System 1 decision models (TypeSafe Jev, Laya, Cua-S1) in the model slot of openJiuwen agents, with browser, tool and rail agents and evals against a chat model on the same agent.
Three tools with mixed quality. list_agents and run_agent have reasonable descriptions (100+ chars) but lack formal input schemas, only run_agent shows parameter details. decide has a vague description (67 chars) and unclear purpose. No tool has documented output schemas. Parameters lack type constraints (enums, ranges). Error handling and recovery guidance absent. Naming is acceptable (verb_noun style) but descriptions do not explain WHEN to use each tool or what downstream tools expect.
One choice question over options you enumerate: the chosen key, a probability per option, a confidence, the latency in ms.
Every S1A agent: its name, its front (tool, browser or rail), what it does and the flags run_agent takes for it (flag, takes_value, required, choices, default, help; a flag with takes_value false is passed alone); an agent that fails to import, a missing optional dependency included, is listed as unavailable with the error instead.
Run one agent with the flags `s1a run <name>` takes, e.g. ["--model", "jev", "--goal", "..."]. A browser agent returns its answer, a tool agent its series summary with the job folder it wrote, a rail its evaluation summary.
list_agents has no visible input schema definition; schema score must be 0 .
No tool documents its output schema. LLMs cannot plan downstream calls or extract required fields (e.g., agent names, job folder paths, decision probabilities).
decide description (67 chars) is vague: 'One choice question over options you enumerate...' does not explain when to use it vs. agent tools, what 'state' and 'rules' contain, or how probabilities/confidence are computed.
run_agent flags parameter is an array of strings with no enum, pattern, or validation guidance. LLMs cannot know valid flag names (--model, --goal, etc.) or their argument formats.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | <=2025-11-25 | v2 |
No error handling or recovery guidance. If an agent fails to import or a flag is invalid, the tool provides no actionable next step for the LLM.