@mmnto/mcp 1.13.0 → 1.14.0

package/dist/tools/search-knowledge.test.js

@@ -3,13 +3,28 @@ import { beforeEach, describe, expect, it, vi } from 'vitest';
 // Mocks — must be declared before imports that reference them
 // ---------------------------------------------------------------------------
 let capturedHandler;
-/** Tracks the mock store behaviour for each test. */
+/**
+ * Tracks the mock store behaviour for each test. `absoluteFilePath` is
+ * optional — when omitted the mock store fills it in from the fake
+ * `projectRoot` (mirroring real `LanceStore` which constructs it via
+ * `path.join(sourceContext.absolutePathRoot, filePath)`).
+ */
 let mockSearchResults = [];
 let mockHealthCheckResult = { healthy: true };
 let mockHealthCheckThrows = false;
 let mockSearchThrows = false;
 let mockSearchThrowsOnce = false;
 let mockReconnectCalled = false;
+/**
+ * mmnto/totem#1295 CR minor: number of upcoming `getContext()` calls
+ * that should throw before getContext starts returning normally.
+ * Decremented on each throw. The test for the one-shot flag fix sets
+ * this to a high enough number to cover every getContext call within
+ * a single handle() invocation, then expects subsequent calls to work.
+ */
+let mockGetContextFailuresRemaining = 0;
+let mockLinkedStores = new Map();
+let mockLinkedStoreInitErrors = new Map();
 vi.mock('@modelcontextprotocol/sdk/server/mcp.js', () => ({
     McpServer: class {
     },
@@ -20,35 +35,67 @@ vi.mock('@mmnto/totem', () => ({
     },
 }));
 vi.mock('../context.js', () => ({
-    getContext: vi.fn(async () => ({
-        projectRoot: '/fake/project',
-        config: {
-            totemDir: '.totem',
-            lanceDir: '.totem/.lance',
-            contextWarningThreshold: 50_000,
-            partitions: { core: ['packages/core/'] },
-        },
-        store: {
-            search: vi.fn(async () => {
-                if (mockSearchThrows) {
-                    throw new Error('LanceDB search failed');
-                }
-                if (mockSearchThrowsOnce) {
-                    mockSearchThrowsOnce = false;
-                    throw new Error('Stale handle error');
-                }
-                return mockSearchResults;
-            }),
-            healthCheck: vi.fn(async () => {
-                if (mockHealthCheckThrows) {
-                    throw new Error('Health check exploded');
-                }
-                return mockHealthCheckResult;
-            }),
-        },
-    })),
+    getContext: vi.fn(async () => {
+        if (mockGetContextFailuresRemaining > 0) {
+            mockGetContextFailuresRemaining -= 1;
+            throw new Error('Transient init failure');
+        }
+        return {
+            projectRoot: '/fake/project',
+            config: {
+                totemDir: '.totem',
+                lanceDir: '.totem/.lance',
+                contextWarningThreshold: 50_000,
+                partitions: { core: ['packages/core/'] },
+            },
+            store: {
+                search: vi.fn(async () => {
+                    if (mockSearchThrows) {
+                        throw new Error('LanceDB search failed');
+                    }
+                    if (mockSearchThrowsOnce) {
+                        mockSearchThrowsOnce = false;
+                        throw new Error('Stale handle error');
+                    }
+                    // Fill in absoluteFilePath from the fake projectRoot when the
+                    // test didn't set one explicitly — mirrors real LanceStore.
+                    return mockSearchResults.map((r) => ({
+                        ...r,
+                        absoluteFilePath: r.absoluteFilePath ?? `/fake/project/${r.filePath}`,
+                    }));
+                }),
+                reconnect: vi.fn(async () => { }),
+                healthCheck: vi.fn(async () => {
+                    if (mockHealthCheckThrows) {
+                        throw new Error('Health check exploded');
+                    }
+                    return mockHealthCheckResult;
+                }),
+            },
+            // mmnto/totem#1294 Phase 2: linked store fields. Tests that want
+            // to exercise the federation path assign to `mockLinkedStores`
+            // before calling the handler; most tests leave it as an empty Map.
+            linkedStores: mockLinkedStores,
+            linkedStoreInitErrors: mockLinkedStoreInitErrors,
+        };
+    }),
     reconnectStore: vi.fn(async () => {
         mockReconnectCalled = true;
+        // mmnto/totem#1295 CR minor: mirror the real reconnectStore — iterate
+        // linked stores and call each one's reconnect spy so tests can assert
+        // per-store reconnect actually fired.
+        for (const linkedStore of mockLinkedStores.values()) {
+            try {
+                await linkedStore.reconnect();
+            }
+            catch (err) {
+                // Best-effort, mirror the real reconnectStore behavior.
+                // Suppression is intentional — discard the error explicitly so
+                // the bare-catch lint rule doesn't fire and unexpected mock
+                // breakage stays grep-able.
+                void err;
+            }
+        }
     }),
 }));
 vi.mock('../xml-format.js', () => ({
@@ -92,6 +139,9 @@ describe('search_knowledge', () => {
         mockSearchThrows = false;
         mockSearchThrowsOnce = false;
         mockReconnectCalled = false;
+        mockGetContextFailuresRemaining = 0;
+        mockLinkedStores = new Map();
+        mockLinkedStoreInitErrors = new Map();
         // Reset modules to clear the firstHealthCheckDone flag
         vi.resetModules();
         // Re-apply all mocks after module reset
@@ -105,35 +155,62 @@ describe('search_knowledge', () => {
             },
         }));
         vi.doMock('../context.js', () => ({
-            getContext: vi.fn(async () => ({
-                projectRoot: '/fake/project',
-                config: {
-                    totemDir: '.totem',
-                    lanceDir: '.totem/.lance',
-                    contextWarningThreshold: 50_000,
-                    partitions: { core: ['packages/core/'] },
-                },
-                store: {
-                    search: vi.fn(async () => {
-                        if (mockSearchThrows) {
-                            throw new Error('LanceDB search failed');
-                        }
-                        if (mockSearchThrowsOnce) {
-                            mockSearchThrowsOnce = false;
-                            throw new Error('Stale handle error');
-                        }
-                        return mockSearchResults;
-                    }),
-                    healthCheck: vi.fn(async () => {
-                        if (mockHealthCheckThrows) {
-                            throw new Error('Health check exploded');
-                        }
-                        return mockHealthCheckResult;
-                    }),
-                },
-            })),
+            getContext: vi.fn(async () => {
+                if (mockGetContextFailuresRemaining > 0) {
+                    mockGetContextFailuresRemaining -= 1;
+                    throw new Error('Transient init failure');
+                }
+                return {
+                    projectRoot: '/fake/project',
+                    config: {
+                        totemDir: '.totem',
+                        lanceDir: '.totem/.lance',
+                        contextWarningThreshold: 50_000,
+                        partitions: { core: ['packages/core/'] },
+                    },
+                    store: {
+                        search: vi.fn(async () => {
+                            if (mockSearchThrows) {
+                                throw new Error('LanceDB search failed');
+                            }
+                            if (mockSearchThrowsOnce) {
+                                mockSearchThrowsOnce = false;
+                                throw new Error('Stale handle error');
+                            }
+                            // Fill in absoluteFilePath from the fake projectRoot when
+                            // the test didn't set one explicitly — mirrors LanceStore.
+                            return mockSearchResults.map((r) => ({
+                                ...r,
+                                absoluteFilePath: r.absoluteFilePath ?? `/fake/project/${r.filePath}`,
+                            }));
+                        }),
+                        reconnect: vi.fn(async () => { }),
+                        healthCheck: vi.fn(async () => {
+                            if (mockHealthCheckThrows) {
+                                throw new Error('Health check exploded');
+                            }
+                            return mockHealthCheckResult;
+                        }),
+                    },
+                    // mmnto/totem#1294 Phase 2: linked store fields
+                    linkedStores: mockLinkedStores,
+                    linkedStoreInitErrors: mockLinkedStoreInitErrors,
+                };
+            }),
             reconnectStore: vi.fn(async () => {
                 mockReconnectCalled = true;
+                // mmnto/totem#1295 CR minor: mirror reconnectStore — iterate
+                // linked stores so tests can assert per-store reconnect fired.
+                for (const linkedStore of mockLinkedStores.values()) {
+                    try {
+                        await linkedStore.reconnect();
+                    }
+                    catch (err) {
+                        // Best-effort, mirror the real reconnectStore behavior.
+                        // Suppression is intentional — discard the error explicitly.
+                        void err;
+                    }
+                }
             }),
         }));
         vi.doMock('../xml-format.js', () => ({
@@ -214,6 +291,83 @@ describe('search_knowledge', () => {
         expect(result.isError).toBe(true);
         expect(result.content[0].text).toContain('DIMENSION MISMATCH');
     });
+    it('dimension mismatch warning persists across queries until the index is fixed (CR MAJOR)', async () => {
+        // mmnto/totem#1295 CR MAJOR: the round-7 fix moved
+        // `firstHealthCheckDone = true` to after the healthCheck() await, but
+        // still consumed the flag on UNHEALTHY results. That meant a persistent
+        // dimension mismatch showed the actionable "rm -rf .lancedb &&
+        // totem sync --full" guidance ONCE, then the next query skipped the
+        // gate and fell back to the cryptic LanceDB "vector dimension mismatch"
+        // error — exactly what this gate exists to prevent.
+        //
+        // The fix: don't consume the flag on dimension mismatch. The friendly
+        // diagnostic fires on EVERY query until the state is actually fixed.
+        mockHealthCheckResult = {
+            healthy: false,
+            dimensionMatch: false,
+            storedDimensions: 768,
+            expectedDimensions: 1536,
+        };
+        const first = (await handle({ query: 'anything' }));
+        expect(first.isError).toBe(true);
+        expect(first.content[0].text).toContain('DIMENSION MISMATCH');
+        expect(first.content[0].text).toContain('rm -rf .lancedb');
+        // Second query: WITHOUT the fix, this would return a regular search
+        // (the flag would have been consumed, runFirstQueryHealthCheck would
+        // return null, the outer catch would see a cryptic LanceDB error from
+        // performSearch OR worse, fall into the success path silently).
+        // WITH the fix, the dimension mismatch warning fires again and the
+        // search is blocked again.
+        const second = (await handle({ query: 'anything' }));
+        expect(second.isError).toBe(true);
+        expect(second.content[0].text).toContain('DIMENSION MISMATCH');
+        expect(second.content[0].text).toContain('rm -rf .lancedb');
+        // Simulate the user fixing the index: healthCheck now returns healthy.
+        // After the fix-then-retry, the warning should stop firing (normal
+        // one-shot semantics resume).
+        mockHealthCheckResult = { healthy: true };
+        mockSearchResults = [
+            {
+                label: 'Post-fix hit',
+                type: 'code',
+                filePath: 'src/foo.ts',
+                score: 0.8,
+                content: 'recovered',
+            },
+        ];
+        const third = (await handle({ query: 'anything' }));
+        expect(third.isError).toBeUndefined();
+        expect(third.content[0].text).not.toContain('DIMENSION MISMATCH');
+        expect(third.content[0].text).toContain('Post-fix hit');
+    });
+    it('non-fatal health warnings stay one-shot even after dimension-mismatch fix', async () => {
+        // Sanity check for the companion rule: non-fatal health issues (stale
+        // rows, missing partitions) still consume the flag after one warning.
+        // Only dimension mismatch is special-cased to persist.
+        mockHealthCheckResult = {
+            healthy: false,
+            dimensionMatch: true, // dim is fine — other issues
+            issues: ['Partition "core" has no rows'],
+        };
+        mockSearchResults = [
+            {
+                label: 'Hit',
+                type: 'code',
+                filePath: 'src/foo.ts',
+                score: 0.7,
+                content: 'content',
+            },
+        ];
+        const first = (await handle({ query: 'test' }));
+        expect(first.isError).toBeUndefined();
+        expect(first.content[0].text).toContain('Index health issues detected');
+        expect(first.content[0].text).toContain('Hit');
+        // Second query: the non-fatal warning should NOT repeat (one-shot)
+        const second = (await handle({ query: 'test' }));
+        expect(second.isError).toBeUndefined();
+        expect(second.content[0].text).not.toContain('Index health issues detected');
+        expect(second.content[0].text).toContain('Hit');
+    });
     it('does not block search when health check itself fails', async () => {
         mockHealthCheckThrows = true;
         mockSearchResults = [
@@ -284,7 +438,9 @@ describe('search_knowledge', () => {
         expect(text).toContain('### 1. First result (lesson)');
         expect(text).toContain('### 2. Second result (spec)');
         expect(text).toContain('### 3. Third result (code)');
-        expect(text).toContain('**File:** lessons/first.md | **Score:** 0.990');
+        // mmnto/totem#1295 CR MAJOR: File line uses absolute path (mock fills
+        // absoluteFilePath from /fake/project + filePath when not set explicitly)
+        expect(text).toContain('**File:** /fake/project/lessons/first.md | **Score:** 0.990');
         expect(text).toContain('---');
     });
     it('passes type_filter and max_results to store.search', async () => {
@@ -313,5 +469,694 @@ describe('search_knowledge', () => {
         expect(result.isError).toBeUndefined();
         expect(result.content[0].text).toContain('No results found');
     });
+    // ─── Cross-Repo Context Mesh (mmnto/totem#1294 Phase 2) ───────
+    describe('federated search (linkedIndexes)', () => {
+        function makeLinkedStore(results) {
+            return {
+                search: vi.fn(async () => results),
+                // mmnto/totem#1295 CR minor: every linked-store mock now exposes
+                // a reconnect spy so tests can assert per-store reconnect actually
+                // fired during the reconnect path.
+                reconnect: vi.fn(async () => { }),
+            };
+        }
+        it('federates across primary + linked stores via fair RRF rank merge', async () => {
+            // mmnto/totem#1295 GCA CRITICAL: federation must merge by
+            // rank-within-store (RRF), not raw score, because LanceStore returns
+            // scores in incompatible scales (hybrid RRF ~0.03 vs vector-only
+            // ~0.85). This test sets up scores where the OLD raw-score sort
+            // would give the wrong answer:
+            //
+            //   Primary: P1 (0.95), P2 (0.94)         — both very high (vector scale)
+            //   Linked:  S1 (0.04), S2 (0.03)         — both very low  (RRF scale)
+            //
+            //   Old raw-score sort: P1, P2, S1, S2     ← linked store starved
+            //   New RRF sort:       P1, S1, P2, S2     ← rank 0s interleave fairly
+            //
+            // The "rank 0 from each store should beat rank 1 from any store"
+            // property is the architectural guarantee RRF provides.
+            mockSearchResults = [
+                {
+                    label: 'P1 primary top',
+                    type: 'code',
+                    filePath: 'src/foo.ts',
+                    score: 0.95,
+                    content: 'primary top content',
+                },
+                {
+                    label: 'P2 primary second',
+                    type: 'code',
+                    filePath: 'src/bar.ts',
+                    score: 0.94,
+                    content: 'primary second content',
+                },
+            ];
+            mockLinkedStores.set('strategy', makeLinkedStore([
+                {
+                    label: 'S1 strategy top',
+                    type: 'spec',
+                    filePath: 'adr/adr-001.md',
+                    absoluteFilePath: '/abs/totem-strategy/adr/adr-001.md',
+                    sourceRepo: 'strategy',
+                    score: 0.04, // raw-score sort would put this BEHIND both P1 and P2
+                    content: 'strategy top content',
+                },
+                {
+                    label: 'S2 strategy second',
+                    type: 'spec',
+                    filePath: 'adr/adr-002.md',
+                    absoluteFilePath: '/abs/totem-strategy/adr/adr-002.md',
+                    sourceRepo: 'strategy',
+                    score: 0.03,
+                    content: 'strategy second content',
+                },
+            ]));
+            const result = (await handle({ query: 'architecture', max_results: 10 }));
+            expect(result.isError).toBeUndefined();
+            const text = result.content[0].text;
+            // All four results present
+            expect(text).toContain('P1 primary top');
+            expect(text).toContain('S1 strategy top');
+            expect(text).toContain('P2 primary second');
+            expect(text).toContain('S2 strategy second');
+            // RRF interleaving: S1 (rank 0 in its store) must appear BEFORE
+            // P2 (rank 1 in primary), even though P2 has 23x the raw score.
+            // This is the architectural fix — without RRF, S1 would be last.
+            const p1Idx = text.indexOf('P1 primary top');
+            const s1Idx = text.indexOf('S1 strategy top');
+            const p2Idx = text.indexOf('P2 primary second');
+            const s2Idx = text.indexOf('S2 strategy second');
+            // Stable sort within ties: bucket order is primary-first then linked,
+            // so among rank-0 results P1 comes before S1; among rank-1 results
+            // P2 comes before S2.
+            expect(p1Idx).toBeLessThan(s1Idx);
+            expect(s1Idx).toBeLessThan(p2Idx); // ← THE KEY ASSERTION
+            expect(p2Idx).toBeLessThan(s2Idx);
+        });
+        it('federation displays normalized RRF scores, not raw store scores', async () => {
+            // mmnto/totem#1295 GCA CRITICAL: the visible `score` field is
+            // overwritten with the RRF score during federation so the displayed
+            // order matches the displayed numbers. Otherwise users would see
+            // results sorted by an invisible secondary key, which is confusing.
+            mockSearchResults = [
+                {
+                    label: 'P1',
+                    type: 'code',
+                    filePath: 'src/foo.ts',
+                    score: 0.99, // raw vector-distance scale
+                    content: 'primary',
+                },
+            ];
+            mockLinkedStores.set('strategy', makeLinkedStore([
+                {
+                    label: 'S1',
+                    type: 'spec',
+                    filePath: 'adr/adr-001.md',
+                    absoluteFilePath: '/abs/strategy/adr/adr-001.md',
+                    sourceRepo: 'strategy',
+                    score: 0.04, // raw RRF scale
+                    content: 'strategy',
+                },
+            ]));
+            const result = (await handle({ query: 'test' }));
+            const text = result.content[0].text;
+            // The original raw scores (0.990 / 0.040) must NOT appear — they
+            // would mislead the user about cross-store comparability.
+            expect(text).not.toContain('0.990');
+            expect(text).not.toContain('0.040');
+            // The RRF score for rank 0 with k=60 is 1/61 ≈ 0.016
+            expect(text).toContain('0.016');
+        });
+        it('boundary matching a linked store name routes ONLY to that store', async () => {
+            mockSearchResults = [
+                {
+                    label: 'Primary hit',
+                    type: 'code',
+                    filePath: 'src/foo.ts',
+                    score: 0.99,
+                    content: 'primary content',
+                },
+            ];
+            const strategyStore = makeLinkedStore([
+                {
+                    label: 'Strategy ADR',
+                    type: 'spec',
+                    filePath: 'adr/adr-001.md',
+                    absoluteFilePath: '/abs/totem-strategy/adr/adr-001.md',
+                    sourceRepo: 'strategy',
+                    score: 0.5,
+                    content: 'adr content',
+                },
+            ]);
+            mockLinkedStores.set('strategy', strategyStore);
+            const result = (await handle({
+                query: 'architecture',
+                boundary: 'strategy',
+            }));
+            expect(result.isError).toBeUndefined();
+            expect(strategyStore.search).toHaveBeenCalled();
+            // Only strategy results should appear — primary is not queried
+            expect(result.content[0].text).toContain('Strategy ADR');
+            expect(result.content[0].text).not.toContain('Primary hit');
+        });
+        it('boundary matching a partition name still routes to primary (partitions win over links)', async () => {
+            mockSearchResults = [
+                {
+                    label: 'Core file',
+                    type: 'code',
+                    filePath: 'packages/core/src/foo.ts',
+                    score: 0.9,
+                    content: 'core content',
+                },
+            ];
+            const linkedStore = makeLinkedStore([
+                {
+                    label: 'Should not appear',
+                    type: 'spec',
+                    filePath: 'other.md',
+                    absoluteFilePath: '/abs/other/other.md',
+                    sourceRepo: 'core',
+                    score: 0.99,
+                    content: 'linked content',
+                },
+            ]);
+            // Collision: link name "core" matches a partition name
+            mockLinkedStores.set('core', linkedStore);
+            const result = (await handle({
+                query: 'anything',
+                boundary: 'core', // partition "core" wins
+            }));
+            expect(result.isError).toBeUndefined();
+            expect(result.content[0].text).toContain('Core file');
+            expect(result.content[0].text).not.toContain('Should not appear');
+            // Linked store was never queried
+            expect(linkedStore.search).not.toHaveBeenCalled();
+        });
+        it('unknown boundary falls back to raw prefix on primary only', async () => {
+            mockSearchResults = [];
+            const linkedStore = makeLinkedStore([
+                {
+                    label: 'Should not appear',
+                    type: 'spec',
+                    filePath: 'other.md',
+                    absoluteFilePath: '/abs/other/other.md',
+                    sourceRepo: 'strategy',
+                    score: 0.99,
+                    content: 'linked content',
+                },
+            ]);
+            mockLinkedStores.set('strategy', linkedStore);
+            const result = (await handle({
+                query: 'test',
+                boundary: 'some/random/prefix/',
+            }));
+            expect(result.isError).toBeUndefined();
+            expect(linkedStore.search).not.toHaveBeenCalled();
+        });
+        it('linked store runtime failure surfaces per-query warning (non-blocking)', async () => {
+            // mmnto/totem#1295 GCA/CR fix: previously this path silently dropped
+            // linked failures with no user-visible signal (Tenet 4 violation).
+            // The new per-query runtime-warning architecture surfaces the failure
+            // inline on every query it occurred on, without mutating global state
+            // (so transient issues don't cause permanent session drift).
+            mockSearchResults = [
+                {
+                    label: 'Primary hit',
+                    type: 'code',
+                    filePath: 'src/foo.ts',
+                    score: 0.8,
+                    content: 'primary content',
+                },
+            ];
+            // Use mockRejectedValue to avoid an inline `throw new Error` that
+            // trips the over-broad "error normalization" and "[Totem Error]
+            // prefix" lint rules on test fixtures (tracked for refinement in
+            // mmnto/totem#1286). The rejected-Promise shape is semantically
+            // equivalent for the federation-failure path under test.
+            const brokenLink = {
+                search: vi.fn().mockRejectedValue(new Error('Connection refused')),
+                reconnect: vi.fn().mockRejectedValue(new Error('Reconnect also broken')),
+            };
+            mockLinkedStores.set('broken', brokenLink);
+            const result = (await handle({ query: 'test' }));
+            // Federation is non-blocking — primary results still land
+            expect(result.isError).toBeUndefined();
+            expect(brokenLink.search).toHaveBeenCalled();
+            expect(result.content[0].text).toContain('Primary hit');
+            // Runtime failure is surfaced as a per-query system warning so the
+            // agent sees the drift in-context (Tenet 4: Fail Loud)
+            expect(result.content[0].text).toContain('[SYSTEM WARNING]');
+            expect(result.content[0].text).toContain('Federated search');
+            expect(result.content[0].text).toContain('broken');
+            expect(result.content[0].text).toContain('Connection refused');
+        });
+        it('linked store literally named "primary" does not collide with actual primary failure slot (CR MAJOR)', async () => {
+            // mmnto/totem#1295 CR MAJOR: `deriveLinkName` strips leading dots
+            // from the basename, so a linked repo at `.primary/` would derive
+            // to the link name `'primary'`. The earlier implementation stored
+            // primary store failures under `runtimeFailures.set('primary', ...)`,
+            // which would have either overwritten or been overwritten by the
+            // legitimate linked store named 'primary'.
+            //
+            // The fix splits primary into a dedicated `failures.primary` slot
+            // (string | null), keeping the linked-store map free of reserved
+            // keys. This test exercises the collision scenario:
+            //
+            //   1. The actual primary store throws (so failures.primary is set).
+            //   2. A linked store literally named 'primary' returns results
+            //      successfully — its results must NOT be misreported as the
+            //      primary store, AND the warning copy must distinguish them.
+            mockSearchThrows = true;
+            const linkedNamedPrimary = makeLinkedStore([
+                {
+                    label: 'Linked-primary hit',
+                    type: 'spec',
+                    filePath: 'adr/adr-001.md',
+                    absoluteFilePath: '/abs/.primary/adr/adr-001.md',
+                    sourceRepo: 'primary',
+                    score: 0.8,
+                    content: 'linked content from a repo named primary',
+                },
+            ]);
+            mockLinkedStores.set('primary', linkedNamedPrimary);
+            const result = (await handle({ query: 'test' }));
+            // Federation succeeds with the linked-named-primary results
+            expect(result.isError).toBeUndefined();
+            expect(result.content[0].text).toContain('Linked-primary hit');
+            // The warning surfaces the actual primary store failure WITHOUT
+            // collision. With the bug, the linked store's success would have
+            // either overwritten the primary failure (no warning) or the
+            // primary failure would have overwritten the linked store's entry.
+            const text = result.content[0].text;
+            expect(text).toContain('[SYSTEM WARNING]');
+            // The warning must reference primary store failure
+            expect(text).toContain('primary store');
+            // The warning must NOT reference any linked store failure (the
+            // 'primary' linked store succeeded)
+            expect(text).not.toContain('1 linked index(es) failed');
+            // Linked store named 'primary' got its tag prefix
+            expect(text).toContain('[primary] Linked-primary hit');
+        });
+        it('entire federation down returns isError (CR MAJOR — do not mask outage as "no results")', async () => {
+            // mmnto/totem#1295 CR MAJOR catch: when primary AND every linked
+            // store fail, results.length === 0 but the previous code fell
+            // through to a success-shaped "No results found" body with the
+            // warning prepended. The agent reads that as "no relevant knowledge
+            // in the index" when actually the entire search plane is broken.
+            // Tenet 4 violation (silent degradation).
+            //
+            // The fix: detect the federated case where `failures.primary !== null`
+            // AND `failures.linked.size === linkedStores.size` AND results are
+            // empty, and return isError: true with the runtime warning as the
+            // error text so the agent sees the breakdown of what's down.
+            mockSearchThrows = true;
+            const brokenLinkA = {
+                search: vi.fn().mockRejectedValue(new Error('Linked A down')),
+                reconnect: vi.fn().mockRejectedValue(new Error('A reconnect failed')),
+            };
+            const brokenLinkB = {
+                search: vi.fn().mockRejectedValue(new Error('Linked B down')),
+                reconnect: vi.fn().mockRejectedValue(new Error('B reconnect failed')),
+            };
+            mockLinkedStores.set('strategy', brokenLinkA);
+            mockLinkedStores.set('playground', brokenLinkB);
+            const result = (await handle({ query: 'test' }));
+            // Must be isError — not a silent "no results found" body
+            expect(result.isError).toBe(true);
+            const text = result.content[0].text;
+            // The runtime warning (breakdown of what failed) is the error text
+            expect(text).toContain('[SYSTEM WARNING]');
+            expect(text).toContain('primary store');
+            expect(text).toContain('2 linked index(es) failed');
+            // Each broken store is named in the detail lines
+            expect(text).toContain('strategy');
+            expect(text).toContain('playground');
+            // Critically: the "No results found." body must NOT appear
+            expect(text).not.toContain('No results found');
+        });
+        it('partial-failure federation (some linked OK, some broken) still returns success-shape "no results"', async () => {
+            // mmnto/totem#1295 CR MAJOR follow-up: verify the all-failed check
+            // is narrow enough. When at least ONE store successfully returned
+            // zero results, the zero-ness is authoritative and the response
+            // should be a normal "No results found" body with the warning
+            // prepended — NOT isError. The agent can see the warning and
+            // decide whether to retry or accept the zero-result answer.
+            const brokenLink = {
+                search: vi.fn().mockRejectedValue(new Error('Linked down')),
+                reconnect: vi.fn().mockRejectedValue(new Error('Reconnect failed')),
+            };
+            const healthyLink = makeLinkedStore([]); // healthy, empty results
+            mockLinkedStores.set('strategy', brokenLink);
+            mockLinkedStores.set('playground', healthyLink);
+            mockSearchResults = []; // primary also empty but healthy
+            const result = (await handle({ query: 'test' }));
+            // Not isError — at least one store answered authoritatively
+            expect(result.isError).toBeUndefined();
+            const text = result.content[0].text;
+            // Runtime warning surfaces the partial failure
+            expect(text).toContain('[SYSTEM WARNING]');
+            expect(text).toContain('strategy');
+            // "No results found" body IS present because primary + healthy linked
+            // both returned zero
+            expect(text).toContain('No results found');
+        });
+        it('primary store failure does not block linked-store results (GCA HIGH)', async () => {
+            // mmnto/totem#1295 GCA HIGH catch: previously the primary store
+            // search bubbled out of `Promise.all` and killed the entire
+            // federation, even when linked stores were healthy. Now primary
+            // failures are caught inside `federatedSearch` and routed through
+            // the same per-query runtime-warning path as linked failures —
+            // populated under the reserved `'primary'` key.
+            mockSearchThrows = true;
+            mockLinkedStores.set('strategy', makeLinkedStore([
+                {
+                    label: 'Linked still works',
+                    type: 'spec',
+                    filePath: 'adr/adr-001.md',
+                    absoluteFilePath: '/abs/strategy/adr/adr-001.md',
+                    sourceRepo: 'strategy',
+                    score: 0.8,
+                    content: 'strategy content',
+                },
+            ]));
+            const result = (await handle({ query: 'test' }));
+            // Federation is non-blocking — linked results still land
+            expect(result.isError).toBeUndefined();
+            expect(result.content[0].text).toContain('Linked still works');
+            // Primary failure is surfaced as a per-query runtime warning under
+            // the reserved 'primary' key
+            expect(result.content[0].text).toContain('[SYSTEM WARNING]');
+            expect(result.content[0].text).toContain('primary');
+        });
+        it('per-query runtime warnings do not mutate global init errors (transient failures stay transient)', async () => {
+            // mmnto/totem#1295 CR/GCA architectural fix: the original Phase 2
+            // implementation mutated linkedStoreInitErrors on runtime failure,
+            // which permanently degraded the session for transient issues like
+            // a file lock during a parallel `totem sync`. The fix: runtime
+            // failures populate a per-query Map that is discarded after the
+            // response is built. A subsequent successful query sees zero warning.
+            const flaky = {
+                search: vi
+                    .fn()
+                    .mockRejectedValueOnce(new Error('Transient lock'))
+                    .mockResolvedValueOnce([
+                    {
+                        label: 'Recovered',
+                        type: 'spec',
+                        filePath: 'adr/adr-001.md',
+                        absoluteFilePath: '/abs/flaky/adr/adr-001.md',
+                        sourceRepo: 'flaky',
+                        score: 0.8,
+                        content: 'recovered content',
+                    },
+                ]),
+                // First-query reconnect attempt fails (so the runtime warning fires);
+                // second query bypasses the catch path entirely.
+                reconnect: vi.fn().mockRejectedValue(new Error('Reconnect lock')),
+            };
+            mockLinkedStores.set('flaky', flaky);
+            mockSearchResults = [
+                {
+                    label: 'Primary',
+                    type: 'code',
+                    filePath: 'src/foo.ts',
+                    score: 0.7,
+                    content: 'primary',
+                },
+            ];
+            // First query: runtime failure → warning present
+            const first = (await handle({ query: 'test' }));
+            expect(first.content[0].text).toContain('[SYSTEM WARNING]');
+            expect(first.content[0].text).toContain('flaky');
+            // Second query: linked store recovers → NO warning carried over
+            const second = (await handle({ query: 'test' }));
+            expect(second.content[0].text).not.toContain('[SYSTEM WARNING]');
+            expect(second.content[0].text).toContain('Recovered');
+        });
+        it('targeted Case 2 returns isError when the linked store fails completely (GCA HIGH)', async () => {
+            // mmnto/totem#1295 GCA HIGH: when the user explicitly names a
+            // boundary and the targeted linked store fails (initial AND
+            // reconnect+retry), the response must signal isError: true rather
+            // than falling through to a "no results found" body. The agent
+            // should not misinterpret a real outage as an absence of relevant
+            // knowledge. This is symmetric with Case 3 (boundary matches a
+            // failed-init linked store).
+            // The search throws on EVERY call (initial + post-reconnect retry).
+            // Reconnect succeeds so the retry actually fires — proves we go all
+            // the way through the retry path before returning isError.
+            const targeted = {
+                search: vi.fn().mockRejectedValue(new Error('Connection refused')),
+                reconnect: vi.fn(async () => { }),
+            };
+            mockLinkedStores.set('strategy', targeted);
+            // Primary results are deliberately set so we can prove they are
+            // NOT leaked into the response (a regression would show them).
+            mockSearchResults = [
+                {
+                    label: 'Bogus primary hit',
+                    type: 'code',
+                    filePath: 'src/strategy.ts',
+                    score: 0.9,
+                    content: 'unrelated local',
+                },
+            ];
+            const result = (await handle({ query: 'test', boundary: 'strategy' }));
+            expect(result.isError).toBe(true);
+            // Both reconnect and the second search were attempted
+            expect(targeted.reconnect).toHaveBeenCalledOnce();
+            expect(targeted.search).toHaveBeenCalledTimes(2);
+            // Error message names the targeted boundary and includes both errors
+            const text = result.content[0].text;
+            expect(text).toContain('strategy');
+            expect(text).toContain('Connection refused');
+            // Critically: NO bogus primary results leaked into the response
+            expect(text).not.toContain('Bogus primary hit');
+        });
+        it('boundary matching a name-collision error keyed under the bare derived name routes via Case 3', async () => {
+            // mmnto/totem#1295 GCA HIGH catch: the collision detection in
+            // initContext used to key the error under a descriptive composite
+            // (e.g., `strategy (collision at .strategy2)`), so a user typing
+            // `boundary: 'strategy'` could not find the entry via
+            // `linkedStoreInitErrors.has('strategy')` and would fall through to
+            // raw-prefix search on the primary — exactly the silent drift this
+            // PR is supposed to prevent.
+            //
+            // The fix in context.ts now keys collisions under the BARE derived
+            // name. This test asserts the contract from the consumer side:
+            // a collision-style error message keyed under the bare name MUST
+            // route via Case 3 (explicit isError) rather than Case 4 (raw
+            // prefix). This is the integration guarantee the bare-name keying
+            // change unlocks.
+            mockLinkedStoreInitErrors.set('strategy', 'Another linked index already claims the name "strategy". ' +
+                'Path "./strategy2" also derives the link name "strategy". ' +
+                'Rename one of the linked directories or remove the duplicate from config.linkedIndexes.');
+            mockSearchResults = [
+                {
+                    label: 'Bogus primary hit that happens to match "strategy"',
+                    type: 'code',
+                    filePath: 'src/strategy-pattern.ts',
+                    score: 0.9,
+                    content: 'some code',
+                },
+            ];
+            const result = (await handle({ query: 'test', boundary: 'strategy' }));
+            expect(result.isError).toBe(true);
+            // The collision-style message should be surfaced via Case 3 wrapping
+            expect(result.content[0].text).toContain('strategy');
+            expect(result.content[0].text).toContain('not available');
+            expect(result.content[0].text).toContain('already claims');
+            // Bogus primary hit MUST NOT appear (no raw-prefix fallback)
+            expect(result.content[0].text).not.toContain('Bogus primary hit');
+        });
+        it('boundary matching a failed-init linked store returns explicit error (no silent primary fallback)', async () => {
+            // Shield AI catch: if a linked store name is in linkedStoreInitErrors
+            // but not in linkedStores (e.g., init failed, or reconnect blew up),
+            // the previous implementation silently fell through to querying the
+            // primary store with the name as a raw path prefix — returning
+            // unrelated local hits. Tenet 4 violation (silent drift).
+            mockLinkedStoreInitErrors.set('strategy', 'Linked index is empty (0 rows).');
+            mockSearchResults = [
+                {
+                    label: 'Bogus primary hit that happens to match "strategy"',
+                    type: 'code',
+                    filePath: 'src/strategy-pattern.ts',
+                    score: 0.9,
+                    content: 'some code',
+                },
+            ];
+            const result = (await handle({ query: 'test', boundary: 'strategy' }));
+            // Must return isError with a clear message naming the broken link,
+            // NOT bogus primary results from the raw-prefix fallback
+            expect(result.isError).toBe(true);
+            expect(result.content[0].text).toContain('strategy');
+            expect(result.content[0].text).toContain('not available');
+            // Explicitly check the primary results are NOT in the response
+            expect(result.content[0].text).not.toContain('Bogus primary hit');
+        });
+        it('one-shot linked-stores warning is NOT consumed by a transient getContext failure (CR minor)', async () => {
+            // mmnto/totem#1295 CR minor: previously `firstLinkedStoresCheckDone`
+            // was set BEFORE awaiting `getContext()`, so a transient init error
+            // on the very first call permanently suppressed the startup warning
+            // for the rest of the session. The fix moves the flag write to
+            // AFTER getContext resolves successfully.
+            //
+            // This test exercises the bug-fixed contract end-to-end:
+            //   1. First handle() call: getContext throws on every call (4+ times
+            //      to cover setLogDir, runFirstQueryHealthCheck, runFirstLinkedStoresCheck,
+            //      and performSearch). The handler returns isError but neither
+            //      first-query flag is consumed.
+            //   2. Second handle() call: getContext succeeds. The first-query gates
+            //      run for real, the linkedStoreInitErrors warning is surfaced,
+            //      and the search returns normally.
+            mockLinkedStoreInitErrors.set('strategy', 'Linked index is empty (0 rows).');
+            mockSearchResults = [
+                {
+                    label: 'Primary result',
+                    type: 'code',
+                    filePath: 'src/foo.ts',
+                    score: 0.9,
+                    content: 'content',
+                },
+            ];
+            // First call: 8 failures is enough to cover every getContext
+            // invocation in a single handle() call (setLogDir, healthCheck,
+            // linkedStores, performSearch, plus the outer-retry path).
+            mockGetContextFailuresRemaining = 8;
+            const first = (await handle({ query: 'test' }));
+            // The first call hard-fails (no context, no results)
+            expect(first.isError).toBe(true);
+            // Critically: the warning is NOT surfaced in the failure response
+            // (linkedStoresWarning was null because getContext threw)
+            expect(first.content[0].text).not.toContain('Linked index is empty');
+            // Second call: getContext now succeeds. With the fix, the first-query
+            // flag was NOT consumed on the first attempt, so the warning surfaces.
+            mockGetContextFailuresRemaining = 0;
+            const second = (await handle({ query: 'test' }));
+            expect(second.isError).toBeUndefined();
+            // The startup warning is now surfaced — the bug fix is observable
+            expect(second.content[0].text).toContain('[SYSTEM WARNING]');
+            expect(second.content[0].text).toContain('strategy');
+            expect(second.content[0].text).toContain('empty (0 rows)');
+            // And the actual search result is also present
+            expect(second.content[0].text).toContain('Primary result');
+        });
+        it('first-query-warn-block surfaces linkedStoreInitErrors', async () => {
+            mockLinkedStoreInitErrors.set('strategy', 'Linked index at /abs/strategy is empty (0 rows).');
+            mockSearchResults = [
+                {
+                    label: 'Primary result',
+                    type: 'code',
+                    filePath: 'src/foo.ts',
+                    score: 0.9,
+                    content: 'content',
+                },
+            ];
+            const result = (await handle({ query: 'test' }));
+            // Not blocking — primary results still return
+            expect(result.isError).toBeUndefined();
+            // But the init error is surfaced in the response via system warning
+            expect(result.content[0].text).toContain('[SYSTEM WARNING]');
+            expect(result.content[0].text).toContain('strategy');
+            expect(result.content[0].text).toContain('empty (0 rows)');
+            // Primary result still present
+            expect(result.content[0].text).toContain('Primary result');
+        });
+        it('linked store stale-handle recovers via per-store reconnect+retry (Shield AI catch)', async () => {
+            // Shield AI finding on the Phase 2 review: if a linked index gets
+            // rebuilt by a concurrent `totem sync`, its table handle goes stale
+            // and queries fail. The fix is the per-linked-store catch+reconnect+
+            // retry inside `federatedSearch`: when a linked store's search
+            // throws, federatedSearch calls its `.reconnect()` and retries the
+            // search before recording a runtime failure.
+            //
+            // mmnto/totem#1295 CR minor: this test asserts the linked-store
+            // reconnect spy was actually invoked AND that the retry produces
+            // recovered results. Without the spy assertion, the test would still
+            // pass if federatedSearch stopped reconnecting linked stores entirely
+            // — the exact regression Shield AI was guarding against.
+            mockSearchResults = [
+                {
+                    label: 'Primary',
+                    type: 'code',
+                    filePath: 'src/foo.ts',
+                    score: 0.7,
+                    content: 'content',
+                },
+            ];
+            const strategyStore = {
+                // First search call throws (stale handle), second call (after
+                // reconnect) returns the recovered result.
+                search: vi
+                    .fn()
+                    .mockRejectedValueOnce(new Error('Stale handle'))
+                    .mockResolvedValueOnce([
+                    {
+                        label: 'Linked recovered',
+                        type: 'spec',
+                        filePath: 'adr/adr-001.md',
+                        absoluteFilePath: '/abs/strategy/adr/adr-001.md',
+                        sourceRepo: 'strategy',
+                        score: 0.8,
+                        content: 'strategy content after reconnect',
+                    },
+                ]),
+                reconnect: vi.fn(async () => { }),
+            };
+            mockLinkedStores.set('strategy', strategyStore);
+            const result = (await handle({ query: 'test' }));
+            // The linked store's reconnect spy was actually invoked
+            expect(strategyStore.reconnect).toHaveBeenCalledOnce();
+            // The search spy was called twice: once that threw, once after reconnect
+            expect(strategyStore.search).toHaveBeenCalledTimes(2);
+            expect(result.isError).toBeUndefined();
+            // Both primary and the recovered linked result are in the merged response
+            expect(result.content[0].text).toContain('Primary');
+            expect(result.content[0].text).toContain('Linked recovered');
+            // No runtime warning because the recovery succeeded
+            expect(result.content[0].text).not.toContain('[SYSTEM WARNING]');
+        });
+        it('all results display absolute paths so the agent can Read/Edit unambiguously', async () => {
+            // mmnto/totem#1295 CR MAJOR: `formatResult` previously fell back to
+            // relative `filePath` for primary hits, reintroducing repo-root
+            // ambiguity in the common case — exactly the bug `absoluteFilePath`
+            // was added to fix. Both primary and linked hits must now display
+            // absolute paths in the File line.
+            mockSearchResults = [
+                {
+                    label: 'Primary',
+                    type: 'code',
+                    filePath: 'src/primary.ts',
+                    score: 0.7,
+                    content: 'primary content',
+                },
+            ];
+            mockLinkedStores.set('strategy', makeLinkedStore([
+                {
+                    label: 'Linked',
+                    type: 'spec',
+                    filePath: 'adr/linked.md',
+                    absoluteFilePath: '/abs/totem-strategy/adr/linked.md',
+                    sourceRepo: 'strategy',
+                    score: 0.6,
+                    content: 'linked content',
+                },
+            ]));
+            const result = (await handle({ query: 'test' }));
+            const text = result.content[0].text;
+            // Primary: absolute path (constructed by the mock from projectRoot
+            // + filePath, mirroring real LanceStore behavior)
+            expect(text).toContain('**File:** /fake/project/src/primary.ts');
+            // Primary's relative path must NOT appear in the File line — the
+            // regression would be the bare `src/primary.ts` (without the
+            // `/fake/project/` prefix) appearing where the absolute path goes
+            expect(text).not.toContain('**File:** src/primary.ts ');
+            // Linked: absolute path in the File field (unchanged)
+            expect(text).toContain('**File:** /abs/totem-strategy/adr/linked.md');
+            // Linked result has the [sourceRepo] tag prefix
+            expect(text).toContain('[strategy] Linked');
+            // Primary has no tag prefix (uses bare label)
+            expect(text).toMatch(/### \d+\. Primary \(code\)/);
+        });
+    });
 });
 //# sourceMappingURL=search-knowledge.test.js.map