@womp/spark-agent 5.1.0 → 5.2.0

package/README.md

@@ -117,6 +117,9 @@ next model request. Consume the run's stream while awaiting its result; the resu
 the terminal checkpoint. A host run owner must keep consuming when a view detaches.
 A failed barrier aborts the loop. The default model-step limit is 100; a limit returns
 `status: "incomplete"`. A host must not automatically restart to bypass that limit.
+A generation that ends on any other non-`stop` reason (`length`, `content-filter`, `error`,
+`other`) rejects the result and ends the run `failed`: its error names the unified reason and the
+provider's raw one, and its diagnostics carry the unified reason as `code`.
 
 Spark does not estimate or limit input tokens or automatically shorten conversation history.
 The provider enforces its context limit; rejections follow the normal provider-error path.

package/dist/agent.js

@@ -7,6 +7,29 @@ import { contextWindowFor, fitMessagesToContext } from "./context-budget.js";
 function errorMessage(error) {
     return sparkErrorDiagnostics(error).message;
 }
+/**
+ * `stop` completes a run and `tool-calls` means the step budget ran out
+ * (`stopWhen` is the only other way the loop ends), which leaves it
+ * incomplete. Any other final reason (`length`, `content-filter`, `error`,
+ * `other`) cut the model's last response short, and the run fails with it.
+ */
+function cutShort(finishReason) {
+    return finishReason !== "stop" && finishReason !== "tool-calls";
+}
+/**
+ * The failure for a generation cut short, in the browser loop's words. The
+ * AI SDK runs a step's tool calls only when the step ends with `stop` or
+ * `tool-calls`, so the calls of the step cut short were not executed. The
+ * unified reason goes in `code` and the provider's raw reason in the message:
+ * the unified one alone hides which provider reason it stands for (Anthropic's
+ * `max_tokens` and `model_context_window_exceeded` are both `length`).
+ */
+function finishReasonError(finishReason, rawFinishReason, responseHeaders) {
+    const raw = rawFinishReason === undefined || rawFinishReason === finishReason
+        ? ""
+        : ` (provider reason: ${rawFinishReason})`;
+    return Object.assign(new Error(`Generation ended with ${finishReason}${raw}; its partial response was not executed.`), { name: "SparkFinishReasonError", code: finishReason, responseHeaders });
+}
 export function runSparkAgent(options) {
     const maxSteps = options.maxSteps ?? 100;
     const contextWindow = options.contextWindow ?? contextWindowFor(options.modelId);
@@ -178,20 +201,30 @@ export function runSparkAgent(options) {
                         if (part.type === "start")
                             return metadata("running");
                         if (part.type === "finish") {
-                            const completed = part.finishReason === "stop";
-                            return metadata(completed ? "completed" : "incomplete", part.totalUsage);
+                            // A run cut short is marked failed here already; its error and
+                            // diagnostics follow once the generation has settled.
+                            const status = cutShort(part.finishReason)
+                                ? "failed"
+                                : part.finishReason === "stop"
+                                    ? "completed"
+                                    : "incomplete";
+                            return metadata(status, part.totalUsage);
                         }
                         return undefined;
                     },
                 }));
-                const [text, usage, finishReason] = await Promise.all([
+                const [text, usage, finishReason, rawFinishReason, response] = await Promise.all([
                     generation.text,
                     generation.totalUsage,
                     generation.finishReason,
+                    generation.rawFinishReason,
+                    generation.response,
                 ]);
                 controller.signal.throwIfAborted();
                 if (generationError !== undefined)
                     throw generationError;
+                if (cutShort(finishReason))
+                    throw finishReasonError(finishReason, rawFinishReason, response.headers);
                 const status = finishReason === "stop" ? "completed" : "incomplete";
                 completedResult = { text, steps: stepCount, usage, status, finishReason };
             }

package/dist/error-diagnostics.js

@@ -30,17 +30,48 @@ function text(value, limit = 1024) {
     // Bound work as well as storage. Embedded JSON/request dumps are not diagnostics.
     return String(value)
         .slice(0, 8192)
-        .replace(/\{[\s\S]*|\[(?=\s*["{0-9])[\s\S]*/u, "[payload omitted]")
-        .replace(/\b(?:request(?:\s*body)?|response\s*body|payload|headers|image|prompt)\s*[:=][\s\S]*/giu, "[payload omitted]")
+        // An object, or an array of strings, objects or several numbers. An index
+        // such as `messages[2].content[0]` is not an array. `[[` is left to the
+        // inner bracket, so a placeholder after a `[` never matches on a rerun.
+        .replace(/\{[\s\S]*|\[(?=\s*(?:["'{]|-?\d[\d.eE+-]*\s*,))[\s\S]*/u, "[payload omitted]")
+        // A prompt, body or headers after its label, whatever follows.
+        .replace(/\b(?:request\s*body|response\s*body|headers|prompt)\s*[:=][\s\S]*/giu, "[payload omitted]")
+        // A request, payload or image only when a quoted value follows its label:
+        // "400 Bad Request: model not found" and "Invalid image: the image exceeds
+        // 5 MB" are the error, not a copy of the input. Objects and arrays after
+        // a label are already gone above.
+        .replace(/\b(?:request|payload|image)\s*[:=]\s*["'`\u201c\u201d\u2018\u2019][\s\S]*/giu, "[payload omitted]")
         .replace(/data:[^\s]+/giu, "[data omitted]")
         .replace(/https?:\/\/[^\s]+/giu, "[url omitted]")
-        .replace(/\b(?:authorization|proxy-authorization|cookie|set-cookie|x-api-key|api[_ -]?key|access[_ -]?token|refresh[_ -]?token|token|password|(?:client[_ -]?)?secret)\b\s*[:=]\s*[^\r\n,;]+/giu, "[credential omitted]")
+        // Userinfo on any other scheme (postgres, redis, mongodb+srv, ssh...);
+        // the scheme, host and path stay for staff. The last `@` before the host
+        // ends it, so a password containing `@` goes with it.
+        .replace(/\b([a-z][a-z0-9+.-]*:\/\/)[^\s/?#]+@/giu, "$1[credential omitted]@")
+        // A `user:password` whose password contains `/`, `?` or `#` (AWS-generated
+        // passwords can). The colon is required, so `s3://bucket/key@v2` stays, and
+        // up to five digits then `/` after it is a port and path, so
+        // `docker://registry:5000/repo@sha256:...` stays. A password of up to five
+        // digits then a `/` therefore survives.
+        .replace(/\b([a-z][a-z0-9+.-]*:\/\/)[^\s/?#:@]+:(?!\d{1,5}\/)[^\s@]*@/giu, "$1[credential omitted]@")
+        .replace(/\b(?:authorization|proxy-authorization|cookie|set-cookie|x-api-key|api[_ -]?key|access[_ -]?token|refresh[_ -]?token|(?:aws[_ -]?)?session[_ -]?token|x-amz-security-token|token|password|(?:aws[_ -]?)?secret[_ -]?access[_ -]?key|(?:client[_ -]?)?secret)\b\s*[:=]\s*[^\r\n,;]+/giu, "[credential omitted]")
         .replace(/\b(?:Bearer|Basic)\s+[^\s,;]+/giu, "[credential omitted]")
         .replace(/\b(?:sk-|sk_|AIza)[A-Za-z0-9_-]+/gu, "[credential omitted]")
+        // GitHub, Slack, Google OAuth access tokens and AWS access key ids.
+        .replace(/\b(?:gh[pousr]_|github_pat_)[A-Za-z0-9_]+/gu, "[credential omitted]")
+        .replace(/\bxox[abprs]-[A-Za-z0-9-]+/gu, "[credential omitted]")
+        .replace(/\bya29\.[A-Za-z0-9_-]+/gu, "[credential omitted]")
+        // No trailing `\b`: the length limit below can cut a longer run to exactly
+        // 16 characters, which a rerun would then match.
+        .replace(/\b(?:AKIA|ASIA)[A-Z0-9]{16}/gu, "[credential omitted]")
         .replace(/\beyJ[A-Za-z0-9_-]+\.[A-Za-z0-9_-]+\.[A-Za-z0-9_-]+/gu, "[credential omitted]")
         .replace(/[A-Za-z0-9+/=_-]{200,}/gu, "[data omitted]")
         .slice(0, limit);
 }
+const PLACEHOLDER = /\[(?:payload|data|url|credential) omitted\]/gu;
+/** Whether redacted text still says anything once its placeholders are removed. */
+function informative(value) {
+    return value !== undefined && /[\p{L}\p{N}]/u.test(value.replace(PLACEHOLDER, ""));
+}
 export function sparkErrorDiagnostics(error) {
     const seen = new Set();
     const visit = (value, depth) => {
@@ -48,14 +79,25 @@ export function sparkErrorDiagnostics(error) {
         const source = record(value);
         const data = record(source.data);
         const body = responseData(source.responseBody);
+        const nested = record(data.error ?? source.error ?? body.error);
         const provider = record(data.error ?? source.error ?? body.error ?? body);
         const headers = record(source.responseHeaders);
         const status = source.statusCode ?? source.status ?? provider.status;
+        // A message redacted down to placeholders (a provider error whose message
+        // is its JSON body, say) gives way to the provider's own message, which is
+        // redacted the same way.
+        const own = text(source.message);
+        const fallback = text(provider.message ?? (source.message === undefined ? value : undefined));
         const result = {
-            message: text(source.message ?? provider.message ?? value) || "An error occurred.",
+            message: (informative(own) ? own : informative(fallback) ? fallback : (own ?? fallback)) ||
+                "An error occurred.",
         };
+        // The provider's error type (Anthropic's `overloaded_error`, OpenAI's
+        // `invalid_request_error`) stands in for a missing code. The AI SDK puts
+        // a mid-stream provider error's type on the error itself.
+        const type = nested.type ?? (source.name === "AI_StreamProviderError" ? source.type : undefined);
         const name = text(source.name, 128);
-        const code = text(source.code ?? provider.code ?? data.code, 128);
+        const code = text(source.code ?? provider.code ?? data.code ?? (typeof type === "string" ? type : undefined), 128);
         const requestId = text(source.requestId ??
             source.request_id ??
             provider.request_id ??

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@womp/spark-agent",
-  "version": "5.1.0",
+  "version": "5.2.0",
   "description": "Portable skills-based AI agent with local and durable chat transports for Womp",
   "type": "module",
   "main": "./dist/index.js",