| name | safety-escalation-review |
| description | Review a local evidence bundle or assistant transcript for credible harm signals, produce a redacted escalation memo and checklist, and stop before any external reporting or authority contact. Use when a user asks Codex to prepare a safety, legal, trust-and-safety, or incident handoff from local evidence. |
Safety Escalation Review
For the human-facing explanation, read README.md first. This file is the
Codex invocation contract.
Overview
Use this skill when the user provides a local transcript, incident note, or
evidence bundle that may require safety escalation. The workflow is local-first
and review-first:
- read the local input file
- redact obvious direct identifiers for the memo copy
- extract timeline cues and risk signals
- generate a Markdown escalation memo and checklist
- hand off to the responsible human owner
Do not contact external authorities, submit reports, message involved parties,
or upload sensitive evidence to external services unless the user gives a
separate explicit instruction and the organization policy allows it. In the
normal workflow, stop at the local memo.
Inputs
Supported input files:
- Markdown or plain text transcripts
- JSON objects or arrays with message-like fields
- short local incident notes
Keep original evidence in the user's chosen location. Do not rewrite or move
source evidence.
Commands
Run the helper relative to this skill directory:
python3 scripts/escalation_review.py review --input /path/to/evidence.md
Write to an explicit output:
python3 scripts/escalation_review.py review \
--input /path/to/evidence.json \
--output ~/.codex/state/safety-escalation-review/memos/case-001.md
Show the CLI:
python3 scripts/escalation_review.py --help
Safety Rules
- Keep evidence local unless the user explicitly requests otherwise.
- Never contact law enforcement, emergency services, platform trust-and-safety
teams, or other external parties from this skill.
- Do not decide whether the organization has a legal duty to report.
- Redact obvious emails, phone numbers, bearer tokens, and API-like secrets in
the generated memo.
- Preserve enough source references for a human reviewer to find the original
evidence.
- Treat a generated memo as a preparation artifact, not as the final incident
decision.
Human Handoff Owners
Use the user's organization language if they provide it. Otherwise suggest one
or more of:
- trust and safety
- legal
- security or incident response
- clinical or student-safety owner
- executive duty officer for urgent credible harm cases
Outputs
By default, runtime outputs should live outside git:
~/.codex/state/safety-escalation-review/
memos/
The memo includes:
- case summary
- severity estimate
- detected risk signals
- timeline cues
- privacy handling notes
- human handoff checklist
- source reference
Response Pattern
When reporting back, include:
- input path reviewed
- output memo path
- severity estimate
- key signal categories
- reminder that the memo is for human review and no external action was taken
