feat(scx-ai): rename scx provider to scx-ai, add GLM-5.2 and Qwen3.8-Max

providers/scx-ai/models/GLM-5.2.toml

@@ -0,0 +1,26 @@
+# Cost is USD per 1M tokens, matching the SCX rates carried in
+# theopenco/llmgateway (packages/models/src/models/zai.ts).
+# Limits and capabilities verified directly against https://api.scx.ai/v1:
+# max_tokens is hard-capped at 131072 by the API ("Range of max_tokens should be
+# [1, 131072]") and context is 1M, which already matches the zhipuai/glm-5.2
+# base entry, so no [limit] override. The enforced input ceiling is
+# "Range of input length should be [1, 1048576]", which is above the published
+# 1M context, so the inherited limits stand.
+# Wire control: `reasoning_effort` in the request body; thinking is returned as
+# `reasoning_content` on the message. SCX's validator accepts none/minimal/low/
+# medium/high/xhigh/max, but only the values below are distinct here (n=4 per
+# level, same prompt, measured as reasoning_content length): `none` and
+# `minimal` both return 0 reasoning characters every time, so only `none` is
+# listed as the off control; low/medium/high are indistinguishable (medians
+# 1515/969/1181), so the Z.AI effective set high/max is kept rather than
+# inventing extra levels.
+base_model = "zhipuai/glm-5.2"
+reasoning_options = [{ type = "effort", values = ["none", "high", "max"] }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.55
+output = 1.784
+cache_read = 0.111

providers/scx-ai/models/MiniMax-M2.7.toml

@@ -0,0 +1,24 @@
+# Cost is USD per 1M tokens, matching the SCX rates carried in
+# theopenco/llmgateway (packages/models/src/models/minimax.ts).
+# Limits and capabilities verified directly against https://api.scx.ai/v1:
+# SCX enforces a single context budget (prompt + max_tokens <= context) with no
+# separate output cap, so output equals context. Probed on the live endpoint:
+# max_tokens 196540 -> 200 OK, 196608 -> 400 "This model's maximum context
+# length is 196608 tokens". Do not trust max_output_length from /v1/models --
+# it reports a placeholder 4096 for most models.
+# Reasoning is returned on a side channel, but this host names the field
+# `reasoning` rather than `reasoning_content`/`reasoning_details`, so
+# `interleaved` is the bare boolean instead of a field name.
+base_model = "minimax/MiniMax-M2.7"
+structured_output = true
+reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
+interleaved = true
+
+[cost]
+input = 0.48
+output = 1.79
+cache_read = 0.05
+
+[limit]
+context = 196_608
+output = 196_608

providers/scx-ai/models/Qwen3.8-Max.toml

@@ -0,0 +1,43 @@
+# Cost is USD per 1M tokens, matching the SCX rates carried in
+# theopenco/llmgateway (packages/models/src/models/alibaba.ts).
+# Limits and capabilities verified directly against https://api.scx.ai/v1:
+# max_tokens is hard-capped at 131072 by the API ("Range of max_tokens should be
+# [1, 131072]"). Input is separately capped below the 1M context inherited from
+# the base entry ("Range of input length should be [1, 983616]"), so limit.input
+# is published as that measured ceiling. The API canonicalises the model id to
+# "Qwen3.8-Max" (lowercase is also accepted and normalises to the same id).
+# Modalities: image and video input confirmed with live requests (a short frame
+# sequence is accepted and described, and an under-length one is rejected with
+# "the range of sequence images should be (4, 8000)"). PDF is inherited from the
+# base entry but is NOT served here -- both the file_url and file_data forms are
+# rejected with "The current model does not support PDF file input" -- so
+# modalities.input is overridden to drop pdf.
+# Wire control: `reasoning_effort` in the request body; thinking is returned as
+# `reasoning_content` on the message. SCX's validator accepts none/minimal/low/
+# medium/high/xhigh/max, but only the values below are distinct here (n=4 per
+# level, same prompt, measured as reasoning_content length): `none` returns 0
+# reasoning characters, minimal/low/medium are one indistinguishable band
+# (medians 339/361/380), and xhigh is clearly higher (1128). This keeps the
+# Alibaba effective set low/medium/xhigh plus the verified `none` off control.
+# No toggle or budget_tokens option is declared: SCX accepts a `thinking` object
+# in the body but there is no evidence it forwards it.
+base_model = "alibaba/qwen3.8-max"
+structured_output = true
+reasoning_options = [
+  { type = "effort", values = ["none", "low", "medium", "xhigh"] },
+]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 1.815
+output = 5.4461
+cache_read = 0.17
+cache_write = 2.5
+
+[limit]
+input = 983_616
+
+[modalities]
+input = ["text", "image", "video"]

providers/scx-ai/models/gpt-oss-120b.toml

@@ -0,0 +1,20 @@
+# Cost is USD per 1M tokens, matching the SCX rates carried in
+# theopenco/llmgateway (packages/models/src/models/openai.ts).
+# Limits and capabilities verified directly against https://api.scx.ai/v1:
+# SCX enforces a single context budget (prompt + max_tokens <= context) with no
+# separate output cap, so output equals context. Do not trust max_output_length
+# from /v1/models -- it reports a placeholder 4096 for most models.
+# Thinking is returned as `reasoning_content` on the message.
+base_model = "openai/gpt-oss-120b"
+reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.17
+output = 0.55
+
+[limit]
+context = 131_072
+output = 131_072