@avocadostudio-ai/orchestrator-core 0.15.0 → 0.17.0

package/dist/chat/chat-pipeline.js

@@ -1338,6 +1338,15 @@ export async function runChatPipeline(ctx, body, options) {
          * true sentence written about the turn is the one the user never sees.
          */
         let strippedPropsNote;
+        /*
+         * The change_log coverage counts, hoisted for the same reason: they are
+         * computed inside the resolve block below, and the result row that should
+         * carry them is written far past it. They have been written to the log
+         * since the validator was added but never to telemetry, so there was no
+         * way to ask how often a planner under-describes what it is about to do —
+         * only to read about one instance at a time.
+         */
+        let changelogTelemetryFields = {};
         const usageFields = planUsage ? {
             inputTokens: planUsage.inputTokens,
             outputTokens: planUsage.outputTokens,
@@ -1416,6 +1425,11 @@ export async function runChatPipeline(ctx, body, options) {
                 // Not an op narration, so not something to trim for over-describing.
                 notes: hallucinationResult.note ? [hallucinationResult.note] : undefined
             });
+            changelogTelemetryFields = {
+                changelogMissingCount: changelogResult.missingCount || undefined,
+                changelogExtraCount: changelogResult.extraCount || undefined,
+                changelogFieldMislabelCount: changelogResult.fieldMislabelCount || undefined
+            };
             if (changelogResult.missingCount > 0) {
                 ctx.log.warn({
                     event: "planner_incomplete_changelog",
@@ -1775,6 +1789,7 @@ export async function runChatPipeline(ctx, body, options) {
                 id: chatRequestId,
                 at: new Date().toISOString(),
                 phase: "result",
+                ...changelogTelemetryFields,
                 session: body.session,
                 requestedSlug,
                 effectiveSlug,
@@ -1833,6 +1848,7 @@ export async function runChatPipeline(ctx, body, options) {
                 id: chatRequestId,
                 at: new Date().toISOString(),
                 phase: "result",
+                ...changelogTelemetryFields,
                 session: body.session,
                 requestedSlug,
                 effectiveSlug,
@@ -1936,6 +1952,7 @@ export async function runChatPipeline(ctx, body, options) {
                     id: chatRequestId,
                     at: new Date().toISOString(),
                     phase: "result",
+                    ...changelogTelemetryFields,
                     session: body.session,
                     requestedSlug,
                     effectiveSlug,
@@ -1996,6 +2013,7 @@ export async function runChatPipeline(ctx, body, options) {
                 id: chatRequestId,
                 at: new Date().toISOString(),
                 phase: "result",
+                ...changelogTelemetryFields,
                 session: body.session,
                 requestedSlug,
                 effectiveSlug,
@@ -2084,6 +2102,7 @@ export async function runChatPipeline(ctx, body, options) {
                 id: chatRequestId,
                 at: new Date().toISOString(),
                 phase: "result",
+                ...changelogTelemetryFields,
                 session: body.session,
                 requestedSlug,
                 effectiveSlug,
@@ -2724,6 +2743,7 @@ export async function runChatPipeline(ctx, body, options) {
                 id: chatRequestId,
                 at: new Date().toISOString(),
                 phase: "result",
+                ...changelogTelemetryFields,
                 session: body.session,
                 requestedSlug,
                 effectiveSlug,
@@ -2819,6 +2839,7 @@ export async function runChatPipeline(ctx, body, options) {
                     id: chatRequestId,
                     at: new Date().toISOString(),
                     phase: "result",
+                    ...changelogTelemetryFields,
                     session: body.session,
                     requestedSlug,
                     effectiveSlug,
@@ -3431,7 +3452,21 @@ export async function runChatPipeline(ctx, body, options) {
                 }
             })
         : planner.generatePlan;
-    const maxPlanningAttempts = 3;
+    /*
+     * Two, not three.
+     *
+     * A retry re-sends the same prompt — the only thing that differs is the
+     * output budget after a truncation — so what a retry buys is a fresh sample,
+     * and the second sample is worth taking: across 5,632 real turns in
+     * chat-telemetry, attempt 2 ended productively 53% of the time.
+     *
+     * The third did so 3.7% of the time. 437 of 454 turns that reached it paid a
+     * full extra planning pass — a median 7.5s, three times the planner tokens,
+     * and a "Retrying plan generation (3/3)" the user watches — to arrive at the
+     * same error attempt 2 had already produced. Raising it back is a decision
+     * about that ratio, so re-measure before you do.
+     */
+    const maxPlanningAttempts = 2;
     let initialPlan = null;
     let routerDetectedInfo = false;
     const planningErrors = [];

package/dist/nlp/deterministic-planner-suggestions.js

@@ -651,7 +651,9 @@ export function blockContractsSummary(manifest) {
  * they are the phrasings this planner *can* execute without a key.
  */
 export const KEYLESS_PLANNER_NOTICE = "No API key is configured, so this is the built-in demo planner, which handles only simple, literal edits. " +
-    "Set ANTHROPIC_API_KEY or OPENAI_API_KEY in your project's .env.local and restart the dev server for full chat.";
+    "Visual editing is not affected: click any block to select it, edit its fields in the property panel, and publish as usual. " +
+    "It is AI editing \u2014 describing a change in a sentence and having it planned for you \u2014 that needs a model. " +
+    "Set ANTHROPIC_API_KEY or OPENAI_API_KEY in your project's .env.local and restart the dev server to turn it on.";
 /**
  * The same notice, without the *path*.
  *

package/dist/nlp/plan-normalizer.js

@@ -659,6 +659,27 @@ function remapListItemPatchKeys(args) {
     }
     return out;
 }
+/**
+ * A model asked for a string sometimes answers with a JSON-encoded array, and
+ * sometimes wraps that in an array of one: `["[\"did X\", \"did Y\"]"]`.
+ * Returns the decoded entries, or null when this is prose that merely opens
+ * with a bracket. Deliberately narrow — a change_log entry is text the user
+ * reads before approving, and rewriting it on a guess is worse than leaving it.
+ */
+function decodeJsonStringArray(value) {
+    if (typeof value !== "string" || !/^\s*\[/.test(value))
+        return null;
+    let parsed;
+    try {
+        parsed = JSON.parse(value);
+    }
+    catch {
+        return null;
+    }
+    if (!Array.isArray(parsed) || !parsed.every((entry) => typeof entry === "string"))
+        return null;
+    return parsed;
+}
 // ---------------------------------------------------------------------------
 // Plan candidate normalisation
 // ---------------------------------------------------------------------------
@@ -999,6 +1020,17 @@ export function normalizePlanCandidate(input, args) {
         if (!raw.patch) {
             raw.patch = raw.props ?? raw.changes;
         }
+        // `update_item` takes `patch`; `add_item` takes `item`. Both names are
+        // offered side by side in the flattened op schema the planner is handed,
+        // so a planner editing one list entry reaches for the noun that names the
+        // thing it is editing and puts the fields under `item`. Unrepaired, that
+        // reaches Zod as "expected record, received undefined at ops.0.patch" and
+        // the whole plan is thrown away. Scoped to update_item — on add_item,
+        // `item` is the real field and must survive untouched.
+        if (!raw.patch && raw.op === "update_item" && raw.item && typeof raw.item === "object" && !Array.isArray(raw.item)) {
+            raw.patch = raw.item;
+            delete raw.item;
+        }
         // LLMs sometimes put prop values directly on the op object instead of
         // nesting them under "patch". Extract non-structural keys as the patch.
         if (!raw.patch && raw.op === "update_props") {
@@ -1651,6 +1683,18 @@ export function normalizePlanCandidate(input, args) {
                 .map((s) => s.replace(/^\s*[-•*\d.)\]]+\s*/, "").trim())
                 .filter((s) => /\w/.test(s));
     }
+    // Unwrap a JSON-encoded change_log before the branches below read it. An
+    // array holding one encoded array satisfies neither of them — it is already
+    // an array — so it survives to the approval UI as a single line of escaped
+    // JSON, and the coverage validator then pads it with synthesized duplicates
+    // of the very changes it already describes.
+    const decodedChangeLog = decodeJsonStringArray(root.change_log);
+    if (decodedChangeLog) {
+        root.change_log = decodedChangeLog;
+    }
+    else if (Array.isArray(root.change_log)) {
+        root.change_log = root.change_log.flatMap((entry) => decodeJsonStringArray(entry) ?? [entry]);
+    }
     if (typeof root.change_log === "string") {
         // Models sometimes emit the literal string "[]" (or "[ ]") for an empty
         // change log; treat that as an empty array instead of a one-item "[]" list.

package/dist/telemetry/chat-telemetry.d.ts

@@ -41,6 +41,19 @@ export type ChatTelemetryEntry = {
     imagesRequested?: number;
     imagesResolved?: number;
     planningAttempts?: number;
+    /**
+     * change_log coverage drift, from `validateChangelogCoverage`: ops the
+     * planner left undescribed, changes it described but will not make, and
+     * entries naming a field the op does not touch.
+     *
+     * The validator has warned about all three since it was written, and a
+     * warning can be read one incident at a time but never counted — which is
+     * why nobody could say how often the approval UI shows a user a list that
+     * does not match the edit they are approving.
+     */
+    changelogMissingCount?: number;
+    changelogExtraCount?: number;
+    changelogFieldMislabelCount?: number;
     contextPackBytes?: number;
     contractMode?: "minimal" | "targeted" | "full";
     contractBytes?: number;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@avocadostudio-ai/orchestrator-core",
-  "version": "0.15.0",
+  "version": "0.17.0",
   "type": "module",
   "exports": {
     "./package.json": "./package.json",
@@ -22,8 +22,8 @@
     "openai": "^4.87.1",
     "sharp": "^0.35.4",
     "zod": "^4.3.6",
-    "@avocadostudio-ai/migration-sdk": "^0.15.0",
-    "@avocadostudio-ai/shared": "^0.15.0"
+    "@avocadostudio-ai/migration-sdk": "^0.17.0",
+    "@avocadostudio-ai/shared": "^0.17.0"
   },
   "devDependencies": {
     "@anthropic-ai/claude-agent-sdk": "^0.3.220",