@oh-my-pi/pi-catalog 18.1.13 → 18.1.14

package/CHANGELOG.md

@@ -2,6 +2,22 @@
 
 ## [Unreleased]
 
+## [18.1.14] - 2026-09-07
+
+### Fixed
+
+- Bills Astra API requests above 272K input at the documented 2x input / 1.5x output long-context tier; the Codex subscription route stays exempt with free cache writes ([#11157](https://github.com/can1357/oh-my-pi/pull/11157) by [@H4vC](https://github.com/H4vC)).
+- Fixed Astra's extended window over-advertising input by 128K; it now uses the documented 922K input cap inside the 1.05M total context ([#11157](https://github.com/can1357/oh-my-pi/pull/11157) by [@H4vC](https://github.com/H4vC)).
+- Fixed explicit Codex context-window overrides widening past the server-honored maximum; they now clamp to the documented ceiling like upstream Codex ([#11157](https://github.com/can1357/oh-my-pi/pull/11157) by [@H4vC](https://github.com/H4vC)).
+- Fixed Codex Astra using its larger window without opt-in; its default is 272K and Extended Context enables at least the documented 1.05M window ([#11126](https://github.com/can1357/oh-my-pi/pull/11126) by [@H4vC](https://github.com/H4vC)).
+- Fixed GitHub Copilot enterprise-only model ids inheriting another provider's wire routing (e.g. `gpt-5.6-sol-fast` pinning every request to the `-none` sibling id regardless of thinking level) ([#11128](https://github.com/can1357/oh-my-pi/pull/11128) by [@H4vC](https://github.com/H4vC)).
+
+## [18.1.13] - 2026-09-07
+
+### Fixed
+
+- Fixed GPT-6 Astra compacting early at a 272K-token window with its full window gated behind `/extended-context`: it now defaults to the documented 1.05M-token window.
+
 ## [18.1.12] - 2026-09-06
 
 ### Added
@@ -15,7 +31,6 @@
 - Fixed OpenCode Go/Zen live model discovery (`GET /v1/models`) missing `x-opencode-session` and omp's `User-Agent`: discovery requests now attribute with the stable install id so the requests OpenCode flags as `Bun fetch` carry the required session header.
 	- Fixed GPT-6 Astra requests through GitHub Copilot failing with an unsupported endpoint error ([#10874](https://github.com/can1357/oh-my-pi/pull/10874) by [@xpcmdshell](https://github.com/xpcmdshell)).
 	- Fixed GPT-6 Astra showing as free with a 272K-token window in the OpenAI Codex catalog by applying its documented pricing; `/extended-context` enables the wire-advertised 872K-token maximum ([#10980](https://github.com/can1357/oh-my-pi/pull/10980) by [@H4vC](https://github.com/H4vC)).
-	- Fixed GPT-6 Astra compacting early at a 272K-token window with its full window gated behind `/extended-context`: it now defaults to the documented 1.05M-token window.
 	- Made extended-context catalog rebuilds faster by resolving each model's maximum window once per process ([#11039](https://github.com/can1357/oh-my-pi/pull/11039) by [@H4vC](https://github.com/H4vC)).
 
 ## [18.1.9] - 2026-09-04

package/dist/types/compat/context-window.d.ts

@@ -1,7 +1,32 @@
 import type { Model } from "../types.js";
 /**
- * Maximum prompt window for extended context. Live discovery takes precedence
- * over rule-owned fallbacks for older bundled or cached model metadata. Resolve
- * at catalog composition time so frozen bundled rows need no runtime mutation.
+ * Extended-context capacity. Curated maxima correct stale lower discovery
+ * values; a higher live maximum still wins. The registry applies this capacity
+ * only when extended context is enabled, before explicit user overrides.
  */
 export declare function resolveMaxContextWindow(model: Model): number | undefined;
+/**
+ * Override ceiling for Codex models. Upstream clamps `model_context_window`
+ * to `min(override, max_context_window)` (`with_config_overrides` in
+ * `models-manager/src/model_info.rs`); the ceiling here is stale-aware — the
+ * curated maximum corrects a lower server value (Astra reports a stale 872K
+ * maximum; OpenAI documents at most 922K input) while a higher live maximum
+ * still wins. No curated or live maximum means no ceiling: overrides pass
+ * through, matching upstream's unclamped branch.
+ */
+export declare function codexOverrideCeiling(model: Model): number | undefined;
+/**
+ * Whether explicit context-window overrides for this model clamp to the
+ * server-honored ceiling. KDL-owned (`clamp-context-override`): branching on
+ * it here keeps provider deployment contracts out of TypeScript.
+ */
+export declare function clampsContextOverride(model: Model): boolean;
+/**
+ * Clamp a requested Codex context window to the override ceiling, mirroring
+ * upstream. `model` is the pre-override model: the ceiling never shrinks the
+ * request below the window that already works, so a stale-low live maximum
+ * (e.g. 128K base with a 64K advertised maximum) cannot punish an explicit
+ * override. Returns the request unchanged when no ceiling applies or it
+ * already fits.
+ */
+export declare function clampCodexContextWindow(model: Model, requested: number): number;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "18.1.13",
+	"version": "18.1.14",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://omp.sh",
 	"author": "Stencil Labs, Inc.",
@@ -35,12 +35,12 @@
 		"gen:proto": "bun scripts/generate-protocols.ts"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.1.13",
-		"@oh-my-pi/pi-utils": "18.1.13"
+		"@oh-my-pi/omptype": "18.1.14",
+		"@oh-my-pi/pi-utils": "18.1.14"
 	},
 	"devDependencies": {
 		"@bgotink/kdl": "0.4.0",
-		"@oh-my-pi/pi-ai": "18.1.13",
+		"@oh-my-pi/pi-ai": "18.1.14",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/axes.ts

@@ -267,6 +267,7 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
 		shape: "scalar",
 		values: ["freeform", "function"],
 	},
+	"clamp-context-override": { key: "clampContextOverride", set: "catalog", shape: "scalar" },
 	"context-promotion-target": { key: "contextPromotionTarget", set: "catalog", shape: "scalar" },
 	"context-window-floor": { key: "contextWindowFloor", set: "catalog", shape: "scalar" },
 	"cost-patch": { key: "costPatch", set: "catalog", shape: "object" },

package/src/compat/context-window.ts

@@ -3,29 +3,72 @@ import type { Model } from "../types";
 import { resolveModelPolicy } from "./resolve";
 
 /**
- * Rule-owned fallback maxima by `provider/id/api`. The rule table is static
- * per process (compiled into rules.json), so one resolution per model key is
- * enough — composition calls this per model on every rebuild while extended
- * context is enabled, and the full policy resolve (identity classification
- * plus cascade walk) is wasted work per call.
+ * Rule-owned maxima by provider/id/api. Resolve once per process rather than
+ * walking the static policy cascade on every catalog rebuild. Null caches the
+ * absence of a curated maximum; undefined means the key has not been resolved.
  */
-const ruleFallbackCache = new Map<string, number | undefined>();
+const ruleMaximumCache = new Map<string, number | null>();
 
 /**
- * Maximum prompt window for extended context. Live discovery takes precedence
- * over rule-owned fallbacks for older bundled or cached model metadata. Resolve
- * at catalog composition time so frozen bundled rows need no runtime mutation.
+ * Extended-context capacity. Curated maxima correct stale lower discovery
+ * values; a higher live maximum still wins. The registry applies this capacity
+ * only when extended context is enabled, before explicit user overrides.
  */
 export function resolveMaxContextWindow(model: Model): number | undefined {
+	const key = `${model.provider} ${model.id} ${model.api}`;
+	let curated = ruleMaximumCache.get(key);
+	if (curated === undefined) {
+		const maximum = resolveModelPolicy(toModelSpec(model)).catalog.maxContextWindow;
+		curated = typeof maximum === "number" && Number.isFinite(maximum) && maximum > 0 ? maximum : null;
+		ruleMaximumCache.set(key, curated);
+	}
+
 	const maximum = model.maxContextWindow;
 	if (typeof maximum === "number" && Number.isFinite(maximum) && maximum > 0) {
-		return maximum;
+		return Math.max(maximum, curated ?? 0);
 	}
+	return curated ?? undefined;
+}
 
-	const key = `${model.provider} ${model.id} ${model.api}`;
-	if (ruleFallbackCache.has(key)) return ruleFallbackCache.get(key);
-	const fallback = resolveModelPolicy(toModelSpec(model)).catalog.maxContextWindow;
-	const resolved = typeof fallback === "number" && Number.isFinite(fallback) && fallback > 0 ? fallback : undefined;
-	ruleFallbackCache.set(key, resolved);
-	return resolved;
+/**
+ * Override ceiling for Codex models. Upstream clamps `model_context_window`
+ * to `min(override, max_context_window)` (`with_config_overrides` in
+ * `models-manager/src/model_info.rs`); the ceiling here is stale-aware — the
+ * curated maximum corrects a lower server value (Astra reports a stale 872K
+ * maximum; OpenAI documents at most 922K input) while a higher live maximum
+ * still wins. No curated or live maximum means no ceiling: overrides pass
+ * through, matching upstream's unclamped branch.
+ */
+export function codexOverrideCeiling(model: Model): number | undefined {
+	return resolveMaxContextWindow(model);
+}
+
+/**
+ * Whether explicit context-window overrides for this model clamp to the
+ * server-honored ceiling. KDL-owned (`clamp-context-override`): branching on
+ * it here keeps provider deployment contracts out of TypeScript.
+ */
+export function clampsContextOverride(model: Model): boolean {
+	return resolveModelPolicy(toModelSpec(model)).catalog.clampContextOverride === true;
+}
+
+/**
+ * Clamp a requested Codex context window to the override ceiling, mirroring
+ * upstream. `model` is the pre-override model: the ceiling never shrinks the
+ * request below the window that already works, so a stale-low live maximum
+ * (e.g. 128K base with a 64K advertised maximum) cannot punish an explicit
+ * override. Returns the request unchanged when no ceiling applies or it
+ * already fits.
+ */
+export function clampCodexContextWindow(model: Model, requested: number): number {
+	if (!Number.isFinite(requested) || requested <= 0) {
+		return requested;
+	}
+	const ceiling = codexOverrideCeiling(model);
+	if (ceiling === undefined || requested <= ceiling) {
+		return requested;
+	}
+	const current = model.contextWindow;
+	const floor = typeof current === "number" && Number.isFinite(current) && current > 0 ? current : 0;
+	return Math.min(requested, Math.max(ceiling, floor));
 }

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

@@ -9,6 +9,11 @@ provider "openai-codex" {
 	// Harmony-protocol leak detection/mitigation applies to every Codex model,
 	// current and future; replaces the provider check in harmony-leak.ts.
 	harmony-leak-mitigation #true
+	// Every Codex SKU clamps explicit context-window overrides to the
+	// server-honored ceiling (`min(override, max)`, mirroring upstream
+	// `with_config_overrides`); models without a curated or live maximum
+	// pass through. Replaces the provider check in model-registry.ts.
+	clamp-context-override #true
 	// Subscription (Codex) discovery reports no pricing; the curated Daybreak
 	// aliases carry the standard GPT-5.6 Sol/Cyber API list price so cost
 	// display reads as API-equivalent spend. The `-wm` worker sibling bills
@@ -29,16 +34,15 @@ provider "openai-codex" {
 			cache-write 15.625
 		}
 	}
-	// Codex discovery still reports Astra's stale 272K window with
-	// max_context_window=872000 (client 0.153.1, 2026-09-04) — the same stale
-	// pair it reports for Luna/Sol/Terra. Public docs list a 1.05M window and
-	// live Codex traffic succeeds past the wire maximum (895,302 input tokens
-	// observed, including cached input), so floor to the documented window like
-	// the GPT-5.6 1M floor below. Astra carries no long-context surcharge, so
-	// the window must not be gated behind `/extended-context`; the
-	// `maximum > contextWindow` guard then ignores the stale 872K live maximum.
-	// Pricing stays at the documented credit-equivalent rates, with free cache
-	// writes and no >272K multiplier.
+	// Codex reports a 272K default and a stale 872K maximum for Astra. Keep
+	// the default explicit so cached 1.05M rows from #11089 cannot bypass the
+	// opt-in; the curated maximum is applied only with extended context on.
+	// OpenAI documents 1.05M total context with at most 922K input and 128K
+	// output, so the input ceiling is 922K — not the total. A higher live
+	// maximum still wins as the ceiling.
+	// Subscription credit-equivalent pricing: no per-token billing, free cache
+	// writes, exempt from the API long-context multiplier. The first-party API
+	// route bills the documented >272K tier instead — see providers/openai.kdl.
 	models "gpt-6-astra" "gpt-6-astra-wm" {
 		cost-patch {
 			input 10.0
@@ -50,7 +54,10 @@ provider "openai-codex" {
 			flex 0.5
 			priority 2.5
 		}
-		context-window-floor 1050000
+		limits-patch {
+			context-window 272000
+		}
+		max-context-window 922000
 	}
 	class "openai" {
 		revision ">=5.3 <5.7" {
@@ -141,8 +148,4 @@ provider "openai-codex" {
 	models "gpt-5.6-luna" "gpt-5.6-sol" "gpt-5.6-terra" {
 		context-window-floor 1000000
 	}
-	// (No `max-context-window` fallback: the 1.05M floor already exceeds the
-	// stale 872K wire maximum, so offline and cached rows compose to the
-	// documented window and the live maximum is ignored by the
-	// `maximum > contextWindow` guard — same as Luna/Sol/Terra above.)
 }

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

@@ -72,6 +72,19 @@ provider "openai" {
 			cache-write 5.0
 		}
 	}
+	// Astra API route bills the documented long-context tier above 272K input:
+	// 2x input and cache rates, 1.5x output for the full request; cache writes
+	// stay 1.25x the long input rate (25.0). The Codex subscription route stays
+	// exempt (credit-equivalent, free cache writes) — see openai-codex.kdl.
+	models "gpt-6-astra*" {
+		long-context-cost {
+			input-threshold 272000
+			input 20.0
+			output 75.0
+			cache-read 2.0
+			cache-write 25.0
+		}
+	}
 	// residue: the bare Daybreak rolling aliases pin the documented 5.6-gen
 	// wire restrictions without carrying the gpt- prefix identity.
 	models "daybreak-blue-latest" "daybreak-red-latest" {

package/src/compat/rules.json

@@ -10544,7 +10544,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:16",
+				"source": "providers/openai-codex.kdl:21",
 				"providers": [
 					"openai-codex"
 				],
@@ -10568,7 +10568,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:24",
+				"source": "providers/openai-codex.kdl:29",
 				"providers": [
 					"openai-codex"
 				],
@@ -10592,7 +10592,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:42",
+				"source": "providers/openai-codex.kdl:46",
 				"providers": [
 					"openai-codex"
 				],
@@ -10617,11 +10617,14 @@
 						"flex": 0.5,
 						"priority": 2.5
 					},
-					"contextWindowFloor": 1050000
+					"limitsPatch": {
+						"contextWindow": 272000
+					},
+					"maxContextWindow": 922000
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:56",
+				"source": "providers/openai-codex.kdl:63",
 				"class": "openai",
 				"providers": [
 					"openai-codex"
@@ -10641,7 +10644,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:61",
+				"source": "providers/openai-codex.kdl:68",
 				"class": "openai",
 				"providers": [
 					"openai-codex"
@@ -10657,7 +10660,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:64",
+				"source": "providers/openai-codex.kdl:71",
 				"class": "openai",
 				"providers": [
 					"openai-codex"
@@ -10673,7 +10676,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:69",
+				"source": "providers/openai-codex.kdl:76",
 				"class": "openai",
 				"providers": [
 					"openai-codex"
@@ -10692,7 +10695,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:81",
+				"source": "providers/openai-codex.kdl:88",
 				"class": "openai",
 				"providers": [
 					"openai-codex"
@@ -10718,7 +10721,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:87",
+				"source": "providers/openai-codex.kdl:94",
 				"class": "openai",
 				"providers": [
 					"openai-codex"
@@ -10744,7 +10747,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:79",
+				"source": "providers/openai-codex.kdl:86",
 				"class": "openai",
 				"providers": [
 					"openai-codex"
@@ -10761,7 +10764,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:97",
+				"source": "providers/openai-codex.kdl:104",
 				"class": "openai",
 				"providers": [
 					"openai-codex"
@@ -10781,7 +10784,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:103",
+				"source": "providers/openai-codex.kdl:110",
 				"class": "unknown",
 				"providers": [
 					"openai-codex"
@@ -10797,7 +10800,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:106",
+				"source": "providers/openai-codex.kdl:113",
 				"class": "unknown",
 				"providers": [
 					"openai-codex"
@@ -10813,7 +10816,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:112",
+				"source": "providers/openai-codex.kdl:119",
 				"providers": [
 					"openai-codex"
 				],
@@ -10834,7 +10837,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:121",
+				"source": "providers/openai-codex.kdl:128",
 				"providers": [
 					"openai-codex"
 				],
@@ -10859,7 +10862,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:130",
+				"source": "providers/openai-codex.kdl:137",
 				"providers": [
 					"openai-codex"
 				],
@@ -10880,7 +10883,7 @@
 				}
 			},
 			{
-				"source": "providers/openai-codex.kdl:141",
+				"source": "providers/openai-codex.kdl:148",
 				"providers": [
 					"openai-codex"
 				],
@@ -10914,7 +10917,8 @@
 					"serviceTierCost": {
 						"flex": 0.5,
 						"priority": 2
-					}
+					},
+					"clampContextOverride": true
 				}
 			},
 			{
@@ -11157,7 +11161,28 @@
 				}
 			},
 			{
-				"source": "providers/openai.kdl:77",
+				"source": "providers/openai.kdl:79",
+				"providers": [
+					"openai"
+				],
+				"models": [
+					{
+						"kind": "glob",
+						"value": "gpt-6-astra*"
+					}
+				],
+				"catalog": {
+					"longContext": {
+						"inputThreshold": 272000,
+						"input": 20,
+						"output": 75,
+						"cacheRead": 2,
+						"cacheWrite": 25
+					}
+				}
+			},
+			{
+				"source": "providers/openai.kdl:90",
 				"providers": [
 					"openai"
 				],

package/src/models.json

@@ -261375,7 +261375,7 @@
 				"api": "openai-codex-responses",
 				"v2StreamingEnabled": true
 			},
-			"contextWindow": 1050000,
+			"contextWindow": 272000,
 			"maxTokens": 128000,
 			"preferWebsockets": true,
 			"useResponsesLite": true,

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

@@ -92,9 +92,15 @@ export function resolveModelCacheProviderId(providerId: string, options: ModelCa
 			// switching `COPILOT_GITHUB_TOKEN` to a different account misses the
 			// prior endpoint's cache and re-runs discovery instead of hitting the
 			// stale host and 403ing (PR #8510 review).
+			// v2: rows cached before the cross-provider routing strip inherit
+			// Cursor collapsed-family wire ids (e.g. enterprise-only
+			// `gpt-5.6-sol-fast` pinned to `-none-fast`); use a fresh namespace
+			// so they refetch instead of serving the poisoned rows. Listing ids
+			// cannot cover this class — any enterprise-only sibling can carry
+			// another provider's routing — so version the namespace instead.
 			const baseUrl = options.baseUrl ?? PERSONAL_GITHUB_COPILOT_BASE_URL;
 			const scope = `${options.apiKey ?? ""}\u0000${baseUrl}`;
-			return `github-copilot:models-v1:${Bun.hash(scope).toString(36)}`;
+			return `github-copilot:models-v2:${Bun.hash(scope).toString(36)}`;
 		}
 		case "openrouter":
 			return "openrouter:pseudo-api";

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

@@ -6380,6 +6380,22 @@ export function githubCopilotModelManagerOptions(config?: GithubCopilotModelMana
 											}
 										: {}),
 								};
+						// Cross-provider fallback references (e.g. a Cursor
+						// collapsed family for an enterprise-only sibling id)
+						// carry provider-specific wire routing that must not
+						// transfer: the off-tier `requestModelId` pin would send
+						// every Copilot request under the `-none` sibling id
+						// regardless of thinking level.
+						if (reference && reference.provider !== "github-copilot") {
+							delete base.requestModelId;
+							if (base.thinking) {
+								// `base` is a shallow copy of the shared global
+								// reference: clone before deleting or the bundled
+								// entry loses its routing process-wide.
+								base.thinking = { ...base.thinking };
+								delete base.thinking.effortRouting;
+							}
+						}
 						const defaultCost = copilotTierCost(tokenPrices.defaultTier);
 						if (defaultCost) {
 							// Cache writes are not reported per tier; retain the bundled provider rate.