I built this so my 8-year-old could direct his Roblox, Blender, and Daz scenes through natural language. He says "make Torus.1 50% thinner on the Z-axis," and the live environment updates. Under the hood: an AI helper bridge through OpenClaw driving live scene state. Three things crossed here that I cared about. AI work. My long-running 3DCG interest. A kid who'd already started building in Blender, Roblox Studio, and Suno. codex-local is my Codex CLI fork. I tuned it to babysit a local model through a structured task list. It runs multi-turn agentic coding sessions, edits multiple files, completes the work, and does it with just Qwopus-3.5:9B underneath. What makes this work isn't the model. It's everything I built around it: 1. **Trim the tool menu** — Codex has ~120 tools. Small models get confused by big menus; we show them ~10. 1. **Plain-English tool cheat sheet** — We paste simple "here's how each tool works, with an example" text into the system prompt. 1. **Rewrite shell-ish tool names** — Models call `ls`, `cat`, `grep` like tools. We catch that and convert to a proper `shell` call. 1. **Browser user-agent for curl** — Sites block `curl/8.0`. We auto-add a real browser UA to any curl command. 1. **Web search + web fetch** — No built-in web search in local mode, so we added two tools for looking up real API docs. 1. **Fix broken patches** — Small models mangle patches: (a) write git-diff format, (b) leave `@@ -1,6 +1,6 @@` line numbers, (c) forget `+` prefixes and closing markers. We fix all three automatically. 1. **Better patch error messages** — Default errors are cryptic. We rewrote the common ones to explain what to try next. 1. **Better network errors** — Instead of "error sending request", we show the real cause (DNS, TLS, connection refused, etc.). 1. **Catch announce-without-act** — When the model says "Now I'll do X" and stops, a judge model spots it and we re-prompt "take the action." Up to 3 retries. 1. **Stop repetition** — (a) Same tool + same args 3× → STOP block in next prompt. (b) Same file failing 3× with different commands → same STOP block. 1. **Trim old conversation** — Keep the most recent turn intact, summarize older turns, drop stale file reads, pin errors so the model can't forget them. 1. **Log the model's thinking** — Reasoning text goes to debug logs so we can explain weird behavior later. 1. **Reroute wrong picks** — If the router picked "text-only model" but the conversation has tool calls, we upgrade to the tool-capable Coder. 1. **Diagnostic logs** — Extra logging for "which tools did we pass?", "did the STOP fire?", "what did the bail judge decide?" 1. **LM Studio / OpenAI-compat support** — Ollama and OpenAI-style servers disagree on URLs, payload shape, tool-call encoding, etc. A "flavor" switch lets one codebase talk to both. 1. **Token + time budget knobs** — New per-role `max_tokens` and `timeout_seconds` in `config.toml`. Set either to `0` for unlimited (reasoning models need this). 1. **Pin current file contents** — Models forget a file changed and generate patches based on the old version. We pin live on-disk contents at the top of the prompt. 1. **Catch thinking loops** — Some models spiral: "Actually, wait. Hmm. Let me reconsider." We watch the stream for 6+ self-doubt phrases after half the token budget is burned, abort mid-generation, and re-prompt "stop second-guessing." 1. **Stream the coder's output** — To make the anti-spiral work, we switched from "send, wait, parse" to "open stream, watch tokens, abort if needed." Also lets us log reasoning in real time. 1. **Skip extras in local-only mode** — Normally a router, reasoner, and compactor help the main model. Locally we can't afford extras, so the Coder handles everything. Most local-model coding agents stall the moment a task goes past trivial. With these nudges in place, 9B parameters really work. - **Main workstation:**
  • 3080 10GB + 1080 8GB (Monitor hooked up to Intel iGPU)
  • 13th Gen i9 + 64GB DDR5 6800
- **Laptop:**
  • 3080 8GB (Screen running from Intel iGPU)
  • 11th Gen i7 + 32GB DDR4 3200
- **LLM Servers:**
  • Ollama - main install (Using `OLLAMA_FLASH_ATTENTION=1` and `OLLAMA_KV_CACHE_TYPE=q8_0)
  • llama.cpp - used to try MoE models across both GPUs and CPU offloading
  • LM Studio - used for qwopus (doesn't run in Ollama)
- **Tested Models (3080 10GB):**
  • gemma4:e4b
  • gemma4:e4b-q6_k (Impressive runner-up)
  • qwen3.5:9b-opus-openclaw-distilled-iq4_xs
  • qwen3.5:9b-opus-openclaw-distilled-q4_k_m
  • qwen3.5:9b-opus-openclaw-distilled-q6_k
  • qwen3:14b-iq4_xs
  • qwopus3.5:9b-v3-q6_k (WINNER!) - **Tested Models (3080 8GB):**
  • deepseek-r1:8b
  • qwen3.5:9b-uncensored-aggressive-q4_k_m
- **Tested Models (1080 8GB):**
  • qwen3:4b
  • qwen3:8b
  • qwen3.5:9b-iq4_nl
  • qwen3.5:9b-q4_k_m (works great as a classifier!)
- **Tested Models (3080 + 1080 18GB):**
  • devstral-small-2:Q4_K_M (CPU offload)
  • gemma-4-26b-a4b-it:Q4_K_M (CPU offload)
  • qwen3.5:35b-a3b (CPU offload)
{/* A multi-agent CLI router with the same docs-first discipline. It routes prompts across specialized workers that share context. The starting assumption: no single model is the right choice for every step in a long task. It pairs with codex-local on the local side and with cloud providers when something needs the bigger model. */} - **Find the upstream fix.** Mitigations, fallbacks, band-aids are the worst. If we don't own upstream, then mitigate. - **Docs-first.** Then architecture. Docs are load-bearing artifacts the agent reads on every turn. - **Long autonomous runs are normal.** I let agents work overnight regularly. - **State in files, not in-context.** Long runs survive crashes, context exhaustion, model swaps, and OAuth expirations. The agent rehydrates from disk. - **No green-by-inference.** Verify before declaring done. Check the deployed version, inspect the running process, read the Discord channel. - **AI is infrastructure, the same as a database.** I instrument it, test it in tiers (hermetic, e2e, live-canary), and put silent failures on a heartbeat (OAuth expiry, model health) the same way I'd put a dead database on one. - **Platform-engineering reflexes don't go away just because it's AI.** I want redundancy, failover, quick recovery, auto-scaling, and reproducible deployment, with low operational overhead.