rovecode 0.3.2 → 0.4.0-beta.2

package/src/providers/catalog-local.ts

@@ -1,160 +0,0 @@
-/** rovecode's own price table for models the models.dev snapshot does not carry — the API ALIASES a
- *  vendor sells under (DeepSeek's `deepseek-chat` / `deepseek-reasoner` are aliases of a release; the
- *  snapshot lists release ids), and xAI's retired grok-4 slugs, which still resolve but are served and
- *  billed as grok-4.3. Without these a real run on those ids is "unpriced" in /cost and `rovecode model show`.
- *
- *  Rules: an entry here applies ONLY when neither the live models.dev layer nor the snapshot has the model
- *  (catalog.ts findIn) — a later snapshot entry wins by construction. Every entry names the vendor page the
- *  numbers were read from and the day; "fetched" means the page was read on that day, "unverified" means the
- *  page no longer documents the id and the row is the last known value. `rovecode model show` prints
- *  "priced from rovecode's own table" so nobody mistakes them for models.dev's. Prices are USD per million
- *  tokens. Output limits are given only where the vendor publishes one (the runtime caps at MAX_OUTPUT_CAP). */
-
-export interface LocalModel {
-  context: number;
-  output?: number;
-  /** has a reasoning mode (thinking.ts sends the dial only when this is not false) */
-  reasoning: boolean;
-  toolCall: boolean;
-  image?: boolean;
-  cost: { input: number; output: number; cacheRead?: number; cacheWrite?: number };
-  /** where the numbers were read — the reader's warrant */
-  source: string;
-  /** when, and how: "YYYY-MM-DD (fetched)" or "YYYY-MM-DD (unverified: …)" */
-  checked: string;
-  /** Take precedence over a models.dev entry for this model. Off by default and deliberately rare: the
-   *  overlay exists to fill gaps, and a snapshot that gains a model should normally win. Set it ONLY when
-   *  the vendor's own page was read and disagrees with the snapshot, and say so in `source` — a reader has
-   *  to be able to see which number rovecode chose and why. */
-  override?: true;
-}
-
-/** A vendor whose price CHANGES WITH PROMPT SIZE. models.dev has no field for this, so it lives beside the
- *  overlay and augments a row whatever its source — the snapshot keeps owning the base prices.
- *
- *  `mode` is the part that is easy to get wrong and expensive to get wrong:
- *   - "per-request": crossing the threshold reprices THE WHOLE REQUEST, every token, at the upper rate
- *     (xAI). A 201k-token prompt costs 201k × the high rate, not 200k low + 1k high.
- *   - "marginal": only the tokens past the threshold take the upper rate. NO vendor row uses this today —
- *     it exists because the two shapes are not interchangeable and a cost function must branch on which one
- *     it has, rather than assume. Do not classify a vendor as marginal without wording that says so. */
-export interface PriceTier {
-  thresholdTokens: number;
-  mode: "per-request" | "marginal";
-  above: { input: number; output: number; cacheRead?: number; cacheWrite?: number };
-  source: string;
-  checked: string;
-}
-
-/** xAI, docs.x.ai/developers/migration/may-15-retirement (fetched 2026-09-04): grok-4-0709, grok-4-fast-reasoning,
- *  grok-4-fast-non-reasoning, grok-4-1-fast-* and grok-3 were retired on 2026-05-15; the slugs still resolve and
- *  are SERVED BY grok-4.3 — reasoning slugs at low effort, non-reasoning ones at none — and BILLED at grok-4.3
- *  rates ($1.25 in / $2.50 out / $0.20 cached per 1M under 200k context; docs.x.ai/docs/models/grok-4.3, fetched
- *  the same day: 1M context, configurable reasoning none|low|medium|high). So a request to `grok-4` costs this. */
-const GROK_43_REDIRECT = { context: 1_000_000, toolCall: true, image: true, cost: { input: 1.25, output: 2.5, cacheRead: 0.2 },
-  source: "docs.x.ai May-15-2026 retirement: served by grok-4.3, billed at grok-4.3 rates (under 200k context)", checked: "2026-09-04 (fetched)" } as const;
-
-/** Tiered pricing, by provider and model id. Verified 2026-09-05 against the vendors' own pricing pages.
- *  Nothing here changes a base price — `catalog.ts` attaches the tier to whatever row it resolved, so a
- *  models.dev entry keeps its numbers and gains the threshold it has no field for. */
-export const PRICE_TIERS: Readonly<Record<string, Readonly<Record<string, PriceTier>>>> = {
-  // xAI: "requests over 200,000 prompt tokens are billed at the higher rate FOR THE ENTIRE REQUEST"
-  xai: {
-    "grok-4.6": { thresholdTokens: 200_000, mode: "per-request", above: { input: 4, output: 12, cacheRead: 1 },
-      source: "docs.x.ai/developers/pricing — over 200k prompt tokens the whole request bills at the high rate", checked: "2026-09-05 (fetched)" },
-    "grok-4.3": { thresholdTokens: 200_000, mode: "per-request", above: { input: 2.5, output: 5, cacheRead: 0.4 },
-      source: "docs.x.ai/developers/pricing — over 200k prompt tokens the whole request bills at the high rate", checked: "2026-09-05 (fetched)" },
-  },
-  // OpenAI, developers.openai.com/api/docs/models/gpt-5.4 and /gpt-5.5: "Prompts with >272K input tokens
-  // are priced at 2x input and 1.5x output for the full session." The multipliers are ASYMMETRIC — input
-  // doubles, output goes up by half — so these are written out rather than derived from one rate.
-  // The -pro tiers are NOT given this: their own pages do not carry the sentence, and inventing a
-  // threshold for a $30/1M model is the expensive kind of guess.
-  openai: {
-    "gpt-5.4": { thresholdTokens: 272_000, mode: "per-request", above: { input: 5, output: 22.5, cacheRead: 0.5 },
-      source: "developers.openai.com/api/docs/models/gpt-5.4 — >272K input tokens: 2x input, 1.5x output for the full session", checked: "2026-09-05 (fetched)" },
-    "gpt-5.5": { thresholdTokens: 272_000, mode: "per-request", above: { input: 10, output: 45, cacheRead: 1 },
-      source: "developers.openai.com/api/docs/models/gpt-5.5 — >272K input tokens: 2x input, 1.5x output for the full session", checked: "2026-09-05 (fetched)" },
-  },
-  // Google prices by PROMPT SIZE, not by overflow: the page's own columns read "prompts <= 200k tokens"
-  // and "prompts > 200k tokens" for BOTH input and output, so a 201k prompt bills every token — and that
-  // request's output — at the upper rate. Same shape as xAI, which is why neither is "marginal".
-  google: {
-    "gemini-3.1-pro-preview": { thresholdTokens: 200_000, mode: "per-request", above: { input: 4, output: 18 },
-      source: "ai.google.dev/gemini-api/docs/pricing — \"$4.00, prompts > 200k tokens\" / \"$18.00, prompts > 200k\"", checked: "2026-09-05 (fetched)" },
-    "gemini-2.5-pro": { thresholdTokens: 200_000, mode: "per-request", above: { input: 2.5, output: 15 },
-      source: "ai.google.dev/gemini-api/docs/pricing — \"$2.50, prompts > 200k tokens\" / \"$15.00, prompts > 200k\"", checked: "2026-09-05 (fetched)" },
-  },
-} as const;
-
-/** A price that is scheduled to change, or a condition the number depends on. Not a tier and not an
- *  override — just the sentence a reader needs so today's number is not mistaken for a permanent one.
- *  Attached to a row whatever its source. */
-export const PRICE_NOTES: Readonly<Record<string, Readonly<Record<string, string>>>> = {
-  google: {
-    "gemini-3.8-flash": "promotional pricing through 2026-12-31: ai.google.dev/gemini-api/docs/pricing says it doubles to $1.50 in / $7.50 out starting 2027-01-01 (checked 2026-09-05)",
-  },
-  mistral: {
-    // Checked against mistral.ai/pricing/api on 2026-09-05: the table publishes ONLY `-latest` aliases
-    // (mistral-medium-latest $1.5/$7.5, mistral-small-latest $0.15/$0.6, mistral-large-latest $0.5/$1.5).
-    // No dated id appears — not the snapshot's `mistral-medium-2604`, and not `mistral-medium-3-5` or
-    // `ministral-8b-2512` either. The prices match, the ids are unverifiable, so nothing is renamed here.
-    "mistral-medium-latest": "mistral.ai/pricing/api publishes only `-latest` aliases and NO maximum output for any model — the catalog's maxOutput equalling the context window is filler, not a vendor figure (checked 2026-09-05)",
-    "mistral-medium-2604": "this dated id is not on mistral.ai/pricing/api, which lists only `-latest` aliases; the price matches mistral-medium-latest but the id itself is unverified (checked 2026-09-05)",
-  },
-  groq: {
-    // Not "unverified" — UNPUBLISHED. Groq's docs table says "Contact Sales" for the Llama models and
-    // console.groq.com/pricing is a 404, so any number here came from a third party, not from Groq.
-    "llama-3.3-70b-versatile": "Groq does not publish a price for this model — its docs table says \"Contact Sales\" and there is no public pricing page; the catalog's figure is models.dev's, not the vendor's (checked 2026-09-05)",
-  },
-  anthropic: {
-    // true of every Anthropic row; recorded on the one most runs use
-    "claude-mythos-5-1": "invite-only access (platform.claude.com/docs/en/models/mythos-5-1/overview); outside the Batches 300k beta, which covers Opus 5/4.8/4.7/4.6 and Sonnet 5/4.6 only (checked 2026-09-05)",
-    "claude-fable-5-1": "outside the Batches 300k beta, which covers Opus 5/4.8/4.7/4.6 and Sonnet 5/4.6 only (checked 2026-09-05)",
-    "claude-opus-5": "thinking tokens bill as OUTPUT and count inside max_tokens; from Opus 4.5 / Sonnet 4.6 / Fable 5.x earlier thinking blocks are preserved and bill as INPUT on the next turn (platform.claude.com/docs, checked 2026-09-05)",
-  },
-} as const;
-
-export const LOCAL_MODELS: Readonly<Record<string, Readonly<Record<string, LocalModel>>>> = {
-  deepseek: {
-    // api-docs.deepseek.com/quick_start/pricing, fetched 2026-09-05. These three rows OVERRIDE models.dev,
-    // which carries roughly a third of the vendor's own peak numbers for the same ids (flash 0.14/0.28 vs
-    // 0.44/1.32). The page is the seller; the snapshot is a mirror that has drifted, so the page wins here
-    // and says so. Peak rates are used: the page's own off-peak wording is self-contradictory (it labels
-    // "peak" 01:00-04:00 and 06:00-10:00 UTC, which are its cheap hours elsewhere), so rovecode quotes the
-    // HIGHER of the two — an estimate that surprises nobody with a bigger bill. Do not build discount logic
-    // on that wording until the page is unambiguous.
-    "deepseek-v4-flash": { context: 1_000_000, output: 384_000, reasoning: true, toolCall: true, override: true,
-      cost: { input: 0.44, output: 1.32, cacheRead: 0.014 },
-      source: "api-docs.deepseek.com/quick_start/pricing (peak rate — off-peak is half, we do not model time-of-day; models.dev carries ~1/3 of these)", checked: "2026-09-05 (fetched)" },
-    "deepseek-v4-pro": { context: 1_000_000, output: 384_000, reasoning: true, toolCall: true, override: true,
-      cost: { input: 1.32, output: 3.96, cacheRead: 0.044 },
-      source: "api-docs.deepseek.com/quick_start/pricing (peak rate — off-peak is half, we do not model time-of-day; models.dev carries ~1/3 of these)", checked: "2026-09-05 (fetched)" },
-    "deepseek-v4-flash-vision-exp": { context: 1_000_000, output: 384_000, reasoning: true, toolCall: true, image: true, override: true,
-      cost: { input: 0.44, output: 1.32, cacheRead: 0.014 },
-      source: "api-docs.deepseek.com/quick_start/pricing (peak rate — off-peak is half, we do not model time-of-day; models.dev carries ~1/3 of these)", checked: "2026-09-05 (fetched)" },
-    // The two API aliases. The pricing page lists only the v4-* ids, so nothing here is verified: the
-    // numbers are the last ones DeepSeek published for them and are marked as such rather than being
-    // quietly refreshed to a v4 price they may not be billed at.
-    "deepseek-chat": { context: 128_000, output: 8_000, reasoning: true, toolCall: true, cost: { input: 0.28, output: 0.42, cacheRead: 0.028 }, source: "api-docs.deepseek.com pricing (V3.2 era)", checked: "2026-09-05 (unverified: the alias is not on the pricing page, which lists deepseek-v4-* only)" },
-    "deepseek-reasoner": { context: 128_000, output: 64_000, reasoning: true, toolCall: true, cost: { input: 0.28, output: 0.42, cacheRead: 0.028 }, source: "api-docs.deepseek.com pricing (V3.2 era)", checked: "2026-09-05 (unverified: not on the pricing page; DeepSeek now exposes thinking as a MODE of the v4 models, not a separate id)" },
-  },
-  openai: {
-    // developers.openai.com/api/docs/pricing, fetched 2026-09-05. Not in the snapshot at all, so this is a
-    // gap fill, not an override.
-    "gpt-6-astra": { context: 1_050_000, output: 128_000, reasoning: true, toolCall: true, image: true,
-      cost: { input: 10, output: 50, cacheRead: 1 },
-      source: "developers.openai.com/api/docs/pricing", checked: "2026-09-05 (fetched)" },
-  },
-  xai: {
-    "grok-4": { ...GROK_43_REDIRECT, reasoning: true },
-    "grok-4-0709": { ...GROK_43_REDIRECT, reasoning: true },
-    "grok-4-fast": { ...GROK_43_REDIRECT, reasoning: true },
-    "grok-4-fast-reasoning": { ...GROK_43_REDIRECT, reasoning: true },
-    "grok-4-fast-non-reasoning": { ...GROK_43_REDIRECT, reasoning: false },
-    "grok-4-1-fast-reasoning": { ...GROK_43_REDIRECT, reasoning: true },
-    "grok-4-1-fast-non-reasoning": { ...GROK_43_REDIRECT, reasoning: false },
-    "grok-3": { ...GROK_43_REDIRECT, reasoning: false },
-    // grok-3-mini: neither on docs.x.ai/docs/models nor in the retirement list on 2026-09-04 — no row, stays unpriced
-  },
-};

package/src/providers/catalog.ts

@@ -1,421 +0,0 @@
-/** Model catalog: capability/pricing lookups backed by the models.dev database (ADR: port #6).
- *
- *  Source of truth is the offline snapshot bundled in @opencode-ai/models (generated from
- *  https://models.dev/api.json). Live refresh against models.dev itself is entirely OPT-IN
- *  (only when a `fetchFn` is supplied) and layers on top of — never replaces — the offline
- *  data, so a fresh `new ModelCatalog()` works fully offline with zero network access. */
-
-import { readFileSync, writeFileSync, mkdirSync, statSync } from "node:fs";
-import { join } from "node:path";
-import type { ProviderMap, Model } from "@opencode-ai/models";
-import { LOCAL_MODELS, PRICE_NOTES, PRICE_TIERS, type LocalModel, type PriceTier } from "./catalog-local.ts";
-
-// Loaded lazily on the first lookup() so that importing catalog.ts costs nothing — the TUI paints its
-// first frame before buildDef() is called.
-//
-// It reads src/providers/models-index.json, not @opencode-ai/models/snapshot. Same shape, same source,
-// five fields per model instead of the whole models.dev record: the upstream snapshot is 4.26 MB and
-// parsing it costs 55 MB resident, which every session paid on its first status update to answer "what
-// is this model's window and price". The trimmed index is 1.1 MB. scripts/build-model-index.mjs
-// generates it and `--check` fails when it and the installed package disagree, so an upgrade cannot
-// leave the catalog quietly describing the previous release.
-let _snapshot: ProviderMap | null = null;
-function snapshotProviders(): ProviderMap {
-  if (_snapshot === null) {
-    // eslint-disable-next-line @typescript-eslint/no-require-imports
-    _snapshot = (require("./models-index.json") as { providers: ProviderMap }).providers;
-  }
-  return _snapshot;
-}
-
-export interface ModelInfo {
-  provider: string;
-  model: string;
-  contextWindow?: number;
-  maxOutput?: number;
-  pricing?: { inputPerMTok?: number; outputPerMTok?: number; cacheReadPerMTok?: number; cacheWritePerMTok?: number };
-  supportsTools?: boolean;
-  supportsReasoning?: boolean;
-  /** where the numbers come from: models.dev (snapshot or live) or rovecode's own table (catalog-local.ts) for a
-   *  model the snapshot lacks — `rovecode model show` says which, so a local price is never mistaken for models.dev's */
-  source?: "models.dev" | "local";
-  /** local entries only: the vendor page the numbers were read from, and when */
-  sourceNote?: string;
-  /** this model's price changes past a prompt size (xAI, Gemini). Attached whatever the row's source —
-   *  models.dev has no field for it. A cost estimator that ignores it under-quotes a long prompt, and for
-   *  `mode: "per-request"` it under-quotes the WHOLE request, not just the overflow. */
-  tier?: PriceTier;
-  /** a scheduled change or a condition attached to the price ("doubles on 2027-01-01") */
-  priceNote?: string;
-}
-
-export interface CatalogOptions {
-  fetchFn?: typeof fetch | null; // null = offline only; default null (live fetch is OPT-IN)
-  cacheDir?: string;             // when live: cache api.json to <cacheDir>/models.json, ttl 24h
-  /** the local overlay (default catalog-local.ts LOCAL_MODELS); tests inject their own */
-  local?: Readonly<Record<string, Readonly<Record<string, LocalModel>>>>;
-}
-
-const CACHE_TTL_MS = 24 * 60 * 60_000;
-const CACHE_FILE = "models.json";
-const LIVE_URL = "https://models.dev/api.json";
-
-/**
- * rovecode provider id -> models.dev provider key. Derived by loading the offline snapshot
- * (node_modules/@opencode-ai/models/dist/snapshot.js) and inspecting Object.keys(providers)
- * directly rather than guessing:
- *   - together -> "togetherai" and fireworks -> "fireworks-ai" (not the bare names)
- *   - lmstudio -> "lmstudio" DOES exist in the snapshot (3 curated local models), so it maps
- *   - everything else here is an exact 1:1 id match confirmed present in the snapshot
- *
- * Deliberately absent (verified NOT a key in the snapshot -> lookup() returns undefined):
- * kaesra (rovecode's own proxy brand, not a models.dev provider — its vendor-prefixed model ids
- * resolve through VENDOR_PREFIX_MAP below instead), ollama (only "ollama-cloud" exists, not
- * bare "ollama"), moondream, vllm.
- */
-const PROVIDER_MAP: Record<string, string> = {
-  openai: "openai",
-  anthropic: "anthropic",
-  deepseek: "deepseek",
-  groq: "groq",
-  openrouter: "openrouter",
-  lmstudio: "lmstudio",
-  together: "togetherai",
-  mistral: "mistral",
-  cerebras: "cerebras",
-  fireworks: "fireworks-ai",
-  perplexity: "perplexity",
-  xai: "xai",
-  // not built-in providers, but the ids people give `rovecode provider add` for these vendors' own endpoints
-  // (Google's OpenAI layer, Z.ai, Moonshot, Alibaba DashScope, MiniMax) — mapped so their models price and
-  // carry the reasoning flag like the built-ins (2026-09-04: models.dev snapshot 0.0.64 has all five)
-  google: "google",
-  gemini: "google",
-  zai: "zai",
-  moonshot: "moonshotai",
-  moonshotai: "moonshotai",
-  alibaba: "alibaba",
-  dashscope: "alibaba",
-  minimax: "minimax",
-};
-
-/**
- * HuggingFace-style vendor prefix -> models.dev provider key, for aggregator providers
- * (kaesra, or any custom base URL) that serve models under "vendor/model" ids. Lets
- * lookup("kaesra", "zai-org/glm-5.3-flash") price against the zai snapshot entry.
- * Verified against Object.keys(snapshotProviders): "zai", "deepseek", "moonshotai" all
- * exist; there is NO bare "moonshot" key (only moonshotai/moonshotai-cn), hence the
- * identity mapping for moonshotai.
- */
-const VENDOR_PREFIX_MAP: Record<string, string> = {
-  "zai-org": "zai",
-  "deepseek-ai": "deepseek",
-  "moonshotai": "moonshotai",
-};
-
-/** the tiered-price row for a model, if the vendor has one (case-insensitive, like every other lookup) */
-function tierFor(providerId: string, modelId: string): PriceTier | undefined {
-  const table = PRICE_TIERS[providerId.toLowerCase()];
-  if (!table) return undefined;
-  const key = Object.keys(table).find((k) => k.toLowerCase() === modelId.toLowerCase());
-  return key === undefined ? undefined : table[key];
-}
-
-function noteFor(providerId: string, modelId: string): string | undefined {
-  const table = PRICE_NOTES[providerId.toLowerCase()];
-  if (!table) return undefined;
-  const key = Object.keys(table).find((k) => k.toLowerCase() === modelId.toLowerCase());
-  return key === undefined ? undefined : table[key];
-}
-
-/** The four rates one request is billed at. Same field names as `ModelInfo.pricing`, so a caller can pass
- *  either where a rate table is wanted. */
-export interface Rates { inputPerMTok?: number; outputPerMTok?: number; cacheReadPerMTok?: number; cacheWritePerMTok?: number }
-
-/** What `promptTokens` actually costs on this model, split the way the vendor bills it. PURE — no I/O, no
- *  clock, no rounding: a cost function multiplies these and decides its own precision.
- *
- *  Read `promptBase`/`promptAbove` as token counts to multiply by `base`/`above`; `request` is the rate
- *  table for everything that is not the prompt (output, cache) on this request, because both vendors we
- *  have select the OUTPUT price by prompt size too. For a flat model the split is trivial (everything in
- *  `promptBase`) so a caller never needs to branch on whether a tier exists — only on whether it wants the
- *  detail. */
-export interface RateBreakdown {
-  mode: "flat" | "per-request" | "marginal";
-  /** true when the prompt crossed a threshold and the upper rates are in play */
-  tierApplied: boolean;
-  thresholdTokens?: number;
-  promptBase: number;
-  promptAbove: number;
-  base: Rates;
-  above?: Rates;
-  request: Rates;
-}
-
-/** The rate breakdown for a prompt of `promptTokens` on this model. A negative or non-finite count is
- *  treated as 0 — an estimator must not be able to produce a negative bill. */
-export function ratesFor(info: ModelInfo, promptTokens: number): RateBreakdown {
-  const base: Rates = { ...(info.pricing ?? {}) };
-  const n = Number.isFinite(promptTokens) && promptTokens > 0 ? Math.floor(promptTokens) : 0;
-  const t = info.tier;
-  if (!t || n <= t.thresholdTokens) {
-    return { mode: t ? t.mode : "flat", tierApplied: false, ...(t ? { thresholdTokens: t.thresholdTokens } : {}),
-      promptBase: n, promptAbove: 0, base, ...(t ? { above: ratesOf(t) } : {}), request: base };
-  }
-  const above = ratesOf(t);
-  return t.mode === "per-request"
-    // every token of the prompt, and this request's output, at the upper rate
-    ? { mode: "per-request", tierApplied: true, thresholdTokens: t.thresholdTokens, promptBase: 0, promptAbove: n, base, above, request: above }
-    // only the overflow at the upper rate; output follows the upper table, which is the conservative read
-    : { mode: "marginal", tierApplied: true, thresholdTokens: t.thresholdTokens, promptBase: t.thresholdTokens, promptAbove: n - t.thresholdTokens, base, above, request: above };
-}
-
-function ratesOf(t: PriceTier): Rates {
-  return { inputPerMTok: t.above.input, outputPerMTok: t.above.output,
-    ...(t.above.cacheRead !== undefined ? { cacheReadPerMTok: t.above.cacheRead } : {}),
-    ...(t.above.cacheWrite !== undefined ? { cacheWritePerMTok: t.above.cacheWrite } : {}) };
-}
-
-/** The price lines a surface shows — one place, so `model show`, /cost and the market cannot drift apart.
- *  Empty when the model is unpriced. */
-export function describePricing(info: ModelInfo): string[] {
-  const p = info.pricing;
-  if (!p || (p.inputPerMTok === undefined && p.outputPerMTok === undefined)) return [];
-  const money = (n?: number): string => (n === undefined ? "—" : `$${n}`);
-  const lines = [`${money(p.inputPerMTok)} in / ${money(p.outputPerMTok)} out per 1M${p.cacheReadPerMTok !== undefined ? ` · cache read ${money(p.cacheReadPerMTok)}` : ""}${p.cacheWritePerMTok !== undefined ? ` · cache write ${money(p.cacheWritePerMTok)}` : ""}`];
-  if (info.tier) {
-    const t = info.tier;
-    lines.push(t.mode === "per-request"
-      ? `over ${t.thresholdTokens.toLocaleString("en-US")} prompt tokens the WHOLE request bills at ${money(t.above.input)} in / ${money(t.above.output)} out`
-      : `tokens past ${t.thresholdTokens.toLocaleString("en-US")} bill at ${money(t.above.input)} in / ${money(t.above.output)} out`);
-  }
-  if (info.priceNote) lines.push(info.priceNote);
-  return lines;
-}
-
-/** a local overlay row in the snapshot's shape, so one toModelInfo serves both */
-function localAsModel(id: string, m: LocalModel): Model {
-  return {
-    id, name: id, description: `${m.source} — checked ${m.checked}`, attachment: false, reasoning: m.reasoning, tool_call: m.toolCall,
-    release_date: m.checked, modalities: { input: m.image ? ["text", "image"] : ["text"], output: ["text"] },
-    limit: { context: m.context, ...(m.output !== undefined ? { output: m.output } : {}) },
-    cost: { input: m.cost.input, output: m.cost.output, ...(m.cost.cacheRead !== undefined ? { cache_read: m.cost.cacheRead } : {}), ...(m.cost.cacheWrite !== undefined ? { cache_write: m.cost.cacheWrite } : {}) },
-  } as unknown as Model;
-}
-
-function isProviderMap(value: unknown): value is ProviderMap {
-  if (typeof value !== "object" || value === null || Array.isArray(value)) return false;
-  for (const entry of Object.values(value as Record<string, unknown>)) {
-    if (typeof entry !== "object" || entry === null || Array.isArray(entry)) return false;
-    const models = (entry as { models?: unknown }).models;
-    if (typeof models !== "object" || models === null || Array.isArray(models)) return false;
-  }
-  return true;
-}
-
-/** exact match, then case-insensitive, then match after stripping a single leading "vendor/" prefix. */
-function findModelKey(models: Record<string, Model>, modelId: string): string | undefined {
-  if (modelId in models) return modelId;
-  const lower = modelId.toLowerCase();
-  for (const key of Object.keys(models)) {
-    if (key.toLowerCase() === lower) return key;
-  }
-  const slashIndex = modelId.indexOf("/");
-  if (slashIndex === -1) return undefined;
-  const stripped = modelId.slice(slashIndex + 1);
-  if (stripped in models) return stripped;
-  const strippedLower = stripped.toLowerCase();
-  for (const key of Object.keys(models)) {
-    if (key.toLowerCase() === strippedLower) return key;
-  }
-  return undefined;
-}
-
-function toModelInfo(providerId: string, modelId: string, model: Model): ModelInfo {
-  const info: ModelInfo = { provider: providerId, model: modelId };
-  if (model.limit?.context !== undefined) info.contextWindow = model.limit.context;
-  if (model.limit?.output !== undefined) info.maxOutput = model.limit.output;
-  const cost = model.cost;
-  if (cost) {
-    const pricing: NonNullable<ModelInfo["pricing"]> = {};
-    if (cost.input !== undefined) pricing.inputPerMTok = cost.input;
-    if (cost.output !== undefined) pricing.outputPerMTok = cost.output;
-    if (cost.cache_read !== undefined) pricing.cacheReadPerMTok = cost.cache_read;
-    if (cost.cache_write !== undefined) pricing.cacheWritePerMTok = cost.cache_write;
-    if (Object.keys(pricing).length > 0) info.pricing = pricing;
-  }
-  if (model.tool_call !== undefined) info.supportsTools = model.tool_call;
-  if (model.reasoning !== undefined) info.supportsReasoning = model.reasoning;
-  return info;
-}
-
-export class ModelCatalog {
-  private readonly fetchFn: typeof fetch | null;
-  private readonly cacheDir: string | undefined;
-  private readonly local: Readonly<Record<string, Readonly<Record<string, LocalModel>>>>;
-  private liveProviders: ProviderMap | null = null;
-  private triedDiskCache = false;
-
-  constructor(opts: CatalogOptions = {}) {
-    this.fetchFn = opts.fetchFn ?? null;
-    this.cacheDir = opts.cacheDir;
-    this.local = opts.local ?? LOCAL_MODELS;
-  }
-
-  /** offline snapshot first; live cache layered on top when enabled.
-   *  Resolution order: (1) the rovecode provider's own models.dev entry (PROVIDER_MAP), then
-   *  (2) the model id's vendor prefix (VENDOR_PREFIX_MAP) with the prefix stripped — the
-   *  path that makes aggregator providers like kaesra (default model zai-org/glm-5.3-flash)
-   *  priceable. The vendor hit reports the vendor as `provider`, naming the pricing source. */
-  lookup(providerId: string, modelId: string): ModelInfo | undefined {
-    const hit = this.resolve(providerId, modelId);
-    if (!hit) return undefined;
-    const info = toModelInfo(hit.as, hit.key, hit.model);
-    info.source = hit.local ? "local" : "models.dev";
-    if (hit.local) info.sourceNote = `${hit.local.source}, checked ${hit.local.checked}`;
-    const tier = tierFor(providerId, modelId) ?? tierFor(providerId, hit.key);
-    if (tier) info.tier = tier;
-    const note = noteFor(providerId, modelId) ?? noteFor(providerId, hit.key);
-    if (note) info.priceNote = note;
-    return info;
-  }
-
-  /** port #34: does the model accept image input? models.dev `modalities.input` (every entry of
-   *  the bundled snapshot carries it — 7479/7479 at @opencode-ai/models 0.0.61). undefined = the
-   *  model is not in the catalog (custom base URLs, local servers, aggregator ids without a
-   *  vendor prefix): callers treat unknown as "send the image" — a wrong guess then fails loudly
-   *  at the provider (HTTP 400 → error turn) instead of silently turning the user's image into
-   *  text. Same resolution order as lookup(). */
-  supportsImages(providerId: string, modelId: string): boolean | undefined {
-    const hit = this.resolve(providerId, modelId);
-    const input: unknown = hit?.model.modalities?.input;
-    return Array.isArray(input) ? input.includes("image") : undefined;
-  }
-
-  /** models.dev `reasoning` for a model: true/false, undefined for a model the catalog does not know — the same
-   *  resolution as supportsImages(). Callers treat unknown as "may reason" (wire-select.ts: an unlisted
-   *  Responses-only id must not be sent to a wire that 404s it). */
-  supportsReasoning(providerId: string, modelId: string): boolean | undefined {
-    const hit = this.resolve(providerId, modelId);
-    return typeof hit?.model.reasoning === "boolean" ? hit.model.reasoning : undefined;
-  }
-
-  /** The catalog entry behind lookup()/supportsImages(): provider key candidates in order, live
-   *  layer over snapshot. `as` = the provider name reported (the vendor for a prefix hit). */
-  private resolve(providerId: string, modelId: string): { as: string; key: string; model: Model; local?: LocalModel } | undefined {
-    this.loadDiskCacheOnce();
-
-    const candidates: { key: string; as: string; model: string }[] = [];
-    const direct = PROVIDER_MAP[providerId];
-    if (direct) candidates.push({ key: direct, as: providerId, model: modelId });
-    const slash = modelId.indexOf("/");
-    if (slash > 0) {
-      const vendorKey = VENDOR_PREFIX_MAP[modelId.slice(0, slash).toLowerCase()];
-      if (vendorKey && vendorKey !== direct) candidates.push({ key: vendorKey, as: vendorKey, model: modelId.slice(slash + 1) });
-    }
-
-    for (const c of candidates) {
-      const hit = this.findIn(c.key, c.model);
-      if (hit) return { as: c.as, ...hit };
-    }
-    return undefined;
-  }
-
-  /** one provider key: live layer, then the snapshot, then rovecode's own table (catalog-local.ts). The
-   *  overlay does not shadow a models.dev entry, so a later snapshot that adds a model wins by construction
-   *  — EXCEPT for a row marked `override: true`, which is how a vendor page that disagrees with the
-   *  snapshot wins (today: DeepSeek's v4 family, where models.dev carries about a third of the published
-   *  price). An override says which page and which day in its `source`, and `model show` prints it. */
-  private findIn(key: string, modelId: string): { key: string; model: Model; local?: LocalModel } | undefined {
-    const overriding = this.local[key];
-    if (overriding) {
-      const found = findModelKey(overriding as unknown as Record<string, Model>, modelId);
-      const row = found === undefined ? undefined : overriding[found];
-      if (found !== undefined && row?.override === true) return { key: found, model: localAsModel(found, row), local: row };
-    }
-
-    const live = this.liveProviders?.[key];
-    if (live) {
-      const found = findModelKey(live.models, modelId);
-      if (found !== undefined) {
-        const model = live.models[found];
-        if (model) return { key: found, model };
-      }
-    }
-
-    const snap = snapshotProviders()[key];
-    if (snap) {
-      const found = findModelKey(snap.models, modelId);
-      if (found !== undefined) {
-        const model = snap.models[found];
-        if (model) return { key: found, model };
-      }
-    }
-
-    const table = this.local[key];
-    if (!table) return undefined;
-    const found = findModelKey(table as unknown as Record<string, Model>, modelId);
-    if (found === undefined) return undefined;
-    const local = table[found]!;
-    return { key: found, model: localAsModel(found, local), local };
-  }
-
-  /** live fetch https://models.dev/api.json when fetchFn set; false on failure, never throws. */
-  async refresh(): Promise<boolean> {
-    if (!this.fetchFn) return false;
-    try {
-      const res = await this.fetchFn(LIVE_URL);
-      if (!res.ok) return false;
-      const json: unknown = await res.json();
-      if (!isProviderMap(json)) return false;
-      this.liveProviders = json;
-      this.triedDiskCache = true;
-      this.writeDiskCache(json);
-      return true;
-    } catch {
-      return false;
-    }
-  }
-
-  /** Load a still-fresh (<24h) on-disk cache at most once per instance. Corrupted/stale/missing
-   *  cache is silently ignored — offline snapshot remains the fallback. Only consulted when a
-   *  fetchFn is configured, matching "live fetch is OPT-IN" (a purely-offline catalog never
-   *  touches disk). */
-  private loadDiskCacheOnce(): void {
-    if (this.triedDiskCache) return;
-    this.triedDiskCache = true;
-    if (!this.fetchFn || !this.cacheDir) return;
-    try {
-      const path = join(this.cacheDir, CACHE_FILE);
-      const age = Date.now() - statSync(path).mtimeMs;
-      if (age > CACHE_TTL_MS) return;
-      const parsed: unknown = JSON.parse(readFileSync(path, "utf8"));
-      if (isProviderMap(parsed)) this.liveProviders = parsed;
-    } catch {
-      // missing file, unreadable, or malformed JSON/shape -> treated as "no cache"
-    }
-  }
-
-  private writeDiskCache(data: ProviderMap): void {
-    if (!this.cacheDir) return;
-    try {
-      mkdirSync(this.cacheDir, { recursive: true });
-      writeFileSync(join(this.cacheDir, CACHE_FILE), JSON.stringify(data));
-    } catch {
-      // best-effort cache write; a failed write must not fail refresh()
-    }
-  }
-}
-
-let offlineCatalog: ModelCatalog | null = null;
-
-/** Port #34 adapter hook: image capability of a ModelRef from the offline snapshot (no fetchFn →
- *  never touches network or disk). undefined = unknown model (see ModelCatalog.supportsImages). */
-export function supportsImages(ref: { provider: string; model: string }, catalog?: ModelCatalog): boolean | undefined {
-  return (catalog ?? (offlineCatalog ??= new ModelCatalog())).supportsImages(ref.provider, ref.model);
-}
-
-/** the same hook for the reasoning flag (wire-select.ts): true/false from the offline snapshot, undefined = unknown model */
-export function supportsReasoning(ref: { provider: string; model: string }, catalog?: ModelCatalog): boolean | undefined {
-  return (catalog ?? (offlineCatalog ??= new ModelCatalog())).supportsReasoning(ref.provider, ref.model);
-}