Claw Family — Framework Comparison Matrix

01 / Capability Radar

Framework scores across 8 dimensions

02 / Score Matrix

Dimension-by-dimension breakdown — click rows to compare

Dimension	Nanobot	OpenClaw	NanoClaw	OpenFang	IronClaw	Lyra Now	Lyra Target

03 / Category Leaders

Best-in-class per domain

5/5

Memory / Search

OpenClaw + IronClaw

5/5

Context Management

OpenClaw

5/5

Channel Count

OpenFang (40 adapters)

5/5

Channel Abstraction

OpenClaw (7-tier bindings)

5/5

Bus / Events

OpenClaw + OpenFang

5/5

LLM Providers

IronClaw (decorator chain)

5/5

Tool System

IronClaw + OpenFang

5/5

Plugin System

OpenClaw (typed SDK)

5/5

Security

OpenFang (16 layers) + IronClaw

5/5

Orchestration

OpenFang (workflow engine)

5/5

Observability

OpenClaw (diagnostic events)

5/5

Config / Hot-Reload

OpenClaw + ScalyClaw

5/5

Proactive Engagement

ScalyClaw (unique)

5/5

Readability

Nanobot + Lyra

04 / The "Chimera" Strategy

Best of each — ~2,170 LOC gives Lyra production-grade capabilities

Phase 1b~710 LOC

1

Provider Registry

Metadata-driven auto-detection via JSON/TOML

Nanobot#76

2

LLM Decorator Chain

Retry → SmartRoute → Failover → CircuitBreaker → Cache

IronClaw#104

3

Hybrid RRF Search

FTS always works, vector boosts — Reciprocal Rank Fusion

IronClaw#83

4

ContextEngine Protocol

Pluggable token-budget assembly — lazy-load recall

OpenClaw#83

5

Smart Routing Decorator

Complexity classifier → cheap model for trivial, powerful for complex

IronClaw#134

6

Runtime Agent Config

!config command — live-tune temperature/model/style without restart

ScalyClaw#135

Phase 2~760 LOC

5

Lane-Based Queue

Cron lane separate — never blocks user messages

OpenClaw#112

6

Binding Resolution Tiers

Exact match → wildcard → default agent

OpenClaw#112

7

Tool Approval Levels

Never / UnlessAutoApproved / Always

IronClaw#106

8

Credential Proxy

Plugins declare secret names — hub injects at boundary

IronClaw#106

9

Diagnostic Events

Stuck detection, token tracking, per-session observability

OpenClaw#44

10

Prompt Injection Scanner

Block / Warn / Review / Sanitize severity levels

OpenFangfuture

11

Proactive Engagement Engine

2-phase: cron signal scan (no LLM) → eval job (LLM only when warranted)

ScalyClawfuture

12

Command Shield

Deterministic blocklist/allowlist — no LLM, fail-closed, zero latency

ScalyClawfuture

05 / Pattern → Issue Dependency Flow

Which patterns unblock which Lyra issues

flowchart TD
  subgraph FW["Source Frameworks"]
    IC["IronClaw"]
    OC["OpenClaw"]
    NC["NanoClaw"]
    NB["Nanobot"]
    SC["ScalyClaw"]
  end

  subgraph P0["Phase 0 — Bot Core Parity #101"]
    I104["#104 Circuit Breaker"]
    I106["#106 Plugin System"]
  end

  subgraph P1b["Phase 1b — Agent Core #73"]
    I76["#76 Anthropic SDK Agent"]
    I75["#75 Persona System"]
    I134["#134 Smart Routing"]
    I135["#135 Runtime Config"]
    I67["#67 Session Persistence"]
    I83["#83 Memory Integration"]
    I99["#99 Hub Command Sessions"]
    I112["#112 Hub Refactor"]
  end

  subgraph P2["Phase 2"]
    I48["#48 Bus Abstraction"]
    I44["#44 Event Monitoring"]
  end

  IC -->|"decorator chain"| I104
  IC -->|"smart routing"| I134
  IC -->|"approval levels"| I106
  IC -->|"credential proxy"| I106
  IC -->|"RRF fusion"| I83
  OC -->|"append-only JSONL"| I67
  OC -->|"lane-based queue"| I112
  OC -->|"diagnostic events"| I44
  NC -->|"dual cursors"| I67
  NC -->|"self-register factory"| I106
  NB -->|"split queues"| I112
  NB -->|"per-session Task"| I112
  SC -->|"runtime config"| I135

  I76 -->|"unblocks"| I134
  I76 -->|"unblocks"| I135
  I75 -->|"persona path"| I135
  P0 -->|"unblocks"| I67
  P0 -->|"unblocks"| I83
  I83 -->|"unblocks"| I99
  I112 -->|"unblocks"| I48

  classDef fw fill:#0f2a3a44,stroke:#22d3ee,stroke-width:1.5px
  classDef issue fill:#0d1f1a33,stroke:#34d399,stroke-width:1px
  classDef new fill:#1a2a0033,stroke:#facc15,stroke-width:1.5px
  class IC,OC,NC,NB,SC fw
  class I104,I106,I67,I83,I99,I112,I48,I44,I76,I75 issue
  class I134,I135 new

06 / Lyra Today

What we have vs what we're building

✓Already solid

✓RoutingKey (platform, bot_id, user_id)Matches OpenClaw session key pattern

✓asyncio single event loopCorrect foundation — Nanobot validates this

✓Protocol-based adapters (ChannelAdapter)Matches all frameworks' structural approach

✓WebSocket via aiogram + discord.pyBest-practice ingestion for both platforms

✓roxabi-memory (SQLite + FTS5 + vectors)Best-in-class storage stack. Needs RRF fusion.

✓Bounded asyncio.Queue (size 100)Better than Nanobot's unbounded queue

✓Python-first paradigmHighest readability score alongside Nanobot

🔨Being built

→Per-session asyncio.Task (has per-user Lock)No cancellation, no per-turn timeout#112

→Split inbound / outbound queuesSingle queue limits backpressure control#112

→Crash recovery (JSONL + dual cursors)Pool.history is volatile, lost on restart#67

→LLM decorator chain (retry + CB + cache)No resilience on Anthropic API failures#104

→Tool approval levels (Never/Auto/Always)Mandatory for plugin safety#106

→Credential proxy at hub boundaryPlugins should never see raw API keys#106

→RRF fusion in hybrid searchCurrently naive score averaging in roxabi-memory#83