Skip to content

Audit, logs and incident evidence

Operational evidence should identify an operation, actor/scope where appropriate, version and outcome without copying credentials, personal data or full private payloads into broader audiences. Toolora has several evidence stores with different authority and access. A log line, business receipt and operator incident must not be treated as interchangeable.

The working agreement keeps secrets and personal data out of Git, logs and review artifacts. The implementation has targeted controls; it does not establish automatic universal redaction of every exception or imported message.

Evidence Current mechanism Interpretation and limit
Business audit/version/receipt Owning transaction and tenant/resource records Establishes the named decision or result, not every downstream provider outcome
Operator incident Fixed allowlisted event code maps to internal title/severity/service/text No caller-supplied request path, feature name or exception text enters the incident row
Application exception log Access-controlled runtime logger, including global handler details Raw detail can remain sensitive; a safe public error does not sanitize the underlying log
Import-job diagnostics Counts, error entries and bounded traceback excerpt Can contain free text; not suitable for automatic public review export
AI usage and lifecycle records Named usage/event contracts and bounded context Usage metadata is not authority to copy prompts or private source bodies
Frontend diagnostics Optional Sentry setup, trace sampling and disabled replay sampling in inspected bootstrap Configuration is not proof of all downstream telemetry’s data handling
Git/artifact scanning PII classes and pinned secret scanner A passing scan is a bounded detector result, not a data-classification guarantee

incident_reporter accepts only the fixed event-code set. Unknown codes are refused before client creation/allocation. Internal templates supply persisted row values; the operator correlates the incident with separately controlled application logs for occurrence detail.

The current fixed codes include unhandled backend failure, readiness failure, AI-provider failure and a user-reported acute fault. A five-minute in-process deduplication window limits repeated reports in that process. It is not a durable cross-process exactly-once or guaranteed alert-delivery mechanism.

The API returns generic detail for an unhandled 500 while logging the exception internally and reporting the fixed incident class. Complete-read refusal and disabled outbound messaging have their own intentional responses. Error diagnosis distinguishes these from successful or uncertain business outcomes.

The import dispatcher stores returned error entries and, on exception, str(exc) plus the final 2,000 characters of traceback in the job diagnostic fields. That bound limits size, not sensitivity. Before exporting a receipt, select safe source identities/counts and outcome codes rather than copying the full diagnostic payload.

Provider errors can also carry request or source detail. Keep message bodies, identity numbers, bearer tokens, signed query strings and file contents out of new logs. A short hash or internal reference can support correlation where the owning policy permits it, but is not automatically anonymous or authorization to broaden access.

The mail contract explicitly forbids real source content in fixtures, logs and review artifacts. Data lifecycles and AI context describe the adjacent audience and preservation boundaries.

Current/history PII scans and pinned secret scanning run through the shared security target. Exceptions are explicit by path/class in the committed contract; an unreviewed blanket allow marker is not an acceptable way to silence a result. Changing a baseline or scanner rule to hide sensitive output is not correction.

An existing leak needs the owning incident/remediation procedure, including affected credentials and retained copies. Do not paste a suspected secret into a PR, graph or another external service to explain it. This review performs no credential rotation or history rewrite.

A synthetic provider exception includes private input text. The operator incident must still contain only its fixed event template, and the user receives the generic error. Verification checks that no caller detail reached the broad incident rows. It must separately inspect any changed logger/import-diagnostic path; the fixed reporter does not sanitize those paths on its behalf.

Reporter and scanner tests specify allowlists, refusal and detection using synthetic inputs. Security scans establish results for their checked commit/history. They do not certify all runtime logs, complete audit coverage or a successful combined recovery rehearsal. Keep the actual commands, version and limitations in the PR.