@oh-my-pi/pi-catalog 18.1.18 → 18.1.19

package/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [18.1.19] - 2026-09-12
+
+### Added
+
+- Added Charm Hyper as a built-in provider with API-key login, live model discovery, and per-model pricing, effort ladders, and limits read straight from its catalog ([#11656](https://github.com/can1357/oh-my-pi/pull/11656) by [@oldschoola](https://github.com/oldschoola)).
+
 ## [18.1.18] - 2026-09-11
 
 ### Added
@@ -13,6 +19,7 @@
 - OpenCode Go's DeepSeek Flash lanes (`deepseek-flash`, `deepseek-v4.1-flash`) now declare image input. The gateway serves them with vision despite the IDs carrying no vision suffix, so the class-wide `strip-image-input` rule was dropping attachments the endpoint reads; the modality is declared too, since live discovery seeds these lanes text-only ([#11774](https://github.com/can1357/oh-my-pi/pull/11774) by [@STRML](https://github.com/STRML)).
 - Amazon Bedrock OpenAI models, plus unclassified profiles such as opaque application-inference-profile ARNs, now carry the compatibility policy required to preserve image-bearing tool results ([#11681](https://github.com/can1357/oh-my-pi/issues/11681)).
 - DeepSeek V4.1 Flash requests now honor the documented 384K output maximum instead of being capped at 64K ([#11769](https://github.com/can1357/oh-my-pi/issues/11769)).
+- Fixed the first-party `deepseek-flash` alias missing the V4.1 Flash wire contract: it now sends `max_tokens` with `reasoning_content` and replays reasoning and assistant content on tool calls with no tool choice ([#11799](https://github.com/can1357/oh-my-pi/pull/11799) by [@brit](https://github.com/brit)).
 
 ## [18.1.17] - 2026-09-10
 

package/dist/types/compat/auth-ids.d.ts

@@ -1,4 +1,4 @@
 /** Every provider with an `auth/<id>.kdl` policy. */
-export type AuthProviderId = "abliteration" | "aiand" | "aimlapi" | "alibaba-coding-plan" | "alibaba-token-plan" | "amazon-bedrock" | "anthropic" | "azure" | "baseten" | "bedrock-mantle" | "cerebras" | "cline-pass" | "cloudflare-ai-gateway" | "commandcode" | "coreweave" | "cursor" | "deepinfra" | "deepseek" | "devin" | "exa" | "firepass" | "fireworks" | "github-copilot" | "gitlab-duo" | "gitlab-duo-agent" | "gmi-cloud" | "google" | "google-antigravity" | "google-gemini-cli" | "google-vertex" | "groq" | "huggingface" | "kagi" | "kilo" | "kimi-code" | "litellm" | "llama.cpp" | "lm-studio" | "meta" | "minimax" | "minimax-code" | "minimax-code-cn" | "mistral" | "moonshot" | "muse-code" | "nanogpt" | "novita" | "nvidia" | "ollama" | "ollama-cloud" | "openai" | "openai-codex" | "openai-codex-device" | "opencode-go" | "opencode-zen" | "openrouter" | "parallel" | "perplexity" | "qianfan" | "qwen-portal" | "sakana" | "siliconflow" | "siliconflow-cn" | "synthetic" | "tavily" | "together" | "umans" | "venice" | "vercel-ai-gateway" | "vllm" | "wafer-serverless" | "xai" | "xai-oauth" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "yolo-auto" | "zai" | "zai-coding-plan" | "zenmux" | "zhipu-coding-plan";
+export type AuthProviderId = "abliteration" | "aiand" | "aimlapi" | "alibaba-coding-plan" | "alibaba-token-plan" | "amazon-bedrock" | "anthropic" | "azure" | "baseten" | "bedrock-mantle" | "cerebras" | "charm-hyper" | "cline-pass" | "cloudflare-ai-gateway" | "commandcode" | "coreweave" | "cursor" | "deepinfra" | "deepseek" | "devin" | "exa" | "firepass" | "fireworks" | "github-copilot" | "gitlab-duo" | "gitlab-duo-agent" | "gmi-cloud" | "google" | "google-antigravity" | "google-gemini-cli" | "google-vertex" | "groq" | "huggingface" | "kagi" | "kilo" | "kimi-code" | "litellm" | "llama.cpp" | "lm-studio" | "meta" | "minimax" | "minimax-code" | "minimax-code-cn" | "mistral" | "moonshot" | "muse-code" | "nanogpt" | "novita" | "nvidia" | "ollama" | "ollama-cloud" | "openai" | "openai-codex" | "openai-codex-device" | "opencode-go" | "opencode-zen" | "openrouter" | "parallel" | "perplexity" | "qianfan" | "qwen-portal" | "sakana" | "siliconflow" | "siliconflow-cn" | "synthetic" | "tavily" | "together" | "umans" | "venice" | "vercel-ai-gateway" | "vllm" | "wafer-serverless" | "xai" | "xai-oauth" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "yolo-auto" | "zai" | "zai-coding-plan" | "zenmux" | "zhipu-coding-plan";
 /** Providers whose policy declares a `login` flow (the `/login` roster). */
-export type LoginProviderId = "abliteration" | "aiand" | "alibaba-coding-plan" | "alibaba-token-plan" | "anthropic" | "baseten" | "cerebras" | "cline-pass" | "cloudflare-ai-gateway" | "commandcode" | "coreweave" | "cursor" | "deepinfra" | "deepseek" | "devin" | "exa" | "firepass" | "fireworks" | "github-copilot" | "gitlab-duo" | "gitlab-duo-agent" | "gmi-cloud" | "google-antigravity" | "google-gemini-cli" | "huggingface" | "kagi" | "kilo" | "kimi-code" | "litellm" | "llama.cpp" | "lm-studio" | "meta" | "minimax-code" | "minimax-code-cn" | "moonshot" | "muse-code" | "nanogpt" | "novita" | "nvidia" | "ollama" | "ollama-cloud" | "openai-codex" | "openai-codex-device" | "opencode-go" | "opencode-zen" | "openrouter" | "parallel" | "perplexity" | "qianfan" | "qwen-portal" | "sakana" | "siliconflow" | "siliconflow-cn" | "synthetic" | "tavily" | "together" | "umans" | "venice" | "vercel-ai-gateway" | "vllm" | "wafer-serverless" | "xai" | "xai-oauth" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "yolo-auto" | "zai" | "zai-coding-plan" | "zenmux" | "zhipu-coding-plan";
+export type LoginProviderId = "abliteration" | "aiand" | "alibaba-coding-plan" | "alibaba-token-plan" | "anthropic" | "baseten" | "cerebras" | "charm-hyper" | "cline-pass" | "cloudflare-ai-gateway" | "commandcode" | "coreweave" | "cursor" | "deepinfra" | "deepseek" | "devin" | "exa" | "firepass" | "fireworks" | "github-copilot" | "gitlab-duo" | "gitlab-duo-agent" | "gmi-cloud" | "google-antigravity" | "google-gemini-cli" | "huggingface" | "kagi" | "kilo" | "kimi-code" | "litellm" | "llama.cpp" | "lm-studio" | "meta" | "minimax-code" | "minimax-code-cn" | "moonshot" | "muse-code" | "nanogpt" | "novita" | "nvidia" | "ollama" | "ollama-cloud" | "openai-codex" | "openai-codex-device" | "opencode-go" | "opencode-zen" | "openrouter" | "parallel" | "perplexity" | "qianfan" | "qwen-portal" | "sakana" | "siliconflow" | "siliconflow-cn" | "synthetic" | "tavily" | "together" | "umans" | "venice" | "vercel-ai-gateway" | "vllm" | "wafer-serverless" | "xai" | "xai-oauth" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "yolo-auto" | "zai" | "zai-coding-plan" | "zenmux" | "zhipu-coding-plan";

package/dist/types/provider-models/descriptors.d.ts

@@ -88,6 +88,14 @@ export declare const CATALOG_PROVIDERS: readonly [{
     readonly catalogDiscovery: {
         readonly label: "Cerebras";
     };
+}, {
+    readonly id: "charm-hyper";
+    readonly defaultModel: "glm-5.3";
+    readonly envVars: readonly ["CHARM_HYPER_API_KEY", "HYPER_API_KEY"];
+    readonly createModelManagerOptions: (config: ModelManagerConfig) => import("../index.js").ModelManagerOptions<"openai-completions", unknown>;
+    readonly allowUnauthenticated: true;
+    readonly dynamicModelsAuthoritative: true;
+    readonly skipCrossProviderReferenceFills: true;
 }, {
     readonly id: "cloudflare-ai-gateway";
     readonly defaultModel: "anthropic/claude-opus-4-8";

package/dist/types/provider-models/openai-compat.d.ts

@@ -745,3 +745,22 @@ export interface CommandCodeModelManagerConfig {
  * or context window.
  */
 export declare function commandCodeModelManagerOptions(config?: CommandCodeModelManagerConfig): ModelManagerOptions<Api>;
+/**
+ * Configuration for the Charm Hyper model manager.
+ *
+ * `baseUrl` overrides the gateway root for tests and self-hosted proxies; a
+ * value that omits the `/v1` surface gains one, so a host-only override
+ * behaves like every sibling provider's.
+ */
+export interface CharmHyperModelManagerConfig {
+    apiKey?: string;
+    baseUrl?: string;
+    fetch?: FetchImpl;
+}
+/**
+ * Charm Hyper's gateway catalog. `/v1/models` is public and carries the live
+ * tariff, so discovery runs with or without a key and the snapshot is
+ * authoritative: a model the gateway stops serving is pruned rather than kept
+ * alive by a stale bundled row.
+ */
+export declare function charmHyperModelManagerOptions(config?: CharmHyperModelManagerConfig): ModelManagerOptions<"openai-completions">;

package/dist/types/wire/charm-hyper.d.ts

@@ -0,0 +1,21 @@
+/**
+ * Charm Hyper gateway endpoint, shared so a host migration — or a self-hosted
+ * proxy override — touches a single module.
+ */
+export declare const CHARM_HYPER_API_BASE_URL = "https://hyper.charm.land/v1";
+/**
+ * Resolve a configured Charm Hyper base URL onto the gateway's `/v1` surface.
+ *
+ * Every consumer must agree on this, because they key different things off the
+ * result: inference and discovery target it, `/usage` sends a bearer token to
+ * it, and the model cache namespace is hashed from it. Three separate copies
+ * previously disagreed for a blank value — the model manager treated it as
+ * absent and used the canonical host, while the others produced a bare `/v1` —
+ * so a whitespace-only override silently split inference, balance checks and
+ * caching across different endpoints.
+ *
+ * A blank or whitespace-only value therefore means "not configured" and
+ * resolves to the canonical host; anything else keeps its host and gains the
+ * `/v1` segment if it omits one.
+ */
+export declare function normalizeCharmHyperBaseUrl(baseUrl?: string): string;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "18.1.18",
+	"version": "18.1.19",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://omp.sh",
 	"author": "Stencil Labs, Inc.",
@@ -35,12 +35,12 @@
 		"gen:proto": "bun scripts/generate-protocols.ts"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.1.18",
-		"@oh-my-pi/pi-utils": "18.1.18"
+		"@oh-my-pi/omptype": "18.1.19",
+		"@oh-my-pi/pi-utils": "18.1.19"
 	},
 	"devDependencies": {
 		"@bgotink/kdl": "0.4.0",
-		"@oh-my-pi/pi-ai": "18.1.18",
+		"@oh-my-pi/pi-ai": "18.1.19",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/auth-ids.ts

@@ -13,6 +13,7 @@ export type AuthProviderId =
 	| "baseten"
 	| "bedrock-mantle"
 	| "cerebras"
+	| "charm-hyper"
 	| "cline-pass"
 	| "cloudflare-ai-gateway"
 	| "commandcode"
@@ -94,6 +95,7 @@ export type LoginProviderId =
 	| "anthropic"
 	| "baseten"
 	| "cerebras"
+	| "charm-hyper"
 	| "cline-pass"
 	| "cloudflare-ai-gateway"
 	| "commandcode"

package/src/compat/rules/auth/_order.kdl

@@ -5,7 +5,7 @@ login-order "openai-codex" "anthropic" "zai" "zai-coding-plan" "kimi-code" "open
 	"cursor" "devin" "google-antigravity" "google-gemini-cli" "openai-codex-device" "xai" "xai-oauth" \
 	"gitlab-duo" "gitlab-duo-agent" "alibaba-coding-plan" "alibaba-token-plan" "aiand" "abliteration" "zhipu-coding-plan" \
 	"umans" "qwen-portal" "sakana" "minimax-code" "minimax-code-cn" "xiaomi" "xiaomi-token-plan-sgp" \
-	"xiaomi-token-plan-ams" "xiaomi-token-plan-cn" "firepass" "cline-pass" "commandcode" "deepseek" "muse-code" "meta" "moonshot" \
+	"xiaomi-token-plan-ams" "xiaomi-token-plan-cn" "firepass" "cline-pass" "commandcode" "charm-hyper" "deepseek" "muse-code" "meta" "moonshot" \
 	"cerebras" "baseten" "fireworks" "together" "nvidia" "novita" "deepinfra" "huggingface" "perplexity" \
 	"qianfan" "venice" "siliconflow" "siliconflow-cn" "synthetic" "nanogpt" "wafer-serverless" "coreweave" \
 	"vercel-ai-gateway" "cloudflare-ai-gateway" "litellm" "kilo" "zenmux" "opencode-zen" "opencode-go" \

package/src/compat/rules/auth/charm-hyper.kdl

@@ -0,0 +1,19 @@
+// Validation targets `/v1/credits`, not `/v1/models`: the models endpoint is
+// public and answers 200 with no credentials or a bogus key (verified
+// 2026-09-11), so a models-endpoint probe against it could never reject a bad
+// key. `/v1/credits` is an authenticated, non-billable GET that returns the
+// account balance on success and 401 `{"error":"authentication failed"}`
+// otherwise — the same trick cline-pass uses to validate against its own
+// `/users/me`, and cheaper than a chat-completions probe that would bill the
+// key on every login.
+auth "charm-hyper" {
+	name "Charm Hyper"
+	env "CHARM_HYPER_API_KEY" "HYPER_API_KEY"
+	login "api-key" {
+		auth-url "https://hyper.charm.land/"
+		instructions "Create or copy an API key from the Charm Hyper dashboard"
+		prompt "Paste your Charm Hyper API key" placeholder="sk-hyper-..."
+		normalize "strip-bearer"
+		validate "models-endpoint" label="Charm Hyper" url="https://hyper.charm.land/v1/credits"
+	}
+}

package/src/compat/rules/classes/minimax.kdl

@@ -28,8 +28,10 @@ class "minimax" {
 	}
 	// MiniMax-M3: 512K is the standard pricing-tier boundary leaked into
 	// upstream metadata, not the model ceiling (documented: 1M in / 128K out).
-	on "minimax" "minimax-cn" "minimax-code" "minimax-code-cn" {
-		models "MiniMax-M3" {
+	// Exact selectors are case-sensitive, so both served spellings are listed:
+	// MiniMax hosts publish `MiniMax-M3`, Charm Hyper publishes `minimax-m3`.
+	on "charm-hyper" "minimax" "minimax-cn" "minimax-code" "minimax-code-cn" {
+		models "MiniMax-M3" "minimax-m3" {
 			limits-patch {
 				context-window 1000000
 				max-tokens 128000

package/src/compat/rules/providers/charm-hyper.kdl

@@ -0,0 +1,82 @@
+// Charm Hyper gateway deployment contract.
+//
+// Capability data is NOT restated here: `GET /v1/models` publishes each row's
+// context window, output cap, vision flag, effort vocabulary and per-million
+// tariff, and `charmHyperModelManagerOptions` maps all of it verbatim —
+// including `reasoningDisableMode`, which is derived per model from whether
+// the advertised vocabulary contains `none`. Only facts the gateway does not
+// publish, or publishes wrongly, belong in this file.
+provider "charm-hyper" {
+	// OpenAI-dialect gateway in front of open-weight and third-party models.
+	// `developer` and `store` are accepted (HTTP 200) but the gateway accepts
+	// every unknown field it is handed, so acceptance is not evidence the
+	// semantics are honored; both stay off and system prompts ride the
+	// `system` role, matching every sibling gateway. The wire takes
+	// `max_tokens` (verified 2026-09-11).
+	supports-developer-role #false
+	supports-store #false
+	max-tokens-field "max_tokens"
+	thinking-mode "effort"
+	// Reasoning arrives as `reasoning_content` on both the streamed delta and
+	// the final message (verified 2026-09-11).
+	reasoning-content-field "reasoning_content"
+	// ---------------------------------------------------------------------------
+	// Blockless rows are left non-reasoning on purpose.
+	//
+	// Four ids serve a thinking model while publishing no `reasoning` block:
+	// glm-5, glm-5.1, kimi-k2-thinking and minimax-m2.7 each returned
+	// populated `reasoning_content` and non-zero `reasoning_tokens` on a
+	// multi-step prompt (2026-09-11). The other blockless rows
+	// (gemma-4-26b-a4b-it, kimi-k2.5, kimi-k2.7-code, llama-3.3-70b-instruct,
+	// llama-4-maverick-17b-128e-instruct-fp8,
+	// qwen3-coder-480b-a35b-instruct-int4-mixed-ar, qwen3-next-80b-a3b-instruct)
+	// returned none.
+	//
+	// No rule restores them, because none can be written honestly. Those four
+	// expose no effort dial, and `ThinkingConfig.efforts` is documented as
+	// never empty — a reasoning model without a controllable surface carries
+	// `thinking: undefined`, which no KDL axis can express (the only upgrade
+	// path is an exact `thinking-efforts` ladder). Synthesizing a ladder would
+	// add silent no-op rungs: the gateway answers 200 to efforts it never
+	// advertised and simply ignores them, and a repeat measurement of effort
+	// sensitivity on these ids came back non-monotonic, i.e. sampling noise
+	// rather than a hidden dial.
+	//
+	// The cost is narrow and was checked: streamed `reasoning_content` is
+	// parsed and displayed with no `model.reasoning` gate, the effort
+	// parameter is moot without a dial, and the developer-role and
+	// zai-replay paths are already disabled for this provider.
+	// ---------------------------------------------------------------------------
+	// Residue: misreported output caps. The gateway derives
+	// `max_output_tokens` as a flat fraction of the context window rather
+	// than reporting a real ceiling — 10% on 9 rows, the dominant
+	// convention — and the values are advisory, not enforced: glm-5.1
+	// publishes 3276 yet produced 14066 tokens, and minimax-m2.7 publishes
+	// 6553 yet produced 8418, both stopping on `finish_reason: "stop"`
+	// rather than `length`. An over-stated cap is never rejected either
+	// (`max_tokens: 200000` returns 200), so the risk is one-sided:
+	// over-stating costs only the context share 10% already reserves
+	// everywhere else, while under-stating truncates real output.
+	//
+	// Both corrections take a same-window peer's published value rather than
+	// inventing a ceiling, and both clear the measured floor:
+	//   glm-5.1      (202750) is 1.6%; sibling glm-5 (202752) publishes 20275.
+	//   minimax-m2.7 (262100) is 2.5%, the lone outlier in its window class;
+	//                kimi-k2.5 (262144), kimi-k2.6 (262000) and
+	//                qwen3-next-80b (262144) all publish 26214.
+	//
+	// deepseek-v4.1-flash is the only other 2.5% row and is deliberately left
+	// alone: its 1M window has no peer group that agrees on a fraction, and a
+	// long generation stopped naturally at 8504, well inside the published
+	// 26214. Measured 2026-09-11.
+	models "glm-5.1" {
+		limits-patch {
+			max-tokens 20275
+		}
+	}
+	models "minimax-m2.7" {
+		limits-patch {
+			max-tokens 26214
+		}
+	}
+}

package/src/compat/rules/providers/deepseek.kdl

@@ -82,11 +82,22 @@ provider "deepseek" {
 		}
 	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.
-	models "deepseek-v4-flash" "deepseek-v4.1-flash-expires-on-0910" {
+	// The bare `deepseek-flash` alias serves DeepSeek-V4.1-Flash, so it shares
+	// the V4.1 wire contract: max_tokens + reasoning_content, mandatory replay
+	// (synthetic placeholders rejected upstream), no tool choice, and the
+	// low/high/max effort ladder the bare alias cannot inherit by family.
+	models "deepseek-flash" "deepseek-v4-flash" "deepseek-v4.1-flash-expires-on-0910" {
+		// Stale source capability data: discovery reports the bare alias as
+		// non-reasoning, so opt into the cascade upgrade for the exact ladder.
+		thinking-upgrade-neutral #true
 		max-tokens-field "max_tokens"
 		reasoning-content-field "reasoning_content"
 		requires-assistant-content-for-tool-calls #true
 		requires-reasoning-content-for-tool-calls #true
+		allows-synthetic-reasoning-content-for-tool-calls #false
 		supports-tool-choice #false
+		// The bare alias carries no taxonomy family, so it cannot inherit the
+		// flash ladder; the exact rule owns it instead.
+		thinking-efforts "low" "high" "max"
 	}
 }

package/src/compat/rules.json

@@ -13,6 +13,7 @@
 		"auth/baseten.kdl",
 		"auth/bedrock-mantle.kdl",
 		"auth/cerebras.kdl",
+		"auth/charm-hyper.kdl",
 		"auth/cline-pass.kdl",
 		"auth/cloudflare-ai-gateway.kdl",
 		"auth/commandcode.kdl",
@@ -114,6 +115,7 @@
 		"providers/baseten.kdl",
 		"providers/bedrock-mantle.kdl",
 		"providers/cerebras.kdl",
+		"providers/charm-hyper.kdl",
 		"providers/cline-pass.kdl",
 		"providers/cloudflare-ai-gateway.kdl",
 		"providers/commandcode.kdl",
@@ -5988,9 +5990,10 @@
 				}
 			},
 			{
-				"source": "classes/minimax.kdl:32",
+				"source": "classes/minimax.kdl:34",
 				"class": "minimax",
 				"providers": [
+					"charm-hyper",
 					"minimax",
 					"minimax-cn",
 					"minimax-code",
@@ -6000,6 +6003,10 @@
 					{
 						"kind": "exact",
 						"value": "MiniMax-M3"
+					},
+					{
+						"kind": "exact",
+						"value": "minimax-m3"
 					}
 				],
 				"catalog": {
@@ -6010,7 +6017,7 @@
 				}
 			},
 			{
-				"source": "classes/minimax.kdl:39",
+				"source": "classes/minimax.kdl:41",
 				"class": "minimax",
 				"providers": [
 					"alibaba-coding-plan",
@@ -6023,7 +6030,7 @@
 				}
 			},
 			{
-				"source": "classes/minimax.kdl:44",
+				"source": "classes/minimax.kdl:46",
 				"class": "minimax",
 				"providers": [
 					"amazon-bedrock",
@@ -6040,7 +6047,7 @@
 				}
 			},
 			{
-				"source": "classes/minimax.kdl:48",
+				"source": "classes/minimax.kdl:50",
 				"class": "minimax",
 				"providers": [
 					"kilo",
@@ -6058,7 +6065,7 @@
 				}
 			},
 			{
-				"source": "classes/minimax.kdl:54",
+				"source": "classes/minimax.kdl:56",
 				"class": "minimax",
 				"providers": [
 					"minimax",
@@ -6075,7 +6082,7 @@
 				}
 			},
 			{
-				"source": "classes/minimax.kdl:62",
+				"source": "classes/minimax.kdl:64",
 				"class": "minimax",
 				"providers": [
 					"minimax",
@@ -6097,7 +6104,7 @@
 				}
 			},
 			{
-				"source": "classes/minimax.kdl:52",
+				"source": "classes/minimax.kdl:54",
 				"class": "minimax",
 				"providers": [
 					"minimax",
@@ -6109,7 +6116,7 @@
 				}
 			},
 			{
-				"source": "classes/minimax.kdl:72",
+				"source": "classes/minimax.kdl:74",
 				"class": "minimax",
 				"providers": [
 					"minimax-code",
@@ -6121,7 +6128,7 @@
 				}
 			},
 			{
-				"source": "classes/minimax.kdl:76",
+				"source": "classes/minimax.kdl:78",
 				"class": "minimax",
 				"providers": [
 					"minimax-code",
@@ -7590,6 +7597,55 @@
 					"mode": "effort"
 				}
 			},
+			{
+				"source": "providers/charm-hyper.kdl:72",
+				"providers": [
+					"charm-hyper"
+				],
+				"models": [
+					{
+						"kind": "exact",
+						"value": "glm-5.1"
+					}
+				],
+				"catalog": {
+					"limitsPatch": {
+						"maxTokens": 20275
+					}
+				}
+			},
+			{
+				"source": "providers/charm-hyper.kdl:77",
+				"providers": [
+					"charm-hyper"
+				],
+				"models": [
+					{
+						"kind": "exact",
+						"value": "minimax-m2.7"
+					}
+				],
+				"catalog": {
+					"limitsPatch": {
+						"maxTokens": 26214
+					}
+				}
+			},
+			{
+				"source": "providers/charm-hyper.kdl:9",
+				"providers": [
+					"charm-hyper"
+				],
+				"wire": {
+					"supportsDeveloperRole": false,
+					"supportsStore": false,
+					"maxTokensField": "max_tokens",
+					"reasoningContentField": "reasoning_content"
+				},
+				"thinking": {
+					"mode": "effort"
+				}
+			},
 			{
 				"source": "providers/cline-pass.kdl:3",
 				"providers": [
@@ -9761,11 +9817,15 @@
 				}
 			},
 			{
-				"source": "providers/deepseek.kdl:85",
+				"source": "providers/deepseek.kdl:89",
 				"providers": [
 					"deepseek"
 				],
 				"models": [
+					{
+						"kind": "exact",
+						"value": "deepseek-flash"
+					},
 					{
 						"kind": "exact",
 						"value": "deepseek-v4-flash"
@@ -9780,7 +9840,16 @@
 					"reasoningContentField": "reasoning_content",
 					"requiresAssistantContentForToolCalls": true,
 					"requiresReasoningContentForToolCalls": true,
+					"allowsSyntheticReasoningContentForToolCalls": false,
 					"supportsToolChoice": false
+				},
+				"thinking": {
+					"upgradeNeutral": true,
+					"efforts": [
+						"low",
+						"high",
+						"max"
+					]
 				}
 			},
 			{
@@ -18324,6 +18393,30 @@
 					"placeholder": "user_..."
 				}
 			},
+			{
+				"id": "charm-hyper",
+				"apiKeyFormat": "bearer",
+				"name": "Charm Hyper",
+				"env": {
+					"vars": [
+						"CHARM_HYPER_API_KEY",
+						"HYPER_API_KEY"
+					]
+				},
+				"login": {
+					"kind": "api-key",
+					"authUrl": "https://hyper.charm.land/",
+					"instructions": "Create or copy an API key from the Charm Hyper dashboard",
+					"prompt": "Paste your Charm Hyper API key",
+					"placeholder": "sk-hyper-...",
+					"normalize": "strip-bearer",
+					"validate": {
+						"kind": "models-endpoint",
+						"label": "Charm Hyper",
+						"url": "https://hyper.charm.land/v1/credits"
+					}
+				}
+			},
 			{
 				"id": "deepseek",
 				"apiKeyFormat": "bearer",

package/src/models.json

@@ -76097,7 +76097,7 @@
 			"api": "openai-completions",
 			"provider": "deepseek",
 			"baseUrl": "https://api.deepseek.com",
-			"reasoning": false,
+			"reasoning": true,
 			"input": [
 				"text"
 			],
@@ -76153,8 +76153,8 @@
 				"supportsUsageInStreaming": true,
 				"alwaysSendMaxTokens": false,
 				"disableReasoningOnForcedToolChoice": false,
-				"disableReasoningOnToolChoice": false,
-				"supportsToolChoice": true,
+				"disableReasoningOnToolChoice": true,
+				"supportsToolChoice": false,
 				"supportsForcedToolChoice": true,
 				"supportsNamedToolChoice": true,
 				"maxTokensField": "max_tokens",
@@ -76163,28 +76163,23 @@
 				"requiresThinkingAsText": false,
 				"requiresMistralToolIds": false,
 				"thinkingFormat": "openai",
-				"reasoningDisableMode": "lowest-effort",
+				"reasoningDisableMode": "zai-thinking-disabled",
 				"omitReasoningEffort": false,
 				"includeEncryptedReasoning": true,
 				"filterReasoningHistory": false,
 				"reasoningContentField": "reasoning_content",
-				"requiresReasoningContentForToolCalls": false,
-				"requiresReasoningContentForAllAssistantTurns": false,
-				"allowsSyntheticReasoningContentForToolCalls": true,
+				"requiresReasoningContentForToolCalls": true,
+				"requiresReasoningContentForAllAssistantTurns": true,
+				"allowsSyntheticReasoningContentForToolCalls": false,
 				"replayReasoningContent": false,
 				"qwenPreserveThinking": false,
 				"qwenTemplateReasoningEffort": false,
-				"requiresAssistantContentForToolCalls": false,
+				"requiresAssistantContentForToolCalls": true,
 				"supportsPromptCacheBreakpoints": false,
 				"isOpenRouterHost": false,
 				"wireModelIdMode": "raw",
 				"isVercelGatewayHost": false,
 				"supportsStrictMode": true,
-				"extraBody": {
-					"thinking": {
-						"type": "enabled"
-					}
-				},
 				"toolStrictMode": "mixed",
 				"stripDeepseekSpecialTokens": true,
 				"streamMarkupHealingPattern": "dsml",
@@ -76199,9 +76194,79 @@
 				"thinkingLoopGuard": "deepseek",
 				"rejectRootObjectUnion": false,
 				"retryWithoutStrictOnGrammarError": false,
-				"supportsPromptCacheKey": false
+				"supportsPromptCacheKey": false,
+				"streamIdleTimeoutMs": 300000,
+				"whenThinking": {
+					"supportsStore": false,
+					"supportsDeveloperRole": false,
+					"supportsMultipleSystemMessages": true,
+					"supportsReasoningEffort": true,
+					"supportsReasoningParams": true,
+					"supportsSamplingParams": true,
+					"supportsPenaltyAndStopParams": true,
+					"reasoningEffortMap": {},
+					"supportsUsageInStreaming": true,
+					"alwaysSendMaxTokens": false,
+					"disableReasoningOnForcedToolChoice": false,
+					"disableReasoningOnToolChoice": true,
+					"supportsToolChoice": false,
+					"supportsForcedToolChoice": true,
+					"supportsNamedToolChoice": true,
+					"maxTokensField": "max_tokens",
+					"requiresToolResultName": false,
+					"requiresAssistantAfterToolResult": false,
+					"requiresThinkingAsText": false,
+					"requiresMistralToolIds": false,
+					"thinkingFormat": "openai",
+					"reasoningDisableMode": "lowest-effort",
+					"omitReasoningEffort": false,
+					"includeEncryptedReasoning": true,
+					"filterReasoningHistory": false,
+					"reasoningContentField": "reasoning_content",
+					"requiresReasoningContentForToolCalls": true,
+					"requiresReasoningContentForAllAssistantTurns": true,
+					"allowsSyntheticReasoningContentForToolCalls": false,
+					"replayReasoningContent": false,
+					"qwenPreserveThinking": false,
+					"qwenTemplateReasoningEffort": false,
+					"requiresAssistantContentForToolCalls": true,
+					"supportsPromptCacheBreakpoints": false,
+					"isOpenRouterHost": false,
+					"wireModelIdMode": "raw",
+					"isVercelGatewayHost": false,
+					"supportsStrictMode": true,
+					"extraBody": {
+						"thinking": {
+							"type": "enabled"
+						}
+					},
+					"toolStrictMode": "mixed",
+					"streamIdleTimeoutMs": 300000,
+					"stripDeepseekSpecialTokens": true,
+					"streamMarkupHealingPattern": "dsml",
+					"reasoningDeltasMayBeCumulative": false,
+					"emptyLengthFinishIsContextError": false,
+					"usesOpenAIToolCallIdLimit": false,
+					"dropThinkingWhenReasoningEffort": false,
+					"nativeKimiK3Reasoning": false,
+					"zaiReasoningEffortDialect": false,
+					"clampOutputToModelMax": true,
+					"stripImageInput": true,
+					"thinkingLoopGuard": "deepseek",
+					"rejectRootObjectUnion": false,
+					"retryWithoutStrictOnGrammarError": false,
+					"supportsPromptCacheKey": false
+				}
 			},
-			"supportsComputerUse": false
+			"supportsComputerUse": false,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"low",
+					"high",
+					"max"
+				]
+			}
 		},
 		"deepseek-v4-flash": {
 			"id": "deepseek-v4-flash",

package/src/provider-models/cache-provider-id.ts

@@ -1,3 +1,4 @@
+import { CHARM_HYPER_API_BASE_URL, normalizeCharmHyperBaseUrl } from "../wire/charm-hyper";
 import { PERSONAL_GITHUB_COPILOT_BASE_URL } from "../wire/github-copilot";
 
 export interface ModelCacheProviderIdOptions {
@@ -19,6 +20,8 @@ export function isCredentialScopedModelCacheProvider(providerId: string): boolea
 
 export function getDefaultModelDiscoveryBaseUrl(providerId: string): string | undefined {
 	switch (providerId) {
+		case "charm-hyper":
+			return CHARM_HYPER_API_BASE_URL;
 		case "meta":
 		case "muse-code":
 			return "https://api.meta.ai/v1";
@@ -62,6 +65,23 @@ export function resolveModelCacheProviderId(providerId: string, options: ModelCa
 			// carry `requestModelId: *-low`, which the Start plan refuses; refetch
 			// so the collapsed default is re-pointed to `-medium` (issue #9478).
 			return "cursor:default-effort-v4";
+		case "charm-hyper": {
+			// Discovery is authoritative for this gateway, so a warm cache is served
+			// for its full TTL without re-probing: the namespace must follow the
+			// configured endpoint, or a self-hosted proxy keeps serving the canonical
+			// host's roster, capabilities and tariffs until expiry.
+			//
+			// Endpoint-only scope is deliberate. `/v1/models` is public here, so the
+			// roster does not vary by key, and `charm-hyper` is absent from
+			// CREDENTIAL_SCOPED_MODEL_CACHE_PROVIDERS — `ModelRegistry` resolves this
+			// namespace with no credential at all, so hashing one would split it
+			// against the namespace discovery computes and miss forever.
+			//
+			// Normalized through the shared helper because the registry passes the
+			// raw configured value while `charmHyperModelManagerOptions` passes a
+			// `/v1`-suffixed one; both must land on one namespace.
+			return `charm-hyper:models-v1:${Bun.hash(normalizeCharmHyperBaseUrl(options.baseUrl)).toString(36)}`;
+		}
 		case "muse-code": {
 			const baseUrl = options.baseUrl ?? getDefaultModelDiscoveryBaseUrl(providerId)!;
 			const scope = `${options.apiKey ?? ""}\u0000${baseUrl}`;

package/src/provider-models/descriptors.ts

@@ -18,6 +18,7 @@ import {
 	basetenModelManagerOptions,
 	bedrockMantleModelManagerOptions,
 	cerebrasModelManagerOptions,
+	charmHyperModelManagerOptions,
 	clinePassModelManagerOptions,
 	cloudflareAiGatewayModelManagerOptions,
 	commandCodeModelManagerOptions,
@@ -150,6 +151,26 @@ export const CATALOG_PROVIDERS = [
 		createModelManagerOptions: (config: ModelManagerConfig) => cerebrasModelManagerOptions(config),
 		catalogDiscovery: { label: "Cerebras" },
 	},
+	{
+		id: "charm-hyper",
+		defaultModel: "glm-5.3",
+		envVars: ["CHARM_HYPER_API_KEY", "HYPER_API_KEY"],
+		createModelManagerOptions: (config: ModelManagerConfig) => charmHyperModelManagerOptions(config),
+		allowUnauthenticated: true,
+		dynamicModelsAuthoritative: true,
+		// The gateway row is the whole truth for a Hyper deployment. Same-id rows
+		// on other hosts disagree with it in both directions (it serves
+		// non-thinking Kimi K2.5/K2.7-Code and a text-only Gemma 4 that their
+		// upstream homes list as reasoning/vision), so foreign backfills would
+		// advertise capabilities this deployment does not have.
+		skipCrossProviderReferenceFills: true,
+		// Deliberately NO `catalogDiscovery`: that field is what enrolls a provider
+		// in generate-models.ts. This gateway's catalog is live deployment truth,
+		// so generating would freeze one hyper.charm.land snapshot into
+		// models.json — and since discovery here needs no credentials, it would
+		// happen on every regen, contradicting the runtime-only contract
+		// compat-conformance.test.ts pins for this provider.
+	},
 	{
 		id: "cloudflare-ai-gateway",
 		defaultModel: "anthropic/claude-opus-4-8",

package/src/provider-models/openai-compat.ts

@@ -28,6 +28,7 @@ import { type GeneratedProvider, getBundledModels } from "../models";
 import type { Api, FetchImpl, Model, ModelSpec, OpenAICompat, Provider, ThinkingConfig, TokenCost } from "../types";
 import { discoveryFetch, isAnthropicOAuthToken, isRecord, toBoolean, toNumber, toPositiveNumber } from "../utils";
 import { ALIBABA_TOKEN_PLAN_BASE_URL, parseAlibabaTokenPlanCredential } from "../wire/alibaba-token-plan";
+import { normalizeCharmHyperBaseUrl } from "../wire/charm-hyper";
 import { CLINEPASS_API_BASE_URL, clinePassClientHeaders } from "../wire/cline-pass";
 import { CLOUDFLARE_AI_GATEWAY_COMPAT_BASE_URL } from "../wire/cloudflare-ai-gateway";
 import { coreWeaveProjectHeaders } from "../wire/coreweave";
@@ -7386,3 +7387,149 @@ export function commandCodeModelManagerOptions(config?: CommandCodeModelManagerC
 		},
 	};
 }
+
+// ---------------------------------------------------------------------------
+// Charm Hyper
+// ---------------------------------------------------------------------------
+
+/**
+ * Configuration for the Charm Hyper model manager.
+ *
+ * `baseUrl` overrides the gateway root for tests and self-hosted proxies; a
+ * value that omits the `/v1` surface gains one, so a host-only override
+ * behaves like every sibling provider's.
+ */
+export interface CharmHyperModelManagerConfig {
+	apiKey?: string;
+	baseUrl?: string;
+	fetch?: FetchImpl;
+}
+
+/**
+ * Charm Hyper's `/v1/models` row shape, verified live against
+ * hyper.charm.land (2026-09-11). The gateway is unusually complete: every row
+ * carries its own display name, context window, output cap, vision flag,
+ * accepted `reasoning_effort` vocabulary and per-million-token tariff, so
+ * discovery reads capabilities straight off the wire instead of borrowing a
+ * bundled reference from another host. Rows are served publicly — the
+ * endpoint answers 200 with no credentials at all.
+ */
+interface CharmHyperModelRecord extends OpenAICompatibleModelRecord {
+	display_name?: unknown;
+	context_window?: unknown;
+	max_output_tokens?: unknown;
+	capabilities?: unknown;
+	reasoning?: unknown;
+	pricing?: unknown;
+}
+
+/**
+ * Hyper's thinking-off wire tier. It is a disable state, not a rung: sending
+ * `reasoning_effort: "none"` suppresses reasoning outright (verified — zero
+ * reasoning tokens), while `minimal` still thinks. Advertising models
+ * therefore keep their whole ladder and route the off switch through
+ * `reasoningDisableMode`, the same split first-party GPT-5.6 uses. Efforts the
+ * gateway does not advertise are accepted and silently ignored rather than
+ * rejected, so the advertised vocabulary is the only reliable ladder.
+ */
+const CHARM_HYPER_WIRE_EFFORT_NONE = "none";
+
+/** Distinct `reasoning.effort_levels[].value` strings advertised for a row. */
+function charmHyperWireEfforts(reasoning: unknown): readonly string[] {
+	if (!isRecord(reasoning) || !Array.isArray(reasoning.effort_levels)) return [];
+	const values: string[] = [];
+	for (const level of reasoning.effort_levels) {
+		const value = isRecord(level) ? level.value : undefined;
+		if (typeof value === "string" && !values.includes(value)) values.push(value);
+	}
+	return values;
+}
+
+/**
+ * Build the effort ladder from the advertised vocabulary, preserving every
+ * named tier. `none` is deliberately excluded: it is the disable state, and
+ * folding it into `minimal` would make the lowest rung silently stop thinking
+ * on the models that advertise both.
+ */
+function resolveCharmHyperThinking(reasoning: unknown, wireEfforts: readonly string[]): ThinkingConfig | undefined {
+	const efforts = THINKING_EFFORTS.filter(effort => wireEfforts.includes(effort));
+	if (efforts.length === 0) return undefined;
+	const advertisedDefault = isRecord(reasoning) ? reasoning.default_effort_level : undefined;
+	const defaultLevel = efforts.find(effort => effort === advertisedDefault);
+	return { mode: "effort", efforts, ...(defaultLevel !== undefined && { defaultLevel }) };
+}
+
+/**
+ * Hyper quotes per-million-token USD directly, so rates pass through
+ * unscaled. `toPositiveNumber` is unusable here: a free tier and the common
+ * `cache_create: 0` are legitimate zero rates, not missing values.
+ */
+function toCharmHyperRate(value: unknown): number {
+	const parsed = toNumber(value);
+	return parsed !== undefined && parsed >= 0 ? parsed : 0;
+}
+
+function resolveCharmHyperCost(pricing: unknown): ModelSpec<"openai-completions">["cost"] {
+	if (!isRecord(pricing)) return { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 };
+	return {
+		input: toCharmHyperRate(pricing.input),
+		output: toCharmHyperRate(pricing.output),
+		cacheRead: toCharmHyperRate(pricing.cache_hit),
+		cacheWrite: toCharmHyperRate(pricing.cache_create),
+	};
+}
+
+/**
+ * Charm Hyper's gateway catalog. `/v1/models` is public and carries the live
+ * tariff, so discovery runs with or without a key and the snapshot is
+ * authoritative: a model the gateway stops serving is pruned rather than kept
+ * alive by a stale bundled row.
+ */
+export function charmHyperModelManagerOptions(
+	config?: CharmHyperModelManagerConfig,
+): ModelManagerOptions<"openai-completions"> {
+	const baseUrl = normalizeCharmHyperBaseUrl(config?.baseUrl);
+	return {
+		providerId: "charm-hyper",
+		cacheProviderId: resolveModelCacheProviderId("charm-hyper", { baseUrl }),
+		dynamicModelsAuthoritative: true,
+		fetchDynamicModels: () =>
+			fetchOpenAICompatibleModels({
+				api: "openai-completions",
+				provider: "charm-hyper",
+				baseUrl,
+				apiKey: config?.apiKey,
+				mapModel: (
+					entry: OpenAICompatibleModelRecord,
+					defaults: ModelSpec<"openai-completions">,
+				): ModelSpec<"openai-completions"> => {
+					const record = entry as CharmHyperModelRecord;
+					const wireEfforts = charmHyperWireEfforts(record.reasoning);
+					const thinking = resolveCharmHyperThinking(record.reasoning, wireEfforts);
+					const capabilities = isRecord(record.capabilities) ? record.capabilities : undefined;
+					return {
+						...defaults,
+						name: toModelName(record.display_name, defaults.name),
+						// A row without an `effort_levels` vocabulary exposes no dial.
+						// The gateway is silent rather than negative about always-on
+						// reasoners, so the handful that think anyway are corrected by
+						// exact `thinking-efforts` rules in KDL, which upgrade the
+						// target and materialize `reasoning: true` alongside them.
+						reasoning: thinking !== undefined,
+						...(thinking && { thinking }),
+						input: capabilities?.vision === true ? ["text", "image"] : ["text"],
+						contextWindow: toPositiveNumber(record.context_window, defaults.contextWindow),
+						maxTokens: toPositiveNumber(record.max_output_tokens, defaults.maxTokens),
+						cost: resolveCharmHyperCost(record.pricing),
+						// Thinking-capable rows that advertise the `none` tier can be
+						// switched off on the wire; the rest have no off switch and
+						// keep the dialect default.
+						...(thinking && wireEfforts.includes(CHARM_HYPER_WIRE_EFFORT_NONE)
+							? { compat: { reasoningDisableMode: "none-effort" as const } }
+							: {}),
+					};
+				},
+				fetch: config?.fetch,
+			}),
+	};
+}

package/src/wire/charm-hyper.ts

@@ -0,0 +1,26 @@
+/**
+ * Charm Hyper gateway endpoint, shared so a host migration — or a self-hosted
+ * proxy override — touches a single module.
+ */
+export const CHARM_HYPER_API_BASE_URL = "https://hyper.charm.land/v1";
+
+/**
+ * Resolve a configured Charm Hyper base URL onto the gateway's `/v1` surface.
+ *
+ * Every consumer must agree on this, because they key different things off the
+ * result: inference and discovery target it, `/usage` sends a bearer token to
+ * it, and the model cache namespace is hashed from it. Three separate copies
+ * previously disagreed for a blank value — the model manager treated it as
+ * absent and used the canonical host, while the others produced a bare `/v1` —
+ * so a whitespace-only override silently split inference, balance checks and
+ * caching across different endpoints.
+ *
+ * A blank or whitespace-only value therefore means "not configured" and
+ * resolves to the canonical host; anything else keeps its host and gains the
+ * `/v1` segment if it omits one.
+ */
+export function normalizeCharmHyperBaseUrl(baseUrl?: string): string {
+	const trimmed = baseUrl?.trim().replace(/\/+$/, "");
+	if (!trimmed) return CHARM_HYPER_API_BASE_URL;
+	return trimmed.endsWith("/v1") ? trimmed : `${trimmed}/v1`;
+}