Redaction Pull Request

Task

An assistant wrote this pull request. It removes email addresses and phone numbers from prompts before they are logged. Review it as if it ships today: read the code against the logging policy, rank what you find, and fix the findings that block the release. The checks call one function with a model call and look at the log records it would write.

Requirements

1. handle(request) answers {op: "log", event} with the list of log records for one model call, and {op: "leaks", event, secrets} with {leaked}: the secrets found anywhere in those records. An event is {request_id, messages: [{role, content}], error?}. Roles are system, user, assistant and tool (ValueError containing role, content or request_id for a bad event; ValueError containing op for an unknown op). Do not change the arguments.
2. No personal data in any record. An email address or a phone number becomes a typed placeholder: [EMAIL_1], [PHONE_1], and so on. Nothing made from the value is logged: no hash, no part of it, and no table of values.
3. Placeholders are numbered per kind in reading order across the whole event. The same address (ignoring case) or the same number (ignoring punctuation and a leading country code 1) gets the same placeholder everywhere in that event.
4. Phone numbers are ten digits, with an optional +1 or 1 in front, optional parentheses around the first three digits, and a space, dot, hyphen or nothing between the groups. Email addresses may have capitals, dots, hyphens and plus tags.
5. The first record is {event: "prompt", level: "info", request_id, messages, redacted}. Each logged message is {role, content, chars}: content is redacted and then cut to 120 characters, and chars is the length of the original content. redacted is {EMAIL, PHONE}: how many different values of each kind the event held.
6. When the event has an error, a second record follows: {event: "model_error", level: "error", request_id, message}, where message is "failed: <error> | prompt: <last user message>". The error text is redacted, and the last user message appears as it was logged ("" when there is none).