@oh-my-pi/pi-catalog 18.2.7 → 18.2.9

package/src/compat/rules/README.md

@@ -50,14 +50,14 @@ class "anthropic" {
 
 Classification trims and lowercases the full model identifier. The **bare name** is the segment after its final `/`. Matcher tokens are also lowercased while parsing.
 
-| Node | Rank | Match |
-| --- | ---: | --- |
-| `exact "token"` | 4 | The whole bare name equals `token`. |
-| `bounded "token"` | 3 | The bare name equals `token`, or starts with it followed by `-`, `_`, `.`, `:`, or an ASCII digit. |
-| `namespace "token"` | 2 | A non-empty `/`-separated segment of the full identifier equals `token`. |
-| `namespace "token" bounded=#true` | 2 | Split the full identifier on `/`, `.`, and `:`; a segment must satisfy the bounded rule above. This is the only matcher property. |
-| `prefix "token"` | 1 | The bare name starts with `token`. |
-| `glob "pattern"` | 0 | An anchored `*` wildcard match over the bare name. `*` spans any substring; all non-wildcard text remains anchored in order. |
+| Node                              | Rank | Match                                                                                                                             |
+| --------------------------------- | ---: | --------------------------------------------------------------------------------------------------------------------------------- |
+| `exact "token"`                   |    4 | The whole bare name equals `token`.                                                                                               |
+| `bounded "token"`                 |    3 | The bare name equals `token`, or starts with it followed by `-`, `_`, `.`, `:`, or an ASCII digit.                                |
+| `namespace "token"`               |    2 | A non-empty `/`-separated segment of the full identifier equals `token`.                                                          |
+| `namespace "token" bounded=#true` |    2 | Split the full identifier on `/`, `.`, and `:`; a segment must satisfy the bounded rule above. This is the only matcher property. |
+| `prefix "token"`                  |    1 | The bare name starts with `token`.                                                                                                |
+| `glob "pattern"`                  |    0 | An anchored `*` wildcard match over the bare name. `*` spans any substring; all non-wildcard text remains anchored in order.      |
 
 A class match is ranked by `(matcher-kind rank, token byte length)`. The greatest tuple wins. Equal tuples from different classes are an ambiguity error; source order is not a tiebreak. If nothing matches, classification returns class `unknown` with no family or revision.
 
@@ -93,16 +93,16 @@ revision skip-bare "o1" "o3" "o4"
 
 Optional properties are:
 
-| Property | Shape and meaning |
-| --- | --- |
-| `provider` | Exact provider key, compared case-insensitively. A matching provider-specific override wins over a provider-agnostic one. |
-| `logical` | Corrected logical model identifier. |
-| `class` | Corrected class ID; a non-empty string. |
-| `family` | Corrected product-family ID; a non-empty string. |
-| `revision` | One to three unsigned 8-bit components separated by `.` or `-`. |
-| `effort` | `off`, `minimal`, `low`, `medium`, `high`, `xhigh`, or `max`. |
-| `thinking-variant` | Boolean marker for a separately exposed thinking sibling. |
-| `expires-at-ms` | Non-negative Unix time in milliseconds. The override is inactive when the observation time is at or after this value. |
+| Property           | Shape and meaning                                                                                                         |
+| ------------------ | ------------------------------------------------------------------------------------------------------------------------- |
+| `provider`         | Exact provider key, compared case-insensitively. A matching provider-specific override wins over a provider-agnostic one. |
+| `logical`          | Corrected logical model identifier.                                                                                       |
+| `class`            | Corrected class ID; a non-empty string.                                                                                   |
+| `family`           | Corrected product-family ID; a non-empty string.                                                                          |
+| `revision`         | One to three unsigned 8-bit components separated by `.` or `-`.                                                           |
+| `effort`           | `off`, `minimal`, `low`, `medium`, `high`, `xhigh`, or `max`.                                                             |
+| `thinking-variant` | Boolean marker for a separately exposed thinking sibling.                                                                 |
+| `expires-at-ms`    | Non-negative Unix time in milliseconds. The override is inactive when the observation time is at or after this value.     |
 
 The tuple `(provider, selector-kind, selector)` must also be unique, including provider-agnostic selectors. Active exact overrides take precedence over every glob. Within each selector kind, provider-scoped overrides precede provider-agnostic ones; matching globs rank by non-wildcard byte count. Equal-ranked globs are an ambiguity error, never resolved by declaration order. When no observation time is supplied, an expiring override remains active.
 
@@ -124,6 +124,7 @@ collapse {
     provider-alias "devin" "opus" "claude-opus-5"
 }
 ```
+
 `variant-family` declares one reviewed provider-scoped collapsed family: positional provider and logical id, `name=` display name, and a body of `members "a" "b" …` (wire ids in priority order), `route "<tier>" "<wire-id>"` per effort tier (`off` included), and optional `mode`, `efforts`, `default-level`, `default-member`, `retired-members`, `effort-budget "<tier>" <n>`, `requires-effort`, `suppress-when-off`, `no-thinking`, `preserve-absent-effort-routes`, and `extra-aliases`. A `{rev}` placeholder in the logical id makes the node a **template**: it is instantiated once per revision found in live ids (`gemini-{rev}-flash` matches `gemini-3.8-flash-low` → family `gemini-3.8-flash`), every wire id in the body and the `name=` carry the same placeholder, and an optional `revision=` constraint (`">=3.6"`) bounds the generations it applies to. A concrete family with the same instantiated id wins over the template. `provider-alias` maps one provider-scoped selector spelling onto a logical id without making it a family member.
 
 `thinking-suffix` accepts one non-empty suffix and no properties. `pair-token` declares bounded (possibly infix) tokens naming the thinking sibling of a live bare twin (`sonar-reasoning-pro` beside `sonar-pro`); it drives thinking-pair derivation only — never identity suffix collapse — and negated `no-`/`non-` forms never match. `effort-suffix` additionally requires `tier` with one of the effort values above, and may have `except-bare-prefix`. `routing-variant-suffix` takes one non-empty suffix followed by one or more provider IDs: a wire identifier carrying the suffix on one of those providers is a **routing variant** of its plain identifier — discovery derives base-model metadata from the plain bundled SKU while keeping the suffixed wire identifier for requests; routing variants never participate in effort collapse. `effort-lane-suffix` takes one non-empty lane suffix followed by one or more provider IDs, plus an optional `bare-prefix` gate: on a declared provider, an identifier ending in the lane suffix collapses the effort suffix wedged before the lane token while keeping the lane on the logical id. `effort-family` takes a provider, the canonical logical id, and zero or more exact aliases that fold onto it.
@@ -151,7 +152,7 @@ discovery {
 
 ## Cascade grammar
 
-A cascade document starts with `class` or `provider`. Every selector adds a conjunct to the current rule. `on` scopes by deployment provider; `on-api` scopes by request adapter, including custom provider names. Axis directives may appear directly in any permitted scope, and nested selector blocks may appear alongside them.
+A cascade document starts with `class`, `provider`, or `on-api`. Root `on-api` declares a transport contract independent of model identity and provider name. Every selector adds a conjunct to the current rule. `on` scopes by deployment provider; `on-api` scopes by request adapter, including custom provider names. Axis directives may appear directly in any permitted scope, and nested selector blocks may appear alongside them.
 
 ```kdl
 class "gemini" {
@@ -178,15 +179,15 @@ provider "openrouter" {
 
 ### Selectors and nesting
 
-| Selector | Form | Matching semantics |
-| --- | --- | --- |
-| `class` | `class "id" { ... }` | Exact class ID. At document root it may contain `on`, `on-api`, `family`, `revision`, and `models`. Under `provider` it may contain `family`, `revision`, and `models`. |
-| `provider` | `provider "id" { ... }` | Exact provider ID. It is root-only and may contain `class` and `models`. |
-| `on` | `on "provider-a" "provider-b" { ... }` | One or more provider IDs, combined as OR. It is allowed only under a root `class`, and may contain `family`, `revision`, and `models`. |
-| `on-api` | `on-api "adapter-a" "adapter-b" { ... }` | One or more request adapter IDs, combined as OR. It is allowed only under a root `class`, and may contain `family`, `revision`, and `models`. |
-| `family` | `family "id" { ... }` | Exact classified family ID. It may contain `revision` and `models`. A target with no family does not match. |
-| `revision` | `revision ">=2.5 <4" { ... }` | A non-empty, whitespace-separated conjunction of comparisons. It may contain `models`. A target with no revision does not match. |
-| `models` | `models "id" "vendor/*" { ... }` | One or more alternatives, combined as OR. It cannot contain another selector. `token="name"` matches an ASCII-case-insensitive token bounded by non-alphanumerics. |
+| Selector   | Form                                     | Matching semantics                                                                                                                                                             |
+| ---------- | ---------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| `class`    | `class "id" { ... }`                     | Exact class ID. At document root it may contain `on`, `on-api`, `family`, `revision`, and `models`. Under `provider` it may contain `family`, `revision`, and `models`.        |
+| `provider` | `provider "id" { ... }`                  | Exact provider ID. It is root-only and may contain `class` and `models`.                                                                                                       |
+| `on`       | `on "provider-a" "provider-b" { ... }`   | One or more provider IDs, combined as OR. It is allowed only under a root `class`, and may contain `family`, `revision`, and `models`.                                         |
+| `on-api`   | `on-api "adapter-a" "adapter-b" { ... }` | One or more request adapter IDs, combined as OR. At document root it may contain `class` and `models`; under a root `class` it may contain `family`, `revision`, and `models`. |
+| `family`   | `family "id" { ... }`                    | Exact classified family ID. It may contain `revision` and `models`. A target with no family does not match.                                                                    |
+| `revision` | `revision ">=2.5 <4" { ... }`            | A non-empty, whitespace-separated conjunction of comparisons. It may contain `models`. A target with no revision does not match.                                               |
+| `models`   | `models "id" "vendor/*" { ... }`         | One or more alternatives, combined as OR. It cannot contain another selector. `token="name"` matches an ASCII-case-insensitive token bounded by non-alphanumerics.             |
 
 Class, provider/`on`, `on-api`, and family selector values are compared exactly and case-sensitively to the structured resolve target. Revision operators are `>=`, `>`, `<=`, `<`, and `=`; operands have one to three dot-separated unsigned 8-bit components, omitted components zero.
 
@@ -401,17 +402,17 @@ Only `discovery` enrolls a provider in `generate-models.ts`; providers without i
 
 ### Seed rows
 
-A `seed` *defines* bundled rows for providers whose catalog cannot be discovered at generation time — credential-scoped rosters, unauthenticated regens, or models ahead of upstream catalogs. Every other stratum patches rows; this one authors them. Runtime model managers hand the rows to `staticModels` through `seedModels(provider)`; the generator bundles them per the seed's `bundle` policy. Values are literal — a seed never derives from another provider's row, and pricing is never borrowed.
+A `seed` _defines_ bundled rows for providers whose catalog cannot be discovered at generation time — credential-scoped rosters, unauthenticated regens, or models ahead of upstream catalogs. Every other stratum patches rows; this one authors them. Runtime model managers hand the rows to `staticModels` through `seedModels(provider)`; the generator bundles them per the seed's `bundle` policy. Values are literal — a seed never derives from another provider's row, and pricing is never borrowed.
 
 `seed` properties: `api` and `base-url` are per-row defaults (a `model` may override either with the same property names); `bundle` defaults to `always`; `precedence="seed"` is optional. `model` takes the wire id positionally, requires `name=`, and its body MUST declare `reasoning`, `input` (`"text"` and/or `"image"`), `cost` (all four per-million rates), and `limits` (`context=` / `max-tokens=`, an omitted limit is `null`); `supports-tools #true` is optional. Any other directive is an axis from the cascade vocabulary: thinking axes become the row's explicit `thinking` (then `thinking-mode` and `thinking-efforts` are both required), wire axes become its explicit `compat` and must apply to the row's API, and catalog axes are rejected because they stay rule-owned in the cascade block. Explicit `thinking`/`compat` on a seed row win over the cascade exactly as they do for any authored spec.
 
 `bundle` decides when the generator includes the rows:
 
-| Policy | Rows enter the bundle |
-| --- | --- |
-| `always` | Every regeneration. Same-id upstream/discovery rows win dedup. |
+| Policy     | Rows enter the bundle                                                     |
+| ---------- | ------------------------------------------------------------------------- |
+| `always`   | Every regeneration. Same-id upstream/discovery rows win dedup.            |
 | `fallback` | Only when the provider's authoritative catalog discovery did not succeed. |
-| `empty` | Only when no other source produced a row for the provider. |
+| `empty`    | Only when no other source produced a row for the provider.                |
 
 `precedence="seed"` prepends the rows after the previous-snapshot merge and cross-provider reference fills, so the authored row wins dedup and same-id rows on other hosts never overwrite its name or capabilities (QwenCloud Token Plan, Meta). The default `upstream` precedence appends before the snapshot merge, so the current seed — not a stale snapshot copy — is the fallback row.
 

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

@@ -10,4 +10,4 @@ login-order "openai-codex" "anthropic" "zai" "zai-coding-plan" "kimi-code" "open
 	"qianfan" "venice" "siliconflow" "siliconflow-cn" "synthetic" "nanogpt" "wafer-serverless" "coreweave" \
 	"vercel-ai-gateway" "cloudflare-ai-gateway" "litellm" "kilo" "zenmux" "opencode-zen" "opencode-go" \
 	"yolo-auto" "tavily" "kagi" "exa" "parallel" "typesafe" "ollama" "ollama-cloud" "lm-studio" "llama.cpp" "vllm" \
-	"gmi-cloud" "stencil"
+	"gmi-cloud" "stencil" "singularityapi-dev" "singularityapi-tech"

package/src/compat/rules/auth/singularityapi-dev.kdl

@@ -0,0 +1,22 @@
+// SingularityAPI universal inference gateway (api.singularityapi.dev): 300+
+// models behind one OpenAI-compatible endpoint, pay-as-you-go `sk-sapi-...`
+// bearer keys, per-million-token tariffs published per model. SingularityAPI
+// ships two unrelated products behind one brand — this gateway and the
+// slot-reserved DeepSeek lanes on api.singularityapi.tech — with disjoint
+// keys, dashboards, effort ladders, and billing, so each owns its own
+// provider id, env var, and login flow (neither key works on the other host,
+// verified 2026-09-22).
+// Key validation probes the non-billable `GET /v1/models`: it answers 401
+// `invalid_api_key` for a missing, malformed, or unknown key (verified
+// 2026-09-22), so a bad key is rejected without spending.
+auth "singularityapi-dev" {
+	name "SingularityAPI"
+	env "SINGULARITYAPI_DEV_API_KEY"
+	login "api-key" {
+		auth-url "https://app.singularityapi.dev"
+		instructions "Create an API key from the SingularityAPI dashboard, then paste it here"
+		prompt "Paste your SingularityAPI API key" placeholder="sk-sapi-..."
+		normalize "strip-bearer"
+		validate "models-endpoint" label="SingularityAPI" url="https://api.singularityapi.dev/v1/models"
+	}
+}

package/src/compat/rules/auth/singularityapi-tech.kdl

@@ -0,0 +1,23 @@
+// SingularityAPI reserved DeepSeek lanes (api.singularityapi.tech, a LiteLLM
+// front end in front of slot-reserved lanes) — the vendor's other product,
+// separated from api.singularityapi.dev because the two share no key, no
+// billing model, and no effort ladder. Usage here is charged against a booked
+// reservation slot rather than prepaid credit: a valid key with no active
+// slot answers 403 `permission_error` ("No active lane reservation for this
+// key", verified 2026-09-22), so login succeeds while a slot is idle and
+// inference needs one. Keys are lane-scoped `sk-...` bearer tokens issued
+// from the dashboard.
+// Key validation probes the non-billable `GET /v1/models`: it answers 401
+// for a missing, malformed, or unknown key (verified 2026-09-21), so a bad
+// key is rejected without spending.
+auth "singularityapi-tech" {
+	name "SingularityAPI Reserved Lanes"
+	env "SINGULARITYAPI_TECH_API_KEY"
+	login "api-key" {
+		auth-url "https://app.singularityapi.tech/compute/billing"
+		instructions "Create a key from the SingularityAPI lanes dashboard, then paste it here"
+		prompt "Paste your SingularityAPI lanes key" placeholder="sk-..."
+		normalize "strip-bearer"
+		validate "models-endpoint" label="SingularityAPI Reserved Lanes" url="https://api.singularityapi.tech/v1/models"
+	}
+}

package/src/compat/rules/classes/gpt-oss.kdl

@@ -13,4 +13,7 @@ class "gpt-oss" {
 	on "venice" "vercel-ai-gateway" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 	}
+	on "google-antigravity" "google-gemini-cli" {
+		supports-function-part-id #true
+	}
 }

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

@@ -5,7 +5,7 @@ class "xai" {
 	// on both the API and OAuth surfaces; the multiplier form tracks
 	// stencil.so price updates. Zero-cost subscription rows carry no tier.
 	on "xai" "xai-oauth" {
-		models "grok-4.3" "grok-4.5" "grok-4.6" "grok-build-0.1" "grok-4.20*" {
+		models "grok-4.3" "grok-4.5" "grok-4.6" "grok-4.7" "grok-build-0.1" "grok-4.20*" {
 			long-context-cost {
 				input-threshold 200000
 				input-threshold-inclusive #true
@@ -26,10 +26,10 @@ class "xai" {
 			}
 		}
 	}
-	// First-party xhigh-capable SKUs (grok-4.6, multi-agent) advertise xhigh.
+	// First-party xhigh-capable SKUs (grok-4.6, grok-4.7, multi-agent) advertise xhigh.
 	on "xai" "xai-oauth" {
 		family "grok" {
-			models "grok-4.6*" "grok-4.20-multi-agent*" {
+			models "grok-4.6*" "grok-4.7*" "grok-4.20-multi-agent*" {
 				thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 			}
 		}

package/src/compat/rules/providers/amazon-bedrock.kdl

@@ -1,7 +1,7 @@
 // Provider-wire compat for "amazon-bedrock"; regenerated from the frozen census using deployment contract selectors.
 
 provider "amazon-bedrock" {
-	default-model "us.anthropic.claude-opus-4-8"
+	default-model "us.anthropic.claude-opus-5-5"
 
 	// The Converse assembler positions content blocks by wire index
 	// (`amazon-bedrock.ts`), so a block can land above text already rendered.

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

@@ -1,7 +1,7 @@
 // Provider-wire compat for "anthropic"; regenerated from the frozen census using deployment contract selectors.
 
 provider "anthropic" {
-	default-model "claude-opus-4-8"
+	default-model "claude-opus-5-5"
 	env "ANTHROPIC_API_KEY"
 	discovery label="Anthropic"
 	web-search "anthropic"

package/src/compat/rules/providers/cloudflare-ai-gateway.kdl

@@ -1,7 +1,7 @@
 // Provider-wire compat for "cloudflare-ai-gateway"; regenerated from the frozen census using deployment contract selectors.
 
 provider "cloudflare-ai-gateway" {
-	default-model "anthropic/claude-opus-4-8"
+	default-model "anthropic/claude-opus-5"
 	env "CLOUDFLARE_AI_GATEWAY_API_KEY"
 	discovery label="Cloudflare AI Gateway"
 

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

@@ -9,7 +9,7 @@
 // dials to a few further ids — MiniMaxAI/MiniMax-M3 and meta/muse-spark-1.3's
 // `max` — which stay out of this change.
 provider "commandcode" {
-	default-model "claude-sonnet-4-6"
+	default-model "claude-sonnet-5"
 	env "COMMAND_CODE_API_KEY" "COMMANDCODE_API_KEY"
 	allow-unauthenticated #true
 	dynamic-models-authoritative #true

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

@@ -1,7 +1,7 @@
 // Provider-wire compat for "cursor"; regenerated from the frozen census using deployment contract selectors.
 
 provider "cursor" {
-	default-model "claude-4.6-opus-high"
+	default-model "claude-opus-5-high"
 	env "CURSOR_ACCESS_TOKEN"
 	discovery label="Cursor" oauth-provider="cursor" {
 		env "CURSOR_API_KEY"

package/src/compat/rules/providers/google-antigravity.kdl

@@ -17,6 +17,11 @@ provider "google-antigravity" {
 		}
 	}
 
+	// Routed wire profiles overwrite caller caps with fixed output limits.
+	// Conservatively reject requested bounds for this provider, including
+	// logical ids whose effort routing selects a profiled wire id later.
+	preserves-max-output-tokens #false
+
 	class "anthropic" {
 		family "opus" {
 			thinking-mode "budget"

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

@@ -1,7 +1,7 @@
 // Provider-wire compat for "kilo"; regenerated from the frozen census using deployment contract selectors.
 
 provider "kilo" {
-	default-model "anthropic/claude-opus-4.8"
+	default-model "anthropic/claude-opus-5"
 	env "KILO_API_KEY"
 	discovery label="Kilo Gateway" allow-unauthenticated=#true
 

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

@@ -1,7 +1,7 @@
 // Catalog entry for "litellm"; no deployment-contract compat rules are needed.
 
 provider "litellm" {
-	default-model "claude-opus-4-8"
+	default-model "claude-opus-5-5"
 	env "LITELLM_API_KEY"
 	discovery label="LiteLLM" allow-unauthenticated=#true
 }

package/src/compat/rules/providers/native-tools.kdl

@@ -0,0 +1,4 @@
+// Cursor always exposes server-owned native operations, even with an empty tool catalog.
+on-api "cursor-agent" {
+	requires-native-tools #true
+}

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

@@ -4,7 +4,9 @@ provider "openai" {
 	default-model "gpt-5.5"
 	env "OPENAI_API_KEY"
 	kind-apis {
+		embedding "openai-embeddings"
 		image "openai-responses"
+		stt "openai-transcriptions"
 	}
 
 	// Daybreak models are approval-gated first-party Responses models that are
@@ -33,6 +35,62 @@ provider "openai" {
 			cost input=12.5 output=75 cache-read=1.25 cache-write=15.625
 			limits context=400000 max-tokens=128000
 		}
+
+		// Catalog cost fields are per 1M tokens: GPT transcribers map documented
+		// audio-input/output-token rates directly. Whisper's $0.006/min duration
+		// tariff has no token-cost representation, so it remains zero and the
+		// transcription response's provider-reported usage/cost is authoritative.
+		model "whisper-1" name="Whisper 1" api="openai-transcriptions" base-url="https://api.openai.com/v1" {
+			reasoning #false
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits
+			supports-tools #false
+		}
+		model "gpt-4o-transcribe" name="GPT-4o Transcribe" api="openai-transcriptions" base-url="https://api.openai.com/v1" {
+			reasoning #false
+			input "text"
+			cost input=2.5 output=10 cache-read=0 cache-write=0
+			limits context=16000 max-tokens=2000
+			supports-tools #false
+		}
+		model "gpt-4o-mini-transcribe" name="GPT-4o Mini Transcribe" api="openai-transcriptions" base-url="https://api.openai.com/v1" {
+			reasoning #false
+			input "text"
+			cost input=1.25 output=5 cache-read=0 cache-write=0
+			limits context=16000 max-tokens=2000
+			supports-tools #false
+		}
+		// OpenAI publishes embedding prices per million input tokens; embedding
+		// responses have no output-token charge.
+		model "text-embedding-3-small" name="Text Embedding 3 Small" api="openai-embeddings" base-url="https://api.openai.com/v1" {
+			reasoning #false
+			input "text"
+			cost input=0.02 output=0 cache-read=0 cache-write=0
+			limits context=8192
+			supports-tools #false
+		}
+		model "text-embedding-3-large" name="Text Embedding 3 Large" api="openai-embeddings" base-url="https://api.openai.com/v1" {
+			reasoning #false
+			input "text"
+			cost input=0.13 output=0 cache-read=0 cache-write=0
+			limits context=8192
+			supports-tools #false
+		}
+		model "text-embedding-ada-002" name="Text Embedding Ada 002" api="openai-embeddings" base-url="https://api.openai.com/v1" {
+			reasoning #false
+			input "text"
+			cost input=0.1 output=0 cache-read=0 cache-write=0
+			limits context=8192
+			supports-tools #false
+		}
+	}
+
+	models "whisper-1" "gpt-4o-transcribe" "gpt-4o-mini-transcribe" {
+		kind "stt"
+	}
+	models "text-embedding-3-small" "text-embedding-3-large" "text-embedding-ada-002" {
+		kind "embedding"
 	}
 
 	// Replaces the strict-mode provider whitelist entry.

package/src/compat/rules/providers/opencode-zen.kdl

@@ -1,7 +1,7 @@
 // Provider-wire compat for "opencode-zen"; regenerated from the frozen census using deployment contract selectors.
 
 provider "opencode-zen" {
-	default-model "claude-opus-4-8"
+	default-model "claude-opus-5"
 	env "OPENCODE_API_KEY"
 	dynamic-models-authoritative #true
 

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

@@ -5,7 +5,12 @@ provider "openrouter" {
 	env "OPENROUTER_API_KEY"
 	discovery label="OpenRouter" allow-unauthenticated=#true
 	kind-apis {
+		embedding "openai-embeddings"
 		image "openrouter-images"
+		rerank "openrouter-rerank"
+		video "openrouter-video"
+		tts "openai-speech"
+		stt "openai-transcriptions"
 	}
 	web-search "openrouter"
 
@@ -19,10 +24,113 @@ provider "openrouter" {
 			cost input=0.042 output=0 cache-read=0 cache-write=0
 			limits context=32000 max-tokens=28800
 		}
+
+		// OpenRouter's STT roster mixes per-token and duration-based prices.
+		// Token-priced GPT rows map per-token rates to catalog per-million costs;
+		// duration-priced rows remain zero because ModelCost has no seconds axis.
+		model "openai/whisper-1" name="OpenAI: Whisper 1" api="openai-transcriptions" base-url="https://openrouter.ai/api/v1" {
+			reasoning #false
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits
+			supports-tools #false
+		}
+		model "openai/whisper-large-v3" name="OpenAI: Whisper Large V3" api="openai-transcriptions" base-url="https://openrouter.ai/api/v1" {
+			reasoning #false
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits
+			supports-tools #false
+		}
+		model "openai/gpt-4o-transcribe" name="OpenAI: GPT-4o Transcribe" api="openai-transcriptions" base-url="https://openrouter.ai/api/v1" {
+			reasoning #false
+			input "text"
+			cost input=2.5 output=10 cache-read=0 cache-write=0
+			limits context=128000 max-tokens=115200
+			supports-tools #false
+		}
+		model "microsoft/mai-transcribe-1.5" name="Microsoft AI: MAI-Transcribe 1.5" api="openai-transcriptions" base-url="https://openrouter.ai/api/v1" {
+			reasoning #false
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits
+			supports-tools #false
+		}
+		model "microsoft/mai-transcribe-2" name="Microsoft AI: MAI-Transcribe 2" api="openai-transcriptions" base-url="https://openrouter.ai/api/v1" {
+			reasoning #false
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits
+			supports-tools #false
+		}
+
+		// OpenRouter bills reranking per 1,000 searches, which has no catalog cost
+		// axis. Keep token costs at zero; the provider-reported response cost is authoritative.
+		model "cohere/rerank-v3.5" name="Cohere: Rerank v3.5" api="openrouter-rerank" base-url="https://openrouter.ai/api/v1" {
+			reasoning #false
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=4096 max-tokens=3686
+			supports-tools #false
+		}
+		// Bundled fallbacks keep the gateway useful offline; live
+		// `/embeddings/models` discovery refreshes this roster and its per-token
+		// pricing. Catalog token costs are per million input tokens.
+		model "openai/text-embedding-3-small" name="OpenAI: Text Embedding 3 Small" api="openai-embeddings" base-url="https://openrouter.ai/api/v1" {
+			reasoning #false
+			input "text"
+			cost input=0.02 output=0 cache-read=0 cache-write=0
+			limits context=8192
+			supports-tools #false
+		}
+		model "qwen/qwen3-embedding-8b" name="Qwen: Qwen3 Embedding 8B" api="openai-embeddings" base-url="https://openrouter.ai/api/v1" {
+			reasoning #false
+			input "text"
+			cost input=0.01 output=0 cache-read=0 cache-write=0
+			limits context=32768
+			supports-tools #false
+		}
+		// OpenRouter bills generated video by output second and resolution/SKU,
+		// which ModelCost cannot represent. Keep token costs at zero; poll-reported
+		// cost is authoritative. Live `/videos/models` discovery refreshes the roster.
+		model "google/veo-3.1" name="Google: Veo 3.1" api="openrouter-video" base-url="https://openrouter.ai/api/v1" {
+			reasoning #false
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits
+			supports-tools #false
+		}
+		model "minimax/hailuo-3" name="MiniMax: H3" api="openrouter-video" base-url="https://openrouter.ai/api/v1" {
+			reasoning #false
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits
+			supports-tools #false
+		}
+		model "alibaba/wan-2.7" name="Alibaba: Wan 2.7" api="openrouter-video" base-url="https://openrouter.ai/api/v1" {
+			reasoning #false
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits
+			supports-tools #false
+		}
 	}
 	models "~typesafe/*" "typesafe/*" {
 		kind "judge"
 	}
+	models "cohere/rerank-v3.5" {
+		kind "rerank"
+	}
+	models "openai/text-embedding-3-small" "qwen/qwen3-embedding-8b" {
+		kind "embedding"
+	}
+	models "google/veo-3.1" "minimax/hailuo-3" "alibaba/wan-2.7" {
+		kind "video"
+	}
+	models "openai/whisper-1" "openai/whisper-large-v3" "openai/gpt-4o-transcribe" \
+		"microsoft/mai-transcribe-1.5" "microsoft/mai-transcribe-2" {
+		kind "stt"
+	}
 
 	// Replaces the OpenRouter provider wire-model-id dispatch branch.
 	wire-model-id-mode "openrouter"

package/src/compat/rules/providers/output-limits.kdl

@@ -0,0 +1,4 @@
+// These transports do not encode caller output limits, regardless of provider or model name.
+on-api "openai-codex-responses" "cursor-agent" "gitlab-duo-agent" {
+	omit-max-output-tokens #true
+}

package/src/compat/rules/providers/singularityapi-dev.kdl

@@ -0,0 +1,92 @@
+// SingularityAPI universal inference gateway (api.singularityapi.dev):
+// one OpenAI-compatible endpoint for 300+ models (DeepSeek, Kimi, GLM,
+// frontier flagships), `sapi_...` bearer keys issued from the dashboard at
+// app.singularityapi.dev. `GET /v1/models` publishes each row's per-endpoint
+// capabilities — context window, max/default output tokens, and per-million
+// 12-decimal pricing strings — with `cache-control: no-store`, so discovery
+// reads limits and tariffs live and no rows are frozen here.
+//
+// Deliberately no `discovery` node: live discovery needs a user credential,
+// and enrolling would freeze one account's snapshot into models.json on every
+// credential-less regen. The provider stays runtime-only and
+// `dynamic-models-authoritative` prunes stale rows.
+// The vendor's second product — slot-reserved DeepSeek lanes on
+// api.singularityapi.tech — lives in `singularityapi-tech`: separate provider
+// id because the two gateways publish conflicting effort ladders (`max`
+// passes there and 400s here; `medium` is the reverse, both measured live
+// 2026-09-22) and neither key works on the other host.
+provider "singularityapi-dev" {
+	default-model "deepseek-v4-flash"
+	env "SINGULARITYAPI_DEV_API_KEY"
+	dynamic-models-authoritative #true
+
+	// Gateway-wide request shape, from the OpenAI-compatibility contract
+	// (docs.singularityapi.dev/get-started/open-ai-compatibility): chat
+	// completions accept `max_tokens` and `max_completion_tokens`
+	// interchangeably; the gateway normalizes to the model's canonical
+	// field. Unknown/reserved fields 400 instead of dropping, duplicate
+	// JSON keys and NaN/Infinity 400, context overflow 400s before spend.
+	// Rows the `models` rules below do not review inherit exactly this
+	// shape instead of the openai-completions default, because the roster
+	// is account-scoped and unreviewed model ids are expected.
+	max-tokens-field "max_tokens"
+
+	// Reviewed lane rules for the DeepSeek rows the gateway publishes,
+	// keyed as case-insensitive globs (`models` exact selectors compare the
+	// raw id verbatim while the compiler lowercases them, so exact rules
+	// can never match mixed-case wire ids). DeepSeek V4 Flash rows classify
+	// deepseek/flash by taxonomy but publish no reasoning metadata, so the
+	// upgrade rule owns the measured low/medium/high ladder; reasoning arrives
+	// as top-level `reasoning_content` on messages and streamed deltas, and
+	// `none` forces it off. V4 Pro the same, with the Pro ladder.
+	//
+	// Ladder measured against the live gateway 2026-09-22: `none`, `minimal`,
+	// `low`, `medium`, `high` pass admission while `xhigh`, `max`, integer and
+	// unknown efforts answer 400 `invalid_request` — admission validation runs
+	// before the credit reservation, so the probe sorted accepted from rejected
+	// at 402 vs 400 without spending a cent. `max` in particular is a
+	// reserve-lane-only tier (`singularityapi-tech`); sending it here turned
+	// every DeepSeek request into a 400.
+	models "*deepseek-v4-flash*" {
+		thinking-upgrade-neutral #true
+		max-tokens-field "max_tokens"
+		thinking-efforts "low" "medium" "high"
+		reasoning-disable-mode "none-effort"
+		reasoning-content-field "reasoning_content"
+	}
+	models "*deepseek-v4-pro*" {
+		thinking-upgrade-neutral #true
+		max-tokens-field "max_tokens"
+		thinking-efforts "low" "medium" "high"
+		reasoning-disable-mode "none-effort"
+		reasoning-content-field "reasoning_content"
+	}
+	// Responses-capable flagships (gpt-5.6-sol/terra/luna): five-tier
+	// low..max ladder with the wire-exact off switch, matching the
+	// first-party GPT-5.6 rows. `thinking-upgrade-neutral` is what makes the
+	// ladder real: the wire publishes no reasoning metadata, so without it the
+	// rows stay non-reasoning and omp sends no `reasoning_effort` at all —
+	// which the gateway rejects (400 from upstream) on chat completions
+	// whenever `tools` are present, exactly the case a coding agent hits.
+	models "gpt-5.6-sol" "gpt-5.6-terra" "gpt-5.6-luna" {
+		thinking-upgrade-neutral #true
+		thinking-efforts "low" "medium" "high" "xhigh" "max"
+		reasoning-disable-mode "none-effort"
+	}
+	// Gateway image models serve POST /v1/images/generations, never billed
+	// by tokens: FLUX per megapixel, GPT Image per image. The `kind` rules
+	// keep them out of chat; `mapSingularityApiModel` reads the same rows'
+	// `capabilities` and assigns the `openai-images` transport, so a new
+	// image id the roster adds works without a rule edit here.
+	models "flux-1-schnell" "flux-pro-1.1" {
+		kind "image"
+	}
+	models "gpt-image-2" "gpt-image-1.5" {
+		kind "image"
+	}
+	// The gateway's DeepSeek rows reject tool_choice: the API defaults to
+	// auto when tools exist. Matches the first-party DeepSeek transport.
+	models "*deepseek*" {
+		supports-tool-choice #false
+	}
+}