Skip to content

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.

provider(llmtr): trim to the models the catalogue already has

knowhycodata722b6b096dcd
diff --git a/gateway/app.py b/gateway/app.py
index c7cd6739..367963d2 100644
--- a/gateway/app.py
+++ b/gateway/app.py
@@ -4140,22 +4140,30 @@ _VENDOR_MODELS: dict[str, dict[str, str]] = {
# tencent/hy3, nemotron-3-ultra is the 550b-a55b id), because one model arriving in the
# picker twice under two names is worse than a name that reads oddly.
#
- # Admitted from its own published catalogue (GET https://llmtr.com/v1/models, which needs no
- # key), read 2026-08-19, by four rules applied to all 245 published models, which leave the
- # 131 below:
+ # This table is deliberately the INTERSECTION with _MODEL_CATALOG rather than everything
+ # LLMTR serves: connecting the provider adds no name to any picker, it only makes models
+ # that are already there reachable through Turkey. Its published catalogue
+ # (GET https://llmtr.com/v1/models, which needs no key) was read and measured on
+ # 2026-08-19: of 245 published ids, 131 clear the agent-loop bar _MODEL_CATALOG states,
+ # and the 107 of those that only LLMTR serves — including the three Turkey-resident
+ # models, gemma-4, qwen3-6-35b and qwen3-5-4b — are proposed separately, once the
+ # capability test can check this catalogue the way it checks OpenRouter's. The rules
+ # they were measured by, so that follow-up can re-apply them rather than trust a number:
#
- # 1. Text out and `tools` in supported_parameters — the same agent-loop bar _MODEL_CATALOG
- # states. 81 fail it: embedders, image/video/speech models, and the small Turkish chat
- # models (trendyol-asure-12b, magibu-11b-v8, medgemma-4b) that serve no tools.
- # 2. No dated snapshot of an id already listed (qwen-plus-2025-07-14 beside qwen-plus). The
- # snapshot is the same model under a pinned name; two picker rows for it help nobody.
+ # 1. Text out and `tools` in supported_parameters — the same agent-loop bar
+ # _MODEL_CATALOG states. 81 fail it: embedders, image/video/speech models, and the
+ # small Turkish chat models (trendyol-asure-12b, magibu-11b-v8, medgemma-4b) that
+ # serve no tools.
+ # 2. No dated snapshot of an id already listed (qwen-plus-2025-07-14 beside qwen-plus).
+ # The snapshot is the same model under a pinned name; two picker rows for it help
+ # nobody.
# 3. The transport has to be the one the runner will actually use for that id. Both ways
# round, and 17 models fail it: gpt-5.4/5.2/5.1/5/o1/o3 and friends are published here
# for /v1/chat/completions only while hermes drives every gpt-5*/o[1-4] id over
# /v1/responses (_HERMES_RESPONSES_API_MODEL), and the grok ids are published for
- # /v1/responses only while hermes drives them over chat. Offering either is a picker row
- # that fails on send — the same broken promise _TOKENROUTER_NO_CHANNEL records.
- # 4. It has to answer. Every id below was called live on 2026-08-19 — max_tokens=4 on
+ # /v1/responses only while hermes drives them over chat. Offering either is a picker
+ # row that fails on send — the same broken promise _TOKENROUTER_NO_CHANNEL records.
+ # 4. It has to answer. Every id was called live on 2026-08-19 — max_tokens=4 on
# /v1/chat/completions, max_output_tokens=16 on /v1/responses — and seven published as
# available did not, twice: llmtr/muse-glimmer-30b-tr returns no bytes at all (45 s,
# 70 s and 120 s), publicai/apertus-8b-instruct 502s after about a minute,
@@ -4163,168 +4171,53 @@ _VENDOR_MODELS: dict[str, dict[str, str]] = {
# and publicai/apertus-v1.5-8b 400, and openai/gpt-oss-120b 402 while the gpt-oss-20b
# beside it answers. A catalogue entry is an advertisement; an answer is the product.
#
- # A ping is not a long agentic turn, so the harnesses on the test box ran real ones the same
- # day: Claude Code over /v1/messages on claude-sonnet-4.6, and hermes over
- # /v1/chat/completions on claude-haiku-4.5 and on the Turkey-hosted qwen3-6-35b, each
- # completing without substitution. /v1/responses — the transport codex drives — was
- # exercised as a direct call on gpt-5.6-luna rather than as a turn. A model here that turns
- # out to fail a real turn comes off this table the same way those seven did.
+ # A ping is not a long agentic turn, so the harnesses on the test box ran real ones, each
+ # completing without substitution: Claude Code over /v1/messages on claude-sonnet-4.6 and
+ # claude-opus-4.8, hermes over /v1/chat/completions on claude-haiku-4.5, and codex over
+ # /v1/responses on gpt-5.3-codex and gpt-5.5, both of which called the shell tool and came
+ # back with its real output. A model here that turns out to fail a real turn comes off this
+ # table the same way those seven did.
+ #
+ # One thing those runs turned up that is NOT this table's doing, recorded so the next person
+ # does not have to rediscover it: driven by CODEX, gpt-5.6-sol, gpt-5.6-terra and
+ # gpt-5.6-luna finish a turn but are never offered tools, so they cannot act — they answer
+ # that they have no shell and stop. This is the codex CLI, not the provider. Pointed at a
+ # local endpoint that records what it is sent, codex 0.148.0 emits the gpt-5.6-* request with
+ # NO `tools` and NO `instructions` field at all, while the same binary against the same
+ # endpoint sends 11 tools and a 21k-char instructions block for gpt-5.5. Nothing ever reaches
+ # the provider for it to drop. The family first tries wss://api.openai.com/v1/responses and
+ # its request carries x-codex-window-id, so it appears to expect the WebSocket session where
+ # tools and instructions are established on the connection; falling back to plain HTTPS
+ # against a configured provider, it sends neither. HERMES drives the same three ids over the
+ # same /v1/responses transport against the same base_url with tools intact. So it applies to
+ # every non-OpenAI provider, not this one: all three are already in _MODEL_CATALOG["codex"]
+ # and reachable through openrouter, tokenrouter and vercel. Left alone here rather than
+ # worked around in one vendor's table.
"llmtr": {
- # Turkey-resident, on LLMTR’s own infrastructure
- "gemma-4": "llmtr/gemma-4",
- "qwen3-6-35b": "llmtr/qwen3-6-35b",
- "qwen3-5-4b": "llmtr/qwen3-5-4b",
- # openai
- "gpt-5.6-sol": "openai/gpt-5.6-sol",
- "gpt-5.6-sol-pro": "openai/gpt-5.6-sol-pro",
- "gpt-5.6-terra": "openai/gpt-5.6-terra",
- "gpt-5.6-terra-pro": "openai/gpt-5.6-terra-pro",
- "gpt-5.6-luna": "openai/gpt-5.6-luna",
- "gpt-5.6-luna-pro": "openai/gpt-5.6-luna-pro",
- "gpt-5.5": "openai/gpt-5.5",
- "gpt-5.5-pro": "openai/gpt-5.5-pro",
- "gpt-4.1": "openai/gpt-4.1",
- "gpt-4o": "openai/gpt-4o",
- "gpt-5.3-codex": "openai/gpt-5.3-codex",
- "gpt-4.1-mini": "openai/gpt-4.1-mini",
- "gpt-4.1-nano": "openai/gpt-4.1-nano",
- "gpt-4o-mini": "openai/gpt-4o-mini",
- "gpt-oss-20b": "openai/gpt-oss-20b",
- "gpt-oss-safeguard-20b": "openai/gpt-oss-safeguard-20b",
- "gpt-4-turbo": "openai/gpt-4-turbo",
- "gpt-3.5-turbo": "openai/gpt-3.5-turbo",
- # anthropic
- "claude-fable-5": "anthropic/claude-fable-5",
- "claude-opus-5": "anthropic/claude-opus-5",
- "claude-opus-4.8": "anthropic/claude-opus-4.8",
- "claude-opus-4.7": "anthropic/claude-opus-4.7",
- "claude-opus-4.6": "anthropic/claude-opus-4.6",
- "claude-opus-4.5": "anthropic/claude-opus-4.5",
- "claude-sonnet-5": "anthropic/claude-sonnet-5",
- "claude-sonnet-4.6": "anthropic/claude-sonnet-4.6",
- "claude-sonnet-4.5": "anthropic/claude-sonnet-4.5",
- "claude-haiku-4.5": "anthropic/claude-haiku-4.5",
- # google
- "gemini-3.7-flash": "google/gemini-3.7-flash",
- "gemini-2.5-flash": "google/gemini-2.5-flash",
- "gemini-2.5-flash-lite": "google/gemini-2.5-flash-lite",
- "gemini-2.5-pro": "google/gemini-2.5-pro",
- "gemini-3-flash-preview": "google/gemini-3-flash-preview",
- "gemini-3.1-flash-lite": "google/gemini-3.1-flash-lite",
- "gemini-3.1-pro-preview": "google/gemini-3.1-pro-preview",
- "gemini-3.1-pro-preview-customtools": "google/gemini-3.1-pro-preview-customtools",
- "gemini-3.5-flash": "google/gemini-3.5-flash",
- "gemini-3.5-flash-lite": "google/gemini-3.5-flash-lite",
- "gemini-3.6-flash": "google/gemini-3.6-flash",
- # deepseek
- "deepseek-v4-flash": "deepseek/deepseek-v4-flash",
- "deepseek-v4-pro": "deepseek/deepseek-v4-pro",
- "deepseek-chat": "deepseek/deepseek-chat",
- "deepseek-reasoner": "deepseek/deepseek-reasoner",
- "deepseek-v4-pro-0813": "deepseek/deepseek-v4-pro-0813",
- # qwen
- "qwen3.8-max": "qwen/qwen3.8-max",
- "qwen3.7-max": "qwen/qwen3.7-max",
- "qwen3.8-2.4t-a95b": "qwen/qwen3.8-2.4t-a95b",
- "qwen3.7-plus": "qwen/qwen3.7-plus",
- "qwen3.6-flash": "qwen/qwen3.6-flash",
- "qwen-plus": "qwen/qwen-plus",
- "qwen-plus-character-ja": "qwen/qwen-plus-character-ja",
- "qwen-plus-latest": "qwen/qwen-plus-latest",
- "qwen3-max": "qwen/qwen3-max",
- "qwen3-max-preview": "qwen/qwen3-max-preview",
- "qwen3.6-plus": "qwen/qwen3.6-plus",
- "qwen-mt-plus": "qwen/qwen-mt-plus",
- "qwen3-coder-flash": "qwen/qwen3-coder-flash",
- "qwen3-coder-plus": "qwen/qwen3-coder-plus",
- "qwen3-vl-plus": "qwen/qwen3-vl-plus",
- "qwen3.5-plus": "qwen/qwen3.5-plus",
- "qwen3.6-27b-free": "qwen/qwen3.6-27b-free",
- # zai
- "glm-5.3": "zai/glm-5.3",
- "glm-5.2": "zai/glm-5.2",
- "glm-5.1": "zai/glm-5.1",
- "glm-5": "zai/glm-5",
- "glm-5-turbo": "zai/glm-5-turbo",
- "glm-5v-turbo": "zai/glm-5v-turbo",
- "glm-4.7": "zai/glm-4.7",
- "glm-4.7-flashx": "zai/glm-4.7-flashx",
- "glm-4.6": "zai/glm-4.6",
- "glm-4.5": "zai/glm-4.5",
- "glm-4.5-x": "zai/glm-4.5-x",
- "glm-4.5-air": "zai/glm-4.5-air",
- "glm-4.5-airx": "zai/glm-4.5-airx",
- "glm-4-32b-0414-128k": "zai/glm-4-32b-0414-128k",
- "glm-4.6v": "zai/glm-4.6v",
- "glm-4.6v-flashx": "zai/glm-4.6v-flashx",
- "glm-4.5v": "zai/glm-4.5v",
- # moonshot
- "kimi-k3": "moonshot/kimi-k3",
- "kimi-k2.7-code": "moonshot/kimi-k2.7-code",
- "kimi-k2.6": "moonshot/kimi-k2.6",
- "kimi-k2.5": "moonshot/kimi-k2.5",
- # mistral
- "mistral-large-latest": "mistral/mistral-large-latest",
- "mistral-small-latest": "mistral/mistral-small-latest",
- "mistral-medium-latest": "mistral/mistral-medium-latest",
- "devstral-medium-latest": "mistral/devstral-medium-latest",
- "magistral-medium-latest": "mistral/magistral-medium-latest",
- "magistral-small-latest": "mistral/magistral-small-latest",
- "devstral-small-latest": "mistral/devstral-small-latest",
- "codestral-latest": "mistral/codestral-latest",
- "pixtral-12b": "mistral/pixtral-12b",
- "ministral-14b-latest": "mistral/ministral-14b-latest",
- "ministral-8b-latest": "mistral/ministral-8b-latest",
- "ministral-3b-latest": "mistral/ministral-3b-latest",
- "voxtral-small-latest": "mistral/voxtral-small-latest",
- # meta
- "muse-spark-1.2": "meta/muse-spark-1.2",
- "muse-glimmer-30b": "meta/muse-glimmer-30b",
- "muse-spark-1.1": "meta/muse-spark-1.1",
- "muse-spark-1.2-contributor": "meta/muse-spark-1.2-contributor",
- "llama-3.3-70b-instruct-12k": "meta/llama-3.3-70b-instruct-12k",
- # nvidia
- "nemotron-3-ultra": "nvidia/nemotron-3-ultra-550b-a55b",
- "nemotron-3-super-120b-a12b": "nvidia/nemotron-3-super-120b-a12b",
- "nemotron-3-nano-30b-a3b": "nvidia/nemotron-3-nano-30b-a3b",
- "nemotron-nano-12b-v2-vl": "nvidia/nemotron-nano-12b-v2-vl",
- # stepfun
- "step-3.7-flash": "stepfun/step-3.7-flash",
- "step-3.5-flash": "stepfun/step-3.5-flash",
- "step-3.5-flash-2603": "stepfun/step-3.5-flash-2603",
- # minimax
- "minimax-m3": "minimax/minimax-m3",
- # inclusionai
- "ling-3.0-flash": "inclusionai/ling-3.0-flash",
- # tencent
- "hunyuan-3": "tencent/hy3",
- # aion-labs
- "aion-3.0": "aion-labs/aion-3.0",
- "aion-3.0-mini": "aion-labs/aion-3.0-mini",
- "aion-2.0": "aion-labs/aion-2.0",
- "aion-rp-llama-3.1-8b": "aion-labs/aion-rp-llama-3.1-8b",
- # publicai
- "apertus-v1.5-70b": "publicai/apertus-v1.5-70b",
- "apertus-70b-instruct": "publicai/apertus-70b-instruct",
- # thinkingmachines
- "inkling": "thinkingmachines/inkling",
- "inkling-small": "thinkingmachines/inkling-small",
- # mimo
- "mimo-v2.5-pro": "mimo/mimo-v2.5-pro",
- "mimo-v2.5": "mimo/mimo-v2.5",
- # kwaikat
- "kat-coder-pro-v2.5": "kwaikat/kat-coder-pro-v2.5",
- "kat-coder-air-v2.5": "kwaikat/kat-coder-air-v2.5",
- # upstage
- "solar-pro4": "upstage/solar-pro4",
- "solar-pro3": "upstage/solar-pro3",
- "solar-pro2": "upstage/solar-pro2",
- # poolside
- "laguna-xs-2.1": "poolside/laguna-xs-2.1",
- # gemma
- "gemma-4-26b-a4b-it": "gemma/gemma-4-26b-a4b-it",
- # dots-studio
- "dots-3-note-preview": "dots-studio/dots-3-note-preview",
+ "gpt-5.6-sol": "openai/gpt-5.6-sol",
+ "gpt-5.6-terra": "openai/gpt-5.6-terra",
+ "gpt-5.6-luna": "openai/gpt-5.6-luna",
+ "gpt-5.5": "openai/gpt-5.5",
+ "gpt-5.3-codex": "openai/gpt-5.3-codex",
+ "claude-opus-5": "anthropic/claude-opus-5",
+ "claude-fable-5": "anthropic/claude-fable-5",
+ "claude-opus-4.8": "anthropic/claude-opus-4.8",
+ "claude-sonnet-5": "anthropic/claude-sonnet-5",
+ "claude-opus-4.7": "anthropic/claude-opus-4.7",
+ "claude-sonnet-4.6": "anthropic/claude-sonnet-4.6",
+ "claude-haiku-4.5": "anthropic/claude-haiku-4.5",
+ "gemini-3.6-flash": "google/gemini-3.6-flash",
+ "deepseek-v4-pro": "deepseek/deepseek-v4-pro",
+ "deepseek-v4-flash": "deepseek/deepseek-v4-flash",
+ "kimi-k3": "moonshot/kimi-k3",
+ "qwen3.7-max": "qwen/qwen3.7-max",
+ "qwen3.8-max": "qwen/qwen3.8-max",
+ "kimi-k2.7-code": "moonshot/kimi-k2.7-code",
+ "step-3.7-flash": "stepfun/step-3.7-flash",
+ "minimax-m3": "minimax/minimax-m3",
+ "nemotron-3-ultra": "nvidia/nemotron-3-ultra-550b-a55b",
+ "hunyuan-3": "tencent/hy3",
+ "ling-3.0-flash": "inclusionai/ling-3.0-flash",
},
}
@@ -4382,7 +4275,6 @@ _VERCEL_RESLUG = {
_VENDOR_MODELS["vercel"] = {c: _VERCEL_RESLUG.get(c, v)
for c, v in _VENDOR_MODELS["openrouter"].items()}
-
# The chain path (_map_model) maps aggregator ids from the same table.
_AGGREGATOR_SLUGS = _VENDOR_MODELS["openrouter"]
@@ -4464,12 +4356,7 @@ _MODEL_CATALOG: dict[str, dict] = {
"gpt-5.4", "gpt-5.4-mini", "gpt-5.2",
# Codex-optimized line (separate from the general one; 5.3-codex is
# OpenAI's most capable agentic coding model, there is no 5.6-codex).
- "gpt-5.3-codex",
- # Responses-API `pro` variants, served by LLMTR: the same base
- # models above with reasoning.mode=pro defaulted on, so they need
- # no separate wiring.
- "gpt-5.6-sol-pro", "gpt-5.6-terra-pro", "gpt-5.6-luna-pro",
- "gpt-5.5-pro"]},
+ "gpt-5.3-codex"]},
# claude-opus-5 is now wired in both _ANTHROPIC_CLAUDE and _BEDROCK_CLAUDE (ids read off
# Anthropic's models overview and AWS's own model card), so it routes directly instead of
# falling through to the unmapped default it used at launch.
@@ -4494,45 +4381,7 @@ _MODEL_CATALOG: dict[str, dict] = {
"qwen3.7-max", "qwen3.8-max", "kimi-k2.7-code",
"mistral-medium-3.5", "step-3.7-flash", "minimax-m3",
"nemotron-3-ultra", "hunyuan-3", "ling-3.0-flash",
- "qwen3.7-flash",
- # Everything else the LLMTR gateway serves that clears the
- # agent-loop bar; _VENDOR_MODELS["llmtr"] carries the rule this
- # set was derived by and what it excludes. Most are LLMTR-only
- # today, so on an instance without that integration they render
- # as unavailable rather than disappearing.
- "gemma-4", "qwen3-6-35b", "qwen3-5-4b",
- "gpt-5.6-sol-pro", "gpt-5.6-terra-pro", "gpt-5.6-luna-pro", "gpt-5.5-pro",
- "gpt-4.1", "gpt-4o", "gemini-3.7-flash", "gemini-2.5-flash",
- "gemini-2.5-flash-lite", "gemini-2.5-pro", "gemini-3-flash-preview",
- "gemini-3.1-flash-lite", "gemini-3.1-pro-preview",
- "gemini-3.1-pro-preview-customtools", "gemini-3.5-flash",
- "gemini-3.5-flash-lite", "claude-opus-4.6", "claude-opus-4.5",
- "claude-sonnet-4.5", "qwen3.8-2.4t-a95b", "qwen3.7-plus", "qwen3.6-flash",
- "qwen-plus", "qwen-plus-character-ja", "qwen-plus-latest", "qwen3-max",
- "qwen3-max-preview", "qwen3.6-plus", "glm-5.3", "glm-5.2", "glm-5.1",
- "glm-5", "glm-5-turbo", "glm-5v-turbo", "mistral-large-latest",
- "mistral-small-latest", "mistral-medium-latest", "devstral-medium-latest",
- "inkling", "inkling-small", "mimo-v2.5-pro", "mimo-v2.5",
- "apertus-v1.5-70b", "muse-spark-1.2",
- "muse-glimmer-30b", "kimi-k2.6", "kimi-k2.5", "laguna-xs-2.1", "aion-3.0",
- "aion-3.0-mini", "solar-pro4", "gemma-4-26b-a4b-it",
- "gpt-4.1-mini", "gpt-4.1-nano", "gpt-4o-mini",
- "gpt-oss-20b", "gpt-oss-safeguard-20b", "gpt-4-turbo", "gpt-3.5-turbo",
- "qwen-mt-plus", "qwen3-coder-flash",
- "qwen3-coder-plus", "qwen3-vl-plus", "qwen3.5-plus", "qwen3.6-27b-free",
- "glm-4.7", "glm-4.7-flashx", "glm-4.6", "glm-4.5", "glm-4.5-x",
- "glm-4.5-air", "glm-4.5-airx", "glm-4-32b-0414-128k", "glm-4.6v",
- "glm-4.6v-flashx", "glm-4.5v", "deepseek-chat", "deepseek-reasoner",
- "deepseek-v4-pro-0813", "magistral-medium-latest",
- "magistral-small-latest", "devstral-small-latest", "codestral-latest",
- "pixtral-12b", "ministral-14b-latest", "ministral-8b-latest",
- "ministral-3b-latest", "voxtral-small-latest",
- "apertus-70b-instruct", "muse-spark-1.1", "muse-spark-1.2-contributor",
- "llama-3.3-70b-instruct-12k", "nemotron-3-super-120b-a12b",
- "nemotron-3-nano-30b-a3b", "nemotron-nano-12b-v2-vl", "step-3.5-flash",
- "step-3.5-flash-2603", "aion-2.0", "aion-rp-llama-3.1-8b",
- "kat-coder-pro-v2.5", "kat-coder-air-v2.5", "solar-pro3", "solar-pro2",
- "dots-3-note-preview"]},
+ "qwen3.7-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