Add Qwen3.6-VL support via an extensible per-model-family adapter layer

[jiayuww]

Summary

Adds support for the Qwen3.6-VL family (Qwen3.6-27B, Qwen3.6-35B-A3B) and, to do it cleanly, introduces a small per-model-family adapter layer so each model family owns its own response parser and coordinate conversion. Existing configs are unaffected.

Why

Qwen-VL models emit a <tool_call>{...}</tool_call> JSON grammar with coordinates in a 0-1000 space — not UI-TARS's <action>...</action> text with raw pixels. The existing flexible_parser can't read that grammar, and there was no seam to plug in a different one without scattering model-specific logic through utils.py. This PR adds that seam.

What changed

New src/open_apps/agent/adapters/ package

• base.py — Adapter contract: parse(response, viewport) -> {action, displayed_action, think}.
• uitars.py — default family; delegates to the existing flexible_parser (unchanged behavior).
• qwen3vl.py — self-contained <tool_call> JSON parser + 0-1000→viewport coordinate rescale.
• __init__.py — REGISTRY + get_adapter(); select a family with adapter: in the agent yaml (defaults to uitars).

Wiring (additive, back-compatible)

• vLLM_agent.py — new adapter and prompt_sections config fields, threaded to the prompt builder.
• vLLM_prompt.py — _parse_answer dispatches to the selected adapter (falls back to flexible_parser); the user message is split into composable sections so a config can render a minimal message (Qwen keeps its <tools> schema in the system prompt, which it was trained to expect); viewport is derived from the screenshot for rescaling. Default configs render byte-for-byte as before.
• utils.py — JPEG→PNG screenshots (some sglang VLM processors crash on JPEG); empty-content retry guard (empty/None content no longer crashes a trial); hotkey normalization (ctrl a → Control+a).

Configs

• config/agent/Qwen3.6-VL.yaml — the family (adapter + <tools> system prompt + minimal prompt_sections + 5-action set).
• config/agent/Qwen3.6-27B.yaml, config/agent/Qwen3.6-35B-A3B.yaml — thin per-model configs inheriting the family (like UI-TARS-1.5-7B.yaml).

Tests

• tests/test_adapters.py — registry, every Qwen action + robustness cases, UI-TARS regression.

Extensibility

Adding a new model family = one adapter module + one REGISTRY line + one config yaml. No changes to the agent/prompt/parse plumbing. If a family needs a new user-message section, add a _render_* method + a _SECTION_RENDERERS entry.

Testing

• tests/test_adapters.py: 22 passing; full suite green except one pre-existing browser e2e flake unrelated to this change.
• Smoke-tested against live sglang servers on add_call_mom_to_my_todo: Qwen3.6-27B 4/4, Qwen3.6-35B-A3B 2/4 (the 2 misses are model grounding, not parsing), with 0 parse failures across 8 runs and correct 0-1000→pixel rescaling verified on real output.

[marksibrahim]

I quite like the idea of separating action parsers into separate module for each model. Nice idea

[marksibrahim]

nice to include this an argument. I'm just going to rename this arg and corresponding module to action_parser since I find that more intuitive, unless you have any objections.

[marksibrahim]

One other change, I'm appending -computer-use to make it clear these are pure computer-use agents