GripProbe Reports

Shell Behavior (High Level)

GripProbe compares model × shell × backend compatibility. Different shells are expected to behave differently even on the same model and test prompt.

Shell Comparison

Shell Interaction style Formats in GripProbe Model routing Runtime isolation
gptme Explicit tool loop (read/patch/shell/save) markdown, tool CLI -m + OPENAI_BASE_URL from OLLAMA_HOST Per-case HOME/XDG runtime dirs
continue-cli Agent run with explicit --allow tool permissions markdown, tool Per-case generated .continue/config.yaml Per-case HOME + isolated Continue config
opencode JSON event stream execution (opencode run --format json) markdown, tool Per-case generated opencode.json + --model Per-case HOME/XDG runtime dirs
aider File-edit chat loop (no separate tool protocol in output) markdown CLI --model (ollama/<model> for Ollama) Per-case HOME + isolated .aider.conf.yml
codex OSS tool-call loop (codex exec --oss --json) tool CLI provider/model + CODEX_OSS_BASE_URL from OLLAMA_HOST Per-case CODEX_HOME, ephemeral run

Why Results Differ Across Shells

How To Read Test Outcomes

Codex Local Model Metadata

For local OSS models in Codex CLI, missing metadata can cause warnings like: Model metadata for <model> not found. Defaulting to fallback metadata.

GripProbe includes a local catalog file:

Interactive run example:

codex --oss --local-provider ollama \
  -m aravhawk/qwen3.5-opus-4.6-text:9b \
  -c 'model_catalog_json="/absolute/path/to/GripProbe-CLI/configs/codex-model-catalog.local.json"'

Persistent setup (~/.codex/config.toml):

oss_provider = "ollama"
model_catalog_json = "/absolute/path/to/GripProbe-CLI/configs/codex-model-catalog.local.json"