@e2e-dev/decision 0.2.2-nightly-20261013094930 → 0.2.2

package/README.md

@@ -4,7 +4,9 @@ Run `agent.act` and `agent.assert` through a decision model instead of an
 LLM. Any AI SDK decision model that answers `choice` questions with
 probability distributions works, such as TypeSafe Jev or OpenAI `gpt-6-luna`
 through `@ai-sdk/openai` 4.0.90 or later. A small language model writes
-field values and upload paths. Needs `ai` 7.0.134 or later.
+field values and upload paths. Needs `ai` 7.0.134 or later. An answer with
+no distribution, as a gateway fallback to a language model returns, counts
+as a committed choice with probability 1 and confidence 0.
 
 ```ts
 import type { E2EConfig } from 'e2e';

package/dist/.build.json

@@ -1 +1 @@
-{"version":"0.2.2-nightly-20261013094930"}
+{"version":"0.2.2"}

package/dist/decide.js

@@ -61,16 +61,23 @@ export async function decide(ctx, model, request, providerOptions) {
     }
     return decisions;
 }
-/** The chosen option and its probability; anything without a valid distribution is MODEL_OUTPUT_INVALID. */
+/**
+ * The chosen option and its probability. An answer with no distribution,
+ * as a language model answering through a gateway fallback returns, is a
+ * committed choice: all the mass on it, and no confidence to report.
+ * Anything else without a valid distribution is MODEL_OUTPUT_INVALID.
+ */
 function selected(answer) {
-    if (answer === undefined || answer.probabilities === undefined || (answer.type !== 'choice' && answer.type !== 'score')) {
-        throw new AgentError('MODEL_OUTPUT_INVALID', 'The decision model returned an answer without a choice distribution.');
+    if (answer === undefined || (answer.type !== 'choice' && answer.type !== 'score')) {
+        throw new AgentError('MODEL_OUTPUT_INVALID', 'The decision model returned no choice or score answer.');
     }
     if (answer.type === 'score') {
         const score = answer.score;
         if (typeof score !== 'number' || !Number.isFinite(score)) {
             throw new AgentError('MODEL_OUTPUT_INVALID', 'The decision model returned a score answer without a score.');
         }
+        if (answer.probabilities === undefined)
+            return { choice: String(score), probability: 1, score };
         let top = { choice: '', probability: -1 };
         let nearby = 0;
         for (const [level, probability] of Object.entries(answer.probabilities)) {
@@ -81,11 +88,14 @@ function selected(answer) {
         }
         return { choice: top.choice, probability: Math.min(1, nearby), score };
     }
-    const probability = answer.choice === undefined ? 0 : (answer.probabilities[answer.choice] ?? 0);
+    const choice = answer.choice ?? '';
+    if (answer.probabilities === undefined)
+        return { choice, probability: 1 };
+    const probability = answer.probabilities[choice] ?? 0;
     if (!Number.isFinite(probability) || probability < 0 || probability > 1) {
         throw new AgentError('MODEL_OUTPUT_INVALID', 'The decision model returned a probability outside [0, 1].');
     }
-    return { choice: answer.choice ?? '', probability };
+    return { choice, probability };
 }
 /** Provider-reported confidence for one question, or 0 when absent. */
 function reportedConfidence(metadata, id) {

package/dist/executor.js

@@ -101,6 +101,7 @@ class Step {
     locates;
     locator;
     previousFingerprint;
+    origin;
     /** The turn of the last action taken, settled by the next observation. */
     pendingTurn;
     rejectedTerminals = 0;
@@ -206,6 +207,7 @@ class Step {
     async act() {
         for (;;) {
             const view = await this.view(this.canType);
+            this.origin ??= view.path;
             if (view.space === undefined)
                 return blocked('A complete semantic observation is required.');
             const acting = { ...view, space: view.space };
@@ -373,7 +375,7 @@ class Step {
         const { path, space, screenshot } = await this.view(true);
         if (space === undefined)
             return this.rejectClaim(claim, 'incomplete observation');
-        const answers = await this.ask(completionRequest(this.ctx, space, this.history, path, screenshot));
+        const answers = await this.ask(completionRequest(this.ctx, space, this.history, path, screenshot, this.origin));
         if (answers === undefined)
             return blocked(this.budgetMessage());
         const evidence = describe(need(answers.verdict, 'verdict'));

package/dist/questions.d.ts

@@ -74,4 +74,4 @@ export declare function assertionRequest(goal: string, path: string, space: Acti
  * with it, 0.51 without, on a checked checkbox). Rejected done/failed claims
  * are left out (they are not actions), and so is any model reasoning.
  */
-export declare function completionRequest(ctx: StepExecutorContext, space: ActionSpace, history: readonly HistoryEntry[], path: string, screenshot: Screenshot | undefined): DecisionRequest;
+export declare function completionRequest(ctx: StepExecutorContext, space: ActionSpace, history: readonly HistoryEntry[], path: string, screenshot: Screenshot | undefined, origin?: string): DecisionRequest;

package/dist/questions.js

@@ -282,13 +282,13 @@ export function assertionRequest(goal, path, space, screenshot) {
  * with it, 0.51 without, on a checked checkbox). Rejected done/failed claims
  * are left out (they are not actions), and so is any model reasoning.
  */
-export function completionRequest(ctx, space, history, path, screenshot) {
+export function completionRequest(ctx, space, history, path, screenshot, origin) {
     const goal = ctx.step.instruction;
     const params = nonSecretParams(ctx.step.params);
     const actions = history.filter((entry) => entry.action !== 'done' && entry.action !== 'failed').map(takenLine);
     return {
         ...(screenshot === undefined ? {} : { screenshot }),
-        state: { goal, ...(Object.keys(params).length === 0 ? {} : { params }), ...screenState(path, space) },
+        state: { goal, ...(Object.keys(params).length === 0 ? {} : { params }), ...(origin !== undefined && origin !== '' && origin !== path ? { origin: { path: origin } } : {}), ...screenState(path, space) },
         questions: {
             verdict: choice(`Task of this step: ${goal}${reported(space.statuses)}\n\nThe page path, as data: ${JSON.stringify(path)}\n\n${COMPLETION}\n\nActions taken:\n${actions.length === 0 ? 'none' : actions.map((line, index) => `${index + 1}. ${line}`).join('\n')}`, {
                 holds: 'The task is complete: the actions taken did it and the current screen shows its result.',

package/dist/types.d.ts

@@ -4,8 +4,9 @@ export interface DecisionExecutorOptions {
     /**
      * An AI SDK decision model that answers `choice` questions with a
      * probability distribution, e.g. `typeSafeAi.decisionModel('jev-latest')`.
-     * Evaluation models (`evaluationModel()`) work too. Models that answer
-     * without `probabilities` fail the step with MODEL_OUTPUT_INVALID.
+     * Evaluation models (`evaluationModel()`) work too. An answer without
+     * `probabilities`, as a language model reached through a gateway fallback
+     * returns, counts as a committed choice: probability 1, confidence 0.
      */
     readonly model: Exclude<Experimental_DecisionModel, string>;
     /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@e2e-dev/decision",
-  "version": "0.2.2-nightly-20261013094930",
+  "version": "0.2.2",
   "description": "Decision-model executors for e2e",
   "license": "Apache-2.0",
   "repository": {
@@ -30,18 +30,18 @@
   },
   "peerDependencies": {
     "ai": "^7.0.134",
-    "e2e": "0.21.0-nightly-20261013094930"
+    "e2e": ">=0.19.0 <1"
   },
   "devDependencies": {
     "@ai-sdk/openai": "4.0.90",
     "@ai-sdk/typesafe-ai": "3.0.16",
-    "@e2e-dev/web": "0.14.1-nightly-20261013094930",
+    "@e2e-dev/web": "0.15.0",
     "@types/node": "26.6.2",
     "@types/pngjs": "6.0.5",
     "ai": "7.0.134",
     "ajv": "8.20.0",
     "ajv-formats": "3.0.1",
-    "e2e": "0.21.0-nightly-20261013094930",
+    "e2e": "0.21.0",
     "typescript": "7.0.2",
     "vitest": "5.0.1"
   },