**FRAMEWERX vs. The Field**

vs. Cursor, Windsurf, Copilot, Devin, Cline. Every number below is backed by the running codebase.

**What only FRAMEWERX has**

**1. Durable, rewindable agent runtime**
Hash-chained, crash-recoverable event journal with real `rewind_to_sequence` (fsync-truncates to any prior event). Kill the server, reboot, resume from the last checkpoint with full history. None of the others ship a rewindable journal — their runs are ephemeral.
`agent_runtime/journal.py`

**2. Multi-provider failover + provenance**
11 provider backends (NIM, Ollama, OpenAI, OpenRouter, Claude, DeepSeek, Jules, AstraFlow…) behind one `FailoverProvider` that health-checks and falls over per-call. Every generation logged with billing provenance before execution. No competitor has this.
`sword_llm/providers/failover.py`

**3. Broadest tool surface — 202 callables, 206 tool definitions**
Not a marketing round number — it's the runtime count. Goes past file/shell/git: ptrace, GDB, heap walking, shellcode assemble/encode, ROP chains, kernel ops, fuzz, vulnscan, crashdump, exploit ops, crypto, binary, packet, p2p, rpc, race, serial… 36 `*_ops` modules. The only one in the set that can drive a debugger and refactor your Python in the same session.
`sword_llm/tools/registry.py`

**4. Graph context engine — 117,504 nodes, 1,322,728 edges, <25ms**
C99 graph with QIHSE semantic edges + KEYSTONE sorted lookup. Others use embedding similarity over chunked files. FRAMEWERX indexes structural + semantic relationships together, native backend. That's why the 202-tool surface is actually usable.
`native/qihse/`, `api/context_index.py`

**5. ACP v1 — one runtime, every editor**
Full Agent Client Protocol v1 over stdio JSON-RPC 2.0. Zed, JetBrains, Cursor, Gemini CLI, Codex CLI, Claude Agent can all drive the FRAMEWERX runtime as a native agent. No editor lock-in.
`sword_llm/acp/server.py`

**Ships today (tested, not roadmap)**
• Multi-file diff review — before/after snapshots, per-file accept/reject, survives restarts (10 tests)
• Plan Mode UI — plan files, approve/execute flow (7 tests)
• Terminal integration — dedicated terminal, auto-install detection, "Run in Terminal", "Send Terminal to Agent"
• Streaming diff preview — live before/after diff via virtual URIs (3 tests)
• Plugin/extension API — discovery, registry, CLI, REST endpoints (18 tests)
• ACP session persistence — recovers from journal after restart (65 tests)

**On the roadmap (honest)**
• Visual checkpoint timeline — journal + rewind already exist; this is UI on top
• Semantic @-mentioning — graph already exists; needs editor autocomplete wiring (`MentionProvider` skeleton in place)
• Git-native workflow — `git_ops` ships; needs branch viz, PR-from-run, commit msg gen
• Telepathic autocomplete — low priority by design; FRAMEWERX is local-first and security-focused
• Team collaboration — low priority by design; local-first scope choice

**The short version**
Durable journal: only FRAMEWERX ✓
Provider failover + provenance: only FRAMEWERX ✓ (11 providers)
Tool surface: others ~dozens → FRAMEWERX 202/206
Native security/exploit tools: only FRAMEWERX ✓
Graph context (117K/1.3M/<25ms): only FRAMEWERX ✓
Open editor protocol (ACP v1): only FRAMEWERX ✓

FRAMEWERX doesn't try to be the best inline-autocomplete editor. It's the most capable, most durable, most portable agent runtime — bring its own context graph, tool surface, provider failover, and rewindable journal to whatever editor you already use.

The L2CAP UAF exploit is now successfully triggering on the Samsung A12