[Dream Cycle 2026-06-04] swarm: AdaptOrch +22.9% topology gain — ADR-147 adaptive topology selection

[ruvnet]

Dream Cycle — 2026-06-04 | SLOT=4 DEEP=swarm SCAN=ruview-integration,ruvector-integration

Tonight's Rotation

Field	Value
SLOT	4
DEEP	swarm
SCAN	ruview-integration, ruvector-integration
Session commit	844f68dbe5f28c4c2b13c56e8e102528aa63b629

---

ADR Filed

ADR-147 — Adaptive Swarm Topology Selection (v3/docs/adr/ADR-147-adaptive-swarm-topology-selection.md)

Proposes a TopologyClassifier in @claude-flow/hooks that selects parallel / hierarchical / hierarchical-mesh at swarm_init time based on task graph shape, replacing the hard-coded hierarchical default. Evidence: AdaptOrch benchmark (Grade A) shows +22.9% task completion on SWE-bench Verified with adaptive topology vs best single fixed topology.

---

Gist Link

Witness-stamped research gist: v3/docs/dream/dream-gist-2026-06-04.md

• SHA-256: 7deea4832c84d69393f72cae61451aaaabb7002cb9dabf63a0df7934f27bdee4
• Witness: 63e3586d034eda666911edaf972820f8af5aad7dad28e87eb73b621475f2a876

---

Issue Link

Closes #2289

---

Research Summary

DEEP — Swarm:

• SWARM+ (arXiv:2603.19431): 97–98% latency reduction at 990 agents via hierarchical consensus (Grade B)
• AdaptOrch: +22.9% on SWE-bench Verified with adaptive topology; 62% hybrid, 24% parallel, 14% hierarchical selected at runtime (Grade A)
• CINOC (arXiv:2605.25867): zero-shot cardinality transfer across swarm sizes (Grade B)
• Industry coordination failure rate: 36.94% across AutoGen/CrewAI/LangGraph — Ruflo has no telemetry baseline
• Supervisor topology adds +20–40% token overhead vs swarm (LangChain, Grade B)

SCAN — ruview-integration: RuView (51,576⭐ Rust WiFi-CSI) has no Ruflo plugin bridge; edge-swarm sensing gap open. No ADR — implementation-level.

SCAN — ruvector-integration: AgentDB benchmarked at 5k–20k vectors; competitors tested at 100M–1B (Qdrant <100ms p99, Milvus 100K+ QPS). Scale cliff above 20k unknown. No ADR — benchmark gap.

---

Merge Policy

Do not self-merge. Leave for human review. This is a research + ADR proposal — implementation requires a separate PR with code, tests, and coordination-failure telemetry baseline. Squash-merge preferred to keep main clean.

---

Generated by Claude Code

[ruvnet]

Review (from #2324)

Renumber: ADR-147 → ADR-148 before merge. All 6 open dream-cycle PRs claim ADR-147; current highest accepted is ADR-146. Proposed sequence in PR-creation order:

• [Dream Cycle 2026-06-03] memory: VikingMem temporal compression + provenance anchoring (ADR-147) #2278 → 147, [Dream Cycle 2026-06-04] swarm: AdaptOrch +22.9% topology gain — ADR-147 adaptive topology selection #2290 → 148, [Dream Cycle 2026-06-05] performance: LAMaS 38-46% critical-path gap + ADR-147 latency-aware scheduling #2295 → 149, [Dream Cycle 2026-06-06] security: ADR-147 memory write integrity + MCP tool verification #2304 → 150, [Dream Cycle 2026-06-07] intelligence: ADR-147 retrospective harness optimization (+19pp SWE-Bench Pro gap) #2310 → 151, [Dream Cycle 2026-06-08] memory: ADR-147 multi-signal retrieval (semantic+BM25+entity) #2317 → 152

Process concerns (zero merges in 14 nights, self-grading rubric vs. merge-rate signal) tracked in #2324.

Research itself is good — sources, gap analysis, and in-tree citations all check out. Collision is the only thing blocking merge here.