@oh-my-pi/pi-catalog 18.13.3 → 18.14.1

package/CHANGELOG.md

@@ -2,6 +2,41 @@
 
 ## [Unreleased]
 
+## [18.14.1] - 2026-10-31
+
+### Changed
+
+- Clamping a requested thinking effort now resolves through the model's `effortMap` alias first, so the request lands on the tier the model serves instead of the nearest lower one ([#17395](https://github.com/can1357/oh-my-pi/pull/17395) by [@Maziar123](https://github.com/Maziar123))
+- Long image-heavy conversations on the official Anthropic API now keep up to 120 MB of inline images before evicting older ones, following the Messages API request limit increase from 32 MB to 128 MB
+- The official Anthropic API now records its 128 MB request limit (`request-body-byte-limit`), so the newest message's images can use the 8 MB reserve and long text cannot push images past the limit ([#17492](https://github.com/can1357/oh-my-pi/pull/17492) by [@will-bogusz](https://github.com/will-bogusz))
+
+### Fixed
+
+- Fixed Cursor's DeepSeek V4.1 Flash 1017 reporting zero cost; every effort level now bills at Cursor's API rate of $0.30/$1.20 per million tokens, with cache reads at $0.006 ([#17390](https://github.com/can1357/oh-my-pi/pull/17390) by [@eggpeat](https://github.com/eggpeat))
+
+## [18.14.0] - 2026-10-29
+
+### Breaking Changes
+
+- Removed the raw `@oh-my-pi/pi-catalog/models.json` export; use the bundled model registry APIs instead.
+
+### Added
+
+- Added `nextBuildPriceChange(cost, now)` to identify when a model's dated pricing card may change, allowing callers to safely reuse earlier builds until that time.
+- Added the `selector-revision-suffix` catalog axis for Google Vertex, so `google-vertex/claude-opus-4-8` resolves exactly to `claude-opus-4-8@default`.
+
+### Changed
+
+- Reduced the bundled model catalog size and loading overhead.
+- Updated role defaults so Factory Droid image models are preferred for the `image` role, allowing `generate_image` to work with only a Factory login.
+- Updated `smol` and `slow` role defaults to use the currently bundled Cerebras and OpenAI Codex model identifiers.
+
+### Fixed
+
+- Fixed Mistral reasoning models so streamed thinking appears in the transcript and is preserved for subsequent turns.
+- Fixed later-turn HTTP 422 errors in Mistral-hosted DeepSeek conversations when assistant reasoning history is included.
+- Fixed DeepSeek V4.1 Flash image attachments being dropped for publisher-namespaced model IDs such as `ds/deepseek-flash`.
+
 ## [18.13.3] - 2026-10-28
 
 ### Added

package/README.md

@@ -21,7 +21,7 @@ Import from subpaths (`@oh-my-pi/pi-catalog/<module>`) or the root barrel.
 
 ## models.json and rules.json are generated
 
-Never edit `src/models.json` or `src/compat/rules.json` by hand. `models.json` is produced from upstream sources (stencil.so, provider catalog discovery, OpenCode docs) by `scripts/generate-models.ts`; `rules.json` is compiled from the KDL tree in `src/compat/rules/`. Regenerate with:
+Never edit `src/models.json` or `src/compat/rules.json` by hand. `models.json` is a compact generated catalog; use `getBundledProviders()` and `getBundledModels()` instead of reading its internal format. It is produced from upstream sources (stencil.so, provider catalog discovery, OpenCode docs) by `scripts/generate-models.ts`; `rules.json` is compiled from the KDL tree in `src/compat/rules/`. Regenerate with:
 
 ```sh
 bun run gen:compat   # src/compat/rules/**/*.kdl -> src/compat/rules.json
@@ -47,7 +47,7 @@ The `models` subpath (also exported from the root) provides timestamp-aware pric
 | `calculateCost(model, usage, timestamp?)`                         | Updates and returns `usage.cost` using `model.cost`.                                                      |
 | `calculateUsageCost(cost, usage, timestamp?)`                     | Updates and returns `usage.cost` using a `ModelCost`.                                                     |
 | `calculateUncachedInputCost(cost, promptInputTokens, timestamp?)` | Returns the cost of a fully uncached prompt.                                                              |
-| `getEffectiveTokenCost(cost, timestamp?)`                         | Returns the base card in force: the latest due effective rate, else the row's own legs.                  |
+| `getEffectiveTokenCost(cost, timestamp?)`                         | Returns the base card in force: the latest due effective rate, else the row's own legs.                   |
 | `getTimeBasedPricingPeriod(cost, timestamp?)`                     | Returns `"peak"`, `"off-peak"`, or `undefined` without a recurring tariff.                                |
 | `hasRecurringTariff(schedule)`                                    | Whether a `TimeBasedCost` carries a peak/off-peak tariff rather than only dated rate cards.               |
 | `getNextTimeBasedPricingTransition(cost, timestamp?)`             | Returns the next actual peak/off-peak change strictly after the timestamp, or `undefined` if none exists. |

package/dist/types/build.d.ts

@@ -6,7 +6,7 @@
  * Request handlers read fields — they never detect, parse ids, or allocate
  * compat per request.
  */
-import { type Api, type Model, type ModelSpec } from "./types.js";
+import { type Api, type Model, type ModelCost, type ModelSpec } from "./types.js";
 /**
  * Applies reviewed catalog-data value corrections (`cost-patch`,
  * `cache-read-at-input-rate`, `cache-write-rate`, `limits-patch`,
@@ -24,8 +24,16 @@ export declare function applyCatalogCorrections(model: Pick<ModelSpec<Api>, "cos
  */
 export declare function buildDiscoveredModel(spec: ModelSpec<Api>, providerType: string): Model<Api>;
 /**
- * Build one model from an authored spec. Bundled models.json rows are fully
- * materialized by the generator and consumed directly (see `models.ts`), so
- * this only runs for discovered/custom/override specs.
+ * The first instant after `now` at which {@link buildModel} could select a
+ * different dated price card for a spec carrying `cost`: the next
+ * `effectiveFrom` of the spec's own rate schedule or of any dated card in the
+ * compiled rules. A build of an unchanged spec made at or after `now` is
+ * identical to a fresh one until then. `Infinity` when no change is pending.
+ */
+export declare function nextBuildPriceChange(cost: ModelCost, now: number): number;
+/**
+ * Build one model from an authored spec. Bundled models are prebuilt by the
+ * generator and decoded by `models.ts`, so this only runs for discovered,
+ * custom, and override specs.
  */
 export declare function buildModel<TApi extends Api>(spec: ModelSpec<TApi>): Model<TApi>;

package/dist/types/compat/types.d.ts

@@ -369,8 +369,10 @@ export type CompiledAuthValidation = {
     kind: "models-endpoint";
     label?: string;
     url: string;
-    /** Env var holding an alternate base URL; `/models` is appended to it. */
+    /** Env var holding an alternate base URL; `baseUrlPath` is appended to it. */
     baseUrlEnv?: string;
+    /** Path appended to the `baseUrlEnv` value (default `/models`), for bases that exclude the API version. */
+    baseUrlPath?: string;
     /** Hook returning extra request headers (may throw a configuration error). */
     headersHook?: string;
     optional?: boolean;

package/dist/types/discovery/cursor-proto.d.ts

@@ -2514,6 +2514,9 @@ export interface InteractionUpdate extends ProtoMessage {
     } | {
         case: "stepCompleted";
         value: StepCompletedUpdate;
+    } | {
+        case: "toolRequestsListed";
+        value: ToolRequestsListedUpdate;
     };
 }
 export declare const InteractionUpdateSchema: MessageCodec<InteractionUpdate>;
@@ -5318,6 +5321,11 @@ export interface ToolCallStartedUpdate extends ProtoMessage {
     modelCallId: string;
 }
 export declare const ToolCallStartedUpdateSchema: MessageCodec<ToolCallStartedUpdate>;
+/** Cursor agent message agent.v1.ToolRequestsListedUpdate. */
+export interface ToolRequestsListedUpdate extends ProtoMessage {
+    callCount: number;
+}
+export declare const ToolRequestsListedUpdateSchema: MessageCodec<ToolRequestsListedUpdate>;
 /** Cursor agent message agent.v1.TruncatedToolCall. */
 export interface TruncatedToolCall extends ProtoMessage {
     originalStepBlobId: Uint8Array;

package/dist/types/model-cache.d.ts

@@ -48,6 +48,14 @@ export interface CacheEntry<TApi extends Api = Api> {
  * cache was opened under — Windows cannot remove a database that is still open.
  */
 export declare function closeModelCache(): void;
+/**
+ * Keeps one connection to a non-default cache path open until the returned
+ * handle is disposed. The default `models.db` already shares a long-lived
+ * connection; other paths otherwise open, initialize and close the database on
+ * every read and write, which a refresh pass repeats once per provider.
+ * Retains nest: the connection closes when the last one is disposed.
+ */
+export declare function retainModelCacheConnection(dbPath: string | undefined): Disposable;
 export declare function readModelCache<TApi extends Api>(providerId: string, ttlMs: number, now: () => number, dbPath?: string): CacheEntry<TApi> | null;
 /**
  * Persist a provider snapshot.

package/dist/types/model-catalog-format.d.ts

@@ -0,0 +1,37 @@
+import type { Api, Model } from "./types.js";
+type EncodedValue = string | number | boolean | null | [number];
+type ObjectNode = [0, EncodedValue[], number, number[]];
+type ArrayNode = [1, EncodedValue[]];
+type Node = ObjectNode | ArrayNode;
+type CompatDelta = [number[], EncodedValue];
+type ModelRow = [string, EncodedValue, CompatDelta?];
+type ProviderRow = [string, EncodedValue, ModelRow[]];
+interface CatalogData {
+    version: 1;
+    keys: string[];
+    booleanShapes: number[][];
+    nodes: Node[];
+    compat: [string, EncodedValue][];
+    providers: ProviderRow[];
+}
+/**
+ * Parsed compact catalog. Decoded nested objects and arrays can be shared by
+ * multiple models; callers must treat those values as immutable.
+ */
+export interface ModelCatalog {
+    readonly data: CatalogData;
+    readonly values: unknown[];
+    readonly decoded: boolean[];
+    readonly providers: Map<string, Record<string, Model<Api>>>;
+    readonly compatBases: Map<string, EncodedValue>;
+    readonly compatValues: Map<string, Map<number, Map<string, Record<string, unknown>>>>;
+}
+/** Encode a model catalog into a deterministic, compact JSON representation. */
+export declare function encodeModelCatalog(models: Record<string, Record<string, Model<Api>>>): string;
+/** Open a parsed catalog without materializing its provider models. */
+export declare function openModelCatalog(value: unknown): ModelCatalog;
+/** List catalog provider keys in deterministic order. */
+export declare function getCatalogProviders(catalog: ModelCatalog): string[];
+/** Materialize and cache one provider, or return undefined if it is absent. */
+export declare function getCatalogProvider(catalog: ModelCatalog, provider: string): Record<string, Model<Api>> | undefined;
+export {};

package/dist/types/model-thinking.d.ts

@@ -32,6 +32,10 @@ export declare function withoutUnverifiedThinking<T extends {
 /**
  * Clamps a requested thinking level against explicit model metadata.
  *
+ * An unserved level first follows `thinking.effortMap` when the map sends it to
+ * a served tier that reaches the wire unchanged (a gateway answering `medium`
+ * as `high`); otherwise it climbs down to the nearest lower tier.
+ *
  * Non-reasoning models always resolve to `undefined`.
  */
 export declare function clampThinkingLevelForModel<TApi extends Api>(model: ApiModel<TApi> | undefined, requested: Effort | undefined): Effort | undefined;

package/dist/types/models.d.ts

@@ -1,10 +1,13 @@
-import MODELS from "./models.json";
 import type { Api, KnownProvider, Model, ModelCost, ModelPricingStatus, TimeBasedCost, TokenCost, Usage } from "./types.js";
-export type GeneratedProvider = keyof typeof MODELS;
+/** Provider identifier accepted by the bundled model registry. */
+export type GeneratedProvider = string;
 /** Whether `provider` has bundled rows in models.json (e.g. a provider id authored in KDL). */
 export declare function isGeneratedProvider(provider: string): provider is GeneratedProvider;
+/** Look up a bundled model; nested catalog values are shared and must not be mutated. */
 export declare function getBundledModel<TApi extends Api = Api>(provider: GeneratedProvider, modelId: string): Model<TApi>;
-export declare function getBundledProviders(): KnownProvider[];
+/** List providers with bundled rows; the array is shared and frozen. */
+export declare function getBundledProviders(): readonly KnownProvider[];
+/** List a provider's bundled models; nested catalog values are shared and must not be mutated. */
 export declare function getBundledModels(provider: GeneratedProvider): Model<Api>[];
 /** Resolve display semantics without mistaking an absent rate card for a free model. */
 export declare function getModelPricingStatus(model: Pick<Model<Api>, "cost" | "pricingStatus">): ModelPricingStatus;

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "18.13.3",
+	"version": "18.14.1",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"keywords": [
 		"ai",
@@ -28,7 +28,6 @@
 		"README.md",
 		"CHANGELOG.md",
 		"LICENSE",
-		"THIRD-PARTY-NOTICES.txt",
 		"dist/types"
 	],
 	"type": "module",
@@ -39,10 +38,6 @@
 			"types": "./dist/types/index.d.ts",
 			"import": "./src/index.ts"
 		},
-		"./models.json": {
-			"types": "./dist/types/models.json.d.d.ts",
-			"import": "./src/models.json"
-		},
 		"./provider-models": {
 			"types": "./dist/types/provider-models/index.d.ts",
 			"import": "./src/provider-models/index.ts"
@@ -93,12 +88,12 @@
 		"gen:proto": "bun scripts/generate-protocols.ts"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.13.3",
-		"@oh-my-pi/pi-utils": "18.13.3"
+		"@oh-my-pi/omptype": "18.14.1",
+		"@oh-my-pi/pi-utils": "18.14.1"
 	},
 	"devDependencies": {
 		"@bgotink/kdl": "0.4.0",
-		"@oh-my-pi/pi-ai": "18.13.3",
+		"@oh-my-pi/pi-ai": "18.14.1",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/build.ts

@@ -7,6 +7,7 @@
  * compat per request.
  */
 
+import RULES from "./compat/rules.json" with { type: "json" };
 import { resolveDiscoveryApi, resolveModelPolicy } from "./compat/resolve";
 import type { ModelIdentity } from "./compat/types";
 import { resolveModelTokenizer } from "./model-tokenizer";
@@ -463,10 +464,56 @@ export function buildDiscoveredModel(spec: ModelSpec<Api>, providerType: string)
 	return buildModel({ ...spec, api, providerType });
 }
 
+/** Every dated card instant in the compiled rules, ascending; collected on first use. */
+let ruleCardInstants: number[] | undefined;
+
+function collectRuleCardInstants(value: unknown, out: number[]): void {
+	if (Array.isArray(value)) {
+		for (const item of value) collectRuleCardInstants(item, out);
+		return;
+	}
+	if (typeof value !== "object" || value === null) return;
+	for (const key in value) {
+		const child: unknown = Reflect.get(value, key);
+		if (key === "effectiveFrom" && typeof child === "string") {
+			const instant = Date.parse(child);
+			if (Number.isFinite(instant)) out.push(instant);
+		} else {
+			collectRuleCardInstants(child, out);
+		}
+	}
+}
+
+/**
+ * The first instant after `now` at which {@link buildModel} could select a
+ * different dated price card for a spec carrying `cost`: the next
+ * `effectiveFrom` of the spec's own rate schedule or of any dated card in the
+ * compiled rules. A build of an unchanged spec made at or after `now` is
+ * identical to a fresh one until then. `Infinity` when no change is pending.
+ */
+export function nextBuildPriceChange(cost: ModelCost, now: number): number {
+	if (ruleCardInstants === undefined) {
+		const instants: number[] = [];
+		collectRuleCardInstants(RULES, instants);
+		ruleCardInstants = instants.sort((a, b) => a - b);
+	}
+	let next = Number.POSITIVE_INFINITY;
+	for (const instant of ruleCardInstants) {
+		if (instant > now) {
+			next = instant;
+			break;
+		}
+	}
+	for (const rate of cost.timeBased?.effectiveRates ?? []) {
+		if (rate.effectiveFrom > now && rate.effectiveFrom < next) next = rate.effectiveFrom;
+	}
+	return next;
+}
+
 /**
- * Build one model from an authored spec. Bundled models.json rows are fully
- * materialized by the generator and consumed directly (see `models.ts`), so
- * this only runs for discovered/custom/override specs.
+ * Build one model from an authored spec. Bundled models are prebuilt by the
+ * generator and decoded by `models.ts`, so this only runs for discovered,
+ * custom, and override specs.
  */
 export function buildModel<TApi extends Api>(spec: ModelSpec<TApi>): Model<TApi> {
 	const policy = resolveModelPolicy(spec);

package/src/compat/axes.ts

@@ -469,6 +469,12 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 		values: ["gemini", "anthropic", "codex", "xai", "openrouter", "openai"],
 	},
 	"web-search-model": { key: "webSearchModel", set: "catalog", shape: "scalar" },
+	/**
+	 * Revision suffix the host appends to the model's id (Vertex `@default`): a
+	 * `provider/id` selector written without it names this model exactly.
+	 * Unassigned: the id has no implied revision and must be written in full.
+	 */
+	"selector-revision-suffix": { key: "selectorRevisionSuffix", set: "catalog", shape: "scalar" },
 	"hosted-image": { key: "hostedImage", set: "catalog", shape: "scalar", values: [true, false] },
 	/** How the model line bills an input image; shape and formulas in `./image-tokenization`. */
 	"image-tokenization": { key: "imageTokenization", set: "catalog", shape: "object" },
@@ -496,8 +502,9 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 	"image-max-edge": { key: "imageMaxEdge", set: "catalog", shape: "scalar" },
 	"inline-image-byte-budget": { key: "inlineImageByteBudget", set: "catalog", shape: "scalar" },
 	/**
-	 * Largest request body the host accepts. Where set, the inline-image byte clamp keeps the newest user message's
-	 * and newest tool result's images past `inline-image-byte-budget` while they alone fit this limit.
+	 * Largest request body the host accepts. Where set, the inline-image byte clamp keeps images within what the
+	 * request's other bytes leave of it, and keeps the newest user message's and newest tool result's images past
+	 * `inline-image-byte-budget` while they fit that room.
 	 */
 	"request-body-byte-limit": { key: "requestBodyByteLimit", set: "catalog", shape: "scalar" },
 	/**

package/src/compat/rules/README.md

@@ -179,16 +179,16 @@ 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`, `on-api`, and `models`.                                                                                           |
-| `on`       | `on "provider-a" "provider-b" { ... }`   | One or more provider IDs, combined as OR. It is allowed under a root `class`, and may contain `on-api` to conjoin an API scope plus `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` (directly or nested inside `on`) or `provider` 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 may contain only `on-upstream`. `token="name"` matches an ASCII-case-insensitive token bounded by non-alphanumerics.                 |
-| `on-upstream` | `on-upstream "a" "b" { ... }`        | Exact selected upstream IDs, combined as OR. Allowed inside any selector scope once; preserves the containing scope's other permitted children. Absent upstream never matches. |
+| 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`, `on-api`, and `models`.                                                                                                                                           |
+| `on`          | `on "provider-a" "provider-b" { ... }`   | One or more provider IDs, combined as OR. It is allowed under a root `class`, and may contain `on-api` to conjoin an API scope plus `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` (directly or nested inside `on`) or `provider` 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 may contain only `on-upstream`. `token="name"` matches an ASCII-case-insensitive token bounded by non-alphanumerics.                                                             |
+| `on-upstream` | `on-upstream "a" "b" { ... }`            | Exact selected upstream IDs, combined as OR. Allowed inside any selector scope once; preserves the containing scope's other permitted children. Absent upstream never matches.                                                |
 
 Every selector scope may additionally contain `on-upstream`; it cannot replace an already constrained upstream. Class, provider/`on`, `on-api`, `on-upstream`, and family 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.
 
@@ -227,7 +227,7 @@ One object axis carries a computed form: `long-context-cost` accepts either the
 
 `inline-image-byte-budget` caps retained base64 image characters for a deployment's request body. It leaves room for tool schemas and text beneath the provider's body-size limit; the coding agent reads it through `resolveInlineImageByteBudget` before sending live history, and snapcompact gives its archive frames half of it (`frameDataBytesBudget`). The budget applies only while the request reaches the host the rules describe (the Anthropic API, first-party OpenAI, the Codex backend at its default URL, DeepSeek's API, or Factory Droid's own global or EU API host), which callers decide with pi-ai's `reachesBudgetedHost`: `compat.officialEndpoint` for the configured `baseUrl`, for first-party Anthropic models also the `ANTHROPIC_BASE_URL` or Foundry `FOUNDRY_BASE_URL` reroute applied at request time, and for Factory Droid models the `baseUrl` origin. A custom or rerouted endpoint imposes its own body limit.
 
-`request-body-byte-limit` is the largest request body the host accepts (Factory Droid's hosted routes: 4,500,000 bytes), under the same gate. Where it is set, the byte clamp keeps the inline images of the newest user message and then of the newest image-bearing tool result even past `inline-image-byte-budget`, each only while the kept images fit the limit less the request's other serialized bytes, and evicts older images instead; read through `resolveRequestBodyByteLimit`.
+`request-body-byte-limit` is the largest request body the host accepts (Factory Droid's hosted routes: 4,500,000 bytes; Anthropic's Messages API: 128,000,000 bytes), under the same gate. Where it is set, the byte clamp also keeps inline images within what the request's other serialized bytes leave of the limit, and keeps the inline images of the newest user message and then of the newest image-bearing tool result even past `inline-image-byte-budget`, each only while the kept images fit the limit less the request's other serialized bytes, and evicts older images instead; read through `resolveRequestBodyByteLimit`.
 
 `catalog-endpoint-budgets #true` makes these host budgets apply whenever a request reaches the provider's own catalog endpoint (an unset `baseUrl`, or an `endpoint-hosts` host; see `isCatalogEndpointUrl`), for a hosted gateway such as Factory Droid whose compat names no official endpoint, or a first-party provider such as DeepSeek whose `openai-completions` compat resolves no `officialEndpoint`. Unassigned, they apply only on the official endpoint.
 
@@ -415,7 +415,7 @@ auth "anthropic" {
 
 Login kinds:
 
-- `login "api-key" { auth-url "…"; instructions "…"; prompt "…" placeholder="…"; empty-fallback "…"; normalize "strip-bearer"; validate … }` — `validate "chat-completions" base-url= model= tolerate-model-denied= max-tokens-field= max-tokens=`, `validate "anthropic-messages" base-url= model=`, or `validate "models-endpoint" url= base-url-env= headers-hook=`; all accept `label=` (error-message label, defaults to `name`), `optional=#true` (only auth failures reject), and `trust-forbidden=#true` (requires `optional=#true`; a 403 no longer rejects, so only a 401 does).
+- `login "api-key" { auth-url "…"; instructions "…"; prompt "…" placeholder="…"; empty-fallback "…"; normalize "strip-bearer"; validate … }` — `validate "chat-completions" base-url= model= tolerate-model-denied= max-tokens-field= max-tokens=`, `validate "anthropic-messages" base-url= model=`, or `validate "models-endpoint" url= base-url-env= base-url-path= headers-hook=` (`base-url-path`, default `/models`, is appended to the `base-url-env` value and needs it); all accept `label=` (error-message label, defaults to `name`), `optional=#true` (only auth failures reject), and `trust-forbidden=#true` (requires `optional=#true`; a 403 no longer rejects, so only a 401 does).
 - `login "oauth-code" { … }` as above; `token`/`refresh` `params` values may use `{code}`, `{state}`, `{redirect_uri}`, `{code_verifier}`, `{client_id}`, `{client_secret}`, `{refresh_token}`, `{scope}`; the standard grant parameters are sent unless `standard=#false`.
 - `login "device-code" { client-id …; base-url "…"; scopes …; headers-hook "…"; device url="{base}/…" body="form" { params {…} headers {…} }; token url="…" url-hook="…"; response user-code= device-code= verification-uri= verification-uri-complete= interval= expires-in=; instructions "Enter code: {user_code}"; credential {…}; userinfo …; after-exchange hook=… }`.
 - `login "custom" hook="name"` — the whole flow is a named `@oh-my-pi/pi-ai` hook (`src/registry/hooks/custom.ts`).

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

@@ -8,6 +8,7 @@ auth "typesafe" {
 		auth-url "https://console.typesafe.ai/"
 		instructions "Create or copy your API key from the TypeSafe console."
 		prompt "Paste your TypeSafe API key" placeholder="API key"
-		validate "models-endpoint" label="TypeSafe" url="https://api.typesafe.ai/v1/models" base-url-env="TYPESAFE_BASE_URL"
+		// TYPESAFE_BASE_URL is an origin without `/v1` (the SDK appends `/v1/systemone`).
+		validate "models-endpoint" label="TypeSafe" url="https://api.typesafe.ai/v1/models" base-url-env="TYPESAFE_BASE_URL" base-url-path="/v1/models"
 	}
 }

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

@@ -90,10 +90,11 @@ class "deepseek" {
 	}
 	// V4.1 Flash is natively multimodal (vision encoder + projector, image-text
 	// pre-training) but its release id carries no `vision` token to key on.
+	// The leading glob also covers publisher-namespaced ids (`ds/deepseek-flash`).
 	// Same priority as the V4.1 Flash carve-out: the OpenCode Go provider rule
 	// declares the same value for its `deepseek-flash` lane and would otherwise
 	// tie at equal rank.
-	models "deepseek-flash" priority=10 {
+	models "*deepseek-flash" priority=10 {
 		strip-image-input #false
 	}
 	// DSML is the model's own tool-call grammar, not a host quirk: any server

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

@@ -78,12 +78,16 @@ provider "anthropic" {
 	// Code's `CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS` kill switch for gateways
 	// that reject experimental beta fields.
 	honors-claude-code-beta-kill-switch #true
-	// Messages requests cap at 32 MB. Reserve 8 MB for text, tools, and JSON
-	// framing when retaining inline base64 images in a long conversation.
-	// Applied only on the official endpoint; a custom baseUrl has its own limit.
+	// Messages requests cap at 128 MB. Reserve 8 MB for text, tools, and JSON
+	// framing when retaining inline base64 images in a long conversation; the
+	// body limit keeps images within what the other bytes leave, and lets the
+	// newest user message's and tool result's images use the reserve.
+	// Applied only on the official endpoint; a custom baseUrl has its own limit
+	// (Bedrock 20 MB; Vertex, Foundry, and Claude Platform on AWS 30-32 MB).
 	// https://platform.claude.com/docs/en/api/overview#request-size-limits
 	on-api "anthropic-messages" {
-		inline-image-byte-budget 24000000
+		inline-image-byte-budget 120000000
+		request-body-byte-limit 128000000
 	}
 	// The official endpoint carries at most 600 images and documents per
 	// request, 100 on models with a 200k-token window: every Claude line before

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

@@ -387,6 +387,16 @@ provider "cursor" {
 			cache-write 0
 		}
 	}
+	// DeepSeek V4.1 Flash 1017: $0.30/$1.20, cache read $0.006, no write fee.
+	// The bare id and every reasoning-effort wire id share one card.
+	models "deepseek-v41-flash-1017*" {
+		cost-patch {
+			input 0.3
+			output 1.2
+			cache-read 0.006
+			cache-write 0
+		}
+	}
 	// Gemini 3 Flash: $0.50/$3.00, cache read $0.05.
 	models "gemini-3-flash" {
 		cost-patch {

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

@@ -7,8 +7,8 @@ provider "deepseek" {
 
 	// DeepSeek's vision guide documents 600 images per request and a 48 MiB
 	// request body. OMP reserves 8 MiB for text, tools, and the JSON envelope
-	// (48 MiB - 8 MiB = 40 MiB); Anthropic's 24,000,000-byte image budget under
-	// its 32 MB body uses an 8 MB reserve. DeepSeek documents MiB, so this uses
+	// (48 MiB - 8 MiB = 40 MiB); Anthropic's 120,000,000-byte image budget under
+	// its 128 MB body uses an 8 MB reserve. DeepSeek documents MiB, so this uses
 	// an 8 MiB reserve. The split is OMP's conservative margin, not a
 	// DeepSeek-documented figure. A custom baseUrl keeps the conservative
 	// catalog floor because the gateway's own limits are unknown.

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

@@ -4,6 +4,9 @@ provider "google-vertex" {
 	default-model "gemini-3.1-pro-preview"
 	allow-unauthenticated #true
 	image-count-budget 200
+	// Vertex publishes Anthropic ids pinned to a revision (`claude-opus-4-8@default`);
+	// `google-vertex/claude-opus-4-8` names that model exactly.
+	selector-revision-suffix "@default"
 
 	// Replaces google-shared.ts stripping functionCall/functionResponse ids for Vertex.
 	supports-function-part-id #false

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

@@ -29,6 +29,14 @@ provider "mistral" {
 	// without it, requests only hit replica-local implicit cache. Same shape as
 	// the kimi-code rule (#6049).
 	supports-prompt-cache-key #true
+	// The gateway accepts assistant reasoning as text, not DeepSeek's native
+	// reasoning_content field, even when it serves DeepSeek-family models.
+	requires-reasoning-content-for-tool-calls #false
+	requires-reasoning-content-for-all-assistant-turns #false
+	// Reasoning streams and replays as typed `thinking` content parts
+	// (https://docs.mistral.ai/studio/conversations/reasoning) for every
+	// reasoning model the host serves, including GLM and DeepSeek (#17375).
+	mistral-reasoning-content-parts #true
 	class "unknown" {
 		thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 	}

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

@@ -16,7 +16,7 @@ provider "openai-codex" {
 
 	// Policy budgets: the Codex backend documents no request-size or image
 	// limit, and requests with several thousand images fail mid-stream at
-	// random. Inline images keep Anthropic's 24 MB; the official endpoint's
+	// random. Inline images keep a 24 MB budget; the official endpoint's
 	// image count stays well below those failures. A custom baseUrl keeps 200.
 	inline-image-byte-budget 24000000
 	image-count-budget 200

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

@@ -428,7 +428,6 @@ behavior {
 		model "oss-120b"
 		model "cerebras/zai-glm-4.7"
 		model "cerebras/zai-glm-4.6"
-		model "cerebras/zai-glm"
 		model "claude-haiku-4-5"
 		model "claude-haiku-4.5"
 		model "haiku-4-5"
@@ -463,7 +462,6 @@ behavior {
 		model "anthropic/claude-opus-5"
 		model "claude-opus-5"
 		model "opus-5"
-		model "openai-codex/gpt-5.5"
 		model "gpt-5.5"
 		model "claude-opus-4-8"
 		model "claude-opus-4.8"
@@ -479,6 +477,11 @@ behavior {
 		model "openrouter/google/gemini-3-pro-image"
 		model "google/gemini-3-pro-image"
 		model "deepinfra/black-forest-labs/FLUX-2-pro"
+		// Factory Droid last, in the CLI's own preference order: a Factory
+		// login alone can generate, and no other provider's turn moves.
+		model "factory-droid/gpt-image-2.5-sunburst"
+		model "factory-droid/gpt-image-2.5-flare"
+		model "factory-droid/gemini-3-pro-image-preview"
 	}
 	role-defaults "web" {
 		model "web/parallel"