modelrig 0.9.0 → 0.11.0

package/dist/prompts/version-hash.d.ts

@@ -0,0 +1,29 @@
+/**
+ * PSR-2 — the content-hash id for a prompt version (`template_sha`).
+ *
+ * A prompt version is identified by the sha256 of its `template` + `manifest`
+ * (spec §2, the Humanloop content-hash pattern: identical content ⇒ no
+ * duplicate row). This is the ONE definition of that hash, used by:
+ *   • the ingest + create-version writer (dedupe: an identical re-ingest resolves
+ *     to the existing row rather than appending a byte-identical version), and
+ *   • the served-projection check (does routes.bundle carry THIS version? — the
+ *     "served this call" marker compares template_sha, PSR-2b).
+ *
+ * The manifest is CANONICALIZED first (object keys sorted recursively, array
+ * order preserved) so a semantically-identical manifest always hashes the same
+ * regardless of key order in the source artifact — the same discipline
+ * `resolved-bundle.ts` uses for `bundle_sha`. Callers MUST normalize the manifest
+ * ({@link normalizeManifest}) before hashing so incidental unknown keys never
+ * split the dedupe. Pure + dependency-free (node:crypto only).
+ */
+import type { Manifest } from "./manifest";
+/**
+ * The content-hash id of a prompt version — `sha256:<hex>`, self-describing so a
+ * future algorithm change is visible in the stored value rather than silent
+ * (parity with `hashResolvedBundle`). The template is hashed verbatim; the
+ * manifest is canonicalized. The two are joined with a NUL separator that cannot
+ * appear in the template, so no `(template, manifest)` pair can collide with a
+ * different pair by boundary ambiguity.
+ */
+export declare function promptTemplateSha(template: string, manifest: Manifest): string;
+//# sourceMappingURL=version-hash.d.ts.map

package/dist/prompts/version-hash.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"version-hash.d.ts","sourceRoot":"","sources":["../../src/prompts/version-hash.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAGH,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,YAAY,CAAC;AAkB3C;;;;;;;GAOG;AACH,wBAAgB,iBAAiB,CAAC,QAAQ,EAAE,MAAM,EAAE,QAAQ,EAAE,QAAQ,GAAG,MAAM,CAO9E"}

package/dist/prompts/version-hash.js

@@ -0,0 +1,54 @@
+"use strict";
+/**
+ * PSR-2 — the content-hash id for a prompt version (`template_sha`).
+ *
+ * A prompt version is identified by the sha256 of its `template` + `manifest`
+ * (spec §2, the Humanloop content-hash pattern: identical content ⇒ no
+ * duplicate row). This is the ONE definition of that hash, used by:
+ *   • the ingest + create-version writer (dedupe: an identical re-ingest resolves
+ *     to the existing row rather than appending a byte-identical version), and
+ *   • the served-projection check (does routes.bundle carry THIS version? — the
+ *     "served this call" marker compares template_sha, PSR-2b).
+ *
+ * The manifest is CANONICALIZED first (object keys sorted recursively, array
+ * order preserved) so a semantically-identical manifest always hashes the same
+ * regardless of key order in the source artifact — the same discipline
+ * `resolved-bundle.ts` uses for `bundle_sha`. Callers MUST normalize the manifest
+ * ({@link normalizeManifest}) before hashing so incidental unknown keys never
+ * split the dedupe. Pure + dependency-free (node:crypto only).
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.promptTemplateSha = promptTemplateSha;
+const node_crypto_1 = require("node:crypto");
+/** Deterministic JSON: object keys sorted recursively, array order preserved.
+ * The same value always produces the same bytes, so the hash is stable across
+ * processes. Mirrors resolved-bundle.ts's canonicalize (kept local so the two
+ * hashers stay independently auditable). */
+function canonicalize(value) {
+    if (Array.isArray(value))
+        return value.map(canonicalize);
+    if (value !== null && typeof value === "object") {
+        const out = {};
+        for (const key of Object.keys(value).sort()) {
+            out[key] = canonicalize(value[key]);
+        }
+        return out;
+    }
+    return value;
+}
+/**
+ * The content-hash id of a prompt version — `sha256:<hex>`, self-describing so a
+ * future algorithm change is visible in the stored value rather than silent
+ * (parity with `hashResolvedBundle`). The template is hashed verbatim; the
+ * manifest is canonicalized. The two are joined with a NUL separator that cannot
+ * appear in the template, so no `(template, manifest)` pair can collide with a
+ * different pair by boundary ambiguity.
+ */
+function promptTemplateSha(template, manifest) {
+    const canonicalManifest = JSON.stringify(canonicalize(manifest));
+    const hash = (0, node_crypto_1.createHash)("sha256");
+    hash.update(template, "utf8");
+    hash.update("\0", "utf8");
+    hash.update(canonicalManifest, "utf8");
+    return `sha256:${hash.digest("hex")}`;
+}

package/dist/prompts-export-cli.d.ts

@@ -0,0 +1,21 @@
+/**
+ * `modelrig prompts export` — PSR-3 (spec §3, the anti-lock-in gate).
+ *
+ *   modelrig prompts export --route <route> [--out modelrig/routes]
+ *       [--label prod | --version N] [--segment default] [--json]
+ *
+ * Fetches a prompt version from the ModelRig System-of-Record
+ * (`GET /v1/prompts/:route/export`, key-authed — a read needs only an org-scoped
+ * key) and writes its git files: the authored `template` to `<route>.md` and the
+ * assembly `manifest` to a sibling `<route>.manifest.json`, BYTE-FAITHFULLY. The
+ * command then PROVES the round-trip locally — it re-hashes the written files the
+ * way the store would (validate → normalize → hash) and checks the content id
+ * matches what ModelRig returned. ModelRig being your prompt's primary store
+ * never traps it in git.
+ *
+ * Goes through the API boundary (T12: the CLI never holds a DB client); the
+ * server scopes the read to the key's org (T11).
+ */
+import type { Flags } from "./cli";
+export declare function cmdPromptsExport(flags: Flags): Promise<number>;
+//# sourceMappingURL=prompts-export-cli.d.ts.map

package/dist/prompts-export-cli.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"prompts-export-cli.d.ts","sourceRoot":"","sources":["../src/prompts-export-cli.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAMH,OAAO,KAAK,EAAE,KAAK,EAAE,MAAM,OAAO,CAAC;AA0CnC,wBAAsB,gBAAgB,CAAC,KAAK,EAAE,KAAK,GAAG,OAAO,CAAC,MAAM,CAAC,CA4EpE"}

package/dist/prompts-export-cli.js

@@ -0,0 +1,130 @@
+"use strict";
+/**
+ * `modelrig prompts export` — PSR-3 (spec §3, the anti-lock-in gate).
+ *
+ *   modelrig prompts export --route <route> [--out modelrig/routes]
+ *       [--label prod | --version N] [--segment default] [--json]
+ *
+ * Fetches a prompt version from the ModelRig System-of-Record
+ * (`GET /v1/prompts/:route/export`, key-authed — a read needs only an org-scoped
+ * key) and writes its git files: the authored `template` to `<route>.md` and the
+ * assembly `manifest` to a sibling `<route>.manifest.json`, BYTE-FAITHFULLY. The
+ * command then PROVES the round-trip locally — it re-hashes the written files the
+ * way the store would (validate → normalize → hash) and checks the content id
+ * matches what ModelRig returned. ModelRig being your prompt's primary store
+ * never traps it in git.
+ *
+ * Goes through the API boundary (T12: the CLI never holds a DB client); the
+ * server scopes the read to the key's org (T11).
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.cmdPromptsExport = cmdPromptsExport;
+const node_fs_1 = require("node:fs");
+const node_path_1 = require("node:path");
+const config_1 = require("./config");
+const index_1 = require("./index");
+const DEFAULT_OUT = "modelrig/routes";
+function usage() {
+    return [
+        "usage: modelrig prompts export --route <route> [--out modelrig/routes]",
+        "                              [--label prod | --version N] [--segment default] [--json]",
+        "",
+        "Exports a prompt version from ModelRig back to git files (the anti-lock-in guarantee):",
+        "  <route>.md            the authored prompt template, verbatim",
+        "  <route>.manifest.json the assembly manifest (structure only)",
+        "A version re-created from these files carries the SAME content hash — proven after write.",
+        "",
+        "  MODELRIG_API_KEY    a rig_sk_ key in your org (a read needs no special scope)",
+        "  MODELRIG_INGEST_URL optional — the server base; defaults to the hosted API",
+    ].join("\n");
+}
+/** The content id a version re-created from the written files would carry — the
+ * store's write discipline (validate → normalize → hash) run locally, so the
+ * round-trip is proven without a second server call. */
+function recreatedSha(templateFile, manifestFile) {
+    let parsed;
+    try {
+        parsed = JSON.parse(manifestFile);
+    }
+    catch {
+        return null;
+    }
+    const validated = (0, index_1.validateManifest)(parsed);
+    if (!validated.ok)
+        return null;
+    return (0, index_1.promptTemplateSha)(templateFile, (0, index_1.normalizeManifest)(validated.manifest));
+}
+async function cmdPromptsExport(flags) {
+    // The subcommand token ("export") is flags.positional[0]; require it explicitly
+    // so `modelrig prompts` (no verb) prints usage rather than guessing.
+    if (flags.positional[0] !== "export") {
+        console.error(usage());
+        return 1;
+    }
+    const route = flags.named.get("route");
+    const out = flags.named.get("out") ?? DEFAULT_OUT;
+    const label = flags.named.get("label");
+    const versionRaw = flags.named.get("version");
+    const segment = flags.named.get("segment");
+    const asJson = flags.bare.has("json");
+    const exp = (0, config_1.loadConfigFromEnv)().export;
+    const api = exp !== undefined && exp.mode === "api" ? exp : undefined;
+    const base = api !== undefined ? new URL(api.ingestUrl).origin : "";
+    const key = api?.apiKey ?? "";
+    if (route === undefined || route === "" || base === "" || key === "") {
+        console.error(usage());
+        return 1;
+    }
+    const query = [];
+    if (versionRaw !== undefined && versionRaw !== "")
+        query.push(`version=${encodeURIComponent(versionRaw)}`);
+    else if (label !== undefined && label !== "")
+        query.push(`label=${encodeURIComponent(label)}`);
+    if (segment !== undefined && segment !== "")
+        query.push(`segment=${encodeURIComponent(segment)}`);
+    const suffix = query.length > 0 ? `?${query.join("&")}` : "";
+    let res;
+    try {
+        res = await fetch(`${base}/v1/prompts/${encodeURIComponent(route)}/export${suffix}`, {
+            headers: { authorization: `Bearer ${key}` },
+        });
+    }
+    catch (err) {
+        console.error(`prompts export: could not reach ${base}: ${err instanceof Error ? err.message : String(err)}`);
+        return 1;
+    }
+    const payload = (await res.json().catch(() => ({})));
+    if (!res.ok) {
+        console.error(`prompts export ${res.status}: ${[payload.error, payload.fix_hint].filter(Boolean).join(" — ") || "failed"}`);
+        return 1;
+    }
+    const files = payload.files ?? [];
+    if (files.length === 0 || payload.template_sha === undefined) {
+        console.error("prompts export: the server returned no files");
+        return 1;
+    }
+    // Write each file under --out, creating directories as needed.
+    const written = [];
+    for (const f of files) {
+        const dest = (0, node_path_1.join)(out, f.path);
+        (0, node_fs_1.mkdirSync)((0, node_path_1.dirname)(dest), { recursive: true });
+        (0, node_fs_1.writeFileSync)(dest, f.content);
+        written.push(dest);
+    }
+    // Prove the round-trip from the bytes on disk (never trust the transfer): the
+    // template file + the manifest file must re-hash to the server's content id.
+    const md = files.find((f) => f.path.endsWith(".md"));
+    const mf = files.find((f) => f.path.endsWith(".manifest.json"));
+    const rehashed = md !== undefined && mf !== undefined ? recreatedSha(md.content, mf.content) : null;
+    const roundTripped = rehashed !== null && rehashed === payload.template_sha;
+    if (asJson) {
+        console.log(JSON.stringify({ route: payload.route, version: payload.version, template_sha: payload.template_sha, written, round_trip: roundTripped }, null, 2));
+    }
+    else {
+        console.log(`Exported ${route} version ${String(payload.version)} → ${written.join(", ")}`);
+        console.log(roundTripped
+            ? `Round-trip verified: re-creating from these files reproduces the same content hash (${payload.template_sha.slice(0, 12)}…). Your prompt is not locked in.`
+            : "WARNING: the written files did NOT re-hash to the server's content id — do not rely on this export; re-run.");
+    }
+    return roundTripped ? 0 : 1;
+}

package/dist/registry/cache-policy.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"cache-policy.d.ts","sourceRoot":"","sources":["../../src/registry/cache-policy.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAiCG;AAEH,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,mBAAmB,CAAC;AAEpD;;;GAGG;AACH,MAAM,WAAW,WAAW;IAC1B;4CACwC;IACxC,QAAQ,CAAC,eAAe,EAAE,OAAO,CAAC;IAClC;;;2EAGuE;IACvE,QAAQ,CAAC,IAAI,EAAE,sBAAsB,GAAG,kBAAkB,GAAG,YAAY,CAAC;IAC1E,0EAA0E;IAC1E,QAAQ,CAAC,mBAAmB,EAAE,MAAM,GAAG,IAAI,CAAC;IAC5C;4CACwC;IACxC,QAAQ,CAAC,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7C;sFACkF;IAClF,QAAQ,CAAC,mBAAmB,EAAE,OAAO,CAAC;IACtC;8EAC0E;IAC1E,QAAQ,CAAC,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7C;mDAC+C;IAC/C,QAAQ,CAAC,eAAe,EAAE,kBAAkB,GAAG,eAAe,GAAG,eAAe,GAAG,IAAI,CAAC;IACxF;8EAC0E;IAC1E,QAAQ,CAAC,mBAAmB,EAAE,OAAO,CAAC;IACtC,qEAAqE;IACrE,QAAQ,CAAC,oBAAoB,EAAE,OAAO,CAAC;IACvC,qDAAqD;IACrD,QAAQ,CAAC,aAAa,EAClB,WAAW,GACX,gBAAgB,GAChB,oBAAoB,GACpB,iBAAiB,GACjB,mBAAmB,CAAC;IACxB;;8BAE0B;IAC1B,QAAQ,CAAC,0BAA0B,EAAE,OAAO,CAAC;IAC7C,4DAA4D;IAC5D,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;CACjC;AA8KD;;;;eAIe;AACf,wBAAgB,qBAAqB,IAAI,WAAW,CAiBnD;AAED;;;;;GAKG;AACH,wBAAgB,cAAc,CAAC,QAAQ,EAAE,UAAU,GAAG,WAAW,CAEhE;AAED,+EAA+E;AAC/E,wBAAgB,0BAA0B,IAAI,SAAS,UAAU,EAAE,CAElE"}
+{"version":3,"file":"cache-policy.d.ts","sourceRoot":"","sources":["../../src/registry/cache-policy.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAiCG;AAEH,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,mBAAmB,CAAC;AAEpD;;;GAGG;AACH,MAAM,WAAW,WAAW;IAC1B;4CACwC;IACxC,QAAQ,CAAC,eAAe,EAAE,OAAO,CAAC;IAClC;;;2EAGuE;IACvE,QAAQ,CAAC,IAAI,EAAE,sBAAsB,GAAG,kBAAkB,GAAG,YAAY,CAAC;IAC1E,0EAA0E;IAC1E,QAAQ,CAAC,mBAAmB,EAAE,MAAM,GAAG,IAAI,CAAC;IAC5C;4CACwC;IACxC,QAAQ,CAAC,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7C;sFACkF;IAClF,QAAQ,CAAC,mBAAmB,EAAE,OAAO,CAAC;IACtC;8EAC0E;IAC1E,QAAQ,CAAC,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7C;mDAC+C;IAC/C,QAAQ,CAAC,eAAe,EAAE,kBAAkB,GAAG,eAAe,GAAG,eAAe,GAAG,IAAI,CAAC;IACxF;8EAC0E;IAC1E,QAAQ,CAAC,mBAAmB,EAAE,OAAO,CAAC;IACtC,qEAAqE;IACrE,QAAQ,CAAC,oBAAoB,EAAE,OAAO,CAAC;IACvC,qDAAqD;IACrD,QAAQ,CAAC,aAAa,EAClB,WAAW,GACX,gBAAgB,GAChB,oBAAoB,GACpB,iBAAiB,GACjB,mBAAmB,CAAC;IACxB;;8BAE0B;IAC1B,QAAQ,CAAC,0BAA0B,EAAE,OAAO,CAAC;IAC7C,4DAA4D;IAC5D,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;CACjC;AAqMD;;;;eAIe;AACf,wBAAgB,qBAAqB,IAAI,WAAW,CAiBnD;AAED;;;;;GAKG;AACH,wBAAgB,cAAc,CAAC,QAAQ,EAAE,UAAU,GAAG,WAAW,CAEhE;AAED,+EAA+E;AAC/E,wBAAgB,0BAA0B,IAAI,SAAS,UAAU,EAAE,CAElE"}

package/dist/registry/cache-policy.js

@@ -199,6 +199,28 @@ const POLICIES = {
         as_of: AS_OF,
         policy_version: AS_OF,
     },
+    // Z.ai / Zhipu (B4a) — GLM context caching is not verified first-party here
+    // (docs egress-blocked at build time), so fail-closed to cache_supported:false,
+    // identical L5 posture to DeepInfra. The raw-lane `cache.key` still maps to the
+    // OpenAI-dialect `prompt_cache_key` HINT (raw-support.ts) — that is call-shaping,
+    // not a claim that a warm cache exists on our path. Revisit when a first-party
+    // Z.ai caching page can be cited (then flip with a new as_of).
+    zai: {
+        cache_supported: false,
+        mode: "automatic_prefix",
+        default_ttl_seconds: null,
+        extended_ttl_seconds: null,
+        ttl_refresh_on_read: false,
+        min_cacheable_tokens: null,
+        cache_key_param: null,
+        reports_read_tokens: false,
+        reports_write_tokens: false,
+        usage_dialect: "openai_details",
+        provider_affinity_required: false,
+        source: "z.ai — cached rates unverified first-party (fail-closed, L5)",
+        as_of: AS_OF,
+        policy_version: AS_OF,
+    },
 };
 /** The fail-closed policy for any provider not seeded (spec §3): caching is
  * assumed OFF until a first-party source says otherwise — identical posture to

package/dist/registry/pricing-shape.d.ts

@@ -29,11 +29,32 @@ export interface TokenRates {
      * ⇒ the 1h write bucket bills at `cache_write` (fail-closed, dev-rule L5). */
     readonly cache_write_1h?: number | null;
 }
+/**
+ * Per-component tier multipliers (issue #2). A vendor's tier does not always
+ * scale every component by one number: Google's Gemini `flex` (and batch) is
+ * 0.5× input + output while cache read is "Same as Standard" — so a single
+ * scalar over-discounts the cached portion. An ABSENT component here is EXEMPT
+ * (multiplier 1.0 — the standard rate), never a guess: write a component
+ * explicitly when the vendor publishes a different multiplier for it (L5).
+ */
+export interface TierComponentMultipliers {
+    readonly input: number;
+    readonly output: number;
+    /** absent/null ⇒ 1.0 (exempt — bills at the standard cache-read rate). */
+    readonly cache_read?: number | null;
+    readonly cache_write?: number | null;
+    readonly cache_write_1h?: number | null;
+}
+/** A tier multiplier: one scalar over EVERY component (`1.25`), or a
+ * per-component block (`{ input: 0.5, output: 0.5, cache_read: 1.0 }`). */
+export type TierMultiplier = number | TierComponentMultipliers;
 /** MULTIPLIERS on `base` (1.0 = base), NOT absolute rates. */
 export interface ServingTiers {
-    readonly priority?: number | null;
-    readonly fast?: number | null;
-    readonly region_us?: number | null;
+    readonly priority?: TierMultiplier | null;
+    readonly fast?: TierMultiplier | null;
+    readonly region_us?: TierMultiplier | null;
+    /** Gemini flex (issue #2): `{ input: 0.5, output: 0.5, cache_read: 1.0 }`. */
+    readonly flex?: TierMultiplier | null;
 }
 /** A rate change that applies above a token threshold. */
 export interface ContextBand {
@@ -101,12 +122,21 @@ export declare const SNAPSHOT_BASIS = "base \u00B7 standard tier \u00B7 text";
  * the caller supplies provenance. `source` defaults to the snapshot lane.
  */
 export declare function synthesizeFromFlat(input: number, output: number, cache_read: number | null, asOf: string, source?: string): ModelPricingV2;
-export type ServingTierName = "priority" | "fast" | "region_us";
+export type ServingTierName = "priority" | "fast" | "region_us" | "flex";
+/** Resolve a tier multiplier to one multiplier PER component. A scalar scales
+ * every component (the original contract, byte-identical); a per-component
+ * block fills an absent component with 1.0 — EXEMPT, i.e. the standard rate. */
+export declare function tierComponentMultipliers(mult: TierMultiplier): Required<{
+    [K in keyof TierComponentMultipliers]: number;
+}>;
 /**
- * Absolute rates for a serving tier = `base × multiplier`. Returns `null` when
- * the tier's multiplier is absent — fail-closed, so no surface can render a
- * tier price the model never published. The multiplier scales every rate
- * component present on `base`.
+ * Absolute rates for a serving tier = `base × multiplier`, PER COMPONENT.
+ * Returns `null` when the tier's multiplier is absent — fail-closed, so no
+ * surface (and no cost row) can carry a tier price the model never published.
+ * A scalar multiplier scales every rate component present on `base`; a
+ * per-component block scales each component by its own multiplier, with an
+ * absent component exempt (1.0) — issue #2: Gemini flex discounts input and
+ * output but NOT cache read.
  */
 export declare function tierRates(pricing: ModelPricingV2, tier: ServingTierName): TokenRates | null;
 /** A display-ready projection of a price. `basis` is always present, so a

package/dist/registry/pricing-shape.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"pricing-shape.d.ts","sourceRoot":"","sources":["../../src/registry/pricing-shape.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH,MAAM,WAAW,UAAU;IACzB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,+BAA+B;IAC/B,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACpC,uEAAuE;IACvE,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACrC;kFAC8E;IAC9E,QAAQ,CAAC,cAAc,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;CACzC;AAED,8DAA8D;AAC9D,MAAM,WAAW,YAAY;IAC3B,QAAQ,CAAC,QAAQ,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAClC,QAAQ,CAAC,IAAI,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAC9B,QAAQ,CAAC,SAAS,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;CACpC;AAED,0DAA0D;AAC1D,MAAM,WAAW,WAAW;IAC1B,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,KAAK,EAAE,UAAU,CAAC;CAC5B;AAED;;;;;;;;GAQG;AACH,MAAM,MAAM,eAAe,GACvB;IACE,kEAAkE;IAClE,QAAQ,CAAC,IAAI,EAAE,WAAW,CAAC;IAC3B,QAAQ,CAAC,aAAa,EAAE,MAAM,CAAC;IAC/B,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;CACxB,GACD;IACE;iDAC6C;IAC7C,QAAQ,CAAC,IAAI,EAAE,WAAW,CAAC;IAC3B,QAAQ,CAAC,aAAa,EAAE,MAAM,CAAC;IAC/B,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;CACxB,CAAC;AAEN,0CAA0C;AAC1C,MAAM,WAAW,aAAa;IAC5B,QAAQ,CAAC,cAAc,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACxC,QAAQ,CAAC,gBAAgB,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAC1C;;;;4DAIwD;IACxD,QAAQ,CAAC,WAAW,CAAC,EAAE,eAAe,GAAG,IAAI,CAAC;CAC/C;AAED,MAAM,WAAW,cAAc;IAC7B,0DAA0D;IAC1D,QAAQ,CAAC,IAAI,EAAE,UAAU,CAAC;IAC1B,QAAQ,CAAC,KAAK,CAAC,EAAE,YAAY,GAAG,IAAI,CAAC;IACrC,QAAQ,CAAC,aAAa,CAAC,EAAE,SAAS,WAAW,EAAE,GAAG,IAAI,CAAC;IACvD,QAAQ,CAAC,cAAc,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACxC,QAAQ,CAAC,QAAQ,CAAC,EAAE,aAAa,GAAG,IAAI,CAAC;IACzC,iDAAiD;IACjD,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,8EAA8E;IAC9E,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;CACxB;AAED;iFACiF;AACjF,eAAO,MAAM,cAAc,0CAAgC,CAAC;AAE5D;;;;;;;;;GASG;AACH,wBAAgB,kBAAkB,CAChC,KAAK,EAAE,MAAM,EACb,MAAM,EAAE,MAAM,EACd,UAAU,EAAE,MAAM,GAAG,IAAI,EACzB,IAAI,EAAE,MAAM,EACZ,MAAM,SAAqB,GAC1B,cAAc,CAWhB;AAED,MAAM,MAAM,eAAe,GAAG,UAAU,GAAG,MAAM,GAAG,WAAW,CAAC;AAMhE;;;;;GAKG;AACH,wBAAgB,SAAS,CAAC,OAAO,EAAE,cAAc,EAAE,IAAI,EAAE,eAAe,GAAG,UAAU,GAAG,IAAI,CAU3F;AAED;gFACgF;AAChF,MAAM,WAAW,gBAAgB;IAC/B,wEAAwE;IACxE,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,yDAAyD;IACzD,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IAClC,QAAQ,CAAC,UAAU,EAAE,MAAM,GAAG,IAAI,CAAC;IACnC,8EAA8E;IAC9E,QAAQ,CAAC,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;IACrC,wDAAwD;IACxD,QAAQ,CAAC,KAAK,EAAE,SAAS;QAAE,QAAQ,CAAC,IAAI,EAAE,eAAe,CAAC;QAAC,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAA;KAAE,EAAE,CAAC;IACrF,QAAQ,CAAC,YAAY,EAAE,SAAS,MAAM,EAAE,CAAC;IACzC,QAAQ,CAAC,KAAK,EAAE,MAAM,GAAG,IAAI,CAAC;IAC9B,QAAQ,CAAC,QAAQ,EAAE,SAAS,MAAM,EAAE,CAAC;IACrC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,2EAA2E;IAC3E,QAAQ,CAAC,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;CACnC;AAqBD;;;;GAIG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,cAAc,GAAG,gBAAgB,CA+C3E"}
+{"version":3,"file":"pricing-shape.d.ts","sourceRoot":"","sources":["../../src/registry/pricing-shape.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH,MAAM,WAAW,UAAU;IACzB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,+BAA+B;IAC/B,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACpC,uEAAuE;IACvE,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACrC;kFAC8E;IAC9E,QAAQ,CAAC,cAAc,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;CACzC;AAED;;;;;;;GAOG;AACH,MAAM,WAAW,wBAAwB;IACvC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,0EAA0E;IAC1E,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACpC,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACrC,QAAQ,CAAC,cAAc,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;CACzC;AAED;2EAC2E;AAC3E,MAAM,MAAM,cAAc,GAAG,MAAM,GAAG,wBAAwB,CAAC;AAE/D,8DAA8D;AAC9D,MAAM,WAAW,YAAY;IAC3B,QAAQ,CAAC,QAAQ,CAAC,EAAE,cAAc,GAAG,IAAI,CAAC;IAC1C,QAAQ,CAAC,IAAI,CAAC,EAAE,cAAc,GAAG,IAAI,CAAC;IACtC,QAAQ,CAAC,SAAS,CAAC,EAAE,cAAc,GAAG,IAAI,CAAC;IAC3C,8EAA8E;IAC9E,QAAQ,CAAC,IAAI,CAAC,EAAE,cAAc,GAAG,IAAI,CAAC;CACvC;AAED,0DAA0D;AAC1D,MAAM,WAAW,WAAW;IAC1B,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,KAAK,EAAE,UAAU,CAAC;CAC5B;AAED;;;;;;;;GAQG;AACH,MAAM,MAAM,eAAe,GACvB;IACE,kEAAkE;IAClE,QAAQ,CAAC,IAAI,EAAE,WAAW,CAAC;IAC3B,QAAQ,CAAC,aAAa,EAAE,MAAM,CAAC;IAC/B,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;CACxB,GACD;IACE;iDAC6C;IAC7C,QAAQ,CAAC,IAAI,EAAE,WAAW,CAAC;IAC3B,QAAQ,CAAC,aAAa,EAAE,MAAM,CAAC;IAC/B,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;CACxB,CAAC;AAEN,0CAA0C;AAC1C,MAAM,WAAW,aAAa;IAC5B,QAAQ,CAAC,cAAc,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACxC,QAAQ,CAAC,gBAAgB,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAC1C;;;;4DAIwD;IACxD,QAAQ,CAAC,WAAW,CAAC,EAAE,eAAe,GAAG,IAAI,CAAC;CAC/C;AAED,MAAM,WAAW,cAAc;IAC7B,0DAA0D;IAC1D,QAAQ,CAAC,IAAI,EAAE,UAAU,CAAC;IAC1B,QAAQ,CAAC,KAAK,CAAC,EAAE,YAAY,GAAG,IAAI,CAAC;IACrC,QAAQ,CAAC,aAAa,CAAC,EAAE,SAAS,WAAW,EAAE,GAAG,IAAI,CAAC;IACvD,QAAQ,CAAC,cAAc,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACxC,QAAQ,CAAC,QAAQ,CAAC,EAAE,aAAa,GAAG,IAAI,CAAC;IACzC,iDAAiD;IACjD,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,8EAA8E;IAC9E,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;CACxB;AAED;iFACiF;AACjF,eAAO,MAAM,cAAc,0CAAgC,CAAC;AAE5D;;;;;;;;;GASG;AACH,wBAAgB,kBAAkB,CAChC,KAAK,EAAE,MAAM,EACb,MAAM,EAAE,MAAM,EACd,UAAU,EAAE,MAAM,GAAG,IAAI,EACzB,IAAI,EAAE,MAAM,EACZ,MAAM,SAAqB,GAC1B,cAAc,CAWhB;AAED,MAAM,MAAM,eAAe,GAAG,UAAU,GAAG,MAAM,GAAG,WAAW,GAAG,MAAM,CAAC;AAMzE;;gFAEgF;AAChF,wBAAgB,wBAAwB,CAAC,IAAI,EAAE,cAAc,GAAG,QAAQ,CAAC;KACtE,CAAC,IAAI,MAAM,wBAAwB,GAAG,MAAM;CAC9C,CAAC,CAWD;AAED;;;;;;;;GAQG;AACH,wBAAgB,SAAS,CAAC,OAAO,EAAE,cAAc,EAAE,IAAI,EAAE,eAAe,GAAG,UAAU,GAAG,IAAI,CAW3F;AAED;gFACgF;AAChF,MAAM,WAAW,gBAAgB;IAC/B,wEAAwE;IACxE,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,yDAAyD;IACzD,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IAClC,QAAQ,CAAC,UAAU,EAAE,MAAM,GAAG,IAAI,CAAC;IACnC,8EAA8E;IAC9E,QAAQ,CAAC,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;IACrC,wDAAwD;IACxD,QAAQ,CAAC,KAAK,EAAE,SAAS;QAAE,QAAQ,CAAC,IAAI,EAAE,eAAe,CAAC;QAAC,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAA;KAAE,EAAE,CAAC;IACrF,QAAQ,CAAC,YAAY,EAAE,SAAS,MAAM,EAAE,CAAC;IACzC,QAAQ,CAAC,KAAK,EAAE,MAAM,GAAG,IAAI,CAAC;IAC9B,QAAQ,CAAC,QAAQ,EAAE,SAAS,MAAM,EAAE,CAAC;IACrC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,2EAA2E;IAC3E,QAAQ,CAAC,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;CACnC;AAqBD;;;;GAIG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,cAAc,GAAG,gBAAgB,CAwD3E"}

package/dist/registry/pricing-shape.js

@@ -22,6 +22,7 @@
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.SNAPSHOT_BASIS = void 0;
 exports.synthesizeFromFlat = synthesizeFromFlat;
+exports.tierComponentMultipliers = tierComponentMultipliers;
 exports.tierRates = tierRates;
 exports.describePricingV2 = describePricingV2;
 /** The basis string for a snapshot-synthesized (base-only, standard, text)
@@ -52,25 +53,44 @@ function synthesizeFromFlat(input, output, cache_read, asOf, source = "litellm-s
 function scale(rate, mult) {
     return rate === null || rate === undefined ? null : rate * mult;
 }
+/** Resolve a tier multiplier to one multiplier PER component. A scalar scales
+ * every component (the original contract, byte-identical); a per-component
+ * block fills an absent component with 1.0 — EXEMPT, i.e. the standard rate. */
+function tierComponentMultipliers(mult) {
+    if (typeof mult === "number") {
+        return { input: mult, output: mult, cache_read: mult, cache_write: mult, cache_write_1h: mult };
+    }
+    return {
+        input: mult.input,
+        output: mult.output,
+        cache_read: mult.cache_read ?? 1,
+        cache_write: mult.cache_write ?? 1,
+        cache_write_1h: mult.cache_write_1h ?? 1,
+    };
+}
 /**
- * Absolute rates for a serving tier = `base × multiplier`. Returns `null` when
- * the tier's multiplier is absent — fail-closed, so no surface can render a
- * tier price the model never published. The multiplier scales every rate
- * component present on `base`.
+ * Absolute rates for a serving tier = `base × multiplier`, PER COMPONENT.
+ * Returns `null` when the tier's multiplier is absent — fail-closed, so no
+ * surface (and no cost row) can carry a tier price the model never published.
+ * A scalar multiplier scales every rate component present on `base`; a
+ * per-component block scales each component by its own multiplier, with an
+ * absent component exempt (1.0) — issue #2: Gemini flex discounts input and
+ * output but NOT cache read.
  */
 function tierRates(pricing, tier) {
     const mult = pricing.tiers?.[tier];
     if (mult === null || mult === undefined)
         return null;
+    const m = tierComponentMultipliers(mult);
     return {
-        input: pricing.base.input * mult,
-        output: pricing.base.output * mult,
-        cache_read: scale(pricing.base.cache_read, mult),
-        cache_write: scale(pricing.base.cache_write, mult),
-        cache_write_1h: scale(pricing.base.cache_write_1h, mult),
+        input: pricing.base.input * m.input,
+        output: pricing.base.output * m.output,
+        cache_read: scale(pricing.base.cache_read, m.cache_read),
+        cache_write: scale(pricing.base.cache_write, m.cache_write),
+        cache_write_1h: scale(pricing.base.cache_write_1h, m.cache_write_1h),
     };
 }
-const TIER_ORDER = ["priority", "fast", "region_us"];
+const TIER_ORDER = ["priority", "fast", "region_us", "flex"];
 /** USD-per-token → "$X.XX" per Mtok. Adaptive precision: a pricing-authority
  * surface must not misrepresent a rate — `$0.075/M` renders as `$0.075`, not a
  * rounded `$0.07`/`$0.08`. Shows up to 6 decimals, trailing zeros trimmed, with
@@ -94,8 +114,16 @@ function describePricingV2(pricing) {
     const tiers = [];
     for (const name of TIER_ORDER) {
         const rates = tierRates(pricing, name);
-        if (rates !== null)
-            tiers.push({ name, rate: inOut(rates) });
+        if (rates === null)
+            continue;
+        // Every scaled tier that publishes a cache-read rate states it explicitly,
+        // so a reader never infers the cache rate from the in/out multiplier
+        // (gemini flex: in/out 0.5× but cache read NOT discounted). Only a tier whose
+        // input equals base (a 1.0× tier) omits the note.
+        const cacheNote = rates.cache_read != null && pricing.base.cache_read != null && rates.input !== pricing.base.input
+            ? ` · cache read ${perMtok(rates.cache_read)}`
+            : "";
+        tiers.push({ name, rate: `${inOut(rates)}${cacheNote}` });
     }
     const contextBands = (pricing.context_bands ?? []).map((b) => `above ${b.above_tokens.toLocaleString()} tok: ${inOut(b.rates)}`);
     const modality = [];

package/dist/registry/pricing.d.ts

@@ -25,11 +25,12 @@ export interface ModelPricing {
 }
 export declare function getPricing(provider: ProviderId, model: string): ModelPricing | null;
 /**
- * The richer `ModelPricingV2` a vendor overlay/manual entry carries, if any —
- * resolved by the SAME key precedence as {@link getPricing} (snapshot wins on a
- * shared key, and the snapshot never carries `pricing_v2`, so a shared key
- * yields null → the caller synthesizes from the flat rate). Returns null when
- * no richer shape is on file. Never guesses: absence is `base`-only downstream.
+ * The richer `ModelPricingV2` on file for a model, if any — resolved by the SAME
+ * key precedence as {@link getPricing} (snapshot wins on a shared key; a snapshot
+ * entry yields its hand-grafted `pricing_v2` when it has one — issue #2's gemini
+ * tiers — else null → the caller synthesizes from the flat rate; an overlay or
+ * manual entry yields its own block). Returns null when no richer shape is on
+ * file. Never guesses: absence is `base`-only downstream.
  */
 export declare function getPricingV2(provider: ProviderId, model: string): ModelPricingV2 | null;
 export declare function estimateCostUsd(pricing: ModelPricing, tokensIn: number, tokensOut: number, tokensCached: number, 

package/dist/registry/pricing.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"pricing.d.ts","sourceRoot":"","sources":["../../src/registry/pricing.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAIH,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,mBAAmB,CAAC;AACpD,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,iBAAiB,CAAC;AAEtD,MAAM,WAAW,YAAY;IAC3B,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;IAChC,QAAQ,CAAC,eAAe,EAAE,MAAM,CAAC;IACjC,QAAQ,CAAC,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7C;;uEAEmE;IACnE,QAAQ,CAAC,mBAAmB,EAAE,MAAM,GAAG,IAAI,CAAC;IAC5C;;sEAEkE;IAClE,QAAQ,CAAC,qBAAqB,EAAE,MAAM,GAAG,IAAI,CAAC;CAC/C;AAyDD,wBAAgB,UAAU,CAAC,QAAQ,EAAE,UAAU,EAAE,KAAK,EAAE,MAAM,GAAG,YAAY,GAAG,IAAI,CA6BnF;AAED;;;;;;GAMG;AACH,wBAAgB,YAAY,CAAC,QAAQ,EAAE,UAAU,EAAE,KAAK,EAAE,MAAM,GAAG,cAAc,GAAG,IAAI,CAYvF;AAED,wBAAgB,eAAe,CAC7B,OAAO,EAAE,YAAY,EACrB,QAAQ,EAAE,MAAM,EAChB,SAAS,EAAE,MAAM,EACjB,YAAY,EAAE,MAAM;AACpB;;;0CAG0C;AAC1C,gBAAgB,SAAI;AACpB,+EAA+E;AAC/E,kBAAkB,SAAI,GACrB,MAAM,CAsBR"}
+{"version":3,"file":"pricing.d.ts","sourceRoot":"","sources":["../../src/registry/pricing.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAIH,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,mBAAmB,CAAC;AACpD,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,iBAAiB,CAAC;AAEtD,MAAM,WAAW,YAAY;IAC3B,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;IAChC,QAAQ,CAAC,eAAe,EAAE,MAAM,CAAC;IACjC,QAAQ,CAAC,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7C;;uEAEmE;IACnE,QAAQ,CAAC,mBAAmB,EAAE,MAAM,GAAG,IAAI,CAAC;IAC5C;;sEAEkE;IAClE,QAAQ,CAAC,qBAAqB,EAAE,MAAM,GAAG,IAAI,CAAC;CAC/C;AA+DD,wBAAgB,UAAU,CAAC,QAAQ,EAAE,UAAU,EAAE,KAAK,EAAE,MAAM,GAAG,YAAY,GAAG,IAAI,CA6BnF;AAED;;;;;;;GAOG;AACH,wBAAgB,YAAY,CAAC,QAAQ,EAAE,UAAU,EAAE,KAAK,EAAE,MAAM,GAAG,cAAc,GAAG,IAAI,CAYvF;AAED,wBAAgB,eAAe,CAC7B,OAAO,EAAE,YAAY,EACrB,QAAQ,EAAE,MAAM,EAChB,SAAS,EAAE,MAAM,EACjB,YAAY,EAAE,MAAM;AACpB;;;0CAG0C;AAC1C,gBAAgB,SAAI;AACpB,+EAA+E;AAC/E,kBAAkB,SAAI,GACrB,MAAM,CAsBR"}

package/dist/registry/pricing.js

@@ -29,6 +29,9 @@ const KEY_PREFIXES = {
     // Fireworks pricing keys carry the accounts/… wire prefix.
     deepinfra: ["deepinfra/"],
     fireworks: ["fireworks_ai/accounts/fireworks/models/", "fireworks_ai/"],
+    // Z.ai (B4a): GLM models keyed under "zai/" in the snapshot. Registry entries
+    // carry their own declared pricing, so this only gaps-fills a future snapshot.
+    zai: ["zai/"],
 };
 let cachedMap = null;
 /** Host-swept gap-filler (scripts/pricing-sweep.ts): prices for models the
@@ -77,11 +80,12 @@ function getPricing(provider, model) {
     return null;
 }
 /**
- * The richer `ModelPricingV2` a vendor overlay/manual entry carries, if any —
- * resolved by the SAME key precedence as {@link getPricing} (snapshot wins on a
- * shared key, and the snapshot never carries `pricing_v2`, so a shared key
- * yields null → the caller synthesizes from the flat rate). Returns null when
- * no richer shape is on file. Never guesses: absence is `base`-only downstream.
+ * The richer `ModelPricingV2` on file for a model, if any — resolved by the SAME
+ * key precedence as {@link getPricing} (snapshot wins on a shared key; a snapshot
+ * entry yields its hand-grafted `pricing_v2` when it has one — issue #2's gemini
+ * tiers — else null → the caller synthesizes from the flat rate; an overlay or
+ * manual entry yields its own block). Returns null when no richer shape is on
+ * file. Never guesses: absence is `base`-only downstream.
  */
 function getPricingV2(provider, model) {
     if (model === "")

package/dist/routes/destination-loader.d.ts

@@ -0,0 +1,16 @@
+/**
+ * Serving-`destination` route-field validation (Unified Route U2, ruling D-2).
+ * Extracted from loader.ts so loader.ts holds under the 700-LOC cap (same reason
+ * loader-warnings.ts / screen-loader.ts were carved out). Hand-written validation
+ * mirroring the loader's PROVIDERS/TIERS pattern: a route's `destination:` is
+ * checked against the DESTINATIONS vocabulary (the ONE source of truth in
+ * run/destination.ts — no second list to drift); an unknown value pushes a
+ * teaching problem (the caller turns it into a RouteConfigError). Absent ⇒
+ * undefined ⇒ resolves to `direct` at run time ⇒ BYTE-IDENTICAL.
+ */
+import type { Destination } from "../run/destination";
+/** Validate a route bundle's `destination:` value. Absent/null ⇒ undefined
+ * (defaults to `direct`); a known value ⇒ that `Destination`; an unknown value ⇒
+ * a pushed problem and undefined. */
+export declare function parseDestination(raw: unknown, problems: string[]): Destination | undefined;
+//# sourceMappingURL=destination-loader.d.ts.map

package/dist/routes/destination-loader.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"destination-loader.d.ts","sourceRoot":"","sources":["../../src/routes/destination-loader.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAGH,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AAMtD;;qCAEqC;AACrC,wBAAgB,gBAAgB,CAAC,GAAG,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,EAAE,GAAG,WAAW,GAAG,SAAS,CAO1F"}

package/dist/routes/destination-loader.js

@@ -0,0 +1,29 @@
+"use strict";
+/**
+ * Serving-`destination` route-field validation (Unified Route U2, ruling D-2).
+ * Extracted from loader.ts so loader.ts holds under the 700-LOC cap (same reason
+ * loader-warnings.ts / screen-loader.ts were carved out). Hand-written validation
+ * mirroring the loader's PROVIDERS/TIERS pattern: a route's `destination:` is
+ * checked against the DESTINATIONS vocabulary (the ONE source of truth in
+ * run/destination.ts — no second list to drift); an unknown value pushes a
+ * teaching problem (the caller turns it into a RouteConfigError). Absent ⇒
+ * undefined ⇒ resolves to `direct` at run time ⇒ BYTE-IDENTICAL.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.parseDestination = parseDestination;
+const destination_1 = require("../run/destination");
+/** The serving-`destination` vocabulary as a lookup set, mirroring the loader's
+ * PROVIDERS/TIERS ReadonlySets. */
+const DESTINATIONS = new Set(destination_1.DESTINATIONS);
+/** Validate a route bundle's `destination:` value. Absent/null ⇒ undefined
+ * (defaults to `direct`); a known value ⇒ that `Destination`; an unknown value ⇒
+ * a pushed problem and undefined. */
+function parseDestination(raw, problems) {
+    if (raw === undefined || raw === null)
+        return undefined;
+    if (typeof raw !== "string" || !DESTINATIONS.has(raw)) {
+        problems.push(`destination "${String(raw)}" is not one of [${[...DESTINATIONS].join(", ")}]`);
+        return undefined;
+    }
+    return raw;
+}

package/dist/routes/draft.d.ts

@@ -34,6 +34,13 @@ export interface RouteDraftSpec {
     readonly variables: readonly string[];
     /** JSON Schema for the output; null drafts an unstructured route. */
     readonly outputSchema: object | null;
+    /** `output: text` (issue #1) — a prose/markdown route, returned verbatim.
+     * Absent / `"json"` ⇒ nothing is emitted (the loader default). `text` with a
+     * non-null `outputSchema` is refused here, as the loader would refuse it. */
+    readonly outputKind?: "json" | "text";
+    /** `prompt.user` (issue #5, R1) — the per-call user-turn template, written
+     * to `prompts/<name>.user.md` and declared beside `prompt.system`. */
+    readonly userPrompt?: string;
     readonly require?: readonly string[];
     readonly prefer?: readonly string[];
     readonly policy?: {

package/dist/routes/draft.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"draft.d.ts","sourceRoot":"","sources":["../../src/routes/draft.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,MAAM,WAAW,cAAc;IAC7B,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;IAC1B,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,mEAAmE;IACnE,QAAQ,CAAC,IAAI,CAAC,EAAE,MAAM,CAAC;CACxB;AAED,MAAM,WAAW,cAAc;IAC7B,sDAAsD;IACtD,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,UAAU,EAAE,SAAS,cAAc,EAAE,CAAC;IAC/C;wCACoC;IACpC,QAAQ,CAAC,SAAS,CAAC,EAAE;QAAE,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;QAAC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAA;KAAE,GAAG,IAAI,CAAC;IAClF,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,SAAS,EAAE,SAAS,MAAM,EAAE,CAAC;IACtC,qEAAqE;IACrE,QAAQ,CAAC,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;IACrC,QAAQ,CAAC,OAAO,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;IACrC,QAAQ,CAAC,MAAM,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;IACpC,QAAQ,CAAC,MAAM,CAAC,EAAE;QAChB,QAAQ,CAAC,SAAS,CAAC,EAAE,MAAM,CAAC;QAC5B,QAAQ,CAAC,IAAI,CAAC,EAAE,MAAM,CAAC;QACvB,QAAQ,CAAC,IAAI,CAAC,EAAE,MAAM,CAAC;QACvB,QAAQ,CAAC,OAAO,CAAC,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,CAAC;KACrD,CAAC;IACF,QAAQ,CAAC,OAAO,CAAC,EAAE,OAAO,CAAC;IAC3B;6EACyE;IACzE,QAAQ,CAAC,UAAU,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;CACzC;AAED,MAAM,WAAW,SAAS;IACxB,6CAA6C;IAC7C,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;CAC3B;AAED,MAAM,WAAW,UAAU;IACzB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,KAAK,EAAE,SAAS,SAAS,EAAE,CAAC;CACtC;AAmBD,wBAAgB,eAAe,CAAC,IAAI,EAAE,cAAc,GAAG,UAAU,CAgGhE;AAMD,MAAM,WAAW,gBAAgB;IAC/B,QAAQ,CAAC,OAAO,EAAE,SAAS,MAAM,EAAE,CAAC;IACpC,sDAAsD;IACtD,QAAQ,CAAC,OAAO,EAAE,SAAS,MAAM,EAAE,CAAC;CACrC;AAED;;;;;;;GAOG;AACH,wBAAgB,eAAe,CAC7B,KAAK,EAAE,UAAU,EACjB,SAAS,EAAE,MAAM,EACjB,EAAE,EAAE;IACF,UAAU,EAAE,CAAC,CAAC,EAAE,MAAM,KAAK,OAAO,CAAC;IACnC,SAAS,EAAE,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE;QAAE,SAAS,EAAE,OAAO,CAAA;KAAE,KAAK,IAAI,CAAC;IAC1D,aAAa,EAAE,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,KAAK,IAAI,CAAC;CAC1D,EACD,IAAI,EAAE;IAAE,IAAI,EAAE,CAAC,GAAG,KAAK,EAAE,MAAM,EAAE,KAAK,MAAM,CAAC;IAAC,OAAO,EAAE,CAAC,CAAC,EAAE,MAAM,KAAK,MAAM,CAAA;CAAE,GAC7E,gBAAgB,CAclB"}
+{"version":3,"file":"draft.d.ts","sourceRoot":"","sources":["../../src/routes/draft.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,MAAM,WAAW,cAAc;IAC7B,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;IAC1B,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,mEAAmE;IACnE,QAAQ,CAAC,IAAI,CAAC,EAAE,MAAM,CAAC;CACxB;AAED,MAAM,WAAW,cAAc;IAC7B,sDAAsD;IACtD,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,UAAU,EAAE,SAAS,cAAc,EAAE,CAAC;IAC/C;wCACoC;IACpC,QAAQ,CAAC,SAAS,CAAC,EAAE;QAAE,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;QAAC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAA;KAAE,GAAG,IAAI,CAAC;IAClF,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,SAAS,EAAE,SAAS,MAAM,EAAE,CAAC;IACtC,qEAAqE;IACrE,QAAQ,CAAC,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;IACrC;;iFAE6E;IAC7E,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;IACtC;0EACsE;IACtE,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,CAAC;IAC7B,QAAQ,CAAC,OAAO,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;IACrC,QAAQ,CAAC,MAAM,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;IACpC,QAAQ,CAAC,MAAM,CAAC,EAAE;QAChB,QAAQ,CAAC,SAAS,CAAC,EAAE,MAAM,CAAC;QAC5B,QAAQ,CAAC,IAAI,CAAC,EAAE,MAAM,CAAC;QACvB,QAAQ,CAAC,IAAI,CAAC,EAAE,MAAM,CAAC;QACvB,QAAQ,CAAC,OAAO,CAAC,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,CAAC;KACrD,CAAC;IACF,QAAQ,CAAC,OAAO,CAAC,EAAE,OAAO,CAAC;IAC3B;6EACyE;IACzE,QAAQ,CAAC,UAAU,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;CACzC;AAED,MAAM,WAAW,SAAS;IACxB,6CAA6C;IAC7C,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;CAC3B;AAED,MAAM,WAAW,UAAU;IACzB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,KAAK,EAAE,SAAS,SAAS,EAAE,CAAC;CACtC;AAmBD,wBAAgB,eAAe,CAAC,IAAI,EAAE,cAAc,GAAG,UAAU,CA6GhE;AAMD,MAAM,WAAW,gBAAgB;IAC/B,QAAQ,CAAC,OAAO,EAAE,SAAS,MAAM,EAAE,CAAC;IACpC,sDAAsD;IACtD,QAAQ,CAAC,OAAO,EAAE,SAAS,MAAM,EAAE,CAAC;CACrC;AAED;;;;;;;GAOG;AACH,wBAAgB,eAAe,CAC7B,KAAK,EAAE,UAAU,EACjB,SAAS,EAAE,MAAM,EACjB,EAAE,EAAE;IACF,UAAU,EAAE,CAAC,CAAC,EAAE,MAAM,KAAK,OAAO,CAAC;IACnC,SAAS,EAAE,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE;QAAE,SAAS,EAAE,OAAO,CAAA;KAAE,KAAK,IAAI,CAAC;IAC1D,aAAa,EAAE,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,KAAK,IAAI,CAAC;CAC1D,EACD,IAAI,EAAE;IAAE,IAAI,EAAE,CAAC,GAAG,KAAK,EAAE,MAAM,EAAE,KAAK,MAAM,CAAC;IAAC,OAAO,EAAE,CAAC,CAAC,EAAE,MAAM,KAAK,MAAM,CAAA;CAAE,GAC7E,gBAAgB,CAclB"}

package/dist/routes/draft.js

@@ -40,10 +40,15 @@ function buildRouteDraft(spec) {
         throw new Error(`route draft "${spec.route}" has no candidates — a route with nothing allowed to serve it ` +
             "cannot run. Name at least one provider/model.");
     }
+    if (spec.outputKind === "text" && spec.outputSchema !== null) {
+        throw new Error(`route draft "${spec.route}" declares output: text AND a schema — a prose route has no schema. ` +
+            "Drop the schema, or draft a JSON route.");
+    }
     const name = slug(spec.route);
     const files = [];
     const schemaRef = spec.outputSchema === null ? null : `./schemas/${name}.schema.json`;
     const promptRef = `./prompts/${name}.system.md`;
+    const userPromptRef = spec.userPrompt === undefined ? null : `./prompts/${name}.user.md`;
     const lines = [
         "# ModelRig route bundle — DRAFT. Nothing in your app calls this yet.",
         "# Read it, edit it, move it up one directory into modelrig/routes/, then",
@@ -55,6 +60,8 @@ function buildRouteDraft(spec) {
             lines.push(`# ${line}`);
     }
     lines.push("#", "# Candidates are ordered incumbent-first: the model you already use leads,", "# so the first run reproduces today's behavior. The cheaper probed", "# alternatives below it are what a bake-off would put to the test.", "#", "# `prefer` is empty ON PURPOSE. Adding `cost` re-sorts these candidates by", "# price, which would hand the first run to the cheapest model rather than", "# to your incumbent — the opposite of a like-for-like migration. Prove a", "# cheaper model with a bake-off, then switch to it deliberately.", `route: ${yamlString(spec.route)}`, "version: 1");
+    if (spec.outputKind === "text")
+        lines.push("output: text   # prose/markdown, returned verbatim (no schema:)");
     if (schemaRef !== null)
         lines.push(`schema: ${schemaRef}`);
     lines.push("candidates:");
@@ -78,6 +85,8 @@ function buildRouteDraft(spec) {
     lines.push(`prefer: [${(spec.prefer ?? []).map(yamlString).join(", ")}]`);
     lines.push("prompt:");
     lines.push(`  system: ${promptRef}`);
+    if (userPromptRef !== null)
+        lines.push(`  user: ${userPromptRef}   # per-call user turn (the delta next to a cache handle)`);
     lines.push(`  variables: [${spec.variables.map(yamlString).join(", ")}]` +
         (spec.variables.length === 0 ? "   # no {{variables}} detected — the prompt is static" : ""));
     const policy = spec.policy ?? {};
@@ -105,6 +114,9 @@ function buildRouteDraft(spec) {
         });
     }
     files.push({ path: `prompts/${name}.system.md`, contents: ensureTrailingNewline(spec.systemPrompt) });
+    if (spec.userPrompt !== undefined) {
+        files.push({ path: `prompts/${name}.user.md`, contents: ensureTrailingNewline(spec.userPrompt) });
+    }
     return { route: spec.route, files };
 }
 function ensureTrailingNewline(text) {