@lorekit/cli 1.11.0 → 1.12.0

package/README.md

@@ -51,9 +51,11 @@ without needing a marketplace:
 2. **MCP server** (`lorekit`) — the connection to your lessons, merged into the
    MCP config (preserving any other servers).
 3. **Hooks** — the *deterministic* layer: lessons injected on every
-   `SessionStart`, a nudge on tool failure (`PostToolUseFailure`), and a
-   retrospective nudge on `Stop`. These fire the shared `lorekit hook` engine
-   and are merged into `settings.json` (existing hooks preserved).
+   `SessionStart`, and on a tool failure (`PostToolUseFailure`) any lessons that
+   look **relevant to that failure** ("you've hit this before") plus a nudge to
+   record the fix, and a retrospective nudge on `Stop`. These fire the shared
+   `lorekit hook` engine and are merged into `settings.json` (existing hooks
+   preserved).
 
 It first asks **where** to install:
 
@@ -212,10 +214,10 @@ rather than crashing. `--endpoint` / `--token` / `--store` behave as in `list`.
 
 ### `lorekit tree` (alias `resolve`)
 
-Show the scopes the hooks actually **inject** — branch → repo → global, in
-precedence order (most-specific first) — as a resolution hierarchy, and mark for
-any key present at more than one scope which scope's lesson **wins** and which are
-**shadowed**:
+Show the scopes the hooks actually **inject** — project → branch → repo →
+global, in precedence order (most-specific first) — as a resolution hierarchy,
+and mark for any key present at more than one scope which scope's lesson **wins**
+and which are **shadowed**:
 
 ```bash
 lorekit tree                  # the injected hierarchy with ✓ winning / ↳ shadowed marks
@@ -224,12 +226,12 @@ lorekit tree --json           # per-entry { winning, shadowedBy } + a winners[]
 ```
 
 This mirrors the SessionStart hook's resolution **exactly**: it reads the scopes
-in `readOrder` (branch → repo → global) and keeps the first value seen per key, so
-a more-specific scope overrides a broader scope's same-key lesson. It answers
-"which lesson actually applies here, and what is being overridden?". Note that
-`project::` scope is **not** part of the injected set (the hooks never inject
-project lessons), so `tree` doesn't show it — browse those with `lorekit list`.
-Each store is resolved independently, in the same Offline / Remote split.
+in `readOrder` (project → branch → repo → global) and keeps the first value seen
+per key, so a more-specific scope overrides a broader scope's same-key lesson. It
+answers "which lesson actually applies here, and what is being overridden?". The
+`project::` scope **is** part of the injected set (project is the most-specific
+scope), so `tree`, the hooks, and every read command's `scopeList` now share one
+ordering. Each store is resolved independently, in the same Offline / Remote split.
 
 ### `lorekit lint`
 
@@ -273,8 +275,9 @@ range. Cross-**store** divergence is `diff`'s job; `dedupe` looks within a store
 The **shared hook engine** behind the Claude Code / Cursor / Codex plugins.
 It is not run by hand — the plugins wire it into their hook config. It reads
 the host framework's JSON on stdin and prints that host's injection format on
-stdout (lessons at session start; a nudge on failure or at end of turn),
-always exiting 0 so it can never block the host agent.
+stdout (lessons at session start; relevant lessons plus a write-nudge on a tool
+failure; a retrospective nudge at end of turn), always exiting 0 so it can never
+block the host agent.
 
 ```bash
 lorekit hook --adapter <claude|cursor|codex> --event <SessionStart|Stop|…>

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lorekit/cli",
-  "version": "1.11.0",
+  "version": "1.12.0",
   "description": "Install the LoreKit shared-memory skill and run health checks for the LoreKit MCP server.",
   "license": "MIT",
   "repository": {

package/src/core/lessons.mjs

@@ -3,23 +3,52 @@
 // Storage is reached through the resolved store (local | remote), never a
 // backend directly, so the same read path serves every mode.
 import { deriveScope } from '../scope.mjs';
+// The precedence merge and the literal substring matcher come from the
+// dependency-free `lessons-pure.mjs` — the SAME primitives `tree` and `search`
+// use, so the hook can't drift from them, and the hot path never pulls in the
+// `lessons-view.mjs` render/`util` stack.
+import { resolvePrecedence, matchesQuery } from '../lessons-pure.mjs';
 
 const MAX_LESSONS = 15;
+// Cap on lessons injected on a failure — a small, focused "you've seen this
+// before" set, never the whole applicable corpus.
+const MAX_RELEVANT = 3;
+// Cap on the terms distilled from a failure — bounds the relevance scan so a
+// huge error blob can never turn into an unbounded match loop.
+const MAX_TERMS = 12;
+// Terms shorter than this are dropped: too generic to make a match meaningful.
+const MIN_TERM_LEN = 4;
+// Cap on how much failure text is scanned for terms. A tool can dump a
+// multi-megabyte stderr/stdout; the salient error words are always near the
+// front, so we bound the input BEFORE lowercasing/splitting it — otherwise a
+// giant blob would materialise a giant token array (a CPU/memory spike) even
+// though the term COUNT is capped. Generous enough to never clip a real error.
+const MAX_SCAN_CHARS = 4096;
 
-// Read lessons narrow-to-broad through the store and merge; more specific
-// scope wins on key. Any per-scope failure is skipped (memory is best-effort).
+// Read lessons narrow-to-broad through the store and resolve cross-scope
+// precedence via the shared pure `resolvePrecedence` (the SAME first-seen /
+// more-specific-wins merge `tree` renders) — so the hook and `tree` provably
+// can't drift. Any per-scope failure is skipped (memory is best-effort).
 export async function fetchLessons(store, cwd) {
   const scope = deriveScope(cwd);
-  const byKey = new Map();
+  const groups = [];
   for (const s of scope.readOrder) {
     const res = await store.list({ scope: s, limit: 25 });
-    if (!res || !res.ok) continue;
-    const entries = Array.isArray(res.entries) ? res.entries : [];
-    for (const e of entries) {
-      if (e && e.key && !byKey.has(e.key)) byKey.set(e.key, { ...e, scope: s });
-    }
+    if (!res || !res.ok) continue; // best-effort: a failed scope contributes nothing
+    const entries = (Array.isArray(res.entries) ? res.entries : [])
+      .filter((e) => e && e.key)
+      .map((e) => ({ ...e, scope: s }));
+    groups.push({ scope: s, error: null, entries });
   }
-  return { scope, lessons: [...byKey.values()].slice(0, MAX_LESSONS) };
+  // First value per key wins (most-specific scope, since `readOrder` is
+  // narrow→broad) — exactly what the old inline `byKey` merge did, now via the
+  // one shared resolver. The winners, in group order, are the injected set.
+  const { groups: resolved } = resolvePrecedence({ groups });
+  const lessons = [];
+  for (const g of resolved) {
+    for (const e of g.entries) if (e.winning) lessons.push(e);
+  }
+  return { scope, lessons: lessons.slice(0, MAX_LESSONS) };
 }
 
 // Render lessons as a compact markdown block, or null when there are none.
@@ -37,6 +66,60 @@ export function formatLessons(lessons, scope) {
   return `${header}\n${body}`;
 }
 
+// Distil a small set of significant, lowercased search TERMS from a tool
+// failure — the tool name plus the salient words of its error text — used to
+// look up lessons that might already cover this failure. Pure and total: any
+// shape of `toolResponse` (object, string, null, malformed) is handled without
+// throwing. Terms are de-duplicated, stopword- and length-filtered so a match
+// stays meaningful, and the count is capped (`MAX_TERMS`) so a huge error blob
+// can't blow up the downstream scan.
+export function failureQuery(toolName, toolResponse) {
+  const text = `${toolName ? String(toolName) : ''} ${errorText(toolResponse)}`.slice(0, MAX_SCAN_CHARS);
+  const seen = new Set();
+  const terms = [];
+  for (const raw of text.toLowerCase().split(/[^a-z0-9]+/)) {
+    if (raw.length < MIN_TERM_LEN || STOPWORDS.has(raw) || seen.has(raw)) continue;
+    seen.add(raw);
+    terms.push(raw);
+    if (terms.length >= MAX_TERMS) break;
+  }
+  return terms;
+}
+
+// Lessons whose key OR value literally contains ANY of the failure `terms`
+// (case-insensitive, via the shared `matchesQuery` — never a regex), capped at
+// `cap`. Pure and best-effort: no terms or no lessons → empty (the caller then
+// falls back to the write-nudge alone). Preserves `lessons` order, so the
+// most-specific scope's relevant lesson surfaces first.
+export function relevantLessons(lessons, terms, cap = MAX_RELEVANT) {
+  if (!Array.isArray(lessons) || !lessons.length || !Array.isArray(terms) || !terms.length) {
+    return [];
+  }
+  const out = [];
+  for (const l of lessons) {
+    if (terms.some((t) => matchesQuery(l, t))) out.push(l);
+    if (out.length >= cap) break;
+  }
+  return out;
+}
+
+// Render the relevant-lessons block injected alongside the failure nudge, or
+// null when nothing matched. Framed as prior art, not a directive — same
+// "considerations, not rules" stance as `formatLessons`.
+export function formatRelevantLessons(lessons) {
+  if (!lessons || lessons.length === 0) return null;
+  const header =
+    `LoreKit — you've hit something like this before. ${lessons.length} related ` +
+    'lesson(s) (considerations, not rules; trust the current code if they conflict):';
+  const body = lessons
+    .map((l) => {
+      const first = String(l.value || '').split('\n')[0].slice(0, 300);
+      return `- (${l.scope}) ${l.key}: ${first}`;
+    })
+    .join('\n');
+  return `${header}\n${body}`;
+}
+
 // The retrospective nudge emitted at end-of-turn (one-shot per session).
 export function retrospectiveNudge(scope) {
   const writeScope = scope.repoScope || 'global';
@@ -58,3 +141,64 @@ export function failureNudge(toolName, scope) {
     `lorekit-memory (memory.write to ${writeScope}), so the next run avoids it.`
   );
 }
+
+// Pull the human-readable error text out of a tool_response of any shape. Total:
+// null/primitive/string/object all yield a (possibly empty) string, never a
+// throw — the relevance lookup is best-effort and must not break the host.
+function errorText(response) {
+  if (response == null) return '';
+  if (typeof response === 'string') return response;
+  if (typeof response !== 'object') return String(response);
+  const parts = [];
+  for (const f of TEXT_FIELDS) {
+    const v = response[f];
+    if (typeof v === 'string') parts.push(v);
+    else if (v && typeof v === 'object') parts.push(safeStringify(v));
+  }
+  return parts.length ? parts.join(' ') : safeStringify(response);
+}
+
+function safeStringify(v) {
+  try {
+    return JSON.stringify(v) || '';
+  } catch {
+    return '';
+  }
+}
+
+// The tool_response fields that commonly carry error text, across frameworks.
+const TEXT_FIELDS = [
+  'stderr',
+  'error',
+  'message',
+  'stdout',
+  'output',
+  'content',
+  'reason',
+  'detail',
+  'details',
+];
+
+// Generic words dropped from a failure query: too common to make a lesson match
+// meaningful (they'd match almost anything). Kept small and hand-picked.
+const STOPWORDS = new Set([
+  'error',
+  'errors',
+  'failed',
+  'failure',
+  'tool',
+  'call',
+  'code',
+  'exit',
+  'with',
+  'this',
+  'that',
+  'from',
+  'null',
+  'true',
+  'false',
+  'undefined',
+  'command',
+  'response',
+  'status',
+]);

package/src/hook.mjs

@@ -7,7 +7,15 @@ import { resolveProjectRoot } from './config.mjs';
 import { deriveScope } from './scope.mjs';
 import { loadControl } from './control.mjs';
 import { createStore } from './store/index.mjs';
-import { fetchLessons, formatLessons, retrospectiveNudge, failureNudge } from './core/lessons.mjs';
+import {
+  fetchLessons,
+  formatLessons,
+  retrospectiveNudge,
+  failureNudge,
+  failureQuery,
+  relevantLessons,
+  formatRelevantLessons,
+} from './core/lessons.mjs';
 import { isFailure } from './core/failure.mjs';
 import { firstTimeThisSession } from './core/state.mjs';
 import { recordFixture } from './core/record.mjs';
@@ -92,7 +100,22 @@ async function run(args) {
     const known = adapter.guaranteedFailure ? adapter.guaranteedFailure(event) : false;
     if (!known && !isFailure(parsed.toolName, parsed.toolResponse)) return 0;
     if (!firstTimeThisSession(parsed.sessionId, 'failure')) return 0;
-    emit(failureNudge(parsed.toolName, scope));
+    // Best-effort: surface any existing lessons that look relevant to THIS
+    // failure ("you've hit this before"), then the write-nudge. A missing /
+    // unusable store, or no match, silently falls back to the nudge alone.
+    let relevant = null;
+    try {
+      const store = createStore(control);
+      if (store) {
+        const { lessons } = await fetchLessons(store, root);
+        const terms = failureQuery(parsed.toolName, parsed.toolResponse);
+        relevant = formatRelevantLessons(relevantLessons(lessons, terms));
+      }
+    } catch {
+      relevant = null; // never let a lesson lookup break the failure nudge
+    }
+    const nudge = failureNudge(parsed.toolName, scope);
+    emit(relevant ? `${relevant}\n\n${nudge}` : nudge);
     return 0;
   }
 

package/src/lessons-pure.mjs

@@ -0,0 +1,65 @@
+// Pure, DEPENDENCY-FREE lesson primitives shared by the hook engine
+// (`core/lessons.mjs`, the hot path) and the read-command view layer
+// (`lessons-view.mjs`, which re-exports these). Kept in its own module so the
+// hook can import the exact precedence + match logic the read commands use —
+// one source of truth, no drift — WITHOUT dragging in the rendering/`util`
+// stack (`heading`/`log`/`c`, plus the lint/dedupe/diff cores) that the rest of
+// `lessons-view.mjs` carries. Zero imports on purpose.
+
+// ── cross-scope precedence resolution (the `tree` + hook merge core) ──────────
+
+// Given per-scope groups in RESOLUTION ORDER — most-specific first, exactly the
+// order `deriveScope().readOrder` produces — compute which scope's lesson WINS
+// each key and which are shadowed. This is the hook engine's merge: it iterates
+// the scopes narrow-to-broad and keeps the FIRST value seen per key (`if
+// (!winnerScopeByKey.has(key)) set`), so a more-specific scope shadows a broader
+// scope's same-key lesson. Returns the same groups with every entry tagged
+// `{ winning, shadowedBy }` (`shadowedBy` = the scope that won the key, or null
+// for a winner), plus the resolved `winners` list (`{ scope, key }`, one per
+// key) and winning/shadowed counts. A per-scope read error is passed through
+// untouched (empty entries), so one unreadable scope never derails resolution.
+// Pure — consumed by both `tree` (display) and `fetchLessons` (injection).
+export function resolvePrecedence({ groups = [] } = {}) {
+  const winnerScopeByKey = new Map(); // key → the scope that first claimed it
+  const outGroups = [];
+  let winningTotal = 0;
+  let shadowedTotal = 0;
+  for (const g of groups) {
+    if (g.error) {
+      outGroups.push({ scope: g.scope, error: g.error, entries: [] });
+      continue;
+    }
+    const entries = [];
+    for (const e of g.entries || []) {
+      const prior = winnerScopeByKey.get(e.key);
+      if (prior === undefined) {
+        winnerScopeByKey.set(e.key, g.scope);
+        entries.push({ ...e, winning: true, shadowedBy: null });
+        winningTotal += 1;
+      } else {
+        entries.push({ ...e, winning: false, shadowedBy: prior });
+        shadowedTotal += 1;
+      }
+    }
+    outGroups.push({ scope: g.scope, error: null, entries });
+  }
+  const winners = [...winnerScopeByKey.entries()].map(([key, scope]) => ({ scope, key }));
+  return { groups: outGroups, winners, winningTotal, shadowedTotal };
+}
+
+// ── literal, case-insensitive substring matching (the `search` + hook core) ───
+
+// Case-insensitive, LITERAL substring match of `query` against a normalized
+// entry's key OR value — the `search` command's matcher, and the hook's
+// failure-relevance matcher. Deliberately a plain `String.includes` (never
+// `new RegExp(query)`) so a query full of regex metacharacters like `a.*(b)`
+// matches those characters verbatim, never as a pattern. An empty query matches
+// everything (the command guards emptiness as a usage error before ever calling
+// this). Pure — trivially unit-testable.
+export function matchesQuery(entry, query) {
+  const needle = String(query == null ? '' : query).toLowerCase();
+  if (!needle) return true;
+  const key = String(entry?.key ?? '').toLowerCase();
+  const value = String(entry?.value ?? '').toLowerCase();
+  return key.includes(needle) || value.includes(needle);
+}

package/src/scope.mjs

@@ -44,7 +44,11 @@ export function deriveScope(cwd = process.cwd()) {
       : null;
   const projectScope = `project::${projectName}`;
 
-  const readOrder = [branchScope, repoScope, 'global'].filter(Boolean);
+  // Injection / resolution order, most-specific → broadest. Matches the read
+  // commands' `scopeList` exactly (`lessons-view.mjs`) — project is FIRST, so a
+  // project-scoped lesson wins and IS injected. De-duplicated (a project whose
+  // basename collides with a scope, or a repo with no branch, never repeats).
+  const readOrder = [...new Set([projectScope, branchScope, repoScope, 'global'].filter(Boolean))];
 
   return {
     ownerRepo,

package/src/tree.mjs

@@ -5,17 +5,18 @@
 //
 // Precedence is not an assumption — it mirrors the hook engine exactly. The
 // SessionStart hook (`core/lessons.mjs → fetchLessons`) reads the scopes in
-// `deriveScope().readOrder` (branch → repo → global, most-specific first) and
-// keeps the FIRST value seen per key, so a more-specific scope shadows a broader
-// scope's same-key lesson. `tree` resolves over that same `readOrder` set via
-// the pure `resolvePrecedence`, so it shows the same resolution order the agent
-// is injected with (the hook additionally caps the injected set at MAX_LESSONS;
-// `tree` is uncapped, so a large workspace may list more winners than the hook
-// injects).
+// `deriveScope().readOrder` (project → branch → repo → global, most-specific
+// first) and keeps the FIRST value seen per key, so a more-specific scope
+// shadows a broader scope's same-key lesson. `tree` resolves over that same
+// `readOrder` set via the pure `resolvePrecedence`, so it shows the same
+// resolution order the agent is injected with (the hook additionally caps the
+// injected set at MAX_LESSONS; `tree` is uncapped, so a large workspace may list
+// more winners than the hook injects).
 //
-// NOTE on scope coverage: `readOrder` is the injected set, and it deliberately
-// excludes `project::` — the hooks never inject project-scope lessons, so `tree`
-// doesn't either (browse those with `lorekit list`). Both stores are resolved
+// NOTE on scope coverage: `readOrder` is the injected set. As of the smart-hooks
+// PR it INCLUDES `project::` (project is the most-specific scope and now wins /
+// is injected), so `tree` shows it too — the ordering is now unified across the
+// hook, `tree`, and every read command's `scopeList`. Both stores are resolved
 // independently (precedence is per-store — the hook reads one resolved store),
 // in the same Offline / Remote split as `list`. Graceful, read-only, wrapped in
 // `traceCommand` by the bin.
@@ -24,7 +25,8 @@ import { resolveProjectRoot } from './config.mjs';
 import { deriveScope } from './scope.mjs';
 import { resolveDenies } from './control.mjs';
 import { resolveStores, remoteUnavailableReason } from './stores.mjs';
-import { gather, resolvePrecedence, preview, shortDate } from './lessons-view.mjs';
+import { resolvePrecedence } from './lessons-pure.mjs';
+import { gather, preview, shortDate } from './lessons-view.mjs';
 import { log, heading, status, c } from './util.mjs';
 
 export async function tree(args) {