@mjasnikovs/pi-task 0.40.12 → 0.40.14

package/dist/shared/child-output.js

@@ -44,9 +44,18 @@ export function isExcerptInContent(excerpt, content) {
         return false;
     return normaliseWhitespace(content).includes(ne);
 }
-/** The child writes these to mark its own join, so they are not source text and
- *  not evidence of invention either. */
+/**
+ * Tokens that are not source text and not evidence of invention either: the
+ * child's own elision marks, and the provenance tag the PROMPT gave it.
+ *
+ * `buildExtractionPrompt` wraps the identity as `<package>aeson@2.2.5.1</package>`,
+ * and re-run 5 caught a child quoting that back inside an otherwise verbatim
+ * excerpt — the only absent word in 24 unverified excerpts across three runs. A
+ * complete lowercase element is the whole rule; `Vec<String>` does not start with
+ * `<` and `<T>` has no closing tag, so neither is excused.
+ */
 const ELISION = /^(?:\.{3}|\u2026|\/\/\s*\.{3})$/;
+const PROMPT_TAG = /^<([a-z_]+)>[^<]*<\/\1>$/;
 /**
  * Cover the excerpt with the longest runs the source actually contains, greedily.
  *
@@ -67,7 +76,7 @@ function coverBySource(normalisedExcerpt, normalisedContent) {
             run++;
         }
         if (run === 0) {
-            if (!ELISION.test(words[i]))
+            if (!ELISION.test(words[i]) && !PROMPT_TAG.test(words[i]))
                 absent.push(words[i]);
             i++;
         }

package/dist/workers/docs-retrieve.d.ts

@@ -22,14 +22,24 @@ export interface RetrieveOptions {
     contentBudget?: number;
 }
 /**
- * How many chunks a retrieval returns, per corpus. An npm package gets 8 and
- * project source gets 50.
+ * How many chunks a retrieval returns. Both corpora get 50, and that is measured.
  *
- * Nothing in this repo records WHY they differ. They are stated here, together,
- * so the divergence is at least visible — and changing either is a
- * retrieval-policy change, not a tidy-up.
+ * The package limit was 8 and nothing recorded why it differed from project
+ * source's 50. Measured both ways, over every recorded query, on the two things
+ * that can be measured:
+ *
+ *   retrieval — does a chunk DEFINE the queried symbol?   91/101 -> 97/101,
+ *               monotone to 50 and flat at 100, p = 0.0703
+ *   answers   — does the extraction child answer at all?  67/94 -> 79/94,
+ *               only-8 3, only-50 15, McNemar exact p = 0.0075
+ *
+ * Abstention on the replay corpus falls 29% to 16%. The cost is +70% of retrieved
+ * text, and `RETRIEVE_CONTENT_BUDGET` still bounds it — at 8 only 6 of 74 calls
+ * reached that budget, so two thirds of it was never spent.
+ *
+ * Changing either is still a retrieval-policy change, not a tidy-up.
  */
-export declare const PACKAGE_RETRIEVE_LIMIT = 8;
+export declare const PACKAGE_RETRIEVE_LIMIT = 50;
 export declare const PROJECT_RETRIEVE_LIMIT = 50;
 /** Character budget for the assembled chunk text. The same for both corpora. */
 export declare const RETRIEVE_CONTENT_BUDGET = 24000;

package/dist/workers/docs-retrieve.js

@@ -1,12 +1,22 @@
 /**
- * How many chunks a retrieval returns, per corpus. An npm package gets 8 and
- * project source gets 50.
+ * How many chunks a retrieval returns. Both corpora get 50, and that is measured.
  *
- * Nothing in this repo records WHY they differ. They are stated here, together,
- * so the divergence is at least visible — and changing either is a
- * retrieval-policy change, not a tidy-up.
+ * The package limit was 8 and nothing recorded why it differed from project
+ * source's 50. Measured both ways, over every recorded query, on the two things
+ * that can be measured:
+ *
+ *   retrieval — does a chunk DEFINE the queried symbol?   91/101 -> 97/101,
+ *               monotone to 50 and flat at 100, p = 0.0703
+ *   answers   — does the extraction child answer at all?  67/94 -> 79/94,
+ *               only-8 3, only-50 15, McNemar exact p = 0.0075
+ *
+ * Abstention on the replay corpus falls 29% to 16%. The cost is +70% of retrieved
+ * text, and `RETRIEVE_CONTENT_BUDGET` still bounds it — at 8 only 6 of 74 calls
+ * reached that budget, so two thirds of it was never spent.
+ *
+ * Changing either is still a retrieval-policy change, not a tidy-up.
  */
-export const PACKAGE_RETRIEVE_LIMIT = 8;
+export const PACKAGE_RETRIEVE_LIMIT = 50;
 export const PROJECT_RETRIEVE_LIMIT = 50;
 /** Character budget for the assembled chunk text. The same for both corpora. */
 export const RETRIEVE_CONTENT_BUDGET = 24_000;
@@ -210,6 +220,16 @@ function definitionChunk(cache, opts, name) {
  * declaration per chunk, so the short chunk carrying the name IS its declaration and
  * the long ones are the prose that merely mentions it.
  */
+/** The chunk without the `// path` / `-- path` line the indexer prepends. */
+function chunkBody(content) {
+    return content.replace(/^(?:\/\/|--)[^\n]*\n/, '');
+}
+/**
+ * A chunk that only declares a MODULE. `pub mod serve;` is shorter than the
+ * function it contains and wins smallest-first, while being neither the value nor
+ * the type this hop exists to find.
+ */
+const MODULE_DECL_RE = /^(?:#\[[^\n]*\]\s*)*(?:pub(?:\s*\([^)]*\))?\s+|export\s+|declare\s+)*(?:mod|module|namespace)\s+[A-Za-z_][\w']*\s*;?\s*$/;
 function valueChunk(cache, opts, name) {
     const rows = cache.db
         .prepare(`SELECT file_path, kind, content, 0 AS rank FROM chunks
@@ -219,7 +239,14 @@ function valueChunk(cache, opts, name) {
     // LIKE has no word boundary, so `decodeFile` matches `decodeFileStrict` — the
     // wrong declaration, and the exact confusion these runs keep producing.
     const whole = new RegExp(`(?<![A-Za-z0-9_])${escapeRe(name)}(?![A-Za-z0-9_])`);
-    const row = rows.find(r => whole.test(r.content));
+    // Over the BODY, never the `// path` line the indexer prepends. Every chunk
+    // from `src/serve/mod.rs` carries "serve" in that line, so smallest-first
+    // handed axum's `serve` a 50-byte `pub struct TapIo` from a neighbouring file
+    // — the name was in the header and nowhere else.
+    const row = rows.find(r => {
+        const text = chunkBody(r.content);
+        return whole.test(text) && !MODULE_DECL_RE.test(text);
+    });
     if (!row)
         return null;
     return { filePath: row.file_path, kind: row.kind, content: row.content, rank: row.rank };

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@mjasnikovs/pi-task",
-    "version": "0.40.12",
+    "version": "0.40.14",
     "description": "Deterministic task planning and spec-orchestration for local models — crash-safe /task pipelines with verify/enforce gates, a real-time remote web view, and web/docs/fetch/worker subagent tools.",
     "type": "module",
     "main": "./dist/index.js",