Σ₀ Coder in Claude Code — Integration Status

Can the local Σ₀ Ouro-1.4B coder drive Claude Code (and the in-app chat) as a tool-using agent, instead of a frontier cloud model? This is the honest status. For the model itself see Σ₀ Ouro Coder; for the next training step see the retraining handoff; to run it on a second machine see the dev-box setup guide.

The gap, and the bridge that closes it

Claude Code speaks only the Anthropic Messages API (/v1/messages); the Ouro server speaks the Ollama API. scripts/ouro_anthropic_bridge.py translates between them — injecting the tool definitions into the prompt, parsing the model's <tool_call> back into Anthropic tool_use blocks, and supporting forced tool_choice (prefill). The protocol round-trips cleanly [verified: curl + headless claude -p].

What works

Bridge — Anthropic ↔ Ollama with tool injection + parse + forced mode. ✓
Standalone agent loop (scripts/sigma0_coder_agent.py) — forced-ReAct over sandboxed

tools; drove a real task end-to-end and finished grounded (counted.py files). ✓

Tool-aware in-app chat — the garage renders the adapter's <tool_call> as a card and

(gated by CHAT_TOOL_EXEC, operator-gated for shell/mutating) executes it through one canonical tool registry. Merged to master (v1.7.14). ✓

FC-retrained adapter — a QLoRA pass taught the trigger: under auto tool_choice it

now spontaneously emits tool_use, where the stock adapter was 0/3. ✓

The honest blocker — model reliability, not plumbing

The integration is solved; the 1.4B adapter is not reliable enough to drive Claude Code:

It under-triggers / over-refuses — 0/3 tool calls on Claude-Code-shaped requests

(memorized refusal templates), erratic across phrasings [verified: bridge test 2026-06-21].

It is Bash-biased (the harvest was Bash-heavy) and weak at multi-step grounding.
Hard ceiling: Claude Code's ~20k-token system prompt overwhelms a model trained at

seq 1024.

Net vs. the first attempt: stock model couldn't call tools at all; retrained model can, just not reliably yet. Real progress, not done.

Next lever

The retraining handoff: rebalance the corpus (downweight Bash, diversify therefusal strings → 150+, drop negatives 25%→~10%), add Hammer function-masking,epochs — targeting the BFCL irrelevance/relevance axis. A reliable trigger is the gate before a full Claude-Code run is worth doing (with the live session stopped, so two model processes don't crash theGB GPU).