pi-inferhub 0.1.0 → 0.2.0

package/AGENTS.md

@@ -59,18 +59,21 @@ upstream quirk.
    members' `official` rates (conservative upper bound — InferHub bills
    discounts below official). `cost.cacheRead/cacheWrite` = 0 always —
    InferHub's /v1/models pricing carries no cache fields; never fabricate.
-3. **Fallback constants**: `CONTEXT_WINDOW_FALLBACK = 128_000`,
-   `MAX_TOKENS_FALLBACK = 8_192` — used when no resolvable member supplies
-   the field. pi-ai's `Model` type requires both as numbers; never guess a
-   model's ceiling other than via these constants.
+3. **Fallback constants**: `CONTEXT_WINDOW_FALLBACK = 262_144`,
+   `MAX_TOKENS_FALLBACK = 131_072` — used when no resolvable member supplies
+   the field (last resort after the alias sibling join; see 5). pi-ai's
+   `Model` type requires both as numbers; never guess a model's ceiling
+   other than via these constants.
 4. **Image rule**: `input: ["text","image"]` only when ≥1 resolvable member
    AND every resolvable member's `modality` includes `image`. Understating
    input is acceptable; overstating sends requests the upstream rejects.
-5. **Alias-member fallback**: empty resolvable set (alias-only combos, or
-   all member joins missed) → `input:["text"]`, `reasoning:true`, zero
-   cost, fallback limits. `reasoning:true` is deliberate: a false negative
-   hides thinking levels (the worse error); a false positive shows an inert
-   knob.
+5. **Alias-member fallback**: bare alias rows (`owned_by: "alias"`, no
+   data — observed live 2026-10-15) and alias-kind members first resolve
+   through prefixed sibling servings of the same model name (`<prefix>/<name>`
+   rows in the /v1/models catalog, min-of applies). Only when no sibling
+   carries data either → `input:["text"]`, `reasoning:true`, zero cost,
+   fallback limits. `reasoning:true` is deliberate: a false negative hides
+   thinking levels (the worse error); a false positive shows an inert knob.
 6. **thinkingLevelMap**: built from the union of resolvable members'
    `reasoning_levels`; each pi level (`minimal|low|medium|high|xhigh|max`)
    maps to itself when advertised, else `null`; `off:"off"`. Set only when
@@ -100,10 +103,20 @@ npm pack --dry-run    # inspect package contents before publishing
 pi packs the extension from `package.json` `pi.extensions:
 ["./src/index.ts"]`. Local dev: `pi -e <repo path>`. Published install:
 `npm:pi-inferhub` in `~/.pi/agent/settings.json` packages (NOT yet
-published as of 2026-10-07 — v0.1.0).
+	published npm:pi-inferhub (v0.1.1).
 
 ## Environment gotchas (learned the hard way)
 
+- **Pi packages declare host modules as peerDependencies, never
+  dependencies**: `@earendil-works/pi-ai`, `@earendil-works/pi-coding-agent`,
+  `typebox` (and pi-tui if used) go in `peerDependencies` with `"*"` ranges;
+  keep `dependencies` empty. Regular dependencies make pi's package installer
+  warn about duplicate host-module copies — extensions must bind to the
+  copies the pi host already loads. Mirror the same modules in
+  `devDependencies` (with real version pins) for the repo's own
+  tests/typecheck.
+
+
 - **TypeScript 6.0.3**: tsconfig MUST carry `types: ["node"]` — TS 6 dropped
   automatic `@types/*` inclusion; without it `node:module`/`fetch`/`AbortSignal`
   don't resolve and `tsc --noEmit` fails.
@@ -112,9 +125,6 @@ published as of 2026-10-07 — v0.1.0).
   error detail lives on `instancePath` (there is no `path` field). Extra
   payload keys pass through by default; do NOT add
   `additionalProperties: false` to wire schemas.
-- **Git signing**: fresh clones inherit `gpg.format=ssh` globally but need
-  `git config user.signingkey ~/.ssh/michaelmacleod.id_ecdsa.pub` locally
-  before the first commit fails.
 - **pi-ai `AuthEvent`** is a union; non-info variants carry no `message`.
   Anything typing an interaction callback as `{type: string; message:
   string}` breaks under contravariance — use `{type: string; message?:

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "pi-inferhub",
-	"version": "0.1.0",
+	"version": "0.2.0",
 	"description": "Pi extension for the InferHub inference provider",
 	"license": "MIT",
 	"keywords": [
@@ -16,13 +16,22 @@
 		"test": "vitest run",
 		"typecheck": "tsc --noEmit"
 	},
-	"dependencies": {
-		"@earendil-works/pi-ai": "^1.0.4",
-		"@earendil-works/pi-coding-agent": "^1.0.4",
-		"typebox": "^1.3.36"
+	"dependencies": {},
+	"peerDependencies": {
+		"@earendil-works/pi-ai": "*",
+		"@earendil-works/pi-coding-agent": "*",
+		"typebox": "*"
+	},
+	"peerDependenciesMeta": {
+		"@earendil-works/pi-ai": { "optional": false },
+		"@earendil-works/pi-coding-agent": { "optional": false },
+		"typebox": { "optional": false }
 	},
 	"devDependencies": {
+		"@earendil-works/pi-ai": "^1.0.4",
+		"@earendil-works/pi-coding-agent": "^1.0.4",
 		"@types/node": "^24.0.0",
+		"typebox": "^1.3.36",
 		"typescript": "^6.0.3",
 		"vitest": "^5.0.3"
 	}

package/src/catalog.ts

@@ -7,8 +7,8 @@ import { INFERENCE_BASE_URL, MGMT_BASE_URL, PROVIDER_NAME } from "./inferhub.js"
 /** pi thinking levels, in ladder order; `off` is handled separately. */
 const PI_THINKING_LEVELS = ["minimal", "low", "medium", "high", "xhigh", "max"] as const satisfies readonly ThinkingLevel[];
 
-const CONTEXT_WINDOW_FALLBACK = 128_000;
-const MAX_TOKENS_FALLBACK = 8_192;
+const CONTEXT_WINDOW_FALLBACK = 262_144;
+const MAX_TOKENS_FALLBACK = 131_072;
 
 export const MemberSchema = Type.Union([
 	Type.Object({ kind: Type.Literal("model"), model: Type.String() }),
@@ -24,8 +24,12 @@ export const ComboSchema = Type.Object({
 });
 export type Combo = Static<typeof ComboSchema>;
 
+/** Live-pinned 2026-10-15: bare (prefix-less) catalog rows are aliases served by
+ *  prefixed upstream rows — they carry owned_by "alias" and NO limits/pricing/
+ *  reasoning fields at all. Unknown extra fields pass validation regardless. */
 export const CatalogModelSchema = Type.Object({
 	id: Type.String(),
+	owned_by: Type.Optional(Type.String()),
 	modality: Type.Optional(Type.String()),
 	input_token_limit: Type.Optional(Type.Number()),
 	max_output_tokens: Type.Optional(Type.Number()),
@@ -34,6 +38,8 @@ export const CatalogModelSchema = Type.Object({
 		Type.Object({
 			official_in: Type.Number(),
 			official_out: Type.Number(),
+			min_ask_in: Type.Optional(Type.Number()),
+			min_ask_out: Type.Optional(Type.Number()),
 		}),
 	),
 });
@@ -45,6 +51,10 @@ export type CatalogModel = Static<typeof CatalogModelSchema>;
  * validation (no additionalProperties restrictions).
  */
 const CombosEnvelope = Type.Array(ComboSchema);
+/** Bare (prefix-less) alias row shape observed live 2026-10-15:
+ *  {id, object:"model", created, owned_by:"alias", supports_cache} — no limits,
+ *  pricing, modality, or reasoning fields at all. Prefixed servings of the same
+ *  model carry the data; catalog.ts resolves aliases through them (sibling join). */
 const ModelsEnvelope = Type.Object({ data: Type.Array(CatalogModelSchema) });
 
 const CombosValidator = Compile(CombosEnvelope);
@@ -69,18 +79,29 @@ export function combosToModels(
 	catalogModels: CatalogModel[],
 ): Model<"openai-completions">[] {
 	const catalogById = new Map(catalogModels.map((model) => [model.id, model]));
-	return combos.map((combo) => comboToModel(combo, catalogById));
+	return combos.map((combo) => comboToModel(combo, catalogById, catalogModels));
 }
 
 function comboToModel(
 	combo: Combo,
 	catalogById: Map<string, CatalogModel>,
+	catalogModels: CatalogModel[],
 ): Model<"openai-completions"> {
+	// Alias rows (`owned_by: "alias"`, no data) and alias-kind members resolve
+	// through prefixed sibling servings of the same model name; min-of over the
+	// siblings bounds the combo exactly like directly-joined members.
 	const resolvable: CatalogModel[] = [];
 	for (const member of combo.members) {
-		if (member.kind !== "model") continue;
-		const entry = catalogById.get(member.model);
-		if (entry !== undefined) resolvable.push(entry);
+		const name = member.kind === "model" ? member.model : member.alias;
+		const entry = member.kind === "model" ? catalogById.get(member.model) : undefined;
+		if (entry !== undefined && entry.owned_by !== "alias") {
+			resolvable.push(entry);
+			continue;
+		}
+		for (const sibling of catalogModels) {
+			if (sibling.owned_by === "alias") continue;
+			if (sibling.id.endsWith(`/${name}`)) resolvable.push(sibling);
+		}
 	}
 
 	// Alias-only / all-missed-join combos reason optimistically: a false

package/src/index.ts

@@ -55,7 +55,7 @@ export function registerInferhubProvider(pi: InferhubRegistrationPi): void {
 			const status = ctx.modelRegistry.getProviderAuthStatus(PROVIDER_NAME);
 			if (status?.configured) {
 				const source = status.label ? ` (${status.label})` : "";
-				ctx.ui.notify(`InferHub is configured${source}. Run /models to pick a combo.`, "info");
+				ctx.ui.notify(`InferHub is configured${source}. Run /model to pick a combo.`, "info");
 				return;
 			}
 			ctx.ui.notify(

package/test/catalog.test.ts

@@ -20,6 +20,10 @@ const combo = (members: Combo["members"], overrides: Partial<Combo> = {}): Combo
 	id: "c1", name: "My Combo", slug: "my-combo", members, ...overrides,
 });
 
+/** Bare alias row shape observed live 2026-10-15: id without prefix, owned_by "alias",
+ *  no limits/pricing/reasoning fields at all. */
+const aliasRow = (name: string): CatalogModel => ({ id: name, owned_by: "alias" });
+
 describe("combosToModels", () => {
 	it("maps id, name, provider, baseUrl, api", () => {
 		const [m] = combosToModels([combo([{ kind: "model", model: "croc/glm-5.3" }])], catalog);
@@ -49,8 +53,8 @@ describe("combosToModels", () => {
 		expect(aliasOnly[0].input).toEqual(["text"]);
 		expect(aliasOnly[0].reasoning).toBe(true);
 		expect(aliasOnly[0].cost).toEqual({ input: 0, output: 0, cacheRead: 0, cacheWrite: 0 });
-		expect(aliasOnly[0].contextWindow).toBe(128_000);
-		expect(aliasOnly[0].maxTokens).toBe(8_192);
+		expect(aliasOnly[0].contextWindow).toBe(262_144);
+		expect(aliasOnly[0].maxTokens).toBe(131_072);
 
 		const partial = combosToModels([combo([{ kind: "model", model: "ghost/gone" }, { kind: "model", model: "croc/glm-5.3" }])], catalog);
 		expect(partial[0].contextWindow).toBe(200_000);
@@ -66,6 +70,51 @@ describe("combosToModels", () => {
 		expect(combosToModels([], catalog)).toEqual([]);
 	});
 
+	it("raises bare alias rows and alias-kind members via prefixed sibling join (live combo/fast shape)", () => {
+		// Members named like a user-built combo: bare names whose catalog rows are
+		// information-free alias rows, but prefixed servings of the same model exist.
+		const sibCatalog: CatalogModel[] = [
+			aliasRow("qwen3.8-flash"),
+			aliasRow("deepseek-v4.1-flash"),
+			aliasRow("glm-5.3-flash"),
+			{ id: "ocg/qwen3.8-flash", ...base },
+			{ id: "ali/qwen3.8-flash", ...base },
+			{ id: "ocg/deepseek-v4.1-flash", ...base, pricing: { official_in: 2.0, official_out: 10.0 } },
+			{ id: "zai/glm-5.3-flash", ...base },
+		];
+		const [m] = combosToModels([combo([
+			{ kind: "model", model: "qwen3.8-flash" },
+			{ kind: "alias", alias: "deepseek-v4.1-flash" },
+			{ kind: "model", model: "glm-5.3-flash" },
+		])], sibCatalog);
+		expect(m.contextWindow).toBe(200_000); // min over sibling servings, not the 128k fallback
+		expect(m.maxTokens).toBe(32_000);
+		expect(m.reasoning).toBe(true);
+		expect(m.thinkingLevelMap).toMatchObject({ low: "low", high: "high" });
+		expect(m.input).toEqual(["text", "image"]); // every sibling serving is vision-capable
+		expect(m.cost).toEqual({ input: 2.0, output: 10.0, cacheRead: 0, cacheWrite: 0 });
+	});
+
+	it("does not cross-join a siblingless alias and keeps safe defaults (claude fable shape)", () => {
+		// cc/ row exists but is null everywhere and has NO prefixed siblings carrying data.
+		const [m] = combosToModels([combo([{ kind: "model", model: "claude-fable-5-1" }])], [aliasRow("claude-fable-5-1")]);
+		expect(m.contextWindow).toBe(262_144);
+		expect(m.maxTokens).toBe(131_072);
+		expect(m.reasoning).toBe(true); // alias-only optimism invariant
+		expect(m.cost).toEqual({ input: 0, output: 0, cacheRead: 0, cacheWrite: 0 });
+	});
+
+	it("prefers a direct row's real data over sibling join", () => {
+		// A prefixed member resolves directly; siblings must not inflate its set.
+		const sibCatalog: CatalogModel[] = [
+			{ id: "croc/glm-5.3", ...base },
+			{ id: "other/glm-5.3", ...base, input_token_limit: 1_000_000, max_output_tokens: 128_000 },
+		];
+		const [m] = combosToModels([combo([{ kind: "model", model: "croc/glm-5.3" }])], sibCatalog);
+		expect(m.contextWindow).toBe(200_000); // only the directly-joined member counts
+		expect(m.maxTokens).toBe(32_000);
+	});
+
 	it("reasoning false when no member advertises levels", () => {
 		const noLevels: CatalogModel = { id: "x/plain", modality: "text" };
 		const [m] = combosToModels([combo([{ kind: "model", model: "x/plain" }])], [noLevels]);