HarnessRouter Community Edition: the self-hosted, Apache-2.0 edition of the unified interface for agent harnesses. Run Codex, Claude Code, Hermes, PI, DSH, and more through one API, with sessions, streaming, files, cancellation, and failure handling. Implements the Unified Harness Protocol (UHP), an open standard. Your keys, your infrastructure.
Python agent-apiagent-frameworkagent-harnessai-agentsai-infrastructurecoding-agentconformance-testingdeepseek-harnessharness-engineeringhermesllmopen-protocolopen-sourceopen-standardpi-coding-agentself-hosteduhpunified-apiunified-harness-protocolunified-interface
fix(models): restore nine chat models I removed on a bad reading of "not an LLM"
richard-epsillaa26dbe79bcbb
diff --git a/gateway/app.py b/gateway/app.pyindex cfa324ce..e3a42ac8 100644--- a/gateway/app.py+++ b/gateway/app.py@@ -3488,6 +3488,15 @@ _VENDOR_MODELS: dict[str, dict[str, str]] = {"kimi-k3": "moonshotai/kimi-k3","glm-5.2": "z-ai/glm-5.2","qwen3.7-max": "qwen/qwen3.7-max",+ "qwen3.8-max": "qwen/qwen3.8-max",+ "qwen3.7-flash": "qwen/qwen3.7-flash",+ "kimi-k2.7-code": "moonshotai/kimi-k2.7-code",+ "minimax-m3": "minimax/minimax-m3",+ "nemotron-3-ultra": "nvidia/nemotron-3-ultra-550b-a55b",+ "mistral-medium-3.5": "mistralai/mistral-medium-3-5",+ "hunyuan-3": "tencent/hy3",+ "step-3.7-flash": "stepfun/step-3.7-flash",+ "ling-3.0-flash": "inclusionai/ling-3.0-flash",},}# TokenRouter is OpenRouter-slug compatible — one table, not a second copy to drift.@@ -3545,12 +3554,17 @@ def _map_model(conn: dict, friendly: str) -> str | None:# the harness's backend family is NOT run — the harness's authorized fallback (its default) runs# instead, and the substitution is recorded in the response's run metadata so it is auditable._MODEL_CATALOG: dict[str, dict] = {- # THE BAR FOR ADDING A MODEL: a real turn on the harness that will run it, completed against- # the live provider. Not "the vendor lists it", not "the aggregator says it supports tools" —- # both are necessary and neither is sufficient. Nine models were once added on that metadata- # alone; every one technically completed a trivial turn, and none of them belonged in a picker- # for driving a coding agent. Metadata tells you a model COULD be addressed. Only a turn tells- # you it should be offered.+ # THE BAR FOR ADDING A MODEL, both halves required:+ # 1. It is a chat model the agent loop can drive — text in, text out, and tool calling.+ # Image/video generators, embedders and safety classifiers are not candidates however+ # capable: picking one produces a task that dies at the first tool call. Vision input is+ # welcome, so a VLM qualifies; a model that only EMITS pixels does not. Enforced by+ # tests/test_model_catalog_capabilities.py against the live capability data.+ # 2. A real turn on the harness that will run it, completed against the live provider.+ # Neither half substitutes for the other: metadata says a model CAN be addressed, a turn says+ # the wiring actually reaches it. Do not trim this list on taste — "too many open-weight+ # models" is a UI problem, and removing a working chat model to solve it takes a model the+ # user picked out of their hands.## Probe-verified against the live provider before listing (2026-07-19:# gpt-5.6 sol/terra/luna + gpt-5.2 deployed on the Azure resource and probed OK;@@ -3582,7 +3596,10 @@ _MODEL_CATALOG: dict[str, dict] = {# integrations (2026-07-22: each probe-verified through the hermes# CLI on the TokenRouter connection)"gemini-3.6-flash", "deepseek-v4-pro", "deepseek-v4-flash", "kimi-k3",- "glm-5.2", "qwen3.7-max"]},+ "glm-5.2", "qwen3.7-max", "qwen3.8-max", "qwen3.7-flash",+ "kimi-k2.7-code", "minimax-m3", "nemotron-3-ultra",+ "mistral-medium-3.5", "hunyuan-3", "step-3.7-flash",+ "ling-3.0-flash"]},}_BARE_MODELS = {"", "claude", "codex", "anthropic", "bedrock", "openai", "hermes"}# Union of BOTH tables' values — a dict merge would drop the bedrock ids (shared keys, anthropic