fix(dashboard): keep resumed chat alive on network binds

[hehehe0803]

What does this PR do?

Fixes dashboard embedded chat resume when the dashboard is intentionally served on an explicit non-loopback network host, such as a private LAN/VPN address.

The existing merged WebSocket fix covers all-interface binds (0.0.0.0 / ::), but not the common explicit private-interface bind shape:

hermes dashboard --host <private-network-ip> --port <port> --insecure --tui --no-open

In that mode the HTTP dashboard can load, while /api/pty closes before accept because the WebSocket client-IP gate still behaves like the dashboard is loopback-only. This can present as Sessions → Resume in Chat opening a chat page that immediately shows [session ended], reconnect/disconnect UI, or a spawned TUI runtime error even though the underlying session is still valid.

This PR:

• records the explicit --insecure operator opt-in on app.state.allow_public and uses it for dashboard PTY WebSocket admission;
• keeps remote WebSocket clients rejected for default loopback-only dashboard binds;
• adds a collision-resistant resumeNonce so repeated Sessions → Resume in Chat clicks for the same session can restart a stale PTY identity;
• clarifies that resumeNonce is only folded into the opaque dashboard event channel id and is not interpreted as backend resume state;
• treats unsupported Node runtimes as unusable even when HERMES_SKIP_NODE_BOOTSTRAP is set, surfacing a clear error instead of letting the embedded TUI crash later;
• adds regression coverage using documentation/test IP addresses only.

Related context: #18633, #25072, #27801.

Related Issue

Fixes #26264

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• hermes_cli/web_server.py
  • Persist allow_public from start_server() onto app.state.
  • Let _ws_client_is_allowed() accept non-loopback clients only when the dashboard was intentionally exposed with --insecure or bound to all interfaces.
• tests/hermes_cli/test_web_server.py
  • Add regression coverage for explicit non-loopback dashboard binds with and without the --insecure opt-in.
  • Reset bound_host / allow_public in the PTY WebSocket test fixture so tests cannot pass because of leaked global app state.
• web/src/pages/SessionsPage.tsx
  • Generate resumeNonce with crypto.randomUUID() when available, with a timestamp/random fallback.
• web/src/pages/ChatPage.tsx
  • Update comments so they match the actual opaque channel-id behavior.
• hermes_cli/main.py
  • Fail clearly when HERMES_SKIP_NODE_BOOTSTRAP is set but the available Node/npm runtime is unusable for the dashboard-spawned TUI.
• tests/hermes_cli/test_tui_resume_flow.py
  • Cover the unsupported-Node + skip-bootstrap error path.

How to Test

Automated checks run on this branch:

1. Python targeted tests:

   scripts/run_tests.sh tests/hermes_cli/test_web_server.py::TestPtyWebSocket tests/hermes_cli/test_tui_resume_flow.py -q

   Result: 40 passed.

2. Python syntax check:

   python -m py_compile hermes_cli/main.py hermes_cli/web_server.py

   Result: passed.

3. Web production build:

   cd web && npm run build

   Result: passed.

Manual/browser verification shape:

1. Start the dashboard on a host machine with an explicit private-network bind:

   hermes dashboard --host <private-network-ip> --port <port> --insecure --tui --no-open

2. From a separate browser/client on the same private network, open:

   http://<private-network-ip>:<port>/sessions
   

3. Click Resume in Chat for an existing session.
4. Verify the chat route loads a live embedded TUI for the selected session instead of [session ended], reconnect-only UI, or an unsupported Node runtime crash.

Additional check:

• cd web && npm run lint was also run. It still fails on the current repository lint baseline in unrelated files/rules (for example OAuthProvidersCard.tsx, Toast.tsx, i18n files, and existing SessionsPage.tsx effect warnings). No new lint-specific issue was introduced by this change.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: Linux / private-network dashboard bind

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — or N/A
• I've updated cli-config.yaml.example if I added/changed config keys — or N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — or N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A
• I've updated tool descriptions/schemas if I changed tool behavior — or N/A

Screenshots / Logs

scripts/run_tests.sh tests/hermes_cli/test_web_server.py::TestPtyWebSocket tests/hermes_cli/test_tui_resume_flow.py -q
40 passed in 2.15s

python -m py_compile hermes_cli/main.py hermes_cli/web_server.py
# passed

cd web && npm run build
✓ built

[hehehe0803]

Rebased/retested this fix against current main and pushed an updated branch.

What changed in the update:

• Kept the original explicit non-loopback dashboard bind fix for /api/pty.
• Tightened the Node bootstrap predicate to match the documented Vite/runtime floor: Node >=20.19, >=22.12, or >22.
• Strengthened the regression coverage so Starlette TestClient's loopback-like peer name cannot mask the real non-loopback WebSocket guard behavior.
• Kept the test data anonymized with RFC documentation/test addresses only.

Verification run locally on current main plus this branch:

python -m py_compile hermes_cli/main.py hermes_cli/web_server.py tests/hermes_cli/test_web_server.py
scripts/run_tests.sh tests/hermes_cli/test_web_server.py::TestPtyWebSocket -q
  14 passed
scripts/run_tests.sh tests/hermes_cli/test_web_server.py -q
  147 passed
cd web && npm run build
  built successfully

I also checked the diff for private/local addresses and did not include any real LAN/VPN IPs, hostnames, local user paths, or credentials.

One unrelated note from the broader retest: tests/hermes_cli currently has a baseline failure on current main around auxiliary.session_search config/test drift. That is separate from this dashboard PR; I am filing/fixing it separately rather than mixing it into this branch.

[austinpickett]

Please use PULL_REQUEST_TEMPLATE.md

[Copilot]

Pull request overview

This PR fixes the dashboard’s “Sessions → Resume in Chat” behavior when the dashboard is intentionally bound to a specific non-loopback host (LAN/VPN), ensuring the embedded PTY WebSocket can stay connected and that repeat resume attempts can reliably restart a stale terminal.

Changes:

• Allow /api/pty WebSocket clients when the dashboard is bound to an explicit non-loopback host (still guarded by the session token).
• Add a resumeNonce query param so repeated “Resume in Chat” clicks for the same session force a fresh PTY identity in the still-mounted ChatPage.
• Validate that the Node binary used for dashboard-spawned TUI is new enough, so old system Node versions don’t get treated as “good enough”.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
web/src/pages/SessionsPage.tsx	Adds resumeNonce to chat navigation query to force a fresh resume attempt.
web/src/pages/ChatPage.tsx	Incorporates resumeNonce into the PTY “identity” so the WS/terminal is rebuilt without unmounting the page.
tests/hermes_cli/test_web_server.py	Adds regression coverage for explicit network-host binds and the WS client allow-guard behavior.
hermes_cli/web_server.py	Broadens the “public bind” determination to include any non-loopback bind (not only 0.0.0.0/::).
hermes_cli/main.py	Adds Node version probing so unsupported system Node is treated as unusable and triggers bootstrap.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[hehehe0803]

Updated this PR against current main and addressed the review feedback:

• Rebased onto current origin/main.
• Replaced Date.now() resume nonce with crypto.randomUUID() plus fallback.
• Corrected the resumeNonce comment: it is included only in the opaque dashboard event channel id, not interpreted as backend resume state.
• Tightened the WebSocket public-bind gate to use the explicit allow_public startup opt-in instead of treating any non-loopback bound host as public by implication.
• Added tests for explicit non-loopback bind allowed with --insecure and rejected without the opt-in.
• Made HERMES_SKIP_NODE_BOOTSTRAP fail clearly when the available Node/npm runtime is unusable.
• Updated the PR body to the repo template.

Verification run locally:

scripts/run_tests.sh tests/hermes_cli/test_web_server.py::TestPtyWebSocket tests/hermes_cli/test_tui_resume_flow.py -q
40 passed

python -m py_compile hermes_cli/main.py hermes_cli/web_server.py
passed

cd web && npm run build
passed

I also ran cd web && npm run lint; it still fails on the existing repo lint baseline in unrelated files/rules, so I did not include lint as a passing gate for this PR.

[hehehe0803]

Follow-up pushed to address the failing CI test job on commit 2b9d2b2c1.

Root causes from the failed test log:

• tests/plugins/web/test_web_search_provider_plugins.py still asserted an exact seven-provider web plugin list, but the bundled web-xai provider is now discovered too. I converted that to an invariant-style core-provider subset check and added xAI capability / availability coverage instead of keeping a brittle count snapshot.
• tests/hermes_cli/test_update_hangup_protection.py could compare against a stale _UpdateOutputStream class object after another test reloads hermes_cli.main; the assertion now imports the current class at assertion time.

Local verification:

• scripts/run_tests.sh tests/plugins/web/test_web_search_provider_plugins.py tests/hermes_cli/test_update_hangup_protection.py -q → 65 passed

CI has restarted; most checks are already green, with the long test and arm build jobs still running at the time of this comment.

[hehehe0803]

Follow-up after the latest CI failure:

• Fixed the order-sensitive stdout/stderr mirror assertion in test_update_hangup_protection.py by asserting the stable _UpdateOutputStream wrapper protocol instead of Python class identity. This avoids failures when other CLI tests reload hermes_cli.main in the same worker.
• Fixed the Vercel sandbox snapshot persistence test to scope its snapshot store with set_hermes_home_override(...), matching code paths where a context-local Hermes home can override HERMES_HOME.
• Verified locally with:
  • scripts/run_tests.sh tests/hermes_cli/test_update_hangup_protection.py tests/tools/test_vercel_sandbox_environment.py tests/hermes_cli/test_web_server.py::TestPtyWebSocket tests/hermes_cli/test_tui_resume_flow.py -q --tb=short
  • Result: 72 passed
• GitHub Actions is now green on head 856135b80; the previously failing test job passed in 12m26s.

All checks are green; PR merge state is CLEAN.

[teknium1]

Thanks for documenting the explicit private-interface reproduction and adding targeted regression coverage. This is already implemented on current main.

• Automated hermes-sweeper review verified commit 234ac009376daba225525195afca96be8a82634c (fix(dashboard): allow insecure WS peers on explicit non-loopback binds (#35386)), which fixes the same explicit LAN/Tailscale bind case.
• hermes_cli/web_server.py:14265-14281 now permits peers for explicit non-loopback binds while preserving the loopback-only restriction for loopback binds.
• tests/hermes_cli/test_dashboard_auth_ws_auth.py:361-386 covers the explicit non-loopback peer path and retains the Host-header rejection guard.
• The subsequent dashboard hardening in hermes_cli/web_server.py:16920-16938 makes every non-loopback bind authenticated and treats --insecure as a no-op, so this branch's allow_public-based admission model is superseded.