Inspect secret names and metadata only when sufficient; do not print the backing values to learn what they contain.
Do not read environment files or credential stores to display values. Do not paste secrets into source, docs, prompts, command arguments, issue comments, logs, or commits.
When one service needs another service's credential, use the provider's secret reference/identity mechanism rather than reading and copying the resolved value.
For write operations, prefer stdin, a secret manager, or a scoped reference; avoid shell history and process arguments that may expose values.
Redact logs and diagnostic output before sharing them. Check output files and staged diffs for accidental credentials before publishing.

If a secret may have been exposed

Do not repeat or copy the value into another location.
Tell the user what kind of material may have been exposed and where, without reproducing the credential.
Recommend revocation/rotation and assess the exposure window and affected systems.
Remove the value from active files and reachable surfaces when authorized; do not rewrite shared history or perform destructive cleanup without authorization.
Verify remediation using names, status, or fingerprints that do not reveal the secret itself.

Before completing a task

No credential value appears in tool output, logs, diffs, or response text.
No literal secret was written to configuration, examples, scripts, or docs.
Any diagnostics shared with the user are redacted.
Rotation or cleanup that remains is clearly identified rather than claimed complete.