provider: add clf-ai-gateway (OpenAI-compatible, 9 models)

providers/clf-ai-gateway/models/deepseek-v4-flash.toml

@@ -0,0 +1,23 @@
+base_model = "deepseek/deepseek-v4-flash-0731"
+
+# Wire: only OpenAI-style `reasoning_effort` is forwarded, verbatim. No on/off toggle
+# field is forwarded on this surface, so there is no "toggle" option to expose.
+# Host accept-set for deepseek-v4-flash: low, medium, high, xhigh (anything else -> 400 naming this set; public
+# in GET /v1/public/models -> capabilities.reasoning_efforts). Only the levels with a
+# MEASURED distinct effect are advertised — same prompt, 3 samples per level, median
+# reasoning_tokens, 2026-09-02: low 563 / medium 242 / high 199 / xhigh 226 (xhigh samples 1210/226/186) -> no consistent ordering; the accepted levels are not a reliable control.
+# Accepted levels are inert and no toggle is forwarded -> no real caller control here.
+reasoning_options = []
+
+# Current promotional pricing (40% off list), USD/MTok — what the API bills today.
+# Live source: GET /v1/public/models (nano-USD per token; /1000 = USD/MTok).
+[cost]
+input = 0.264
+output = 0.792
+cache_read = 0.008
+
+# This gateway's serving surface, measured (context probed to the boundary;
+# output verified via max_completion_tokens probes).
+[limit]
+context = 1_048_576
+output = 131_072

providers/clf-ai-gateway/models/deepseek-v4-pro.toml

@@ -0,0 +1,24 @@
+base_model = "deepseek/deepseek-v4-pro-0813"
+
+# Wire: only OpenAI-style `reasoning_effort` is forwarded, verbatim. No on/off toggle
+# field is forwarded on this surface, so there is no "toggle" option to expose.
+# Host accept-set for deepseek-v4-pro: low, medium, high, xhigh (anything else -> 400 naming this set; public
+# in GET /v1/public/models -> capabilities.reasoning_efforts). Only the levels with a
+# MEASURED distinct effect are advertised — same prompt, 3 samples per level, median
+# reasoning_tokens, 2026-09-02: low 282 / medium 214 / high 746 / xhigh 4124 -> three graded behaviors; medium is an inert alias of low.
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "xhigh"]
+
+# Current promotional pricing (40% off list), USD/MTok — what the API bills today.
+# Live source: GET /v1/public/models (nano-USD per token; /1000 = USD/MTok).
+[cost]
+input = 0.792
+output = 2.376
+cache_read = 0.026
+
+# This gateway's serving surface, measured (context probed to the boundary;
+# output verified via max_completion_tokens probes).
+[limit]
+context = 1_048_576
+output = 131_072

providers/clf-ai-gateway/models/glm-4.7-flash.toml

@@ -0,0 +1,23 @@
+base_model = "zhipuai/glm-4.7-flash"
+
+# Wire: only OpenAI-style `reasoning_effort` is forwarded, verbatim. No on/off toggle
+# field is forwarded on this surface, so there is no "toggle" option to expose.
+# Host accept-set for glm-4.7-flash: low, medium, high (anything else -> 400 naming this set; public
+# in GET /v1/public/models -> capabilities.reasoning_efforts). Only the levels with a
+# MEASURED distinct effect are advertised — same prompt, 3 samples per level, median
+# reasoning_tokens, 2026-09-02: low 1294 / medium 1451 / high 567 with overlapping spreads (high samples 1778/567/192) -> no measurable graded control; the accepted levels are inert.
+# Accepted levels are inert and no toggle is forwarded -> no real caller control here.
+reasoning_options = []
+
+# Current promotional pricing (40% off list), USD/MTok — what the API bills today.
+# Live source: GET /v1/public/models (nano-USD per token; /1000 = USD/MTok).
+# No cached tier for this model on this surface (public endpoint: cached_input = null).
+[cost]
+input = 0.036
+output = 0.24
+
+# This gateway's serving surface, measured (context probed to the boundary;
+# output verified via max_completion_tokens probes).
+[limit]
+context = 131_072
+output = 131_072

providers/clf-ai-gateway/models/glm-5.2.toml

@@ -0,0 +1,24 @@
+base_model = "zhipuai/glm-5.2"
+
+# Wire: only OpenAI-style `reasoning_effort` is forwarded, verbatim. No on/off toggle
+# field is forwarded on this surface, so there is no "toggle" option to expose.
+# Host accept-set for glm-5.2: low, medium, high, xhigh (anything else -> 400 naming this set; public
+# in GET /v1/public/models -> capabilities.reasoning_efforts). Only the levels with a
+# MEASURED distinct effect are advertised — same prompt, 3 samples per level, median
+# reasoning_tokens, 2026-09-02: low 471 / medium 367 / high 356 / xhigh 666 -> only xhigh is distinct; low/medium/high overlap (medium/high are inert aliases of low).
+[[reasoning_options]]
+type = "effort"
+values = ["low", "xhigh"]
+
+# Current promotional pricing (40% off list), USD/MTok — what the API bills today.
+# Live source: GET /v1/public/models (nano-USD per token; /1000 = USD/MTok).
+[cost]
+input = 0.84
+output = 2.64
+cache_read = 0.156
+
+# This gateway's serving surface, measured (context probed to the boundary;
+# output verified via max_completion_tokens probes).
+[limit]
+context = 262_144
+output = 131_072

providers/clf-ai-gateway/models/glm-5.3-flash.toml

@@ -0,0 +1,29 @@
+base_model = "zhipuai/glm-5.3-flash"
+
+# Wire: only OpenAI-style `reasoning_effort` is forwarded, verbatim. No on/off toggle
+# field is forwarded on this surface, so there is no "toggle" option to expose.
+# Host accept-set for glm-5.3-flash: low, medium, high, xhigh (anything else -> 400 naming this set; public
+# in GET /v1/public/models -> capabilities.reasoning_efforts). Only the levels with a
+# MEASURED distinct effect are advertised — same prompt, 3 samples per level, median
+# reasoning_tokens, 2026-09-02: low 27 / medium 289 / high 31 / xhigh 295 -> two behaviors: {low,high} minimal, {medium,xhigh} full; medium and high are inert aliases.
+[[reasoning_options]]
+type = "effort"
+values = ["low", "xhigh"]
+
+# Current promotional pricing (40% off list), USD/MTok — what the API bills today.
+# Live source: GET /v1/public/models (nano-USD per token; /1000 = USD/MTok).
+[cost]
+input = 0.09
+output = 0.3
+cache_read = 0.018
+
+# This gateway's serving surface, measured (context probed to the boundary;
+# output verified via max_completion_tokens probes).
+[limit]
+context = 1_048_576
+output = 131_072
+
+# Gateway accepts text + image_url parts only (no video/pdf on this surface).
+[modalities]
+input = ["text", "image"]
+output = ["text"]

providers/clf-ai-gateway/models/glm-5.3.toml

@@ -0,0 +1,24 @@
+base_model = "zhipuai/glm-5.3"
+
+# Wire: only OpenAI-style `reasoning_effort` is forwarded, verbatim. No on/off toggle
+# field is forwarded on this surface, so there is no "toggle" option to expose.
+# Host accept-set for glm-5.3: none, low, medium, high, max (anything else -> 400 naming this set; public
+# in GET /v1/public/models -> capabilities.reasoning_efforts). Only the levels with a
+# MEASURED distinct effect are advertised — same prompt, 3 samples per level, median
+# reasoning_tokens, 2026-09-02: low 26 / medium 330 / high 32 / max 368 -> two behaviors: {low,high} minimal, {medium,max} full; medium and high are inert aliases; "none" is accepted but does not disable thinking, so it is not advertised.
+[[reasoning_options]]
+type = "effort"
+values = ["low", "max"]
+
+# Current promotional pricing (40% off list), USD/MTok — what the API bills today.
+# Live source: GET /v1/public/models (nano-USD per token; /1000 = USD/MTok).
+[cost]
+input = 0.84
+output = 2.64
+cache_read = 0.156
+
+# This gateway's serving surface, measured (context probed to the boundary;
+# output verified via max_completion_tokens probes).
+[limit]
+context = 1_048_576
+output = 131_072

providers/clf-ai-gateway/models/kimi-k2.6.toml

@@ -0,0 +1,28 @@
+base_model = "moonshotai/kimi-k2.6"
+
+# Wire: only OpenAI-style `reasoning_effort` is forwarded, verbatim. No on/off toggle
+# field is forwarded on this surface, so there is no "toggle" option to expose.
+# Host accept-set for kimi-k2.6: low, medium, high (anything else -> 400 naming this set; public
+# in GET /v1/public/models -> capabilities.reasoning_efforts). Only the levels with a
+# MEASURED distinct effect are advertised — same prompt, 3 samples per level, median
+# reasoning_tokens, 2026-09-02: low 528 / medium 577 / high 456 -> flat; thinking is always on and the accepted levels are inert (matches the lab: toggle-only, and no toggle is forwarded here).
+# Accepted levels are inert and no toggle is forwarded -> no real caller control here.
+reasoning_options = []
+
+# Current promotional pricing (40% off list), USD/MTok — what the API bills today.
+# Live source: GET /v1/public/models (nano-USD per token; /1000 = USD/MTok).
+[cost]
+input = 0.57
+output = 2.4
+cache_read = 0.096
+
+# This gateway's serving surface, measured (context probed to the boundary;
+# output verified via max_completion_tokens probes).
+[limit]
+context = 262_144
+output = 131_072
+
+# Gateway accepts text + image_url parts only (no video/pdf on this surface).
+[modalities]
+input = ["text", "image"]
+output = ["text"]

providers/clf-ai-gateway/models/kimi-k2.7-code.toml

@@ -0,0 +1,28 @@
+base_model = "moonshotai/kimi-k2.7-code"
+
+# Wire: only OpenAI-style `reasoning_effort` is forwarded, verbatim. No on/off toggle
+# field is forwarded on this surface, so there is no "toggle" option to expose.
+# Host accept-set for kimi-k2.7-code: low, medium, high (anything else -> 400 naming this set; public
+# in GET /v1/public/models -> capabilities.reasoning_efforts). Only the levels with a
+# MEASURED distinct effect are advertised — same prompt, 3 samples per level, median
+# reasoning_tokens, 2026-09-02: low 141 / medium 155 / high 283 with overlapping spreads (high 207/283/426 vs medium 210/155/136) -> no reliable graded control; accepted levels are inert.
+# Accepted levels are inert and no toggle is forwarded -> no real caller control here.
+reasoning_options = []
+
+# Current promotional pricing (40% off list), USD/MTok — what the API bills today.
+# Live source: GET /v1/public/models (nano-USD per token; /1000 = USD/MTok).
+[cost]
+input = 0.57
+output = 2.4
+cache_read = 0.114
+
+# This gateway's serving surface, measured (context probed to the boundary;
+# output verified via max_completion_tokens probes).
+[limit]
+context = 262_144
+output = 131_072
+
+# Gateway accepts text + image_url parts only (no video/pdf on this surface).
+[modalities]
+input = ["text", "image"]
+output = ["text"]

providers/clf-ai-gateway/models/qwen3.8-27b.toml

@@ -0,0 +1,33 @@
+base_model = "alibaba/qwen3.8-27b"
+
+# Wire: only OpenAI-style `reasoning_effort` is forwarded, verbatim. No on/off toggle
+# field is forwarded on this surface, so there is no "toggle" option to expose.
+# Host accept-set for qwen3.8-27b: low, medium, xhigh (anything else -> 400 naming this set; public
+# in GET /v1/public/models -> capabilities.reasoning_efforts). Only the levels with a
+# MEASURED distinct effect are advertised — same prompt, 3 samples per level, median
+# reasoning_tokens, 2026-09-02: low 138 / medium 166 / xhigh 245 (every xhigh sample above every low sample) -> xhigh distinct; medium is an inert alias of low.
+[[reasoning_options]]
+type = "effort"
+values = ["low", "xhigh"]
+
+# Current promotional pricing (40% off list), USD/MTok — what the API bills today.
+# Live source: GET /v1/public/models (nano-USD per token; /1000 = USD/MTok).
+# cache_read = input is intentional, not a typo: this model has no verified cached
+# tier on this surface, so cached tokens bill at the input rate. The public endpoint
+# shows cached_input == input for both promo and list price (270/270 nano now,
+# 450/450 list) — we do not advertise a cache discount we cannot prove.
+[cost]
+input = 0.27
+output = 1.92
+cache_read = 0.27
+
+# This gateway's serving surface, measured (context probed to the boundary;
+# output verified via max_completion_tokens probes).
+[limit]
+context = 262_144
+output = 131_072
+
+# Gateway accepts text + image_url parts only (no video/pdf on this surface).
+[modalities]
+input = ["text", "image"]
+output = ["text"]

providers/clf-ai-gateway/provider.toml

@@ -0,0 +1,10 @@
+name = "CLF AI Gateway"
+env = ["CLF_AI_GATEWAY_API_KEY"]
+npm = "@ai-sdk/openai-compatible"
+# OpenAI-compatible gateway reselling Cloudflare Workers AI open-weight models
+# (GLM, Kimi, DeepSeek, Qwen) with prepaid credits. Runs on Workers AI upstream
+# by design and says so publicly: https://clfaigateway.dev/docs
+# Model IDs are the gateway's canonical names (glm-5.3, kimi-k2.7-code, ...);
+# the upstream @cf/... id is returned per model by GET /v1/public/models.
+doc = "https://clfaigateway.dev/docs"
+api = "https://api.clfaigateway.dev/v1"