of the venv Scripts\pythonw.exe shim, which re-execs a
console python.exe and flashes a conhost the desktop backend can't suppress.
Both backend creators put the venv site-packages on PYTHONPATH so imports
still resolve under the base interpreter. (main's commit said this Electron
leg "needs a Windows-tested change of its own".)
- tools/tts_tool.py, tools/transcription_tools.py, plugins/platforms/discord:
ffmpeg conversions (voice notes / TTS / STT) via windows_hide_flags().
- plugins/platforms/whatsapp: netstat + taskkill bridge-port cleanup via
windows_hide_flags().
All no-ops on POSIX. Tests assert the base-pythonw preference and the ffmpeg
legs pass CREATE_NO_WINDOW.
The Windows desktop GUI runs its backend headless via pythonw.exe. Several
auxiliary subprocess sites that run inside that windowless backend spawned
console-subsystem children (git, gh, wmic, powershell, bash, rg, taskkill)
WITHOUT CREATE_NO_WINDOW, so Windows allocated a fresh conhost per call and
flashed a black window on screen — sometimes continuously (the dashboard
Projects-tree git probe alone fired ~118 spawns in 60s on startup).
The terminal tool, cron, browser, code_execution, and gateway-spawn paths
already carry windows_hide_flags(); these auxiliary probe/scan/launcher legs
were missed. Wire the existing helper into them:
- tui_gateway/git_probe.py: run_git (+ encoding=utf-8/errors=replace, fixes the
cp950 UnicodeDecodeError on CJK paths from the same site)
- agent/coding_context.py: _git (per-turn git status/log/diff)
- agent/context_references.py: _run_git + _rg_files (@file/@ref resolution)
- hermes_cli/copilot_auth.py: gh auth token probe (auxiliary provider:auto)
- hermes_cli/gateway.py: wmic + PowerShell Get-CimInstance PID scan
- hermes_cli/main.py: wmic stale-dashboard PID scan
- gateway/status.py: taskkill /T /F force-kill
windows_hide_flags() returns 0 on POSIX, so every changed call is a no-op on
Linux/macOS (verified: real git/rg probes still work; Windows-simulated calls
all pass creationflags=CREATE_NO_WINDOW).
Scoped to the windowless-backend paths that cause the reported flashing. The
Electron updater-handoff leg (main.cjs windowsHide:false) and the
interactive-CLI banner probes (cli.py) are intentionally NOT touched here —
the former needs a Windows-tested change of its own, the latter runs in a
visible console anyway.
Tracking: #54220
Refs: #53178#53631#53781#53957#49602#52982#53424#53053#53016
OAuth-protected MCP servers (e.g. Hospitable) return 200 text/html on an
unauthenticated HEAD probe — a login/landing page the server cannot substitute
for a real MCP response without a Bearer token. The preflight cannot
distinguish this from a misconfigured URL, so it raises NonMcpEndpointError
before the OAuth browser flow has a chance to run.
Add `and self._auth_type != "oauth"` to the preflight condition in
MCPServerTask.run(). The probe is inapplicable to OAuth servers: their URL
legitimacy is established by .well-known/oauth-protected-resource during the
OAuth handshake, not by a GET content-type check.
Concrete repro: Hospitable (https://mcp.hospitable.com/mcp) returns
`200 text/html` to an unauthenticated httpx HEAD. Without the guard:
✗ NonMcpEndpointError at `hermes mcp test`
With the guard:
✓ Connected (1487ms) — 63 tools discovered
Relation to open PRs:
- #37598 adds a POST probe fallback for POST-only non-OAuth servers (e.g.
DocuSeal), but only passes when POST returns 2xx + MCP content-type.
Hospitable returns 401 on the POST probe (Bearer challenge), so #37598
does not cover this case.
- #49463 extends the POST probe to also pass on non-2xx auth challenges
(making it OAuth-aware), but is labeled duplicate of #37598 and may not
land independently.
This fix is complementary: it handles OAuth servers with zero extra
round-trips rather than adding a POST probe step.
Tests:
- test_oauth_server_html_response_raises_without_skip: documents that
_preflight_content_type raises NonMcpEndpointError for 200 text/html
(the underlying issue), with an OAuth-server docstring.
- test_run_skips_preflight_for_oauth: verifies that run() does NOT invoke
_preflight_content_type when auth_type=="oauth", using class-level
monkeypatching so the gate is exercised without a live MCP transport.
23 passed tests/tools/test_mcp_preflight_content_type.py
When a stale lock file survives a gateway crash, `acquire_scoped_lock()`
may return `(False, existing_dict)` even after detecting and deleting
the stale lock (e.g. if unlink fails or a race condition occurs).
Previously, `_acquire_platform_lock()` called
`_set_fatal_error(..., retryable=False)`, which permanently killed the
platform — the reconnect watcher never retries a non-retryable fatal
error.
Change to `retryable=True` so the platform enters the "retrying"
state and the reconnect watcher can attempt acquisition again after the
standard backoff delay.
Fixes#54167
_append_model_switch_marker() appended the post-/model-switch context marker
to session history as {"role": "system"}. The cached system prompt is
prepended to the API message list (conversation_loop.py), so this marker
became a SECOND system message mid-array after prior user/assistant turns.
Strict OpenAI-compatible providers (vLLM, Qwen) reject any system message
that is not at the beginning of the array, returning HTTP 400 and killing
the conversation on the next turn.
Flip the marker to role="user" (history entry + both session-DB persist
sites), matching the existing personality-overlay marker which already uses
role="user". repair_message_sequence() then coalesces it with adjacent user
turns as needed.
Co-authored-by: liuhao1024 <sunsky.lau@gmail.com>
Co-authored-by: Lucas Nicolas <lucas.nicolas@proton.me>
On macOS app.quit() closes windows but window-all-closed deliberately keeps
the process alive (Dock convention). Every detached hand-off (update swap,
relaunch, Windows bootstrap recovery, uninstall cleanup) waits for the
desktop PID to exit before replacing/removing the bundle — so the process
never dying means the script spins its full PID-wait and the user sees a
blank app, or an uninstall that appears to do nothing.
Add a module-level isQuittingForHandoff flag, set before every hand-off
app.quit(); window-all-closed then quits on all platforms when it's set.
Covers all five hand-off sites including the Linux relaunch path.
launchd restart can leave the gateway job stopped but still registered after
update-time drain logic, so a direct bootstrap hits exit 5 and falls back to a
detached process. Booting the stale registration out before bootstrap keeps the
launchd-managed restart path intact and locks it with a regression test.
Constraint: Keep upstream-facing conventional commit style while preserving local decision context
Rejected: Treat bootstrap exit 5 as expected | Leaves macOS launchd restart outside launchd supervision after update
Confidence: high
Scope-risk: narrow
Directive: Keep launchd start/restart recovery flows aligned when changing launchctl handling
Tested: pytest -q tests/hermes_cli/test_gateway_service.py -k "launchd_restart_boots_out_stale_registration_before_bootstrap or launchd_restart_falls_back_to_detached_on_error_5 or launchd_restart_drains_running_gateway_before_kickstart or launchd_restart_self_requests_graceful_restart_without_kickstart"
Tested: pytest -q tests/hermes_cli/test_gateway_service.py -k launchd
Not-tested: Manual macOS launchctl restart after hermes update
test_spinner_elapsed_format_is_fixed_width_to_reduce_wrap_jitter derived
_tool_start_time from the live time.monotonic() clock (now - 65.2 / now - 9.2).
monotonic()'s epoch is arbitrary — on a host where monotonic() < 65.2 (fresh
subprocess on a freshly-booted CI runner) the start time went negative, the
(t0 > 0) guard in _render_spinner_text() dropped the '(elapsed)' suffix, and
short.split('(',1)[1] raised IndexError: list index out of range. Deterministic
given a small clock, so it would keep flaking, not clear on rerun.
Pin time.monotonic to a fixed 1000.0 and offset _tool_start_time from it so both
the <60s and >=60s paths always render the elapsed suffix regardless of the
runner's monotonic epoch.
Pre-existing main flake (surfaced in CI test slice 1/8).
The README line enumerated 11 providers inline, which dilutes the point
and goes stale as providers come and go. Replace with Nous Portal,
OpenRouter, OpenAI, your own endpoint, and a 'many others' link to the
canonical AI Providers docs page that already lists them all.
The desktop app and dashboard chat reach the agent through the /api/ws
JSON-RPC sidecar (tui_gateway.ws.handle_ws), NOT through
tui_gateway.entry.main() — the stdio-TUI path that spawns the background
MCP discovery thread. In the WS process discovery was therefore never
started: _make_agent only *waits* (wait_for_mcp_discovery), which no-ops
when the thread was never created, so the agent snapshotted an MCP-less
tool list. The only discovery trigger reachable was a manual /reload-mcp,
which is why tools appeared after a reload but vanished on restart.
Start the shared, idempotent, config-gated background discovery in
handle_ws right after accept() and before gateway.ready, so the first
agent build picks up already-spawning servers (and the existing
late-binding refresh handles slow ones).
Fixes#38945.
The salvaged #35519 regression guard asserted that default (non-file_read)
mode keeps a head/tail `ghp_S1...Pn2T` mask for a `token: <key>` line. On
current main the YAML config pass (`_YAML_ASSIGN_RE`, key `token`) re-masks
the already-prefix-masked value to `***`, so the assertion was stale. Switch
to a bare-token context so the guard isolates what it claims (prefix-mask
head/tail shape in default mode) without depending on the YAML collapse.
When security.redact_secrets is on (default), read_file/search_files/cat
applied redact_sensitive_text(code_file=True) to file content, which still
ran prefix masking. An API key in config.yaml (ghp_..., sk-..., xai-..., etc.)
came back as a head/tail mask like `ghp_S1...Pn2T` — a plausible-looking
truncated key. When an agent read that and wrote it back to config, the masked
value replaced the real credential, silently breaking auth (401). Production
evidence: a config.yaml found containing the exact 13-char masked GitHub PAT.
The two community PRs (#35529, #35534) fixed the corruption by NOT redacting
prefixes for config reads — but that exposes the user's real keys to the agent
context, model, and logs (a security regression). This takes the safer route:
keep redacting, but for file content emit a NON-REUSABLE sentinel.
- New `_mask_token_nonreusable`: prefix secrets -> `«redacted:ghp_…»` (vendor
label preserved for debuggability; zero secret bytes; angle-bracket/ellipsis
wrapper is syntactically invalid as a token so it can't be mistaken for or
written back as a usable key).
- New `redact_sensitive_text(file_read=True)` routes prefix matches through it
(implies code_file=True). Default/log/display mode is UNCHANGED — `_mask_token`
still keeps head/tail (fine for logs, never written back).
- Wired the 3 file_tools.py call sites (read_file / search_files / cat) to
file_read=True.
Fixes both the corruption AND avoids the secret-exposure of the un-redact
approach. 6 new tests (sentinel shape, no-leak, not-a-plausible-key, default
mode unchanged, file_read implies code_file, sk- prefix); 88 redact tests pass;
mutation-verified (reverting to the old mask fails the sentinel/leak tests).
Co-authored-by: liuhao1024 <sunsky.lau@gmail.com>
Co-authored-by: adammatski1972 <289282750+adammatski1972@users.noreply.github.com>
Closes#35519. Supersedes #35529, #35534.
The zombie status probe referenced psutil.Process/NoSuchProcess/Error
unconditionally, which raised AttributeError when psutil is a partial
stub that only defines pid_exists (as in test_windows_native_support's
fallback tests). Guard the probe so any failure to read status degrades
to the authoritative pid_exists() instead of raising.
Under systemd Restart=always, the old gateway becomes a zombie (in the
process table, awaiting reap) when the replacement starts. _pid_exists()
reported the zombie as alive, so --replace waited on a PID that never
dies, then aborted with exit 1 — a silent crash loop. Standalone runs are
unaffected because nothing respawns the gateway into a zombie.
The live path is psutil.pid_exists(), which returns True for zombies, so
the check is added there (Process.status() == STATUS_ZOMBIE -> dead). The
psutil-less POSIX fallback also reads /proc/<pid>/stat (state Z) with a ps
state= fallback for macOS/BSD, before the os.kill(pid, 0) liveness probe.
Diagnosis and the /proc + ps POSIX fallback by MorAlekss (PR #44898);
extended to cover the psutil hot path so the fix applies on normal installs.
Co-authored-by: MorAlekss <mor.aleksandr@yahoo.com>
Folds in PR #42124 (kyssta-exe): systemd_install gained a non_interactive
flag so the 'Remove the legacy unit(s)?' prompt — the second hidden prompt
not guarded by --start-now/--start-on-login — is also skipped in headless
contexts. Updates systemd_install test mocks to accept the new kwarg and
adds coverage for the legacy-unit-skip path.
When running `hermes gateway install` on Linux/systemd, the command
unconditionally prompts with two `prompt_yes_no` questions, breaking
headless installs (SSH, CI, provisioning scripts) and ignoring the
existing --start-now / --start-on-login CLI flags that the Windows
branch already respects.
The fix mirrors the Windows path: read CLI flags first, prompt only
when flags are not provided AND stdin is a TTY, and fall back to True
defaults for non-TTY contexts. The argparse help strings are promoted
from SUPPRESS to visible so users can discover the flags.
Fixes#42065
When the previous frame's stdout.write has not drained (the outer terminal
parser is overwhelmed by a wide CR+LF burst — CJK + ANSI tool output on a
high-context session), the renderer kept writing a new frame every tick. That
piled writes onto an already-backed-up pipe and kept the macrotask queue hot,
starving the stdin 'readable' callback — the observed stdin freeze where the
agent loop keeps running but keystrokes/Ctrl-C are dead.
onRender now coalesces: while pendingWriteStart is non-null (prior write's
drain callback hasn't fired) it skips the frame and retries on the drain tick
instead of writing. A MAX_COALESCED_BACKPRESSURE_FRAMES ceiling forces a write
through after N skips so a terminal whose drain callback never fires (OSError
EIO on flush) self-heals once the pipe recovers rather than wedging forever.
TTY-only; piped stdout has no flow control. Coalesce counter resets on every
real write.
This is the stdout-backpressure strand left open after #54046 fixed the
swallowed-exception strand.
The /api/pty handler only closed the PtyBridge in the writer loop's finally.
On child EOF the reader task closes the WebSocket, but if the handler task is
cancelled the instant the socket closes, the writer's finally can be skipped
and the PTY fds leak (#54028) — the FD-leak the regression test guards. Under
dashboard auto-reconnect this stacks orphaned PTYs until fds are exhausted.
Reap the bridge in the reader's EOF finally too (close() is idempotent), so
the PTY is reaped independently of the writer-loop cancellation race. Harden
the regression test to poll for teardown instead of asserting on the same
tick. Was flaky on main (2/20); now 25/25.
PR #39219 split run_browser_install_with_timeout into a thin wrapper that
delegates to a new run_with_timeout helper (and parameterized the timeout
binary as $timeout_bin for macOS gtimeout support), but did not update
tests/test_install_sh_browser_install.py. The behavioral harness extracted
only the now-empty wrapper, so the install command never ran (runs==[]),
failing all 8 behavioral cases; two text assertions also still expected the
old literal 'timeout' invocation.
Fix the stale test: extract run_with_timeout alongside the wrapper, and match
the $timeout_bin-parameterized GNU-timeout strings. Behavior unchanged.
The desktop install step ran npm ci / npm run pack with no wall-clock cap, and
the sibling browser-tools / TUI / agent-browser dependency installs had the same
gap. The Electron binary (~150MB) is fetched from GitHub during the pack; on a
throttled or region-blocked link that download can *stall* rather than fail —
npm never errors and never exits, so the installer sits on "Build desktop app"
(step 9/11) indefinitely with only harmless 'npm warn deprecated' lines visible.
The existing self-heal escalation (cache purge -> dist restore -> npmmirror
fallback) only fires when pack returns non-zero, so a stall bypassed it.
- run_with_timeout (generalized from run_browser_install_with_timeout): GNU
timeout --foreground -k 10 (Ctrl+C-aware, #35166) / gtimeout for external
commands, else a pure-shell process-group watchdog so stock macOS (neither
binary present) is protected. Shell functions (_desktop_pack) always take the
pure-shell path — the timeout binary can't exec a function. Integer-normalized
budget + a boundary recheck so a command finishing in the final poll second
isn't mislabeled 124. The internal wait is guarded so set -e can't abort
mid-function before the real exit code is computed.
- Wrap the desktop npm ci/install (sharing ONE budget via a computed deadline so
a stall can't cost 2x DESKTOP_BUILD_TIMEOUT) + all three _desktop_pack attempts
(DESKTOP_BUILD_TIMEOUT, default 900s), and the browser-tools / TUI / agent-
browser registry installs (NODE_DEPS_TIMEOUT, default 600s).
A stall now converts to a bounded non-zero exit that feeds the existing mirror
self-heal instead of hanging the whole install.
* fix(security): redact secrets in background process + foreground env-dump output
Terminal-output redaction was incomplete (#43025):
- Gap 1: process(action=poll/log/wait) returned background stdout verbatim —
no redaction at all. A background printenv/server/test emitting a key leaked
raw to the model, session.db, and CLI display. Same for the gateway
background-process watcher's completion/progress notifications.
- Gap 2: the foreground terminal path hardcoded code_file=True, which skips the
ENV-assignment pass, so an opaque token (no vendor prefix) from env/printenv
leaked even there.
Adds agent.redact.redact_terminal_output(output, command) as the single policy
for ALL terminal-output surfaces: env-dump commands (env/printenv/set/export/
declare) get the ENV-assignment pass (code_file=False) to mask opaque tokens;
other commands stay on code_file=True to avoid false positives on source dumps.
Wired into terminal_tool, process_registry (_handle_process boundary), and the
gateway watcher. Respects security.redact_secrets (no force) — opt-out preserved.
* docs: add infographic for #43025 terminal-output redaction fix
The merged CLOSE-WAIT heartbeat (#52744) only probes get_me(), which uses the
general request path and stays healthy while PTB's getUpdates consumer is
silently wedged (updater.running=True but the long-poll task is stuck, observed
on WSL2). DMs then queue in the Bot API and never reach handlers (#42909).
Augment the existing _polling_heartbeat_loop to also probe
get_webhook_info().pending_update_count. After two consecutive probes that see a
non-draining queue while the updater claims to be running, escalate into the
existing _handle_polling_network_error recovery ladder — no new restart
machinery. No-ops in webhook mode, when the updater is not running, or when a
reconnect is already in flight.
Credit to @gazzumatteo, whose PR #42959 identified the pending_update_count
signal as the missing liveness probe. This reuses the existing heartbeat +
recovery path rather than adding a parallel watchdog.
Fixes#42909.
Two related redaction bugs from #43083:
1. build_assistant_message redacted tool-call arguments in-memory. That dict
feeds both the replayed conversation history and state.db (which is itself
replayed verbatim on session resume), so the model read back its own
PGPASSWORD='***' psql call and copied the placeholder, breaking every
credential-dependent command on the second turn. The masking gave no real
protection either — the same secret still leaks through tool OUTPUT. Remove
it. Keeping secrets out of the replayable store is a separate
tokenization/vault concern (security.redact_secrets still governs
storage-time redaction elsewhere).
2. _AUTH_HEADER_RE's greedy \S+ credential class ate a closing quote when the
token sat flush against it (Authorization: Bearer sk-.."), turning value
corruption into syntax corruption (unterminated quote -> shell EOF /
SyntaxError). Exclude " and ' from the token class; real credentials never
contain them.
Closes#43083.
Config-backed quick_commands bypassed the admin-only slash gate. The
early gate in _handle_message only fires for registry-known commands
(is_gateway_known_command), but quick_commands are never in the gateway
registry, so they reached the type:exec dispatch sink unchecked. An
allowlisted non-admin gateway user could invoke admin-only quick
commands — including shell exec in the gateway process — even when the
operator set allow_admin_from / user_allowed_commands to lock them out.
Apply _check_slash_access(source, command) at the quick_commands
dispatch site (the single exec chokepoint, cold-path only) using the
raw typed name. Admins and users with the command in
user_allowed_commands still run it; backward-compat (no policy set)
is unaffected.
Fixes#44727.
Co-authored-by: maxpetrusenko <max.petrusenko.agent@gmail.com>
Co-authored-by: zapabob <1920071390@campus.ouj.ac.jp>
* fix(dashboard): close PTY WebSocket on child EOF to stop FD leak
The /api/pty handler's reader task returns on child EOF, but the writer
loop stayed blocked on ws.receive() until the browser sent a disconnect.
When the browser socket is half-open (no FIN delivered — common on
macOS/launchd), that disconnect never arrives, so the handler never
reaches its finally and the PTY master fd + child process leak. With
dashboard auto-reconnect (#52962), every dropped socket then spawns a
fresh PTY on top of the orphaned one, exhausting file descriptors within
hours (EMFILE / Errno 24).
Fix: the reader task now closes the WebSocket in a finally when the child
EOFs or the send side breaks, which unblocks ws.receive() so the existing
finally runs bridge.close(). The writer loop also guards ws.receive()
against the RuntimeError Starlette raises once the socket is closed.
Reported by @fifteenzhang.
Fixes#54028
* docs: add infographic for #54028 PTY FD leak fix
The snapshot/vision guards re-check the page URL before returning content,
but browser_console(expression=...) -> _browser_eval returns arbitrary JS
results directly, leaving two same-class bypasses open:
1. Direct fetch: fetch('http://127.0.0.1/secret').then(r=>r.text()) reads
a private endpoint and returns the body — the page URL stays public so
the post-eval recheck never sees it.
2. Navigate-then-read: location.href='http://127.0.0.1/' then a later eval
reads document.body.innerText.
Guard _browser_eval on the same condition as navigate/snapshot/vision
(not local backend, not local sidecar, not allow_private_urls):
- pre-scan the expression for private/always-blocked URL literals
- re-check window.location.href after the eval at both success-return
sites (supervisor fast-path + subprocess fallback)
Probe failures fail-open (matching the snapshot/vision guards).
The private-network guard in browser_snapshot() and browser_vision()
blocked all private URLs, including those accessed via local sidecar
sessions (hybrid routing). Local sidecar sessions intentionally access
private URLs — the cloud provider never sees the URL in that case.
Add `_is_local_sidecar_key(effective_task_id)` check to both guards,
matching the existing pattern in browser_navigate().
Fixes#45101 review feedback from egilewski.
The SSRF bypass in #44731 was only patched for browser_snapshot(), but
browser_vision() exposes the same vulnerability — it takes a screenshot
and sends it to the vision model without checking if eval-driven
navigation moved the page to a private/internal URL.
Add the same current-page URL safety check to browser_vision() before
any screenshot is captured, encoded, or forwarded to the vision model.
This covers both the normal screenshot path and the Lightpanda Chrome
fallback path.
7 new tests: blocks private URL, allows public URL, skips in local
backend, skips when private URLs allowed, handles eval failure/empty/exception.
