About 150 unlogged paid AI calls worth $8 spur a single budget gate

The unlogged calls were not the largest problem in this account; the fact that no one noticed was. The gate described here counts usage as the larger of the ledger total and the API key's actual usage, so a missing entry in the ledger is still caught. It also stops rather than reading an unreadable ledger as zero, a design choice tied directly to the "nothing reported" failure.

Other fixes in the account follow the same shape. Before text goes to an external model, a redaction function masks payment numbers, API keys, and token-like values, on the principle that over-masking is safer than a leak. Keys and workspace IDs are read only from a file, never from environment variables, and their values are kept out of logs, exception messages, and return values. Framework accidents — a command chain that continued after a failure and pushed from the wrong working tree — led to set -e for write sequences and checks of the current tree and branch before a push, plus bans on force push and on bypassing hooks with --no-verify.

The throughline is that attention is placed in the mechanism, not in the operator. Agents do not get tired, but they also do not carry over carefulness, so the account favors gates that must be passed before anything proceeds and settings that stop on failure. Money is handled the same way: real funding, charges, and identity checks are done only by people, while AI operates within a pre-approved range such as free tiers or a fixed budget.