dream(intelligence): #4249 agentdb-backend.ts calls a native HNSW controller that has never existed in the real agentdb package (evaluated, ACCEPT)

[ruvnet]

1. Hypothesis

Given AgentDBBackend's storeInAgentDB(), searchWithAgentDB(), and getStats() (v3/@claude-flow/memory/src/agentdb-backend.ts) each call this.agentdb.getController('hnsw') inside a try/catch, expecting an optional native HNSW controller from the agentdb npm package, when the real agentdb package's AgentDB.getController() (confirmed directly from agentdb@3.0.0-alpha.20's installed dist/src/core/AgentDB.js) recognizes only 'memory'/'reflexion', 'skills', 'causal'/'causalGraph' and throws Error('Unknown controller: hnsw') for anything else — then every store/search/getStats call pays for a thrown-and-caught exception for a controller that can never exist, and searchWithAgentDB additionally logs a misleading console.error('HNSW search failed:', ...) on every search even though the code unconditionally falls back to the already-correct bruteForceSearch() either way. Removing the three dead call sites should (a) eliminate the false error log with zero change to search results, and (b) remove a real, isolated, measurable per-call exception-handling cost. Frozen before evaluation; not modified after.

2. Candidate

v3/@claude-flow/memory/src/agentdb-backend.ts — removed the three dead this.agentdb.getController('hnsw') call sites in storeInAgentDB(), searchWithAgentDB(), and getStats(). searchWithAgentDB() now calls bruteForceSearch() directly (the only path that ever actually ran); getStats()'s hnswStats is now explicitly undefined (what it always evaluated to anyway); storeInAgentDB() drops the dead add-to-index block entirely. Diff: 2 files, 80 insertions / 62 deletions (142 changed lines, well under 300), no behavior change to search results.

This is the third instance of a recurring pattern already documented twice in this repo's own ledger — aspirational "native acceleration" integration code checking for an export/controller that was never real in the installed dependency (2026-08-30 diskann-backend.ts removal, 2026-10-10 graph-analyzer.ts native-acceleration correction).

3. Evaluation Receipt

evaluated: accepted. Real evaluator: Vitest 4.1.8 against a real (non-mocked) agentdb@3.0.0-alpha.20 instance (:memory: db), after a full pnpm install --filter @claude-flow/memory....

• Direct reproduction against the unmodified source: console.error called with ["HNSW search failed:", {}] during a normal store() + search() — the bug fires in practice.
• New test 1 (false-error regression guard): stash-isolated — fails on baseline (expected [] to have length 0 but got 1), passes on candidate.
• New test 2 (tripwire): asserts the real agentdb package's getController('hnsw') throws /Unknown controller/ — so a future agentdb release adding real HNSW support fails this test loudly instead of the dead code silently staying dead.
• agentdb-backend.test.ts: 21/21 (baseline) → 23/23 (candidate, +2 new), 0 regressions.
• Full @claude-flow/memory suite: 614/616 both baseline and candidate — the 2 failures are pre-existing/unrelated (a read-only-file permission test that doesn't work running as root; an N=5000 perf test timing out at 34s vs. a 30s ceiling, neither touching this file; first confirmed failing identically on baseline via stash isolation).
• tsc --noEmit: 0 errors, both ways.
• Consumer regression check: hybrid-backend.test.ts + database-provider.test.ts 41/41 pass on candidate.

4. Baseline Comparison

	Baseline	Candidate
console.error("HNSW search failed:", ...) on a normal search	fires every call	never fires
Search results returned	via bruteForceSearch() (same either way)	via bruteForceSearch() (same either way)
getStats().hnswStats	always undefined (via swallowed throw)	always undefined (explicit)
agentdb-backend.test.ts	21/21	23/23 (+2 new)
Full @claude-flow/memory suite	614/616	614/616 (same 2 pre-existing failures)
tsc --noEmit	clean	clean

Isolated microbenchmark (explicitly NOT an end-to-end store()/search() timing claim — a synthetic probe of only the mechanism removed, flagged by the adversarial critic as needing this exact qualifier): a tight loop calling a throw-then-catch path 200k times vs. a direct-return path 200k times measured ~5.66µs/call vs. ~0.008µs/call, a ~700x difference — consistent with V8's documented stack-trace-capture cost for thrown exceptions. Real-world store()/search() wall-clock is dominated by SQLite I/O and embedding work untouched by this fix; this number supports "removes a real per-call cost," not an end-to-end speedup claim.

5. Darwin Lineage

Skipped — binary dead-code-removal/log-suppression fix, no continuous or categorical parameter for Darwin's real interface (npx ruvector harness darwin <config> --execute, confirmed available) to search over. Same skip class as nearly every accepted fix in this repo's history (#3929, #3977, #4053, #3395, ...).

6. Flywheel Evidence

No .claude-flow/flywheel/ state exists in this checkout (same disclosed gap as every recent night). Evidence retained as: 2 new discriminating tests, the direct reproduction probe, the full-suite comparison, the isolated microbenchmark, issue #4249, the gist, and the witness below.

7. Reward Hack Check

Clean. Independently re-verified by an adversarial-critic subagent (separate context, no authoring bias): no weakened assertions, no gold-data edits, no cherry-picking (full suite disclosed both ways), no evaluator exploitation, no cost hiding ($0, deterministic), no seed manipulation. The critic specifically chased a possible null-safety regression (whether the dynamically-imported HNSWIndex global might be perpetually falsy, silently swallowing real results via the unrelated if (!this.agentdb || !HNSWIndex) guard) and confirmed it is truthy and search still returns real results end-to-end — no guard the removed try/catch depended on.

8. Security Review

Not security-sensitive — pure dead-code removal plus removing a misleading log line. No new I/O/network/credential/filesystem surface; no change to what gets stored, searched, or returned.

9. Regression Analysis

Zero regressions: agentdb-backend.test.ts full pass (+2 new), full @claude-flow/memory suite byte-identical failure set (614/616 both ways, both failures pre-existing/unrelated), hybrid-backend.test.ts/database-provider.test.ts (consumers) 41/41, tsc --noEmit clean both ways.

10. ADR

Not created — scoped dead-code-removal fix matching an already-decided repo convention (2026-08-30 diskann-backend.ts, 2026-10-10 graph-analyzer.ts corrections — neither created an ADR either).

11. Research Gist

docs/dream-cycle/dream-gist-2026-10-17.md (this branch) — full SOTA report: 4-role parallel research fan-out (deep researcher, 2 scan researchers, competitor analyst), 9-row competitor comparison (LangGraph/AutoGen/CrewAI/OpenAI Agents SDK + 5 vector DBs), 2026-dated sources (SLEUTH, "When Routing Collapses", cheap reward-hack detection, HNSW tombstone repair, SLoRA), and 3 specific future-night recommendations. No gist-publishing tool available this session (no gh CLI, no gist-equivalent MCP tool) — same disclosed FALLBACK as every night since 2026-09-30; full report embedded in issue #4249 instead.

12. Issue

Closes the evidence trail documented in #4249 (full 15-section report, including the capabilities-surface and memory-surface scan findings).

13. Witness

Session commit:  1882006a02487590c8c66a7eec8d7504f6af33c3
Gist SHA-256:    cbc84de295ffb01ce5ddbe2e6fa5a78b544228dabdda7129cdf18f8aaa19dba4
Witness stamp:   e2abed9831542617a74b3b0254ff4e775beea19199327a159dbd6138bf3d6914

Verifier: fetch docs/dream-cycle/dream-gist-2026-10-17.md from this branch, SHA-256 it, concatenate with the session commit above, SHA-256 again — must equal the witness stamp.

14. Merge Policy

Human review required. Do not self-merge. Do not autonomously promote Flywheel state. Independent adversarial critic verdict: CONFIRMED-WITH-CAVEATS (no blocking issues; caveat was solely to ensure the microbenchmark stays labeled as isolated, which it is above).

---

🤖 Generated with RuFlo

https://claude.ai/code/session_01KmvMMVh9xNQdRGXfk9vwjt

---

Generated by Claude Code

[ruvnet]

CI: "Plugin package install-safety" was failing — fixed, pushed.

The audit-leaf-package-publish.mjs check correctly flagged that this PR changed publishable source in @claude-flow/memory (agentdb-backend.ts) — a non-bundled leaf dependency that resolves from the npm registry rather than shipping with the release tarball — without bumping its version, which would mean the fix never reaches users on the next release.

Fix pushed in 73c0bc6: bumped @claude-flow/memory 3.0.5 → 3.0.6 (patch: bug fix only, no API change) and updated @claude-flow/cli's exact pin for it to match, in both dependencies and optionalDependencies, per CLAUDE.md's Publishing Rules and the exact-pin discipline from #3414. Verified locally: audit-leaf-package-publish.mjs, audit-plugin-packages.mjs, audit-package-dep-overlap.mjs, audit-wrapper-dep-ranges.mjs, and audit-umbrella-version-lockstep.mjs all pass against origin/main.

---

Generated by Claude Code

[ruvnet]

CI: 5 more checks were failing — 4 fixed (this PR's), 1 is not this PR's.

Fixed, pushed in 88b90fd: "Smoke (no better-sqlite3)" (Node 22 & 24), "Plugin hooks smoke", "statusline generator delegation smoke", and "Memory import smoke" were all failing with ERR_PNPM_OUTDATED_LOCKFILE — my previous commit (73c0bc6) bumped @claude-flow/memory's version in package.json but didn't sync v3/pnpm-lock.yaml's specifier field for it, so every job running pnpm install --frozen-lockfile failed immediately on install, before any of their actual test logic ran.

Fixed by hand-editing the single specifier: 3.0.5 -> 3.0.6 line under @claude-flow/cli's optionalDependencies for @claude-flow/memory in v3/pnpm-lock.yaml, rather than regenerating the whole lockfile via pnpm install: a full regeneration in this environment pulled in ~175 lines of unrelated incidental drift (ruvector 0.2.27→0.2.41, @opentelemetry/api 1.9.1→1.9.0, stale dev: false annotation cleanup) that has nothing to do with this PR, since @claude-flow/memory is a workspace link: dependency with no real registry resolution to re-fetch — the targeted edit is exactly what pnpm install --frozen-lockfile needs and nothing more. Verified locally: pnpm install --frozen-lockfile now succeeds from the checked-out v3/ workspace with only this one line changed.

Not this PR's — pin-drift (metaharness-pin-drift check): failing because @metaharness/darwin is pinned at ~0.10.2 in v3/@claude-flow/cli/package.json while npm's current latest is 0.12.0. This PR never touches any @metaharness/* pin — confirmed the last scheduled run of this same check on main (2026-10-12) passed, meaning @metaharness/darwin@0.12.0 was published sometime after that and nothing has re-run the check against main since; it would fail identically on main right now. Bumping that pin is out of scope for this intelligence/memory dead-code-removal PR — flagging it as a separate, pre-existing drift item rather than widening this PR to fix it.

---

Generated by Claude Code