fix(tui): keep model switch markers out of API history

[markoub]

What does this PR do?

Fixes TUI model-switch metadata so it cannot become a mid-conversation system message in provider API history.

When /model changes the active runtime, Hermes already refreshes and persists the session system prompt. This PR keeps the additional marker as transcript-only session_meta and filters both new session_meta rows and legacy model-switch system markers before handing history to AIAgent.

The merge path then reconciles the agent's append-style result back onto the raw TUI transcript so transcript-only rows are preserved locally without being sent to the provider.

This covers new switches and already-persisted markers from older versions that could trigger strict OpenAI-compatible backends such as vLLM/Qwen to return HTTP 400.

Related Issue

Fixes #48338

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• Tests (adding or improving test coverage)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Documentation update
• Performance improvement
• Code refactoring

Changes Made

• tui_gateway/server.py: store model-switch markers as session_meta, not system.
• tui_gateway/server.py: filter transcript-only session_meta rows and legacy TUI model-switch system markers before prompt.submit calls AIAgent.
• tui_gateway/server.py: merge append-style agent results back onto the raw TUI transcript so filtered local metadata is not lost.
• tests/test_tui_gateway_server.py: cover new session_meta markers, legacy system model-switch markers, preserved ordinary system messages, and transcript merge behavior.

How to Test

1. Switch model in an active TUI session, then send another prompt through a strict OpenAI-compatible backend such as vLLM/Qwen.
2. Confirm the provider request contains the rebuilt leading system prompt but no model-switch marker as a later system message.
3. Run:
   • uv run scripts/run_tests.sh tests/test_tui_gateway_server.py -- -k 'model_switch or config_set_model or preserves_transcript_meta' --tb=long
   • uv run scripts/run_tests.sh tests/run_agent/test_session_meta_filtering.py -- --tb=long
   • git diff --check && uv run ruff check tui_gateway/server.py tests/test_tui_gateway_server.py

Checklist

Code

• I've read the Contributing Guide and this PR follows the project's coding standards
• My commit messages follow the conventional commits format
• I searched for existing PRs/issues to avoid duplicates
• My PR contains only changes related to the described issue
• I've run pytest tests/ -q and all tests pass
• I've added tests for new functionality or bug fixes
• I've tested on my platform: macOS 15 / Python 3.11 via uv run scripts/run_tests.sh

Documentation

• I've updated relevant documentation where needed, or this change does not require documentation updates
• I've updated cli-config.yaml.example if adding config options, or this change does not add config options
• I've updated CONTRIBUTING.md or AGENTS.md if changing development workflow, or this change does not affect development workflow
• I've considered cross-platform impact: no process, shell, or path behavior changed
• I've updated tool descriptions/schemas if adding or changing tools, or this change does not affect tool schemas

For New Skills

N/A

Screenshots / Logs

uv run scripts/run_tests.sh tests/test_tui_gateway_server.py -- -k 'model_switch or config_set_model or preserves_transcript_meta' --tb=long
passed 12 tests.

uv run scripts/run_tests.sh tests/run_agent/test_session_meta_filtering.py -- --tb=long
passed 5 tests.

git diff --check && uv run ruff check tui_gateway/server.py tests/test_tui_gateway_server.py
passed.

Note: I also ran the full tests/test_tui_gateway_server.py file. It passed 277/278 tests; the only failure was the unrelated local browser-launch assertion test_browser_manage_connect_default_local_reports_launch_hint, which also fails when run alone in this checkout.

[teknium1]

Thanks for separating transcript-only model-switch metadata from provider-visible history. The current-main premise remains real: tui_gateway/server.py:2448 persists a switch marker as a user message, and _run_prompt_submit passes raw history to AIAgent at tui_gateway/server.py:9048.

Problems

• The new legacy filter at tui_gateway/server.py:4663 only removes markers whose role is system. Current main changed new markers to user in commit 61622bb (tui_gateway/server.py:2444-2470), so already-persisted user-role model-switch markers remain in the provider history after this change.

Suggested changes

• Recognize the exact model-switch marker independent of its stored role, then filter legacy user and system rows plus new session_meta rows. Add a test covering a persisted user-role marker without filtering ordinary user turns.

Automated hermes-sweeper review.

[teknium1]

Current main has persisted model-switch markers as role="user" since 61622bb. This condition removes only older system rows, so those user-role markers still reach conversation_history; recognize the exact marker content independently of role.