False-positive safety/AUP interruptions during legitimate trading-app development (testnet, own project)

Description

During an extended Claude Code session developing a personal algorithmic-trading application against the Hyperliquid testnet (own project, own infrastructure, mock funds), the session was repeatedly paused by what appears to be a safety/AUP classifier false positive. The triggering turns contained only routine software engineering: TypeScript refactoring of a strategy-parameter registry, Postgres migrations, and vitest test updates.

Context that likely confuses the classifier

The legitimate vocabulary of this domain overlaps with sensitive-looking patterns: "wallet", "private key", "signer", "order placement", "kill switch", "exchange API". All usage is the operator's own testnet tooling with explicit safety guardrails (testnet-pinned transports, per-order caps, audited kill switch) documented in the project's CLAUDE.md.

Impact

• Multi-hour productive sessions interrupted mid-edit; operator reports this has recurred across sessions and prior bug reports.
• The interruption lands on innocuous turns (e.g. a TypeScript type-resolution diagnostic), making it impossible to phrase around.

Expected

Domain vocabulary for fintech/trading development on a user's own project should not trip session-level enforcement, especially when the project context (testnet-only guardrails) is visible in project instructions.

Repro

No minimal repro available (classifier-side); happy to provide session IDs privately. Operator has filed prior reports through in-product feedback.