@tangle-network/agent-app 0.44.11 → 0.44.13

package/dist/work-product/index.js

@@ -1,7 +1,7 @@
 import {
   parseReviewQueueItem,
   projectReviewQueue
-} from "../chunk-ZRSUCTST.js";
+} from "../chunk-GEYACSFW.js";
 import {
   isWorkProductStatus,
   parseAgentCheckInput,
@@ -11,7 +11,7 @@ import {
   persistedPartToWorkProduct,
   unresolvedBlockingExceptions,
   workProductToPersistedPart
-} from "../chunk-V3HO43PG.js";
+} from "../chunk-ZVEEWGDK.js";
 import {
   ToolInputError,
   defineAppTool
@@ -403,6 +403,45 @@ function sourceContainsQuote(sourceText, quote) {
   if (normalizedQuote.length === 0) return false;
   return normalizeQuoteText(sourceText).includes(normalizedQuote);
 }
+function lineAround(text, index) {
+  let start = text.lastIndexOf("\n", index);
+  start = start < 0 ? 0 : start + 1;
+  let end = text.indexOf("\n", index);
+  if (end < 0) end = text.length;
+  if (end > start && text[end - 1] === "\r") end -= 1;
+  return { start, end };
+}
+function findSourceLine(sourceText, needle, occurrence = 1) {
+  const trimmed = needle.trim();
+  if (trimmed.length === 0) return { ok: false, failure: { reason: "blank_needle" } };
+  const positions = [];
+  for (let at = sourceText.indexOf(trimmed); at >= 0; at = sourceText.indexOf(trimmed, at + 1)) {
+    positions.push(at);
+  }
+  if (positions.length === 0) {
+    const wanted = normalizeQuoteText(trimmed);
+    if (wanted.length > 0) {
+      let cursor = 0;
+      while (cursor <= sourceText.length) {
+        const bound2 = lineAround(sourceText, cursor);
+        const line = sourceText.slice(bound2.start, bound2.end);
+        if (normalizeQuoteText(line).includes(wanted) || normalizeQuoteText(line.replace(/,/gu, "")).includes(wanted)) {
+          positions.push(bound2.start);
+        }
+        if (bound2.end >= sourceText.length) break;
+        cursor = bound2.end + 1;
+      }
+    }
+  }
+  if (positions.length === 0) return { ok: false, failure: { reason: "not_found" } };
+  if (occurrence < 1 || occurrence > positions.length) {
+    return { ok: false, failure: { reason: "occurrence_out_of_range", found: positions.length } };
+  }
+  const bound = lineAround(sourceText, positions[occurrence - 1]);
+  const quote = sourceText.slice(bound.start, bound.end);
+  if (quote.trim().length === 0) return { ok: false, failure: { reason: "not_found" } };
+  return { ok: true, span: bound, quote, occurrences: positions.length };
+}
 function sliceSourceSpan(sourceText, span) {
   for (const field of ["start", "end"]) {
     const value = span[field];
@@ -480,6 +519,16 @@ function spanErrorDetail(failure) {
       return "that range is only whitespace. Widen it to the characters that carry the value.";
   }
 }
+function findErrorDetail(failure, needle) {
+  switch (failure.reason) {
+    case "blank_needle":
+      return "the value to locate is empty.";
+    case "not_found":
+      return `${JSON.stringify(needle)} does not occur in that document. Read it again and cite a value it actually contains, or \u2014 if this figure is COMPUTED rather than read \u2014 omit the locator and state the computation in claim.`;
+    case "occurrence_out_of_range":
+      return `that value occurs ${failure.found} time(s) in the document; findOccurrence is out of range.`;
+  }
+}
 async function resolveEvidenceQuotes(config, entries, ctx) {
   const readSourceText = config.readSourceText;
   const texts = /* @__PURE__ */ new Map();
@@ -489,8 +538,36 @@ async function resolveEvidenceQuotes(config, entries, ctx) {
   };
   for (let index = 0; index < entries.length; index += 1) {
     const entry = entries[index];
+    const find = entry.locator.find;
     const span = entry.locator.span;
     const quote = entry.locator.quote;
+    if (find !== void 0) {
+      if (!readSourceText) {
+        throw new ToolInputError(
+          "span_unsupported",
+          `entries[${index}].locator.find: this deployment cannot read source text, so a value cannot be located. Record the entry without locator.find.`,
+          500
+        );
+      }
+      const text2 = await readText(entry.sourceRef);
+      if (text2 === null) {
+        throw new ToolInputError(
+          "unverifiable_quote",
+          `entries[${index}].locator.find: "${entry.sourceRef}" has no readable text, so the value cannot be located. Record the entry without a locator and state the basis in claim.`
+        );
+      }
+      const located = findSourceLine(text2, find, entry.locator.findOccurrence ?? 1);
+      if (!located.ok) {
+        throw new ToolInputError(
+          "value_not_found",
+          `entries[${index}].locator.find into "${entry.sourceRef}": ${findErrorDetail(located.failure, find)}`
+        );
+      }
+      entry.locator.span = located.span;
+      entry.locator.quote = located.quote;
+      entry.locator.quoteBasis = "span";
+      continue;
+    }
     if (span) {
       if (!readSourceText) {
         throw new ToolInputError(
@@ -529,7 +606,7 @@ async function resolveEvidenceQuotes(config, entries, ctx) {
     if (!sourceContainsQuote(text, quote)) {
       throw new ToolInputError(
         "quote_not_found",
-        `entries[${index}].locator.quote: ${JSON.stringify(quote)} does not occur in "${entry.sourceRef}". Cite locator.span {start, end} instead \u2014 the character range you read it at, which this platform slices out of the document itself so the text is right by construction. If the value is COMPUTED rather than read, omit both and state the computation in claim.`
+        `entries[${index}].locator.quote: ${JSON.stringify(quote)} does not occur in "${entry.sourceRef}". Cite locator.find instead \u2014 the value as it appears in the document \u2014 and this platform locates it and writes the quote itself, so the text is right by construction. If the value is COMPUTED rather than read, omit both and state the computation in claim.`
       );
     }
     entry.locator.quoteBasis = "model";
@@ -577,7 +654,7 @@ function buildWorkProductTools(config) {
   });
   const upsertEvidence = defineAppTool({
     name: "upsert_evidence",
-    description: "Record source\u2192field lineage for the current work product, incrementally as you find it. Each entry links a source document (sourceRef + locator) to one artifact target and states the claim it supports. Cite by locator.span {start, end} \u2014 the character range you read the value at \u2014 and the platform slices the supporting quote out of the document for you. Re-emitting an entry id replaces that entry. Address the work product by scopeKey; the first call creates the draft.",
+    description: "Record source\u2192field lineage for the current work product, incrementally as you find it. Each entry links a source document (sourceRef + locator) to one artifact target and states the claim it supports. Cite by locator.find \u2014 the value exactly as it appears in the document \u2014 and the platform locates it and writes the supporting quote for you. Re-emitting an entry id replaces that entry. Address the work product by scopeKey; the first call creates the draft.",
     parameters: {
       type: "object",
       properties: {
@@ -596,9 +673,18 @@ function buildWorkProductTools(config) {
                 properties: {
                   page: { type: "number" },
                   range: { type: "string", description: "Free-form location: 'L120-L134' | 'B7' | '\xB64'." },
+                  find: {
+                    type: "string",
+                    description: 'PREFERRED \u2014 use this whenever the value was READ from a document. The value exactly as it appears in the source (for example "128,450.00"), or a short distinctive phrase from the supporting line. The platform LOCATES it, cites the whole line it sits on, and returns that line to you. You do not retype the quote and you do not compute character offsets \u2014 so the citation can neither be invented nor land on the wrong line. If the value is not in the document, the entry is refused.'
+                  },
+                  findOccurrence: {
+                    type: "integer",
+                    minimum: 1,
+                    description: "Which occurrence of `find` to cite when the value appears more than once. Defaults to the first."
+                  },
                   span: {
                     type: "object",
-                    description: "PREFERRED. The character range you read the value at, as absolute offsets into the whole document: start is the first character, end is one past the last. If you read with an offset, add that offset to the position within the text you got back. The platform slices the quote out of the document itself and stores it, and returns it to you \u2014 you never retype source text, so the citation cannot be wrong. Use this whenever the value was READ from a document.",
+                    description: "Only when you have exact character offsets from a tool that computed them. Absolute offsets into the whole document: start is the first character, end is one past the last. Prefer `find` \u2014 offsets computed by hand land on the wrong line.",
                     properties: {
                       start: { type: "integer", minimum: 0 },
                       end: { type: "integer", minimum: 1 }
@@ -832,7 +918,7 @@ function buildWorkProductTools(config) {
           await unwrap(() => service.recordChecks(draft.id, checks), "checks_rejected");
           throw new ToolInputError(
             "evidence_not_anchored",
-            `Cannot submit: these material targets have an evidence row but no source anchor: ${unanchored.join(", ")}. Re-emit each with locator.span {start, end} \u2014 the character range in the document you read the value at. The platform slices the quote out of the document itself, so you never retype source text.`
+            `Cannot submit: these material targets have an evidence row but no source anchor: ${unanchored.join(", ")}. Re-emit each with locator.find \u2014 the value exactly as it appears in the document. The platform locates it and writes the quote itself, so you never retype source text or count characters.`
           );
         }
       }
@@ -954,6 +1040,7 @@ export {
   createWorkProductRoutes,
   createWorkProductService,
   finalizeWorkProductProvenance,
+  findSourceLine,
   isWorkProductStatus,
   isWorkProductTerminal,
   normalizeQuoteText,