akm-cli 0.10.261016-beta → 0.10.261017-beta

package/dist/indexer/usage/usage-events.js

@@ -69,20 +69,28 @@ export function insertUsageEventStrict(db, event) {
 // ── Query ────────────────────────────────────────────────────────────────────
 /**
  * User feedback counts for many assets at once, keyed by durable `entry_ref`.
- * Refs with no user feedback are absent from the map. Read in one query per
- * chunk of refs.
+ * A negative that carried an exact fix does not count once an accepted
+ * `feedback` proposal for the asset exists at or after it (the rule `akm
+ * improve` uses for a fix already acted on); reverting that proposal makes it
+ * count again. Refs with no user feedback are absent from the map. Read in one
+ * query per chunk of refs.
  */
 export function countFeedbackSignalsByRef(db, refs) {
     const counts = new Map();
     for (let i = 0; i < refs.length; i += SQLITE_CHUNK_SIZE) {
         const chunk = refs.slice(i, i + SQLITE_CHUNK_SIZE);
         const rows = db
-            .prepare(`SELECT entry_ref,
-           SUM(CASE WHEN signal = 'positive' THEN 1 ELSE 0 END) AS pos,
-           SUM(CASE WHEN signal = 'negative' THEN 1 ELSE 0 END) AS neg
-         FROM usage_events
-         WHERE event_type = 'feedback' AND source = 'user' AND entry_ref IN (${chunk.map(() => "?").join(",")})
-         GROUP BY entry_ref`)
+            .prepare(`SELECT u.entry_ref,
+           SUM(CASE WHEN u.signal = 'positive' THEN 1 ELSE 0 END) AS pos,
+           SUM(CASE WHEN u.signal = 'negative' AND NOT (
+             json_extract(u.metadata, '$.fix') IS NOT NULL AND EXISTS (
+               SELECT 1 FROM proposals p
+               WHERE p.ref = u.entry_ref AND p.source = 'feedback' AND p.status = 'accepted'
+                 AND julianday(p.created_at) >= julianday(u.created_at))
+           ) THEN 1 ELSE 0 END) AS neg
+         FROM usage_events u
+         WHERE u.event_type = 'feedback' AND u.source = 'user' AND u.entry_ref IN (${chunk.map(() => "?").join(",")})
+         GROUP BY u.entry_ref`)
             .all(...chunk);
         for (const row of rows)
             counts.set(row.entry_ref, { pos: row.pos, neg: row.neg });
@@ -150,3 +158,17 @@ export function purgeOldUsageEvents(db, retentionDays) {
         /* Table may not exist yet */
     }
 }
+/**
+ * Rewrite every `usage_events.entry_ref` naming `oldBundleId` to `newBundleId`
+ * (`akm bundle rename`). Search ranking counts feedback by `entry_ref`, so an
+ * event left under the old prefix stops counting for the renamed bundle's
+ * assets. `entry_id` is untouched: the rename keeps every entry's row id. A
+ * legacy unqualified ref is not matched by the `<old>//` prefix and is left alone.
+ * Returns the number of rows rewritten.
+ */
+export function renameUsageEventRefs(db, oldBundleId, newBundleId) {
+    const prefix = `${oldBundleId}//`;
+    return Number(db
+        .prepare(`UPDATE usage_events SET entry_ref = ? || substr(entry_ref, length(?) + 1) WHERE substr(entry_ref, 1, length(?)) = ?`)
+        .run(newBundleId, oldBundleId, prefix, prefix).changes);
+}

package/dist/indexer/walk/walker.js

@@ -112,7 +112,7 @@ function walkStashGit(stashRoot, options) {
         ? ignoredResult.stdout.toString("utf8")
         : String(ignoredResult.stdout ?? "");
     // Data-hygiene filename skips: the legacy metadata sidecar (never indexed as
-    // content — Chunk-8 folds it into the bundle format) plus git dot-files.
+    // content) plus git dot-files.
     const SKIP_FILES = new Set([".stash.json", ".gitignore", ".gitattributes"]);
     const stdout = Buffer.isBuffer(result.stdout) ? result.stdout.toString("utf8") : String(result.stdout ?? "");
     const files = [...stdout.split("\0"), ...ignoredStdout.split("\0")]

package/dist/output/shapes/helpers.js

@@ -71,7 +71,7 @@ export function shapeProposalEntry(entry, detail) {
     if (detail === "normal") {
         // `confidence` and `gateDecision` (#577) explain why a proposal is pending,
         // so they are projected at `normal` for `akm proposal list/show` when present.
-        // `retirement` (alpha.9) is the same kind of field for a consolidate
+        // `retirement` is the same kind of field for a consolidate
         // `retire` proposal: without it, "list/show" carries no reason a reviewer
         // could act on (its `payload.content` is empty by design — a retire
         // proposal deletes, it does not write).
@@ -278,7 +278,7 @@ export function shapeSearchHit(hit, detail) {
     }
     // Stash hit (local or remote)
     // `ref` is included at `brief` so agents can run `akm show <ref>` without
-    // needing --detail full or --detail agent (REC-03).
+    // needing --detail full or --detail agent.
     if (detail === "brief") {
         return pickFields(hit, ["type", "name", "ref", "action", "estimatedTokens", "keys"]);
     }
@@ -376,7 +376,7 @@ export function shapeShowOutput(result, detail, shape = "human") {
     const base = pickFields(result, [
         "type",
         "name",
-        // ref is always projected, same as path/editable below (R-020).
+        // ref is always projected, same as path/editable below.
         "ref",
         "origin",
         "action",

package/dist/output/text/command-format.js

@@ -7,8 +7,7 @@
  * events/history, search/curate, and the stash-lifecycle verbs (init,
  * index, list, add, remove, update, upgrade, clone).
  *
- * Split out of `helpers.ts` (formerly 1418 lines / 59 fns) as its own
- * sibling module: unlike `show`/`workflow`/`proposal` (each a stateful,
+ * Unlike `show`/`workflow`/`proposal` (each a stateful,
  * multi-branch renderer with its own module), these are flat
  * field-extraction formatters with no shared state.
  */
@@ -308,7 +307,7 @@ export function formatSearchPlain(r, detail) {
         if (parts.length > 0)
             lines.push(`timing: ${parts.join(", ")}`);
     }
-    // REC-02: When stash hits exist, tell the agent the next required step so it
+    // When stash hits exist, tell the agent the next required step so it
     // doesn't skip `akm show` and write from training memory instead.
     if (hits.length >= 1) {
         // Prefer skill/command/agent type hits for the "Next:" ref — knowledge docs are
@@ -517,7 +516,7 @@ export function formatAddPlain(r) {
     // writes a desired locator to config — nothing is synced until a later
     // `akm bundle update`. Force-fitting the declarative shape's `added`/`entry`/`message`
     // fields into the "Installed <ref> (N scanned...)" wording produced
-    // "Installed undefined (0 directories scanned...)" (R-014) because it
+    // "Installed undefined (0 directories scanned...)" because it
     // implied a sync that never happened. `AddResponse` always carries an
     // `index` block (even when zero entries were scanned); `SourceAddResult`
     // never does — that is the reliable shape discriminator used below to
@@ -630,11 +629,11 @@ export function formatBundleRenamePlain(r) {
     return lines.join("\n");
 }
 export function formatUpdatePlain(r) {
-    // R-015: `processed` alone conflated three cases into one empty array —
+    // `processed` alone would conflate three cases into one empty array —
     // a true no-op (--all with nothing configured), a source this call flatly
     // never looked at (a plain source under --all), and a plain git/website
     // source that WAS just synced successfully but has no `UpdateResultItem`
-    // shape to report (no version/lock to diff). All three rendered as the
+    // shape to report (no version/lock to diff). All three would render as the
     // same misleading "nothing to update". `plainSynced` and `skipped` (see
     // sources/types.ts) cover the other two so this can report accurately.
     const processed = r.processed;

package/dist/scripts/akm-migrate-node.js

@@ -34568,6 +34568,27 @@ function isMarkdownFenceClosing(line, fence) {
     return false;
   return /^[\t ]*$/.test(line.slice(cursor));
 }
+function fencedCodeIdentifiers(line) {
+  const cleaned = line.replace(/(["'`])((?:\\.|(?!\1).)*)\1(\s*:)?/g, (match, _quote, body, colon, offset) => {
+    if (colon)
+      return /^\s*["'`]/.test(line.slice(offset + match.length)) ? ` ${body} ` : ` ${body}: `;
+    return /^\/[\w./-]+$/.test(body) ? ` ${body} ` : " ";
+  }).replace(/([A-Za-z_][A-Za-z0-9_.-]*\s*[:=]\s*)((?:[Bb]earer|[Bb]asic|[Tt]oken|[Dd]igest)\s+\S+|\S+)/g, (_match, name, value) => /^[!(]*[A-Za-z_][\w$]*(?:\.[A-Za-z_][\w$]*)+[()\]}]*[,;]?$/.test(value) ? `${name} ${value}` : name).replace(/(?<![\w.-])[A-Za-z0-9+/=]{20,}(?![\w.-])/g, (run) => /[a-z]/.test(run) && /[A-Z]/.test(run) && /\d/.test(run) ? " " : run);
+  const words = [];
+  for (let token of cleaned.match(/--?[A-Za-z][A-Za-z0-9-]*|[A-Za-z_][A-Za-z0-9_.-]*/g) ?? []) {
+    token = token.replace(/[.-]+$/, "");
+    const bare = token.replace(/^-+/, "");
+    if (bare.length < 2)
+      continue;
+    if (bare.length >= TOKEN_LIKE_MIN_LENGTH && /\d/.test(bare))
+      continue;
+    words.push(token);
+    const parts = bare.split(/[_.-]|(?<=[a-z0-9])(?=[A-Z])/).filter((part) => part.length > 1);
+    if (parts.length > 1)
+      words.push(...parts);
+  }
+  return words.join(" ");
+}
 function stripMarkdownHtmlComments(line, state) {
   let cursor = 0;
   let visible = "";
@@ -34733,6 +34754,7 @@ function projectMarkdownContent(body, truncationInfo) {
   const innerBlock = findInnerFrontmatterBlock(lines);
   const start = innerBlock && isFrontmatterShaped(lines, innerBlock) ? innerBlock.close + 1 : 0;
   const projected = [];
+  const fenced = [];
   forEachVisibleMarkdownLine(lines, start, (_index, visible) => {
     let trimmed = visible.trim();
     if (!trimmed || /^(-{3,}|\*{3,}|_{3,}|=+)$/.test(trimmed))
@@ -34750,15 +34772,19 @@ function projectMarkdownContent(body, truncationInfo) {
     trimmed = trimmed.replace(/\s+/g, " ").trim();
     if (trimmed)
       projected.push(trimmed);
+  }, (_index, fencedLine) => {
+    const identifiers = fencedCodeIdentifiers(fencedLine);
+    if (identifiers)
+      fenced.push(identifiers);
   });
-  const text2 = projected.join(" ").replace(/\s+/g, " ").trim();
+  const text2 = [...projected, ...fenced].join(" ").replace(/\s+/g, " ").trim();
   if (!text2)
     return;
   if (truncationInfo)
     truncationInfo.truncated = text2.length > MARKDOWN_CONTENT_MAX_CHARS;
   return truncateUnicodeSafe(text2, MARKDOWN_CONTENT_MAX_CHARS);
 }
-function forEachVisibleMarkdownLine(lines, start, visit3) {
+function forEachVisibleMarkdownLine(lines, start, visit3, visitFenced) {
   let fence;
   const htmlComment = { inComment: false };
   for (let index = start;index < lines.length; index += 1) {
@@ -34766,6 +34792,8 @@ function forEachVisibleMarkdownLine(lines, start, visit3) {
     if (fence) {
       if (isMarkdownFenceClosing(rawLine, fence))
         fence = undefined;
+      else
+        visitFenced?.(index, rawLine);
       continue;
     }
     const leadingFence = htmlComment.inComment ? undefined : parseMarkdownFenceOpening(rawLine);
@@ -34936,7 +34964,7 @@ function extractDirTagsFromName(name) {
   }
   return Array.from(tags);
 }
-var SCOPE_KEYS, KNOWN_QUALITY_VALUES, warnedUnknownQualityValues, COMMENT_TAGS, MARKDOWN_CONTENT_MAX_CHARS = 1e6, MAX_MARKDOWN_LINK_NESTING = 32, markdownFragmentContentByEntry, markdownFragmentProjectionEntries;
+var SCOPE_KEYS, KNOWN_QUALITY_VALUES, warnedUnknownQualityValues, COMMENT_TAGS, MARKDOWN_CONTENT_MAX_CHARS = 1e6, TOKEN_LIKE_MIN_LENGTH = 20, MAX_MARKDOWN_LINK_NESTING = 32, markdownFragmentContentByEntry, markdownFragmentProjectionEntries;
 var init_metadata = __esm(() => {
   init_asset_ref();
   init_frontmatter();

package/dist/scripts/akm-migrate.js

@@ -34567,6 +34567,27 @@ function isMarkdownFenceClosing(line, fence) {
     return false;
   return /^[\t ]*$/.test(line.slice(cursor));
 }
+function fencedCodeIdentifiers(line) {
+  const cleaned = line.replace(/(["'`])((?:\\.|(?!\1).)*)\1(\s*:)?/g, (match, _quote, body, colon, offset) => {
+    if (colon)
+      return /^\s*["'`]/.test(line.slice(offset + match.length)) ? ` ${body} ` : ` ${body}: `;
+    return /^\/[\w./-]+$/.test(body) ? ` ${body} ` : " ";
+  }).replace(/([A-Za-z_][A-Za-z0-9_.-]*\s*[:=]\s*)((?:[Bb]earer|[Bb]asic|[Tt]oken|[Dd]igest)\s+\S+|\S+)/g, (_match, name, value) => /^[!(]*[A-Za-z_][\w$]*(?:\.[A-Za-z_][\w$]*)+[()\]}]*[,;]?$/.test(value) ? `${name} ${value}` : name).replace(/(?<![\w.-])[A-Za-z0-9+/=]{20,}(?![\w.-])/g, (run) => /[a-z]/.test(run) && /[A-Z]/.test(run) && /\d/.test(run) ? " " : run);
+  const words = [];
+  for (let token of cleaned.match(/--?[A-Za-z][A-Za-z0-9-]*|[A-Za-z_][A-Za-z0-9_.-]*/g) ?? []) {
+    token = token.replace(/[.-]+$/, "");
+    const bare = token.replace(/^-+/, "");
+    if (bare.length < 2)
+      continue;
+    if (bare.length >= TOKEN_LIKE_MIN_LENGTH && /\d/.test(bare))
+      continue;
+    words.push(token);
+    const parts = bare.split(/[_.-]|(?<=[a-z0-9])(?=[A-Z])/).filter((part) => part.length > 1);
+    if (parts.length > 1)
+      words.push(...parts);
+  }
+  return words.join(" ");
+}
 function stripMarkdownHtmlComments(line, state) {
   let cursor = 0;
   let visible = "";
@@ -34732,6 +34753,7 @@ function projectMarkdownContent(body, truncationInfo) {
   const innerBlock = findInnerFrontmatterBlock(lines);
   const start = innerBlock && isFrontmatterShaped(lines, innerBlock) ? innerBlock.close + 1 : 0;
   const projected = [];
+  const fenced = [];
   forEachVisibleMarkdownLine(lines, start, (_index, visible) => {
     let trimmed = visible.trim();
     if (!trimmed || /^(-{3,}|\*{3,}|_{3,}|=+)$/.test(trimmed))
@@ -34749,15 +34771,19 @@ function projectMarkdownContent(body, truncationInfo) {
     trimmed = trimmed.replace(/\s+/g, " ").trim();
     if (trimmed)
       projected.push(trimmed);
+  }, (_index, fencedLine) => {
+    const identifiers = fencedCodeIdentifiers(fencedLine);
+    if (identifiers)
+      fenced.push(identifiers);
   });
-  const text2 = projected.join(" ").replace(/\s+/g, " ").trim();
+  const text2 = [...projected, ...fenced].join(" ").replace(/\s+/g, " ").trim();
   if (!text2)
     return;
   if (truncationInfo)
     truncationInfo.truncated = text2.length > MARKDOWN_CONTENT_MAX_CHARS;
   return truncateUnicodeSafe(text2, MARKDOWN_CONTENT_MAX_CHARS);
 }
-function forEachVisibleMarkdownLine(lines, start, visit3) {
+function forEachVisibleMarkdownLine(lines, start, visit3, visitFenced) {
   let fence;
   const htmlComment = { inComment: false };
   for (let index = start;index < lines.length; index += 1) {
@@ -34765,6 +34791,8 @@ function forEachVisibleMarkdownLine(lines, start, visit3) {
     if (fence) {
       if (isMarkdownFenceClosing(rawLine, fence))
         fence = undefined;
+      else
+        visitFenced?.(index, rawLine);
       continue;
     }
     const leadingFence = htmlComment.inComment ? undefined : parseMarkdownFenceOpening(rawLine);
@@ -34935,7 +34963,7 @@ function extractDirTagsFromName(name) {
   }
   return Array.from(tags);
 }
-var SCOPE_KEYS, KNOWN_QUALITY_VALUES, warnedUnknownQualityValues, COMMENT_TAGS, MARKDOWN_CONTENT_MAX_CHARS = 1e6, MAX_MARKDOWN_LINK_NESTING = 32, markdownFragmentContentByEntry, markdownFragmentProjectionEntries;
+var SCOPE_KEYS, KNOWN_QUALITY_VALUES, warnedUnknownQualityValues, COMMENT_TAGS, MARKDOWN_CONTENT_MAX_CHARS = 1e6, TOKEN_LIKE_MIN_LENGTH = 20, MAX_MARKDOWN_LINK_NESTING = 32, markdownFragmentContentByEntry, markdownFragmentProjectionEntries;
 var init_metadata = __esm(() => {
   init_asset_ref();
   init_frontmatter();

package/dist/sources/providers/git-provider.js

@@ -154,10 +154,8 @@ function hasExtractedRepo(repoDir) {
     try {
         if (!fs.statSync(repoDir).isDirectory())
             return false;
-        // WI-3.1: the "any type dir present" test is now sourced from the `akm`
-        // adapter's directoryList() — behavior-identical to the placement
-        // stash-subdir names, with `stashDirNames()` kept live as the fallback. The
-        // content/ subdir check above is preserved verbatim.
+        // The "any type dir present" test is sourced from the `akm`
+        // adapter's directoryList(), with `stashDirNames()` as the fallback.
         const ownedDirs = akmAdapter.directoryList?.({ id: "akm", adapter: "akm", root: repoDir, writable: false }) ?? stashDirNames();
         return ownedDirs.some((dirName) => fs.existsSync(path.join(repoDir, dirName)));
     }

package/dist/sources/providers/provider-utils.js

@@ -37,22 +37,16 @@ function looksLikeAnyBundleRoot(root) {
 }
 /**
  * Find the directory inside `extractedDir` that should be treated as the
- * bundle/stash root. Probes the top-level dir via the ordered §1.2 registry
+ * bundle/stash root. Probes the top-level dir via the ordered registry
  * probe; if nothing claims it, BFS for the shallowest nested candidate.
  */
 export function detectStashRoot(extractedDir) {
     const root = path.resolve(extractedDir);
-    // WI-3.1 → wired to the static registry: the top-level root check now routes
-    // through the ordered §1.2 `looksLikeRoot` probe over ALL built-in adapters
-    // (`registry.getAdapters()`), not the single hardcoded `akm` probe. This is
-    // BEHAVIOR-IDENTICAL for akm roots — `akm.looksLikeRoot` reproduces the former
-    // `.stash`-marker + immediate-stash-subdir (`hasStashDirs`) checks VERBATIM
-    // (akm-adapter.looksLikeRoot) and `akm` is in the probe set — so those two
-    // local checks are now subsumed and were removed. It additionally recognizes
-    // NON-akm bundle roots (okf root index / llm-wiki / `.claude` / …) at the top
-    // level, which the single-adapter probe missed but the old final `return root`
-    // fallback also happened to yield. The shallowest-BFS nested-layout discovery
-    // is unchanged.
+    // The top-level root check routes through the ordered `looksLikeRoot` probe
+    // over ALL built-in adapters (`registry.getAdapters()`). `akm.looksLikeRoot`
+    // covers the `.stash` marker and immediate stash subdirs; the other adapters
+    // recognize NON-akm bundle roots (okf root index / llm-wiki / `.claude` / …).
+    // Otherwise fall through to the shallowest-BFS nested-layout discovery.
     if (looksLikeAnyBundleRoot(root)) {
         return root;
     }
@@ -161,7 +155,8 @@ function countStashDirs(dirPath) {
  * (scripts/, skills/, etc.), so nested layouts like `project/my-stash/scripts/`
  * are discovered even without a `.stash` marker.
  *
- * Skips `root` itself since the caller already checked it via `hasStashDirs`.
+ * Skips `root` itself since the caller already checked it via
+ * `looksLikeAnyBundleRoot`.
  */
 const BFS_MAX_DEPTH = 5;
 function findShallowestStashRoot(root) {

package/dist/sources/snapshot-fetchers/website-ingest.js

@@ -90,7 +90,7 @@ export async function ensureWebsiteMirror(config, options) {
                 crawlTimeoutMs: resolveCrawlTimeoutMs(config.options?.crawlTimeoutMs),
                 resolveSecret: options?.resolveSecret,
                 // As-supplied, pre-normalization start URL (see crawlWebsite's
-                // `rawStartUrl` doc comment): threaded through purely for the C-02
+                // `rawStartUrl` doc comment): threaded through purely for the start-URL
                 // robots.txt check, which must see the trailing slash the user
                 // actually typed before `normalizeSiteUrl` strips it.
                 rawStartUrl: rawUrl,
@@ -461,7 +461,7 @@ async function resolveCrawlRobotsDecision(robots, normalizedUrl, rawUrl) {
 /**
  * Fail fast, before any page fetch, when the start URL itself is
  * off-limits per an actually-parsed robots.txt rule. A 5xx on robots.txt
- * itself no longer reaches this as `disallowAll`: `loadRobotsTxt` reports it the same way a
+ * itself does not reach this as `disallowAll`: `loadRobotsTxt` reports it the same way a
  * 4xx is reported (`ALLOW_ALL_RULES`), with its own warning, rather than
  * treating the site's own transient error as a Disallow. The check below
  * stays as defense-in-depth for a `RobotsPolicy` built over some other
@@ -594,7 +594,7 @@ async function crawlWebsite(startUrl, options) {
     }
     // Counts actual `fetchWebsitePage` invocations (regardless of outcome) so
     // Crawl-delay pacing skips the first fetch and never charges a delay slot
-    // to a URL that robots.txt skipped without ever being fetched (C-11).
+    // to a URL that robots.txt skipped without ever being fetched.
     let fetchAttempts = 0;
     // URLs pushed back because their Crawl-delay would not fit in the remaining
     // budget, and URLs that ran out of retries entirely. Reported at the end so
@@ -827,8 +827,7 @@ async function fetchWebsiteResponse(pageUrl, redirectCount = 0, options) {
                 // origin is the literal string "null", and re-resolving against it
                 // throws an unhandled TypeError that aborts the whole crawl. Refuse
                 // the hop outright instead of ever handing such a URL to the policy —
-                // `fetchWebsitePage` maps this to a graceful skip. Regression
-                // introduced by a67412c, which fell back to the raw `nextUrl` here.
+                // `fetchWebsitePage` maps this to a graceful skip.
                 warnVerbose("[akm] website crawl: skipping redirect to %s (not an http(s) URL)", nextUrl);
                 throw new RobotsDisallowedRedirectError(nextUrl);
             }
@@ -990,7 +989,7 @@ export function parseLlmsTxtLinks(text, baseUrl) {
  * content legitimately contains `##` headings too — so recovered page
  * boundaries can't be trusted, whereas per-page fetches through the existing
  * pipeline are cheap, bounded by this author-curated list, and produce
- * cleanly addressable assets. See the issue's "alternatives considered".
+ * cleanly addressable assets.
  */
 async function fetchLlmsManifest(start, robots, options) {
     const manifestUrl = new URL("/llms.txt", start.origin).toString();
@@ -1043,7 +1042,7 @@ function normalizeCrawlUrl(rawUrl) {
     }
 }
 /**
- * D-R6: `index.md`/`log.md` are OKF reserved structural filenames at every
+ * `index.md`/`log.md` are OKF reserved structural filenames at every
  * depth — no adapter indexes them, so a crawled page must never land on one.
  * Same remap convention as the content migration (`index.md` →
  * `index-content.md`). Segments are already lowercased by slugifySegment.

package/dist/storage/repositories/events-repository.js

@@ -147,3 +147,18 @@ export function purgeOldEvents(db, retentionDays = 90) {
     const changes = result.changes ?? 0;
     return typeof changes === "bigint" ? Number(changes) : changes;
 }
+/**
+ * Rewrite every `events.ref` naming `oldBundleId` to `newBundleId` (`akm bundle
+ * rename`). Improve reads an asset's feedback events by this ref, and
+ * `insertEventOnce` finds a journal event by it under the proposal or workflow
+ * ref the rename also rewrites, so an event left under the old prefix stops
+ * counting for the renamed bundle. A legacy unqualified ref is not matched by
+ * the `<old>//` prefix and is left alone; refs inside `metadata_json` are
+ * history and stay as written. Returns the number of rows rewritten.
+ */
+export function renameEventRefs(db, oldBundleId, newBundleId) {
+    const prefix = `${oldBundleId}//`;
+    return Number(db
+        .prepare(`UPDATE events SET ref = ? || substr(ref, length(?) + 1) WHERE substr(ref, 1, length(?)) = ?`)
+        .run(newBundleId, oldBundleId, prefix, prefix).changes);
+}

package/dist/storage/repositories/improve-ledger-repository.js

@@ -269,3 +269,16 @@ export function listImproveLedgerRows(db, stashDir, sources) {
         .all(stashDir, ...(sources && sources.length > 0 ? sources : []));
     return rows.map((row) => toRow(withHash ? row : { ...row, content_hash: null }));
 }
+/**
+ * Rewrite every `improve_ledger.ref` naming `oldBundleId` to `newBundleId`
+ * (`akm bundle rename`), so improve keeps the renamed bundle's attempts and
+ * back-off windows next to its feedback events. A row already at the new key
+ * can only be an orphan of a removed bundle by that name; the renamed bundle's
+ * row replaces it. Returns the number of rows rewritten.
+ */
+export function renameImproveLedgerRefs(db, oldBundleId, newBundleId) {
+    const prefix = `${oldBundleId}//`;
+    return Number(db
+        .prepare(`UPDATE OR REPLACE improve_ledger SET ref = ? || substr(ref, length(?) + 1) WHERE substr(ref, 1, length(?)) = ?`)
+        .run(newBundleId, oldBundleId, prefix, prefix).changes);
+}

package/dist/storage/repositories/index-entries-repository.js

@@ -18,6 +18,7 @@ import { getStateDbPath, withStateDb } from "../../core/state-db.js";
 import { warn } from "../../core/warn.js";
 import { getMarkdownFragmentContent, hasMarkdownFragmentContent, } from "../../indexer/passes/metadata.js";
 import { buildSearchText } from "../../indexer/search/search-fields.js";
+import { capEmbeddingText, DEFAULT_MAX_INPUT_TOKENS } from "../../llm/embedders/remote.js";
 import { sha256Hex } from "../../runtime.js";
 import { ENTRY_COLUMNS, rowToIndexedEntry } from "./index-entry-mapper.js";
 import { deleteFtsEntries, replaceFtsEntry } from "./index-fts-repository.js";
@@ -26,13 +27,21 @@ import { SQLITE_CHUNK_SIZE } from "./index-sql.js";
 import { deleteEntryVectors } from "./index-vec-repository.js";
 // ── Entry operations ────────────────────────────────────────────────────────
 /**
- * Hash of the text an entry's vector is embedded from (`buildSearchText`),
- * stored as `entries.embed_hash`: a change deletes the entry's vector, and the
- * next embedding pass re-embeds it.
+ * Hash of an entry's whole search text (`buildSearchText`), stored as
+ * `entries.embed_hash`: text that hashes the same cannot change what the
+ * entry's vector is embedded from.
  */
 function embedHash(entry) {
     return sha256Hex(buildSearchText(entry));
 }
+/**
+ * The text an entry's vector is embedded from: its search text cut to the
+ * embedding cap (`embedding.maxInputTokens`). Text past the cut never reaches
+ * the embedder, so a change there leaves the vector as it is.
+ */
+function embeddedInput(entry, maxInputTokens) {
+    return capEmbeddingText(buildSearchText(entry), maxInputTokens).text;
+}
 /**
  * Insert or update one canonical entry and all synchronously derived search
  * state. Returns the stable row id.
@@ -40,8 +49,13 @@ function embedHash(entry) {
  * The entries row, FTS projection, declared links, and stale-vector
  * invalidation commit as one SQLite transaction. Callers therefore cannot
  * publish an entry and forget a second FTS or links maintenance step.
+ *
+ * The entry's vector is deleted when the text it is embedded from changes,
+ * that is when the search text cut to `maxInputTokens` (the embedding cap,
+ * `embedding.maxInputTokens`) differs from the cut of the stored document. A
+ * change past the cut rewrites the keyword rows and keeps the vector.
  */
-export function upsertEntry(db, filePath, entry, provenance, contentHash) {
+export function upsertEntry(db, filePath, entry, provenance, contentHash, maxInputTokens = DEFAULT_MAX_INPUT_TOKENS) {
     // Hot path during indexing — cache prepared statements per database
     // connection so we don't pay the SQL parse/compile cost on every call.
     const stmts = getUpsertStmts(db);
@@ -54,10 +68,14 @@ export function upsertEntry(db, filePath, entry, provenance, contentHash) {
     // writer's current value.
     const apply = () => {
         const previous = stmts.findByItemRef.get(provenance.itemRef);
+        // Decided before the upsert overwrites the stored document it compares against.
+        const vectorStale = previous != null &&
+            previous.embed_hash !== hash &&
+            !storedDocumentEmbedsAs(stmts, previous, embeddedInput(entry, maxInputTokens), maxInputTokens);
         const result = stmts.upsert.get(provenance.itemRef, provenance.bundleId, provenance.componentId, provenance.conceptId, provenance.adapterId, entry.type, filePath, contentHash ?? null, JSON.stringify(entry), derivedFrom, hash);
         if (!result)
             throw new Error("upsertEntry: item_ref not found after upsert");
-        if (previous?.id === result.id && previous.embed_hash !== hash)
+        if (previous?.id === result.id && vectorStale)
             deleteEntryVectors(db, result.id);
         replaceFtsEntry(db, result.id, entry, hasMarkdownFragmentContent(entry) ? (getMarkdownFragmentContent(entry) ?? null) : undefined);
         replaceEntryLinks(db, result.id, entry, provenance);
@@ -69,6 +87,25 @@ export function upsertEntry(db, filePath, entry, provenance, contentHash) {
     // partial entries row through its outer transaction.
     return db.transaction(apply)();
 }
+/**
+ * Whether the stored document of `previous` embeds as `input`. The stored
+ * document is what the entry's current vector was embedded from, so the
+ * previous input is derived from it and needs no column of its own. The stored
+ * hash vouches for that: a document that cannot be read, or whose search text
+ * no longer hashes to `embed_hash`, has no known input, and its vector goes.
+ */
+function storedDocumentEmbedsAs(stmts, previous, input, maxInputTokens) {
+    const row = stmts.documentById.get(previous.id);
+    if (!row)
+        return false;
+    try {
+        const text = buildSearchText(JSON.parse(row.document_json));
+        return sha256Hex(text) === previous.embed_hash && capEmbeddingText(text, maxInputTokens).text === input;
+    }
+    catch {
+        return false;
+    }
+}
 const upsertStmtsByDb = new WeakMap();
 // item_ref is the sole durable conflict target. `content_hash` COALESCEs so
 // an upsert that omits it (`contentHash` is optional) cannot wipe a scan hash.
@@ -101,6 +138,7 @@ function getUpsertStmts(db) {
       RETURNING id
     `),
         findByItemRef: db.prepare("SELECT id, embed_hash FROM entries WHERE item_ref = ?"),
+        documentById: db.prepare("SELECT document_json FROM entries WHERE id = ?"),
     };
     upsertStmtsByDb.set(db, stmts);
     return stmts;

package/dist/storage/repositories/outcome-repository.js

@@ -118,3 +118,15 @@ export function countAssetOutcomeMissing(db) {
     const row = db.prepare(`SELECT COUNT(*) AS n FROM asset_outcome WHERE missing_since IS NOT NULL`).get();
     return row.n;
 }
+/**
+ * Rewrite every `asset_outcome.asset_ref` naming `oldBundleId` to
+ * `newBundleId` (`akm bundle rename`). Same prefix match and orphan-replacing
+ * collision policy as `renameAssetSalienceRefs`. Returns the number of rows
+ * rewritten.
+ */
+export function renameAssetOutcomeRefs(db, oldBundleId, newBundleId) {
+    const prefix = `${oldBundleId}//`;
+    return Number(db
+        .prepare(`UPDATE OR REPLACE asset_outcome SET asset_ref = ? || substr(asset_ref, length(?) + 1) WHERE substr(asset_ref, 1, length(?)) = ?`)
+        .run(newBundleId, oldBundleId, prefix, prefix).changes);
+}

package/dist/storage/repositories/proposals-repository.js

@@ -504,8 +504,8 @@ export function listStateProposalIdsByPrefix(db, stashDir, idPrefix) {
  * rename --dry-run`: the same two `WHERE` predicates, read-only.
  */
 export function countProposalsForBundleRename(db, oldBundleId) {
-    const prefix = `${escapeLikePattern(oldBundleId)}//`;
-    const refs = db.prepare(`SELECT COUNT(*) AS n FROM proposals WHERE ref LIKE ? ESCAPE '\\'`).get(`${prefix}%`).n;
+    const prefix = `${oldBundleId}//`;
+    const refs = db.prepare(`SELECT COUNT(*) AS n FROM proposals WHERE substr(ref, 1, length(?)) = ?`).get(prefix, prefix).n;
     const targets = db
         .prepare(`SELECT COUNT(*) AS n FROM proposals WHERE json_extract(metadata_json, '$.proposedTarget.source') = ?`)
         .get(oldBundleId).n;
@@ -522,11 +522,11 @@ export function countProposalsForBundleRename(db, oldBundleId) {
  * rewritten.
  */
 export function renameProposalsBundleRef(db, oldBundleId, newBundleId) {
-    const prefix = `${escapeLikePattern(oldBundleId)}//`;
+    const prefix = `${oldBundleId}//`;
     return db.transaction(() => {
         const refs = Number(db
-            .prepare(`UPDATE proposals SET ref = ? || substr(ref, length(?) + 1) WHERE ref LIKE ? ESCAPE '\\'`)
-            .run(newBundleId, oldBundleId, `${prefix}%`).changes);
+            .prepare(`UPDATE proposals SET ref = ? || substr(ref, length(?) + 1) WHERE substr(ref, 1, length(?)) = ?`)
+            .run(newBundleId, oldBundleId, prefix, prefix).changes);
         const targets = Number(db
             .prepare(`UPDATE proposals SET metadata_json = json_set(metadata_json, '$.proposedTarget.source', ?)
            WHERE json_extract(metadata_json, '$.proposedTarget.source') = ?`)

package/dist/storage/repositories/salience-repository.js

@@ -153,3 +153,16 @@ export function countAssetSalienceMissing(db) {
     const row = db.prepare(`SELECT COUNT(*) AS n FROM asset_salience WHERE missing_since IS NOT NULL`).get();
     return row.n;
 }
+/**
+ * Rewrite every `asset_salience.asset_ref` naming `oldBundleId` to
+ * `newBundleId` (`akm bundle rename`). A legacy unqualified ref is not matched
+ * by the `<old>//` prefix and is left alone. A row already at the new ref can only
+ * be an orphan of a removed bundle by that name; the renamed bundle's row
+ * replaces it. Returns the number of rows rewritten.
+ */
+export function renameAssetSalienceRefs(db, oldBundleId, newBundleId) {
+    const prefix = `${oldBundleId}//`;
+    return Number(db
+        .prepare(`UPDATE OR REPLACE asset_salience SET asset_ref = ? || substr(asset_ref, length(?) + 1) WHERE substr(asset_ref, 1, length(?)) = ?`)
+        .run(newBundleId, oldBundleId, prefix, prefix).changes);
+}

package/dist/storage/repositories/task-history-repository.js

@@ -11,7 +11,6 @@
  */
 import { isRecord } from "../../core/common.js";
 import { warnOnce } from "../../core/warn.js";
-import { escapeLikePattern } from "../like-pattern.js";
 function metadataError(message) {
     throw new Error(`invalid task_history metadata_json: ${message}`);
 }
@@ -208,19 +207,21 @@ export function queryCompletedTaskIntervals(db, since, until) {
  * rename --dry-run`: the same `WHERE` predicate, read-only.
  */
 export function countTaskHistoryTargetRefs(db, oldBundleId) {
-    const prefix = `${escapeLikePattern(oldBundleId)}//`;
-    return db.prepare(`SELECT COUNT(*) AS n FROM task_history WHERE target_ref LIKE ? ESCAPE '\\'`).get(`${prefix}%`).n;
+    const prefix = `${oldBundleId}//`;
+    return db
+        .prepare(`SELECT COUNT(*) AS n FROM task_history WHERE substr(target_ref, 1, length(?)) = ?`)
+        .get(prefix, prefix).n;
 }
 /**
  * Rewrite every `task_history.target_ref` naming `oldBundleId` (a workflow
  * run's fully-qualified `<bundle>//conceptId` target) to `newBundleId`
  * (`akm bundle rename`, D6). A row whose `target_ref` is unset or not
- * bundle-qualified is not matched by the `LIKE` prefix and is left alone.
+ * bundle-qualified is not matched by the `<old>//` prefix and is left alone.
  * Returns the number of rows rewritten.
  */
 export function renameTaskHistoryTargetRefs(db, oldBundleId, newBundleId) {
-    const prefix = `${escapeLikePattern(oldBundleId)}//`;
+    const prefix = `${oldBundleId}//`;
     return Number(db
-        .prepare(`UPDATE task_history SET target_ref = ? || substr(target_ref, length(?) + 1) WHERE target_ref LIKE ? ESCAPE '\\'`)
-        .run(newBundleId, oldBundleId, `${prefix}%`).changes);
+        .prepare(`UPDATE task_history SET target_ref = ? || substr(target_ref, length(?) + 1) WHERE substr(target_ref, 1, length(?)) = ?`)
+        .run(newBundleId, oldBundleId, prefix, prefix).changes);
 }