@oh-my-pi/pi-catalog 18.2.8 → 18.2.9

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
+	}
+}

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

@@ -0,0 +1,72 @@
+// SingularityAPI reserved DeepSeek lanes (api.singularityapi.tech, a LiteLLM
+// front end for slot-reserved lanes). This is the vendor's other product —
+// see `singularityapi-dev` for the pay-as-you-go universal gateway — split
+// out because the two gateways publish conflicting effort ladders: `max` is
+// valid here and400s there, `medium` is rejected here and accepted there
+// (both measured live 2026-09-22). Usage bills against a booked reservation
+// slot, so a valid key with no active slot answers 403 `permission_error`
+// ("No active lane reservation for this key", verified 2026-09-22): login
+// succeeds while a slot is idle, inference waits for one.
+//
+// The live `/v1/models` roster owns membership: keys are lane-scoped `sk-...`
+// bearer tokens issued from the dashboard, and a bogus key is rejected with
+// 401 `token_not_found_in_db` (verified 2026-09-21), so no rows are frozen
+// here. Deliberately no `discovery` node: that node enrolls a provider in
+// generate-models.ts, which would freeze one account's lane snapshot into
+// models.json on every credential-less regen. The roster is per-key
+// credential-scoped, so `singularityApiTechModelManagerOptions` keys its
+// authoritative cache on both the resolved credential and the endpoint;
+// `dynamic-models-authoritative` prunes stale rows.
+provider "singularityapi-tech" {
+	default-model "deepseek-ai/DeepSeek-V4.1-Flash"
+	env "SINGULARITYAPI_TECH_API_KEY"
+	dynamic-models-authoritative #true
+
+	// Gateway-wide request shape. The deployment fronts DeepSeek lanes only and
+	// its front end takes `max_tokens`, returning reasoning as top-level
+	// `reasoning_content` on the final message and on streamed deltas (verified
+	// live 2026-09-22). A lane row that no `models` rule below matches inherits
+	// exactly this shape instead of the openai-completions default
+	// (`max_completion_tokens`), because the roster is per-key and unreviewed
+	// lanes are expected.
+	max-tokens-field "max_tokens"
+	reasoning-content-field "reasoning_content"
+
+	// Model guide (dashboard): the lane runs DeepSeek V4.1 Flash. Send
+	// `deepseek-ai/DeepSeek-V4.1-Flash`; the short alias
+	// `deepseek-v4.1-flash` is accepted too, and the account also serves the
+	// dated V4 Flash row. Context is 262,144 tokens prompt+completion (not the
+	// documented 1M); an overflowing max_tokens is reduced automatically, and a
+	// prompt within ~1K of the limit 400s. Both ids classify deepseek/flash by
+	// taxonomy, but bare `/v1/models` rows carry no reasoning metadata, so
+	// neutral discovery would leave the effort dial dark.
+	// Glob, not `exact`: one lane ships as a `deepseek-ai/…` id, a dated `-0731`
+	// row and a short alias, and cascade globs match the lowercased id, so one
+	// pattern covers each spelling where `exact` compares the raw id verbatim.
+	models "*deepseek-v4-flash-0731" "*deepseek-v4.1-flash" {
+		thinking-upgrade-neutral #true
+		// Effort vocabulary measured against the live gateway 2026-09-22:
+		// `none`, `minimal`, `low`, `high`, `xhigh`, `max` pass admission;
+		// `medium` and integer efforts 400 even though the gateway's own
+		// rejected-variant text lists `medium` as valid, so the guide's
+		// integer 1-100 scale does not survive the forward path. The ladder
+		// below is the accepted named set, `none` owns the off position.
+		thinking-efforts "low" "high" "xhigh" "max"
+		reasoning-disable-mode "none-effort"
+		// Guide: an overflowing max_tokens is reduced automatically, so clamp the
+		// request to the lane's ceiling rather than let the gateway answer 400.
+		// Lane-scoped, not provider-wide: for a row whose ceiling discovery leaves
+		// unknown the clamp falls back to the OpenAI default ceiling, which is not
+		// something an unreviewed lane should inherit.
+		clamp-output-to-model-max #true
+		// Guide: up to 8 image_url parts per request (PNG/JPEG, data URL or
+		// public https). Declared here as well as clearing the class strip:
+		// live discovery seeds input ["text"], and the wire guard requires the
+		// declared modality.
+		input-modalities "text" "image"
+		strip-image-input #false
+		limits-patch {
+			context-window 262144
+		}
+	}
+}

package/src/compat/rules/providers/tool-free-history.kdl

@@ -0,0 +1,6 @@
+// Bedrock injects a callable sentinel when tool-free requests retain tool history, but the
+// sentinel is model-visible and calling it ends the turn: a genuine tool opt-out needs
+// history without prior tool calls or results. Applies to every routed Bedrock deployment.
+on-api "bedrock-converse-stream" {
+	requires-tool-free-history-for-tool-opt-out #true
+}

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

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

package/src/compat/rules/providers/xai-oauth.kdl

@@ -58,6 +58,13 @@ provider "xai-oauth" {
 			cost input=0 output=0 cache-read=0 cache-write=0
 			limits context=500000 max-tokens=500000
 		}
+		// 500K context per docs.x.ai/developers/models/grok-4.7.
+		model "grok-4.7" name="Grok 4.7" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=500000 max-tokens=500000
+		}
 		// Text-only per the bundled catalog.
 		model "grok-4.20-multi-agent-0309" name="Grok 4.20 (Multi-Agent)" {
 			reasoning #true
@@ -144,7 +151,7 @@ provider "xai-oauth" {
 		supports-reasoning-effort #false
 	}
 	// xhigh-capable SKUs keep xhigh unmapped.
-	models "grok-4.6*" "grok-4.20-multi-agent*" priority=1 {
+	models "grok-4.6*" "grok-4.7*" "grok-4.20-multi-agent*" priority=1 {
 		reasoning-effort-map {
 			minimal "low"
 		}

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

@@ -49,14 +49,14 @@ provider "xai" {
 			thinking-mode "effort"
 		}
 	}
-	// Dormant until a catalog snapshot ships grok-4.6; xhigh is native there.
+	// Dormant until a catalog snapshot ships grok-4.7; xhigh is native there.
 	// The effort-capable allowlist accepts the wire reasoning.effort dial;
-	// xhigh-capable SKUs (4.6, multi-agent) keep xhigh unmapped.
-	models "grok-3-mini*" "grok-4.20-multi-agent*" "grok-4.3*" "grok-4.5*" "grok-4.6*" {
+	// xhigh-capable SKUs (4.6, 4.7, multi-agent) keep xhigh unmapped.
+	models "grok-3-mini*" "grok-4.20-multi-agent*" "grok-4.3*" "grok-4.5*" "grok-4.6*" "grok-4.7*" {
 		supports-reasoning-effort #true
 		omit-reasoning-effort #false
 	}
-	models "grok-4.20-multi-agent*" "grok-4.6*" priority=1 {
+	models "grok-4.20-multi-agent*" "grok-4.6*" "grok-4.7*" priority=1 {
 		reasoning-effort-map {
 			minimal "low"
 		}

package/src/compat/rules/providers/xiaomi-token-plan-cn.kdl

@@ -4,6 +4,29 @@ provider "xiaomi-token-plan-cn" {
 	default-model "mimo-v2.5"
 	env "XIAOMI_TOKEN_PLAN_CN_API_KEY"
 
+	// Xiaomi's credential-scoped roster returns bare ids, so these documented
+	// V2.6 capabilities must remain authoritative across online refreshes.
+	seed api="openai-completions" base-url="https://token-plan-cn.xiaomimimo.com/v1" bundle="always" precedence="seed" {
+		model "mimo-v2.6-pro" name="MiMo-V2.6-Pro" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=1048576 max-tokens=131072
+		}
+		model "mimo-v2.6-flash" name="MiMo-V2.6-Flash" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=1048576 max-tokens=131072
+		}
+		model "mimo-v2.6-pro-ultraspeed" name="MiMo-V2.6-Pro-Ultraspeed" {
+			reasoning #true
+			input "text" "image"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=1048576 max-tokens=131072
+		}
+	}
+
 	class "mimo" {
 		family "v2" {
 			thinking-mode "effort"

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

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

package/src/compat/rules/runtime/behavior.kdl

@@ -197,6 +197,7 @@ behavior {
 	// at /responses (#10610, #12030). The Muse prefix covers gateway-first
 	// revisions; exact ids stay as cache-migration identity.
 	api-routes provider="opencode-zen" {
+		route "anthropic-messages" exact="union-alpha"
 		route "openai-responses" prefix="muse-spark-" exact="muse-spark-1.3-contributor-free" exact="gpt-6-astra"
 		route "openai-completions" exact="minimax-m3" exact="minimax-m3-free"
 	}
@@ -204,6 +205,7 @@ behavior {
 	// responses pins are gateway-verified (deepseek-v4-flash 2026-08-08;
 	// muse-spark per opencode.ai/docs/go/#endpoints, #8957, #10610).
 	api-routes provider="opencode-go" {
+		route "anthropic-messages" exact="union-alpha"
 		route "openai-responses" prefix="muse-spark-" exact="deepseek-v4-flash" exact="muse-spark-1.2" \
 			exact="muse-spark-1.2-contributor" exact="muse-spark-1.3" exact="muse-spark-1.3-contributor"
 		route "openai-completions" exact="minimax-m2.7" exact="minimax-m3" exact="minimax-m3-free" \