@oh-my-pi/pi-catalog 18.12.4 → 18.12.6

package/CHANGELOG.md

@@ -2,6 +2,22 @@
 
 ## [Unreleased]
 
+## [18.12.6] - 2026-10-21
+
+### Added
+
+- Added the `prompt-cache-persistent` catalog axis (`isPromptCachePersistent`) for providers whose own endpoint never expires prompt caches; logged-in OpenAI Codex declares it
+
+### Changed
+
+- Updated the bundled Yolo-Auto catalog to its current models, `qwen3.8-27b` (now the default), `yolo`, `yolo-small`, and `yolo-128k`, and removed the retired `deepseek-flash-v4` and `qwen3.8-flash`
+
+## [18.12.5] - 2026-10-21
+
+### Fixed
+
+- Fixed Moonshot `kimi-k3` pricing cache writes at $0; usage cost now includes cache writes at Moonshot's $3/M 5-minute rate ([#16508](https://github.com/can1357/oh-my-pi/issues/16508))
+
 ## [18.12.4] - 2026-10-20
 
 ### Added

package/dist/types/compat/prompt-cache-persistent.d.ts

@@ -0,0 +1,7 @@
+import type { Model } from "../types.js";
+/**
+ * Whether {@link model}'s prompt-cache entries never expire
+ * (`prompt-cache-persistent` axis). Only the provider's own endpoint
+ * qualifies: a gateway or proxy reusing the provider id makes no such promise.
+ */
+export declare function isPromptCachePersistent(model: Model): boolean;

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

@@ -475,8 +475,8 @@ export interface YoloAutoModelManagerConfig {
     fetch?: FetchImpl;
 }
 /**
- * Yolo-Auto model manager: OpenAI-compatible chat completions at the
- * flat-rate `deepseek-flash-v4` endpoint. Live `/v1/models` discovery replaces the
+ * Yolo-Auto model manager: flat-rate OpenAI-compatible chat completions
+ * (`qwen3.8-27b`, `yolo`, ...). Live `/v1/models` discovery replaces the
  * bundled seed once a key is stored or present in `YOLO_AUTO_API_KEY`. Models
  * the provider adds later inherit metadata (reasoning, thinking, context)
  * from the global bundled reference index, so new ids work without a per-id

package/dist/types/wire/codex.d.ts

@@ -78,8 +78,10 @@ export declare function getCodexAccountId(accessToken: string): string | undefin
  * needed: `chatgpt_data_residency` is the authoritative claim, with
  * `chatgpt_compute_residency` as the fallback for tokens that only carry that.
  *
- * Returns undefined for a non-JWT token, or for the (common) accounts whose
- * claims omit residency entirely — those workspaces are not region-pinned.
+ * Returns undefined for a non-JWT token, or for the (common) accounts that are
+ * not region-pinned: their claims omit residency or carry the
+ * {@link CODEX_UNCONSTRAINED_RESIDENCY} sentinel. The backend answers 404 to a
+ * declared `no_constraint`, and codex-rs never sends it.
  */
 export declare function getCodexResidency(accessToken: string): string | undefined;
 /**

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "18.12.4",
+	"version": "18.12.6",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"keywords": [
 		"ai",
@@ -93,12 +93,12 @@
 		"gen:proto": "bun scripts/generate-protocols.ts"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.12.4",
-		"@oh-my-pi/pi-utils": "18.12.4"
+		"@oh-my-pi/omptype": "18.12.6",
+		"@oh-my-pi/pi-utils": "18.12.6"
 	},
 	"devDependencies": {
 		"@bgotink/kdl": "0.4.0",
-		"@oh-my-pi/pi-ai": "18.12.4",
+		"@oh-my-pi/pi-ai": "18.12.6",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/axes.ts

@@ -560,6 +560,13 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	 * Unassigned: no known lookback bound.
 	 */
 	"prompt-cache-lookback": { key: "promptCacheLookback", set: "catalog", shape: "scalar" },
+	/**
+	 * The provider's own endpoint keeps prompt-cache entries indefinitely
+	 * (logged-in Codex subscriptions): an idle gap never cools the prefix, so
+	 * rewriting it while idle only discards a warm cache. Read through
+	 * `compat/prompt-cache-persistent.ts`. Unassigned: entries expire.
+	 */
+	"prompt-cache-persistent": { key: "promptCachePersistent", set: "catalog", shape: "scalar", values: [true, false] },
 	"long-usage-limit-fallback": { key: "longUsageLimitFallback", set: "catalog", shape: "scalar" },
 	"max-context-window": { key: "maxContextWindow", set: "catalog", shape: "scalar" },
 	/**

package/src/compat/prompt-cache-persistent.ts

@@ -0,0 +1,12 @@
+import type { Model } from "../types";
+import { isCatalogEndpointUrl } from "./catalog-endpoint";
+import { resolveCatalogPolicy } from "./catalog-policy";
+
+/**
+ * Whether {@link model}'s prompt-cache entries never expire
+ * (`prompt-cache-persistent` axis). Only the provider's own endpoint
+ * qualifies: a gateway or proxy reusing the provider id makes no such promise.
+ */
+export function isPromptCachePersistent(model: Model): boolean {
+	return resolveCatalogPolicy(model).promptCachePersistent === true && isCatalogEndpointUrl(model, model.baseUrl);
+}

package/src/compat/rules/README.md

@@ -243,6 +243,8 @@ One object axis carries a computed form: `long-context-cost` accepts either the
 
 `prompt-cache-lookback` is how many block positions back from a cache breakpoint the model's prompt cache looks for an earlier request's entry (Claude: 20, counting the breakpoint; a run of consecutive `tool_use` or `tool_result` blocks is one position). The coding agent reads it through `resolvePromptCacheLookback` and keeps warm-cache tool-result pruning inside the window; unassigned means no known bound.
 
+`prompt-cache-persistent #true` marks a provider whose own endpoint keeps prompt-cache entries indefinitely (logged-in Codex accounts). The coding agent reads it through `isPromptCachePersistent` and skips idle compaction, which would only discard a prefix that stays warm; a custom `baseUrl` reusing the provider id does not qualify. Unassigned means entries expire.
+
 `task-sibling-prompt-cache #true` lets Task children use an automatically derived sibling cache key on the matching request model. Missing or `#false` keeps the provider's per-child default; explicit cache keys still win. The session resolves this catalog policy against the actual model for primary turns, handoff, and ephemeral requests that reuse the live context, including model changes, without sharing conversation IDs or response state. The automatic candidate is persisted separately from explicit keys and survives full forks, `/tan`, and session reopening; new and partial sessions discard it. A fork carrying an automatic candidate uses its own conversation ID on unsupported models, and re-enables the candidate when returning to an eligible model. Ordinary forks without an automatic candidate keep their existing inherited cache routing. Conventional compaction keeps its existing cache routing because its serialized summary prompt differs from the primary context. The opt-in covers Muse Code's `muse-spark-1.3`, `muse-spark-1.3-contributor`, `muse-wonder-1.0`, `muse-wonder-1.0-contributor`, `muse-wonder-full-1.0`, and `muse-wonder-full-1.0-contributor` over `openai-responses`. Meta documents shared-prefix `prompt_cache_key` routing for Responses. Warmed sibling cache reuse was measured on Spark 1.3 Contributor; the standard Spark variant shares this scope following maintainer guidance. A bounded synthetic probe observed cache reuse on all four Wonder variants; it does not establish production or latency benefits. Auxiliary-request latency has not been measured. Provider selectors also match custom providers that declare the same `providerType`; no benefit is established for custom endpoints.
 
 `image-max-edge` is the longest image edge, in pixels, a host passes to the model; it downscales larger images to fit first (Cursor: 1024 for Claude and GPT). Snapcompact reads it through `resolveImageMaxEdge` and renders frames no wider, so their glyphs reach the model at the size they were drawn; unassigned means the host keeps the image's size.

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

@@ -33,15 +33,15 @@ provider "moonshot" {
 			input-modalities "text" "image"
 		}
 	}
-	// stencil.so reports a $3 K3 cache write; Moonshot bills only cache hits
-	// ($0.30) and misses (the $3 input rate), with no cache-write charge.
+	// Official K3 rates: cache write is billed separately by TTL ($3 for the
+	// default 5min tier, $6 for 1h). omp sends no TTL, so the 5min rate applies.
 	// https://platform.moonshot.ai/docs/pricing/chat-k3
 	models "kimi-k3" {
 		cost-patch {
 			input 3.0
 			output 15.0
 			cache-read 0.3
-			cache-write 0.0
+			cache-write 3.0
 		}
 	}
 	// The /models roster carries no modality metadata for the named vision/vl

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

@@ -24,6 +24,9 @@ provider "openai-codex" {
 	// The backend decodes zstd request bodies, as the official client sends.
 	// Its rejection body is undocumented, so any 400/415 re-checks uncompressed.
 	request-encodings "zstd"
+	// Logged-in Codex accounts keep prompt-cache entries with no TTL, so
+	// idle compaction would only discard a prefix that stays warm.
+	prompt-cache-persistent #true
 
 	seed api="openai-codex-responses" base-url="https://chatgpt.com/backend-api" bundle="always" {
 		// Label only: the Codex backend ignores `tool.model` and runs gpt-image-2-codex.

package/src/compat/rules/providers/yolo-auto.kdl

@@ -1,37 +1,27 @@
 // Provider-wire compat for "yolo-auto"; wire constraints apply to every discovered model.
 
 provider "yolo-auto" {
-	default-model "qwen3.8-flash"
+	default-model "qwen3.8-27b"
 	env "YOLO_AUTO_API_KEY"
 	dynamic-models-authoritative #true
 	discovery label="Yolo-Auto"
 
-	// Documented Yolo-Auto catalog (yolo-auto.com/models, yolo-auto.com/docs):
-	// `qwen3.8-flash` is the default model on Free and paid plans at a 256K
-	// (262144) context window, and the paid `yolo` route is initially backed by
-	// the same Flash deployment. The flat-rate `/v1/models` response carries
-	// bare ids with no limit fields and is authoritative once discovery runs.
-	// Without provider-local seeds the bare `qwen3.8-flash` slug resolves
-	// through the global reference index, which prefers the largest bundled
-	// window (1M) and misreports the deployment cap. The compat block mirrors
-	// the provider's documented wire surface: the API speaks the Qwen
-	// chat-template dialect with `reasoning_effort` support and rejects the
-	// `developer` role and `store` param.
+	// Documented Yolo-Auto catalog (yolo-auto.com/models, yolo-auto.com/docs,
+	// 2026-10): `qwen3.8-27b`, `yolo`, `yolo-small` (256K) and `yolo-128k`
+	// (128K), all paid-only. `qwen3.8-27b` is the recommended default; the
+	// `yolo` route is currently backed by the same Qwen3.8-27B deployment
+	// (its server-side target can change). The flat-rate `/v1/models`
+	// response is authoritative once discovery runs. Without provider-local
+	// seeds the bare `qwen3.8-27b` slug resolves through the global reference
+	// index, which prefers the largest bundled window (1M) and misreports the
+	// deployment cap. The compat block mirrors the provider's documented wire
+	// surface: the API speaks the Qwen chat-template dialect with
+	// `reasoning_effort` support and rejects the `developer` role and `store`
+	// param. `yolo-small` and `yolo-128k` are documented only by id and
+	// context, and their `/v1/models` rows advertise no thinking ladder, so
+	// they seed as plain text routes until the provider documents more.
 	seed api="openai-completions" base-url="https://yolo-auto.com/v1" bundle="fallback" {
-		model "deepseek-flash-v4" name="DeepSeek Flash V4" {
-			reasoning #true
-			input "text" "image"
-			cost input=0 output=0 cache-read=0 cache-write=0
-			limits context=131072
-			thinking-mode "effort"
-			thinking-efforts "minimal" "low" "medium" "high" "xhigh" "max"
-			thinking-effort-map { minimal "low"; low "low"; medium "high"; high "high"; xhigh "max"; max "max" }
-			supports-developer-role #false
-			supports-store #false
-			supports-reasoning-effort #true
-			thinking-format "chat-template"
-		}
-		model "qwen3.8-flash" name="Qwen3.8 Flash" {
+		model "qwen3.8-27b" name="Qwen3.8-27B" {
 			reasoning #true
 			input "text" "image"
 			cost input=0 output=0 cache-read=0 cache-write=0
@@ -57,6 +47,22 @@ provider "yolo-auto" {
 			thinking-format "qwen-chat-template"
 			template-reasoning-effort #true
 		}
+		model "yolo-small" name="Yolo Small" {
+			reasoning #false
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=262144
+			supports-developer-role #false
+			supports-store #false
+		}
+		model "yolo-128k" name="Yolo 128K" {
+			reasoning #false
+			input "text"
+			cost input=0 output=0 cache-read=0 cache-write=0
+			limits context=131072
+			supports-developer-role #false
+			supports-store #false
+		}
 	}
 
 	supports-developer-role #false
@@ -64,17 +70,7 @@ provider "yolo-auto" {
 	// Live /v1/models advertises `thinking: minimal..xhigh` for the Qwen3.8
 	// rows; the cascade rule carries it into the regenerated bundle, whose
 	// thinking re-bake ignores seed-level ladders.
-	models "qwen3.8-flash" "yolo" "qwen3.8-27b" {
+	models "qwen3.8-27b" "yolo" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 	}
-	models "deepseek-flash-v4" {
-		disable-reasoning-on-tool-choice #true
-		max-tokens-field "max_completion_tokens"
-		requires-reasoning-content-for-all-assistant-turns #true
-		requires-reasoning-content-for-tool-calls #true
-		supports-reasoning-effort #true
-		thinking-efforts "minimal" "low" "medium" "high" "xhigh" "max"
-		thinking-format "chat-template"
-		thinking-mode "effort"
-	}
 }

package/src/compat/rules/taxonomy/qwen.kdl

@@ -12,7 +12,7 @@ class "qwen" {
 
     revision prefix="qwen"
 
-    override id="yolo-auto-yolo-flash-identity" provider="yolo-auto" model="yolo" class="qwen" revision="3.8" rationale="Yolo-Auto's paid yolo route is initially backed by the same Qwen3.8 Flash deployment as qwen3.8-flash; the opaque alias carries no lineage tokens" provenance="yolo-auto.com/models (2026-09)"
+    override id="yolo-auto-yolo-identity" provider="yolo-auto" model="yolo" class="qwen" revision="3.8" rationale="Yolo-Auto's paid yolo route is currently backed by the same Qwen3.8-27B deployment as qwen3.8-27b; the opaque alias carries no lineage tokens" provenance="yolo-auto.com/models (2026-10)"
     override id="prismml-bonsai-27b-qwen-3-6" glob="*bonsai-27b*" class="qwen" revision="3.6" rationale="Reviewed PrismML Bonsai and Ternary-Bonsai 27B lineage is Qwen3.6-27B; the wire basename may carry a prefix or GGUF quantization suffix" provenance="packages/coding-agent/src/config/model-discovery.ts; original reviewed alias e3aa6594e5"
     override id="prismml-bonsai-2-27b-qwen-3-8" glob="*bonsai-2-27b*" class="qwen" revision="3.8" rationale="Reviewed PrismML Bonsai 2 27B lineage is Qwen3.8-27B; the wire basename may carry a prefix or GGUF quantization suffix" provenance="packages/coding-agent/src/config/model-discovery.ts"
     override id="kilo-qwq-32b-family" provider="kilo" model="qwq-32b" logical="qwen/qwq-32b" class="qwen" rationale="The reviewed Kilo QwQ deployment belongs to the Qwen family despite its opaque product spelling" provenance="fixtures/llm-oracle/catalog/models.normalized.json"