Loading ToolBench
Deterministic safety/state helpers for the /vizier Claude Code skill — risk limits, conviction sizing, autonomy ceilings, thesis store, decision log, exposure reconciliation, and data-sufficiency gate for agentic trading
25 tools with complete input schemas and descriptions, but quality is uneven. Most descriptions are adequate (50-150 chars) and parameter types are present. However, output schemas are entirely undocumented, LLMs cannot predict return structures. No error handling guidance, no recovery paths, and no tool annotations (readOnlyHint/destructiveHint). Risk classification is present in metadata but not surfaced in descriptions. Tools like 'size', 'allocate', 'ceiling' perform complex financial calculations but lack worked examples or constraint documentation. Naming is generally clear (verb_noun pattern) but some tools like 'tranches' and 'tranche_sell' could be more explicit about their relationship.
Split a fixed budget across candidates, then cap each leg at the per-asset NAV limit
Append a decision record to the decision log
Assess whether scout responses meet minimum evidence requirements for a decision type
Check circuit breaker status based on VIX and monthly drawdown
Build own sent orders list from decision log within a time window
Check cumulative daily ceiling against baseline NAV and spending limits
Single per-candidate autonomy verdict composing armed state, per-run ceiling, cumulative ceiling, and drawdown kill checks
No output schemas documented. LLMs cannot predict return structures for any of the 25 tools. This violates pattern:tool and forces agents to guess at response fields, increasing errors and wasting tokens on exploratory calls.
No error handling guidance. Tools like 'allocate', 'size', 'ceiling' perform financial calculations but provide no recovery paths if inputs are invalid (e.g., negative NAV, conviction out of range). Agents cannot self-correct.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | 2026-07-28+ | v2 |
Close a thesis record with exit details
Compute drawdown over a specified window
Check if NAV drawdown exceeds the kill threshold
Size a protective stop / any exit so it can never exceed the holding
Check position limits against the risk profile
List all open theses
Record a NAV snapshot for drawdown tracking
Load and return the active risk profile configuration
Read a thesis record from persistent storage
Reconcile real exposure against own sent orders and broker positions
Reduce the quantity held in a thesis due to a partial sale
Compute a performance scorecard for all theses as of a given date
Linear conviction sizing scaled by the conviction-full-size knob, hard-capped at the per-asset NAV limit
Check if a tranche sell is allowed
Get tranche balances for a ticker
Size a tactical trim so it respects position limits and tranche constraints
Update the last reviewed timestamp for a thesis
Write a thesis record to persistent storage
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Risk metadata exists (READ_ONLY vs WRITE) but is not exposed in the MCP tool definition. Agents cannot distinguish safe reads from destructive writes without parsing descriptions.
Complex financial tools lack constraint documentation. 'size' accepts conviction 1-5 but description does not state range. 'allocate' accepts weighting modes but does not enumerate valid values. LLMs will hallucinate invalid inputs.
Parameter descriptions are generic and lack worked examples or context. 'conviction' is described as 'Conviction level (1-5)' but does not explain what conviction means in the trading context or how it affects sizing. Agents cannot reason about appropriate values.