fix(tui): persist model switch marker as user message

[lucas-nicolas]

What does this PR do?

Fixes a gateway/TUI model-switch bug where _append_model_switch_marker() persisted an informational model-switch marker as role: "system" after prior conversation turns.

Strict OpenAI-compatible providers such as vLLM reject message arrays with system messages outside the beginning of the conversation, returning:

HTTP 400: System message must be at the beginning.

The marker is runtime metadata for the next model, not a privileged instruction. This PR persists it as a user message instead, so it can safely appear mid-history and later be merged with adjacent user messages by the existing sanitizer.

Related Issue

Fixes #48338

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

• tui_gateway/server.py
  • Persist model-switch markers as role: "user" instead of role: "system" in both in-memory history and SQLite session persistence.
  • Added a comment explaining why strict OpenAI-compatible providers require this.
• tests/tui_gateway/test_model_switch_marker.py
  • Added a regression test that verifies model-switch markers are appended and persisted as user messages, not mid-history system messages.

How to Test

1. Run the focused regression:

   scripts/run_tests.sh tests/tui_gateway/test_model_switch_marker.py tests/gateway/test_model_switch_persistence.py

2. Run the relevant gateway/TUI slice:

   scripts/run_tests.sh tests/tui_gateway/ tests/gateway/test_model_switch_persistence.py

3. Reproduce against a strict provider by sending messages shaped like:

   [
       {"role": "system", "content": "..."},
       {"role": "user", "content": "hello"},
       {"role": "assistant", "content": "hi"},
       {"role": "system", "content": "[System: The active model ...]"},
       {"role": "user", "content": "say hi again"},
   ]

   Before this fix, vLLM returns HTTP 400: System message must be at the beginning. With this change, the gateway no longer creates that mid-history system role.

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

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

Test Results

Passed:

scripts/run_tests.sh tests/tui_gateway/test_model_switch_marker.py tests/gateway/test_model_switch_persistence.py
# 2 files, 10 tests passed, 0 failed

scripts/run_tests.sh tests/tui_gateway/ tests/gateway/test_model_switch_persistence.py
# 13 files, 143 tests passed, 0 failed

git diff --check
# passed

python3 scripts/check-windows-footguns.py --diff origin/main
# passed; no Windows footguns found in changed diff

Full-suite note:

I attempted a full scripts/run_tests.sh run after installing [all,dev] per CONTRIBUTING. The full suite is very large; I stopped it after the relevant suites had already passed and the run had progressed past 22k passing tests with several failures in unrelated test files. I am not claiming full-suite green for this PR.

Screenshots / Logs

Relevant failure reproduced before the fix:

HTTP 400: {"error":{"message":"System message must be at the beginning.","type":"BadRequestError","param":null,"code":400}}

[alt-glitch]

Duplicate of #48346 — both fix _append_model_switch_marker() in tui_gateway/server.py by persisting the model-switch marker as role: user instead of role: system, so it no longer lands mid-history and triggers HTTP 400: System message must be at the beginning on strict providers (vLLM/Qwen). #48346 is the earlier open PR for the same fix. Both address #48338; related to the broader symptom anchor #20866.

[lucas-nicolas]

Duplicate of #48346 — both fix _append_model_switch_marker() in tui_gateway/server.py by persisting the model-switch marker as role: user instead of role: system, so it no longer lands mid-history and triggers HTTP 400: System message must be at the beginning on strict providers (vLLM/Qwen). #48346 is the earlier open PR for the same fix. Both address #48338; related to the broader symptom anchor #20866.

Sorry did not see the PR at the time I started trying to fix this issue. Was hoping to make my first contribution with this PR :). There is an assertion failing on the previously opened PR that needs to be changed. changed it here and left a comment on the other PR.

Have a nice day and thank you for working on Hermes Agent

[alaamohanad169-ship-it]

Note: This is NOT a duplicate of the prior closed PRs (#48346, #48346, #48338, #20866). The earlier PRs were closed because they were stale (far behind upstream main). This version has been cherry-picked onto the latest origin/main with a clean diff containing only the fix changes. No conflicts. CI green. Ready for review.

[lucas-nicolas]

Rebased this branch onto current upstream/main and re-ran the focused checks:

env -u VIRTUAL_ENV scripts/run_tests.sh tests/tui_gateway/test_model_switch_marker.py tests/test_tui_gateway_server.py
# 2 files, 279 tests passed, 0 failed

git diff --check upstream/main..HEAD
# passed

python3 scripts/check-windows-footguns.py --diff upstream/main
# passed

[teknium1]

Merged via PR #54210 (commit 61622bb). Thanks for the report (#48338) and the fix — your explanatory comment was folded into the salvaged change and you are credited via Co-authored-by. The cleanest minimal version came from the earliest duplicate (#48346); both of you are on the commit.