Bounded TypeSafe Jev decisions for Aside browser agents — pick one app-owned action ID, never invent clicks.
Six tools with explicit schemas and descriptions, but significant gaps in clarity and LLM-friendliness. Tool names lack clear action verbs (jev_system_one, jev_choose, jev_step are domain-specific jargon). Descriptions are lengthy (200-400 chars) and assume deep Jev/Aside knowledge; they do not explain WHEN to use each tool or how they differ. Parameter descriptions exist but are sparse. Output schemas are partially documented in code but not formally declared. Error handling is minimal. The jev_browser_run tool is particularly complex with many interdependent parameters but lacks clear guidance on parameter relationships.
Run observe → Jev → execute in one persistent Aside REPL session. action_rules are explicitly authorized exact role/name/action rows; fill also requires an explicit value. Never include an action needing ungranted user approval. Supports click/focus/fill only. Uses live Jev, never a fallback LLM. A completion_text and optional exact completion_url verify success locally. Includes completion Noul and risk Score assessment after DOM completion.
Choose an app-owned action ID. No action is executed by this tool.
Read the popup toggle state; this does not change settings or call Jev.
Choose, validate, and apply a confidence gate; return an execution payload.
Call TypeSafe Jev with named questions. Each question uses type, instructions, criteria ONLY, never question or rubric. Score requires 2..64 ordered criteria from lowest to highest score. Noul uses true/false criteria descriptions. Example questions: {"completed":{"type":"noul","criteria":{"true":"The requested completion is visible","false":"Completion is absent"}}, "risk":{"type":"score","criteria":["No sensitive action","Reversible action","Sensitive or irreversible action"]}}. Report any error as an incomplete evaluation, never as overall success.
Tool names lack action verbs. 'jev_system_one', 'jev_choose', 'jev_step' are domain jargon; LLMs cannot infer intent from names alone. Should be 'decide_with_system_one', 'choose_action', 'execute_decision_step'.
Descriptions assume Jev/Aside domain expertise. No guidance on WHEN to use jev_choose vs jev_step vs jev_system_one. LLMs cannot disambiguate without explicit comparison.
jev_browser_run has 10 parameters with complex interdependencies (action_rules, completion_text, completion_url, min_confidence) but no documentation of which are required together or how they interact.
Output schemas not formally declared in tool definitions. Return types inferred from code (e.g., jev_choose returns dict with choice_id, confidence, probabilities, candidate, timing_ms) but not documented for LLM consumption.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | 2026-07-28+ | v2 |
Validate a supplied candidate ID and return its original action payload.
Error handling minimal. No recovery guidance. ValueError for invalid provider in jev_choose, but no actionable error messages for LLM self-correction.