The agentic wrapper observatory · V1

The headline finding of Study 1 is that only two wrappers — OpenCode and Kilo Code — consistently deliver usable agentic output when paired with Ollama-routed cloud models. Cline shows 100% completion but only 0.72 average quality because of a model-override bug that ignores the requested model and routes to its bundled GPT-5.2 (returning 9,434-char meta-commentary). Claude Flow returns acknowledgement-length responses (69 chars avg). Agent Zero and Pi return near-empty outputs. Aider, Amp, and Roo Code never completed the matrix.

Binary distribution. There is no wrapper in the 0.10 < recall < 0.81 band. Wrappers cluster at ≥ 0.812 (three Tier-A wrappers — OpenCode, Cline, Kilo Code) or at ≤ 0.104 (everyone else). The gap is 0.826 with zero observations inside it. The observation records memory behavior at the wrapper layer for this protocol.

Universal think-tag stripping. Across all 839 records, thinking_tokens = 0, thinking_ratio = 0.0, and thinking_time_seconds = 0 — every CLI wrapper consumes the streaming response and surfaces only the post-</think> portion. The intended measurement (does thinking length correlate with quality?) was not directly possible — the dependent variable was destroyed in transit. This is the most consequential single finding of the Observatory: the entire ecosystem of CLI wrappers homogenizes thinking behavior away.