exeterminal / Exe-Guard-Dynamic-GGUF

Exe Guard Dynamic

A tiny guardian model for the Extended Workflow feature of the Exe AI Terminal Website to the Main AI Agent Harness: https://exe-hq.net When a tool step fails, it reads the report of that failure and writes the one instruction the user should send next to fix it — a corrected command, the real file, a path inside the shared folder. It suggests; it never acts.

What it does

The guardian watches tool calls and speaks up only when a step failed in a way a person would want a suggestion for. It is built for exactly six kinds of failure:

It answers in one imperative English sentence , no greeting, no explanation.

Intended use

Drop-in as the small background model behind the Exe AI Terminal's Extended Workflow. It is a specialist: it turns a failed-step report into a single corrective instruction.

Out of scope: general chat, code generation, vision, or any use outside the failed-step-repair task. It is not a general assistant.

Files

All builds carry an importance matrix (imatrix) computed from the model's own task data, and were tested on 24 held-out repair cases at temperature 0.1 . "Test" is the number of those 24 cases solved correctly — a task metric, not perplexity.

Note: at low bit-widths the I-quants (IQ3, IQ2) hold up better than the K-quants ( Q3_K drops to 19–20/24). The 1-bit builds are included for the curious but are not recommended.

Prompt and sampling

Two messages only — a fixed system instruction and the failure report — with temperature 0.1 , max_tokens 200 , thinking off, context 4096. No conversation history.

Base model and license

• Base: Qwen/Qwen2.5-Coder-3B-Instruct
• License: Apache-2.0 (base and this derivative). You may use, modify, rebrand and redistribute; the origin of the base model must be named — it is, here.

Training

A thin LoRA adapter on top of the base, trained locally (Apple Silicon, MLX) on synthetic examples of failed tool steps and their one-sentence corrections, built to match the exact report format the terminal produces. The adapter was fused into the base and then quantized.

Evaluation

On 24 held-out repair cases at temperature 0.1 , the untrained base solves 10 / 24 (42%) and Exe Guard Dynamic solves 24 / 24 (100%) . The base fails completely on the two cases that need this training — wrong Python environment and wrong edit target (0%) — which the trained model fixes entirely.

Transparency

This is a fine-tuned derivative of an openly licensed base model, released with its provenance, intended use, limits and evaluation stated above, in line with transparency expectations for shared models (incl. the EU AI Act).