feat: add token market provider

providers/tokensmarket/models/MiniMax-M2.1.toml

@@ -0,0 +1,11 @@
+# Alibaba Cloud (DashScope) channel. For MiniMax models the adapter sets
+# `thinking.type` to adaptive/disabled from the effort value and does not forward
+# an effort ladder upstream, so the surface is a plain toggle.
+base_model = "minimax/MiniMax-M2.1"
+reasoning_options = [{ type = "toggle" }]
+
+[cost]
+input = 0.248
+output = 0.992
+cache_read = 0.024
+cache_write = 0.312

providers/tokensmarket/models/MiniMax-M2.5.toml

@@ -0,0 +1,11 @@
+# Alibaba Cloud (DashScope) channel. For MiniMax models the adapter sets
+# `thinking.type` to adaptive/disabled from the effort value and does not forward
+# an effort ladder upstream, so the surface is a plain toggle.
+base_model = "minimax/MiniMax-M2.5"
+reasoning_options = [{ type = "toggle" }]
+
+[cost]
+input = 0.248
+output = 0.992
+cache_read = 0.024
+cache_write = 0.312

providers/tokensmarket/models/claude-fable-5.toml

@@ -0,0 +1,19 @@
+# Claude Hub channel: raw passthrough to Anthropic /v1/messages, so the upstream
+# reasoning surface survives unchanged. Modalities follow Token Market's own
+# catalog, which lists text and image input only.
+base_model = "anthropic/claude-fable-5"
+structured_output = true
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "xhigh", "max"]
+
+[cost]
+input = 2.5
+output = 12.5
+cache_read = 0.25
+cache_write = 3.125
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

providers/tokensmarket/models/claude-haiku-4-5.toml

@@ -0,0 +1,19 @@
+# Claude Hub channel: raw passthrough to Anthropic /v1/messages, so the upstream
+# reasoning surface survives unchanged. Modalities follow Token Market's own
+# catalog, which lists text and image input only.
+base_model = "anthropic/claude-haiku-4-5"
+structured_output = true
+
+[[reasoning_options]]
+type = "budget_tokens"
+min = 1_024
+
+[cost]
+input = 0.25
+output = 1.25
+cache_read = 0.025
+cache_write = 0.3125
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

providers/tokensmarket/models/claude-opus-4-8.toml

@@ -0,0 +1,19 @@
+# Claude Hub channel: raw passthrough to Anthropic /v1/messages, so the upstream
+# reasoning surface survives unchanged. Modalities follow Token Market's own
+# catalog, which lists text and image input only.
+base_model = "anthropic/claude-opus-4-8"
+structured_output = true
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "xhigh", "max"]
+
+[cost]
+input = 1.25
+output = 6.25
+cache_read = 0.125
+cache_write = 1.5625
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

providers/tokensmarket/models/claude-opus-5.toml

@@ -0,0 +1,21 @@
+# Served over the Claude Hub channel, which is a raw passthrough to Anthropic
+# /v1/messages (the channel adapter only rewrites the URL and forwards the
+# anthropic-version / anthropic-beta headers), so the upstream reasoning surface
+# survives unchanged. Modalities follow Token Market's own catalog, which lists
+# text and image input only.
+base_model = "anthropic/claude-opus-5"
+structured_output = true
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "xhigh", "max"]
+
+[cost]
+input = 1.25
+output = 6.25
+cache_read = 0.125
+cache_write = 1.5625
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

providers/tokensmarket/models/claude-sonnet-5.toml

@@ -0,0 +1,25 @@
+# Claude Hub channel: raw passthrough to Anthropic /v1/messages, so the upstream
+# reasoning surface survives unchanged. Modalities follow Token Market's own
+# catalog, which lists text and image input only.
+# Toggle: `thinking.type = "adaptive"` (on, budget managed by Anthropic) |
+# `thinking.type = "disabled"`. claude-sonnet-5 rejects manually-set extended
+# thinking and sampling controls, so no budget_tokens option is listed.
+base_model = "anthropic/claude-sonnet-5"
+structured_output = true
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high", "xhigh", "max"]
+
+[cost]
+input = 0.5
+output = 2.5
+cache_read = 0.05
+cache_write = 0.625
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

providers/tokensmarket/models/deepseek-v4-flash.toml

@@ -0,0 +1,21 @@
+# Alibaba Cloud (DashScope) channel.
+# Toggle: `enable_thinking = false` (effort "none") | `enable_thinking = true`.
+# Effort: `reasoning_effort = high | max`; low/medium collapse to high and xhigh
+# collapses to max on this channel, so only two levels are distinctly reachable.
+# `thinking_budget` is projected for the Qwen family only, so no budget_tokens here.
+base_model = "deepseek/deepseek-v4-flash"
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["high", "max"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.22
+output = 0.665
+cache_read = 0.022

providers/tokensmarket/models/deepseek-v4-pro.toml

@@ -0,0 +1,21 @@
+# Alibaba Cloud (DashScope) channel.
+# Toggle: `enable_thinking = false` (effort "none") | `enable_thinking = true`.
+# Effort: `reasoning_effort = high | max`; low/medium collapse to high and xhigh
+# collapses to max on this channel, so only two levels are distinctly reachable.
+# `thinking_budget` is projected for the Qwen family only, so no budget_tokens here.
+base_model = "deepseek/deepseek-v4-pro"
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["high", "max"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.665
+output = 1.985
+cache_read = 0.0665

providers/tokensmarket/models/gemini-3.5-flash.toml

@@ -0,0 +1,15 @@
+# Gemini Hub channel. The adapter does not rewrite the reasoning fields, so the
+# upstream Gemini effort ladder applies. Token Market's catalog lists text,
+# image, audio, and video input (no PDF), so modalities are overridden.
+base_model = "google/gemini-3.5-flash"
+reasoning_options = [{ type = "effort", values = ["minimal", "low", "medium", "high"] }]
+
+[cost]
+input = 0.225
+output = 1.35
+cache_read = 0.0225
+cache_write = 0.225
+
+[modalities]
+input = ["text", "image", "audio", "video"]
+output = ["text"]

providers/tokensmarket/models/glm-5.1.toml

@@ -0,0 +1,13 @@
+# Alibaba Cloud (DashScope) channel. The adapter maps effort "none" to
+# `enable_thinking = false`, which is the toggle. GLM-5.1 on Model Studio
+# documents no effort ladder (unlike 5.2), so only the toggle is exposed.
+base_model = "zhipuai/glm-5.1"
+reasoning_options = [{ type = "toggle" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.88
+output = 3.53
+cache_read = 0.19

providers/tokensmarket/models/glm-5.2.toml

@@ -0,0 +1,13 @@
+# Alibaba Cloud (DashScope) channel, which accepts the full
+# reasoning_effort enum for GLM-5.2 ("none" disables reasoning, so no separate
+# toggle is needed).
+base_model = "zhipuai/glm-5.2"
+reasoning_options = [{ type = "effort", values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"] }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.944
+output = 3.296
+cache_read = 0.232

providers/tokensmarket/models/gpt-5.4-mini.toml

@@ -0,0 +1,10 @@
+# GPT Hub channel: the adapter only strips the gateway-canonical `reasoning`
+# object and forwards native `reasoning_effort`. Upstream GPT-5.4 tops out at
+# xhigh.
+base_model = "openai/gpt-5.4-mini"
+reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high", "xhigh"] }]
+
+[cost]
+input = 0.1125
+output = 0.675
+cache_read = 0.01125

providers/tokensmarket/models/gpt-5.4-nano.toml

@@ -0,0 +1,13 @@
+# GPT Hub channel: the adapter only strips the gateway-canonical `reasoning`
+# object and forwards native `reasoning_effort`. Upstream GPT-5.4 tops out at
+# xhigh.
+# Cost: the catalog publishes no discounted rate for this model, so the listed
+# vendor rate applies.
+base_model = "openai/gpt-5.4-nano"
+reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high", "xhigh"] }]
+
+[cost]
+input = 0.2
+output = 1.25
+cache_read = 0.02
+cache_write = 0.2

providers/tokensmarket/models/gpt-5.5.toml

@@ -0,0 +1,16 @@
+# GPT Hub channel: the adapter only strips the gateway-canonical `reasoning`
+# object and forwards native `reasoning_effort`. Upstream GPT-5.5 tops out at
+# xhigh (max arrived with 5.6). Modalities follow Token Market's own catalog,
+# which lists text and image input only.
+base_model = "openai/gpt-5.5"
+reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high", "xhigh"] }]
+
+[cost]
+input = 0.75
+output = 4.5
+cache_read = 0.075
+cache_write = 0.75
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

providers/tokensmarket/models/gpt-5.6-luna.toml

@@ -0,0 +1,16 @@
+# GPT Hub channel: the adapter only strips the gateway-canonical `reasoning`
+# object and forwards native `reasoning_effort`, so the full upstream effort
+# ladder is available. Modalities follow Token Market's own catalog, which lists
+# text and image input only.
+base_model = "openai/gpt-5.6-luna"
+reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high", "xhigh", "max"] }]
+
+[cost]
+input = 0.03
+output = 0.18
+cache_read = 0.003
+cache_write = 0.0375
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

providers/tokensmarket/models/gpt-5.6-sol.toml

@@ -0,0 +1,16 @@
+# GPT Hub channel: the adapter only strips the gateway-canonical `reasoning`
+# object and forwards native `reasoning_effort`, so the full upstream effort
+# ladder is available. Modalities follow Token Market's own catalog, which lists
+# text and image input only.
+base_model = "openai/gpt-5.6-sol"
+reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high", "xhigh", "max"] }]
+
+[cost]
+input = 0.75
+output = 4.5
+cache_read = 0.075
+cache_write = 0.9375
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

providers/tokensmarket/models/gpt-5.6-terra.toml

@@ -0,0 +1,16 @@
+# GPT Hub channel: the adapter only strips the gateway-canonical `reasoning`
+# object and forwards native `reasoning_effort`, so the full upstream effort
+# ladder is available. Modalities follow Token Market's own catalog, which lists
+# text and image input only.
+base_model = "openai/gpt-5.6-terra"
+reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high", "xhigh", "max"] }]
+
+[cost]
+input = 0.3
+output = 1.8
+cache_read = 0.03
+cache_write = 0.46875
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

providers/tokensmarket/models/gpt-image-2.toml

@@ -0,0 +1,3 @@
+# GPT Hub channel. Billed per request ($0.01), a unit the per-token [cost]
+# schema cannot express, so [cost] is omitted rather than misstated.
+base_model = "openai/gpt-image-2"

providers/tokensmarket/models/hy3.toml

@@ -0,0 +1,12 @@
+# Tencent Cloud channel. The adapter is a plain OpenAI-compatible passthrough
+# that forwards `reasoning_effort` and does not project onto Tencent's
+# `thinking.type`, so only the effort ladder is reachable. Aligned to the
+# none|low|high surface established relays use for Hy3.
+# https://cloud.tencent.com/document/product/1823/131208
+base_model = "tencent/hy3"
+reasoning_options = [{ type = "effort", values = ["none", "low", "high"] }]
+
+[cost]
+input = 0.147
+output = 0.588
+cache_read = 0.037

providers/tokensmarket/models/kimi-k2.7-code.toml

@@ -0,0 +1,17 @@
+# Alibaba Cloud (DashScope) channel. The adapter maps effort "none" to
+# `enable_thinking = false` and forwards `reasoning.budget_tokens` as
+# `thinking_budget`; Model Studio documents no bounds for the budget on Kimi.
+# Matches the declared surface of the DashScope peer providers on this same
+# model ID (e.g. alibaba-token-plan/kimi-k2.7-code.toml), which cite
+# https://www.alibabacloud.com/help/en/model-studio/kimi-api and
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking for both fields.
+base_model = "moonshotai/kimi-k2.7-code"
+reasoning_options = [{ type = "toggle" }, { type = "budget_tokens" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.768
+output = 3.176
+cache_read = 0.152

providers/tokensmarket/models/kimi-k3.toml

@@ -0,0 +1,20 @@
+# Alibaba Cloud (DashScope) channel.
+# Toggle: `enable_thinking = false` (effort "none") | `enable_thinking = true`.
+# Effort: `reasoning_effort = high | max`; low/medium collapse to high and xhigh
+# collapses to max on this channel, so only two levels are distinctly reachable.
+base_model = "moonshotai/kimi-k3"
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["high", "max"]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 3.0
+output = 15.0
+cache_read = 0.3

providers/tokensmarket/models/longcat-2.0.toml

@@ -0,0 +1,16 @@
+# SiliconFlow.cn channel. The adapter maps effort "none" to
+# `enable_thinking = false` and otherwise `enable_thinking = true`. First-party
+# LongCat-2.0 documents only `{"thinking": {"type": "enabled"|"disabled"}}` — no
+# reasoning budget for this model. `thinking_budget` (128-32,768) is a channel-wide
+# dialect field the adapter also supports, but no LongCat-specific evidence backs
+# it, so it's left out here.
+base_model = "meituan/longcat-2.0"
+reasoning_options = [{ type = "toggle" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.75
+output = 2.95
+cache_read = 0.015