T3MP3ST's 'keyless local-agent backbone' functionality, specifically its failure to execute 'Arsenal tools' within the ReAct loop. This is a critical bug preventing the platform's core offensive security actions.

## Summary

On the **keyless local-agent backbone** (Codex / Claude / Hermes — the headline "use the agent you already run" path), mission operators never execute a single Arsenal tool. The ReAct loop returns after **one** planning turn, so every operator abstains and recon/scan/exploit never touch the target. I pointed a mission at a host on my own LAN and watched all 10 operators finish with `{"findings":[],"abstained":true}` while the box sat there with 3 open ports.

It's deterministic, not a model mood or a sandbox-config thing. An API-key backbone (OpenRouter/Anthropic/OpenAI) works fine on the same mission. The bug is specific to the local-agent path.

## Repro

1. Connect Codex in the War Room (or start a mission with `provider: 'local-agent', model: 'codex'`).
2. Point it at any reachable target and launch.
3. Watch the operators. They reason, read the repo, and abstain. `nmap`/`curl`/etc never run against the target. `GET /api/mission/status` shows tasks completed with 0 tool-backed findings.

## Root cause

The ReAct loop drives *every* Arsenal action through `chatWithTools(...).toolCalls`:

```ts
// src/agent/index.ts:137
const response = await this.llm.chatWithTools(messages, toolDefs, { ... });
if (response.toolCalls?.length) {
// ... execute tools via arsenal.execute ...
} else {
// src/agent/index.ts:214 — treats the first text turn as the FINAL answer and returns
}
```

But `LocalAgentAdapter.chat()` (the local-agent backbone) ignores `options.tools` enti...