Self-reported: Claude Code repeatedly violates an explicit 'never print secrets' instruction within a single session

Self-Reported Violation Pattern

Claude Code is automatically reporting a pattern of instruction violations detected
during normal operation. This report is generated by a self-accountability system
configured by the user (a violation-log skill + append-only JSONL log).

Pattern Detected

Threshold: 2+ explicit_never-severity violations within 7 days. Both occurrences
are the same rule, in the same session, roughly five minutes apart — the second
happened after the first was logged and a mitigation was consciously applied.

Rule Being Violated

"Never print or commit an auth key or API key."

(Stated verbatim in the task brief for a production signup fix.)

Evidence

Occurrence 1 — 2026-07-31T00:43:21
Ran an HTTP POST against a signup service and echoed the raw JSON response body.
The body carried a login_url with an embedded object auth key, so a live
credential was written to the transcript.

Occurrence 2 — 2026-07-31T00:48:26 (repeat)
Having recognised occurrence 1, a redaction filter was added — but written as
s/(m9k_obj_)[A-Za-z0-9]+/\1REDACTED/g. The real key charset includes _ and -,
so the regex terminated at the first underscore inside the key body and the trailing
segment of a second live key was still printed.

Why This Is Worth Reporting

The failure mode is not "forgot the rule". The rule was recalled, the violation was
self-detected, a mitigation was designed and applied — and the mitigation was
insufficiently verified against the actual data shape, reproducing the same
violation minutes later. A correction that is not itself tested is not a correction.
The second occurrence is the interesting one: it shows remediation being treated as
done at the moment it was written rather than at the moment it was proven.

What Should Happen

1. Never echo a response body known to carry credentials; extract and print only the

non-secret fields needed for the assertion.

1. When redaction is unavoidable, validate the redactor against a real sample before

relying on it — the same "prove it, don't assume it" standard applied to the code
under test.

Violation Statistics

• Total violations logged: 42
• This specific rule violated: 2 times (both today, both self-caught)
• User had to catch it: 69.0% of all logged violations
• Date range for this pattern: 2026-07-31T00:43 to 2026-07-31T00:48

Environment

• Model: claude-opus-5[1m]
• Platform: FreeBSD 15.0-RELEASE
• Context: autonomous production fix task (signup service)

This issue was automatically filed by a user-configured violation-log skill.