llm-output-guard 1.11.1 → 1.11.3

package/README.md

@@ -32,7 +32,7 @@ Your retry layer watches for `429`, `5xx` and timeouts. It cannot see a model th
 looped until `max_tokens`, returned `{}`, stopped mid-sentence, or answered in the
 wrong language — because all of those arrive as a **successful request**.
 
-This produces the signal that layer is missing. Zero dependencies, ~5 KB gzipped,
+This produces the signal that layer is missing. Zero dependencies, ~6 KB min+gzip,
 synchronous, no network.
 
 ```bash
@@ -90,9 +90,15 @@ a boolean** — you pick the line.
 import OpenAI from 'openai';
 import { withOutputGuard } from 'llm-output-guard/openai';
 
-const client = withOutputGuard(new OpenAI(), { ...presets.chat, onDegenerate: 'abort' });
+const client = withOutputGuard(new OpenAI(), { ...presets.chat, onDegenerate: 'throw' });
 ```
 
+`onDegenerate` decides what a failed check does: **`'throw'`** (the default)
+fails the call, and on a stream also cancels the request. `'abort'` is the
+streaming-only alternative — it ends the stream and keeps what arrived, and on a
+non-streaming call it stops nothing. `'ignore'` reports and changes nothing,
+which is how to roll out.
+
 Adapters for the **OpenAI SDK** (both `chat.completions` and `responses`),
 **Anthropic**, **Google Gemini** and the **Vercel AI SDK** — plus anything
 speaking OpenAI's protocol: Groq, Together, OpenRouter, Fireworks, vLLM,
@@ -207,7 +213,9 @@ withOutputGuard(new OpenAI(), { ...presets.chat, checkPromptEcho: true });
 ```
 
 Not for rewrite, translate or summarise endpoints, where copying the input is
-the job. [More →](docs/detectors.md#returning-the-prompt-instead-of-an-answer)
+the job. Abstains under 40 words, so a one-line prompt echoed perfectly still
+scores 0.000.
+[More →](docs/detectors.md#returning-the-prompt-instead-of-an-answer)
 
 ## The verdict
 
@@ -267,6 +275,8 @@ because only one of those is evidence.
 ## Design notes
 
 - **Zero runtime dependencies**, enforced in CI. Node ≥ 18; works on edge, browser, Deno, Bun.
+- **The size claim is a budget, not a memory.** `npm run size` bundles each entry, minifies and gzips it, and fails over budget — enforced in CI beside the zero-dependency check, because a number in prose is the cheapest thing in a repo to go stale.
+- **Types resolve on old and new TypeScript alike.** Every subpath is listed in `typesVersions` as well as `exports`, so `moduleResolution: "node"` — still the default under `module: commonjs` — sees the adapters instead of only the root.
 - **Pure and synchronous.** No network, no clock, no randomness — safe on a hot path, trivial to test.
 - **Scores, not booleans.** Detectors report 0–1 and leave the threshold decision to you.
 - **Abstains rather than guesses.** Samples too short to judge score 0.