Building the QWEN3.6 - Codex Bridge Furthe + Kindergarten Harness Reality Check

I got a bit further with my harness for running Qwen 3.6 model on Codex. While testing, analyzing, and building the harness, I evolved TBG(O)llama-swap into a full forensic UI bridge and LLM analytics tool where every harness finding, modification, correction, tool call, reasoning step, and execution flow is fully visible. This level of transparency was necessary to identify the behavioral differences between native OpenAI models and Qwen 3.6, and to fine-tune the harness accordingly. The video shows a full Codex run on Qwen 3.6 running on a single NVIDIA GeForce RTX 5090. (Codex in VS Code -> tbg(o)llama-swap -> llama.cpp with qwen 3.6 27B) The ongoing work can be checked here [https://github.com/Ltamann/tbg-ollama-swap-prompt-optimizer/tree/qwen3.6](https://github.com/Ltamann/tbg-ollama-swap-prompt-optimizer/tree/qwen3.6) ,[First post](https://www.patreon.com/posts/building-bridge-157050652) , [second post](https://www.patreon.com/posts/building-bridge-158134849?utm_medium=clipboard_copy&utm_source=copyLink&utm_campaign=postshare_creator&utm_content=join_link) Here’s the clearest current status. **Working** * `apply_patch` * `apply_patch` create/update/delete flow * `create_file` requires non-empty `diff` or `content` * `update_file` requires non-empty `diff` or `content` * `delete_file` works without `diff` * `shell` * `web_search` * `web_search` using TBG(O)llama-swap built-in web search * `file_search` * `view_image` * `request_user_input` * `update_plan` * `spawn_agent` * `wait_agent` * `send_input` * `resume_agent` * `close_agent` * `supports_search_tool` catalog inconsistency * `agent_send_input_roundtrip` * `agent_subagent_same_model` * `shell_patch_verify_sequence` * `web_research_then_notes` * `plan_act_switch_impl` * `multi_web_patch_verify` * `skill_create_and_use_local` * `workspace_summary_then_plan` * `skill_read_local` * `direct_plan_no_web` * `web_research_then_plan` * `file_search_then_patch` * `view_image_then_report` * invalid `apply_patch` retry exhaustion no longer finalizes with fake progress prose * safer recovery branch after broken `apply_patch` * false patch-intent/path-hint extraction from instructions * reconnect bug caused by unhealthy or duplicate upstream adoption * long delayed `502` timeout path shortened and improved * native-vs-local contrast harness: * `init` * `compare` * per-scenario `comparison.json` * top-level `comparison_summary.json` * tool-surface diff * item-type diff * stream/completion diff * final visible text diff * grouped UX-summary diff **Implemented in the Bridge Contract** * stricter separation of: * visible assistant text * tool call items * tool outputs * file/code artifacts * explicit continuation-state handling for: * research flow * write-pending flow * verification flow * final-answer handoff **Fixed Enough To Work, But Still Not Native-Perfect** * grouped searches * grouped tool calls * grouped file changes * collapsible internal history These areas are significantly improved in both the UI and harness, but I would still describe them as *partially aligned*, not fully native-identical yet. **Fixed** * `mcp__playwright__browser_navigate` * `mcp__playwright__browser_snapshot` * `mcp__playwright__browser_click` * `mcp__playwright__browser_evaluate` * `mcp__playwright__browser_resize` * `mcp__playwright__browser_take_screenshot` Important nuance: * llama-swap now preserves and exposes these much more accurately * however, the WSL Codex router still rejects Playwright leaf calls as unsupported in this surface * this is now tracked as a known limitation, not an active llama-swap bridge bug **Still Not Fully Closed / Needs More Work** * full native-style grouped worker UX parity * some remaining model-quality quirks during long multi-step runs * continuation/reporting polish around malformed reasoning/text splits