v3.7.0-alpha.8 — agentdb delete tools + substrate upgrades since alpha.3

TL;DR

`npx ruflo@latest` is now 3.7.0-alpha.8. Five alphas since the last GitHub release (alpha.3, daemon-survival fix on Windows). Public CLI surface unchanged — these alphas are substrate improvements that compound on every existing feature.

```bash
npx ruflo@latest --version # → ruflo v3.7.0-alpha.8
```

Package	This release	Tags
`@claude-flow/cli`	3.7.0-alpha.8	latest, alpha, v3alpha
`claude-flow`	3.7.0-alpha.8	latest, alpha, v3alpha
`ruflo`	3.7.0-alpha.8	latest, alpha
`@claude-flow/cli-core`	3.7.0-alpha.5	latest, alpha
`@claude-flow/neural`	3.0.0-alpha.8	latest, alpha, v3alpha

What's new since alpha.3

🆕 `agentdb_*-delete` MCP tools — alpha.8 (#1784, #1789)

Three new MCP tools close the gap that previously made `/adr-index` re-index impossible (the reporter @Tovli flagged that deleted ADR files left stale nodes + dangling edges in the graph forever):

Tool	Backed by
`agentdb_hierarchical-delete`	`ReflexionMemory.deleteEpisode` — graph + vector + SQL all-in-one
`agentdb_causal-edge-delete`	`GraphDatabaseAdapter.deleteEdgesByEndpoints` — Cypher-routed, label-validated upstream
`agentdb_causal-node-delete`	`GraphDatabaseAdapter.deleteNode({cascade: true})` — returns native `{deletedNode, deletedEdges}` audit

All wrapped in MutationGuard (fail-closed) + AttestationLog (audit). Wired through agentdb@3.0.0-alpha.13's new Cypher API (the upstream fix shipped in agentic-flow#150/#151). SQL fallback for pre-alpha.13.

🧠 `@claude-flow/neural@3.0.0-alpha.8` substrate upgrades — alpha.7 (#1773)

Six concrete additions to the neural package, all targeted at the question "is this package SOTA and useful?":

1. Persistence — `serialize()` / `deserialize()` on `SONAManager`, `ReasoningBank`, `PatternLearner`. Process restarts no longer wipe state. Float32Array embeddings + Maps round-trip losslessly via `deepEncode` / `deepDecode` helpers.
2. Seedable PRNG — `Mulberry32` + `setGlobalRng` injection point for reproducible training, A/B mode comparison, and deterministic tests.
3. Self-consistency orchestrator — `selfConsistency(N, op, aggregator)` (Wang et al. 2022). Three aggregators (majority / mean / first). 5–15pp accuracy on reasoning tasks at Nx compute.
4. Flash Attention + MoE migration — `flash-attention.ts` (857 LOC) and `moe-router.ts` (822 LOC) moved from cli into the neural package. Single source of truth; cli imports via package boundary.
5. Retrieval-path observability — `hnswRetrievalCount` vs `bruteForceRetrievalCount` in `ReasoningBank.getStats()`. `PatternLearner` honestly reports `hnswEnabled: 0` (no HNSW yet — separate followup).
6. 80+ npm SEO keywords + complete metadata on the neural package. Now discoverable for `ai-agents`, `multi-agent`, `RL`, `LoRA`, `EWC`, `SONA`, etc.

Validation: 132 neural tests pass (44 algorithms + 39 patterns + 23 sona + 15 persistence + 9 self-consistency + 2 retrieval-path). 781 cli tests still pass.

🎯 Thompson sampling model router — alpha.5 (#1772)

The 3-tier model selector (Haiku / Sonnet / Opus) is now a cost-adjusted multi-armed bandit instead of static thresholds. Every `hooks_model-outcome` call updates Beta(α, β) priors per tier; `hooks_model-route` samples θ ~ Beta(α, β) and picks argmax. The 77% Opus / 0.30 complexity mismatch corrects itself within ~50 routing decisions, automatically.

• Marsaglia-Tsang gamma sampler + Box-Muller normal — pure JS, no new deps
• Cost-adjusted Bernoulli rewards: Haiku-success ≫ Sonnet-success ≫ Opus-success
• Convergence test: 100-trial workload at avg complexity 0.3 — Haiku posterior mean dominates Opus by N=100
• Overhead: 45 µs per route() call (under the file's documented `<100 µs` target)

🏎️ `@claude-flow/cli-core` lite path — alpha.4–alpha.5 (#1760, #1773 Phase 1)

A new sibling package that handles memory commands only (no SQLite, no HNSW, no ONNX). Cold-cache `npx` wall-time drops from ~35s to ~1.5s — measured 22.9× speedup for plugin scripts.

```javascript
// Plugin scripts opt in via env flag:
const cliPkg = process.env.CLI_CORE === '1'
? '@claude-flow/cli-core@alpha' // ~1.5s cold-cache
: '@claude-flow/cli@latest'; // ~35s cold-cache (full features)
```

8 plugin scripts in this repo are already CLI_CORE-aware (cost-tracker × 6, ruflo-adr × 2). The full `@claude-flow/cli` is unchanged for end users; cli now depends on cli-core for 4 foundation modules (1,229 LOC of duplication eliminated).

Reproducible benchmark: `bash v3/@claude-flow/cli-core/scripts/cold-cache-bench.sh`

🩹 Earlier alphas in this series

• alpha.7 (#1773) — neural Phase 2 cli↔neural convergence: cli now depends on `@claude-flow/neural` and lazy-loads through a bridge. 1,215 LOC of byte-identical foundation modules deduplicated.
• alpha.6 (#1773 Phase 1) — `@claude-flow/neural` package wired into cli's dep graph for the first time + neural README cleanup (5 factual errors fixed against live exports).
• alpha.5 (#1772) — Thompson sampling bandit (above).
• alpha.4 (#1770) — fixed two silent intelligence regressions: `neural_status.flashAttention` was hardcoded `false`; `ruvllm.coordinator` reported `"unavailable"` despite `@ruvector/ruvllm` being resolvable (lazy-init timing bug in `getIntelligenceStats()`).

What didn't change

• Public CLI surface (26 commands, 140+ subcommands)
• Agent registry (60+ agent types)
• Plugin marketplace (21 native plugins, IPFS distribution)
• Hooks system (27 hooks + 12 background workers)
• MCP tool surface (314 tools)
• Configuration files (`claude-flow.config.json`, `.env`, plugin schemas)

If you're running 3.6.x, the upgrade is opt-in: most users won't notice anything beyond the version bump. Substrate improvements compound underneath.

Validation

• ✅ `tsc` clean (cli + neural packages)
• ✅ 793 cli tests pass — 769 baseline + 6 router-bandit + 6 neural-package-bridge + 12 agentdb-delete-tools
• ✅ 132 neural tests pass — 44 algorithms + 39 patterns + 23 sona + 15 persistence + 9 self-consistency + 2 retrieval-path
• ✅ All 5 packages publish-verified at the dist-tags shown above

Closed issues / merged PRs since alpha.3

#	Title
#1770	intelligence: two silent regressions — Flash Attention hardcoded false + ruvllm coordinator package-name mismatch
#1771	fix(intelligence): make Flash Attention + ruvllm coordinator stats truthful (PR)
#1772	feat(router): Thompson sampling bandit replaces deterministic argmax
#1773	@claude-flow/neural substrate upgrades (6 items)
#1774	feat(neural): Phase 1 convergence — cli depends on @claude-flow/neural
#1776	feat(neural): substrate upgrades 1-6 + SEO keywords
#1784	agentdb hierarchical-store and causal-edge have no delete tools
#1789	feat(agentdb): hierarchical + causal-edge + causal-node delete MCP tools (PR)

Upgrading

```bash

Just use latest:

npx ruflo@latest

Or pin:

npx ruflo@3.7.0-alpha.8
```

State files are forward-compatible — `schemaVersion: 1` is enforced on all new persistence formats. Existing `.swarm/` and `.claude-flow/` state directories work unchanged.

Next

Loose ends, all queued:

• Tombstone tables for native graph-node storage (recall-time filtering) — would let pre-alpha.13 backends also support delete. Bigger change, separate follow-up.
• Updating `plugins/ruflo-adr/scripts/import.mjs` to invoke the new agentdb delete tools during `/adr-index` re-index — needs sync-semantics design.
• `SONAManager.serialize/deserialize` round-trip test (deferred — `src/modes/balanced.ts` has a load-order bug that blocks instantiation in vitest; production path is correct).
• Promotion of `@claude-flow/cli-core` from `@alpha` to broader visibility — gated on external validation.

🤖 Generated with RuFlo