@ultimat3/ai 23.0.0 → 25.0.0

package/src/fix-line.evals.ts

@@ -1,35 +0,0 @@
-/**
- * The eval attached to `fixLinePrompt`. Cases live here rather than in the test file, matching
- * every app's own prompts — a fixture-driven, deterministic `exact` scorer, no judge, because
- * this is a proof of the eval mechanism itself and a proof that can drift is not one.
- */
-
-import { defineEval } from './evals';
-import { fixLinePrompt } from './fix-line';
-import { exact } from './scorers';
-
-export const fixLineCases = [
-  {
-    name: 'runnable/x-command',
-    vars: { fixLine: 'bun run scripts/verify.ts --only lint,boundaries' },
-    expected: 'runnable',
-  },
-  { name: 'runnable/cli-flag', vars: { fixLine: 'x db migrate' }, expected: 'runnable' },
-  {
-    name: 'vague/check',
-    vars: { fixLine: 'check the configuration and try again' },
-    expected: 'vague',
-  },
-  { name: 'vague/see-docs', vars: { fixLine: 'see the docs for details' }, expected: 'vague' },
-];
-
-export const fixLineEval = defineEval({
-  name: 'ai.fix-line-runnable',
-  prompt: fixLinePrompt,
-  // The gate is the drop from this recorded score, never an absolute number: models drift,
-  // prompts should not. Accepting a new number is a diff in the committed baseline file.
-  baseline: import.meta.resolve('./fix-line.v1.baseline.json'),
-  tolerance: 0.05,
-  scorers: [exact],
-  cases: fixLineCases,
-});

package/src/fix-line.ts

@@ -1,27 +0,0 @@
-/**
- * The prompt behind `fix-line` — `@ultimat3/ai`'s own dogfood eval, and the package's first
- * framework-level `*.eval.test.ts`. Not an app feature: every app that declares a prompt is
- * required to pair it with an eval and a committed baseline (`defineEval`, `X_EVAL_MISSING`,
- * `X_EVAL_BASELINE_MISSING`), and this proves that whole convention actually catches a
- * regression from inside the package that owns it — through the real opt-in suite `x verify`'s
- * `eval` step runs, not only through `evals.test.ts`'s unit fixtures against a temp-dir baseline.
- *
- * The task is small on purpose: classify whether an error's `fix:` line is a runnable command —
- * axiom 4, "errors are instructions" — or vague guidance ("check the config", "see the docs").
- */
-
-import { definePrompt } from './prompt';
-
-export const fixLinePrompt = definePrompt<{ fixLine: string }>({
-  id: 'ai.fix-line-runnable',
-  version: '1',
-  template:
-    'Reply with exactly one word: "runnable" if the fix line below names a command to run, or ' +
-    '"vague" if it only describes what to do without naming one.\nFix line: {{fixLine}}',
-  input: {
-    type: 'object',
-    properties: { fixLine: { type: 'string' } },
-    required: ['fixLine'],
-  },
-  output: { type: 'string', enum: ['runnable', 'vague'] },
-});

package/src/openai-models.ts

@@ -1,84 +0,0 @@
-// The OpenAI-format built-in catalogue: the vendor's own list prices, registered through the same
-// public `registerModel` the Anthropic built-ins use. Here rather than in models.ts because these
-// rows belong to a PROVIDER — models.ts owns the registry mechanism, never one vendor's price list.
-
-import type { Money } from '@ultimat3/money';
-import type { ModelId } from './models';
-import { registerModel } from './models';
-
-/** A price per million tokens, in INTEGER MINOR UNITS. Same rule as models.ts: never a float. */
-const usd = (minor: number): Money => ({ minor, currency: 'USD' });
-
-/**
- * The ids this package prices, in ladder order (most capable first). A provider's `models` list is
- * still its own — an endpoint speaking this format serves whatever ids it was deployed with, and
- * `openAiProvider({ models })` is where those are named.
- */
-export const OPENAI_MODEL_IDS: readonly ModelId[] = [
-  'gpt-5.6-sol',
-  'gpt-5.6-terra',
-  'gpt-5.6-luna',
-];
-
-/**
- * Shared by the whole 5.6 family: a 1.05M context, a 128k output ceiling, and `reasoning_effort`
- * over exactly the five rungs `EFFORTS` declares (the endpoint also takes `none`, which is what
- * `thinking: 'disabled'` maps onto). `adaptive` is false because the format has no
- * adaptive-thinking control at all — depth is `reasoning_effort` and nothing else.
- */
-const FAMILY = {
-  /** One ladder: `moreCapableThan` compares these three with each other and with nothing else. */
-  family: 'openai',
-  contextWindow: 1_050_000,
-  maxOutput: 128_000,
-  /** Automatic caching starts at a 1024-token prefix; a shorter one silently does not cache. */
-  cacheMinimumTokens: 1_024,
-  reasoning: { effort: true, adaptive: false, disableThinkingUpTo: undefined },
-} as const;
-
-/**
- * The three models this package is confident enough to price, and no more.
- *
- * Prices are the vendor's published list, read from developers.openai.com/api/docs/pricing on
- * **2026-08-16**, in USD per million tokens:
- *
- * | id | input | cached input | output |
- * |---|---|---|---|
- * | `gpt-5.6-sol` | $5.00 | $0.50 | $30.00 |
- * | `gpt-5.6-terra` | $2.00 | $0.20 | $12.00 |
- * | `gpt-5.6-luna` | $0.20 | $0.02 | $1.20 |
- *
- * Deliberately not registered: `gpt-4o`, `gpt-4o-mini` and the `o1` family, whose cached input is
- * **0.5x** their input rate rather than the 0.1x `costOf` assumes — a spec that prices them would
- * under-report a cache-heavy workload by four fifths, and `costOf` answers confidently either way.
- * `gpt-5.5-pro` and `o1-pro` are out for the same class of reason: they publish no cached rate.
- * A wrong price is worse than no entry, so an app wanting one of those registers it itself, with
- * the rate its own contract names.
- */
-export function registerOpenAiModels(): void {
-  registerModel({
-    id: 'gpt-5.6-sol',
-    ...FAMILY,
-    inputPerMillion: usd(500),
-    outputPerMillion: usd(3_000),
-  });
-  registerModel({
-    id: 'gpt-5.6-terra',
-    ...FAMILY,
-    inputPerMillion: usd(200),
-    outputPerMillion: usd(1_200),
-  });
-  registerModel({
-    id: 'gpt-5.6-luna',
-    ...FAMILY,
-    inputPerMillion: usd(20),
-    outputPerMillion: usd(120),
-  });
-}
-
-// The same shape models.ts uses for its built-ins: registration is a module side effect, so the
-// default path is the app's path and importing the provider is enough to price what it serves.
-// Exported as well as called, because the registry is module state and a suite that clears it with
-// `resetModels()` otherwise leaves this provider serving ids nothing can price. Re-registering
-// REPLACES, so an app with a negotiated rate calls `registerModel` after this one and wins.
-registerOpenAiModels();