observerclaude

An embedded UX research ethnographer that passively observes everything users do in OpenClaw, automatically logs all interactions, tracks use cases and their...

What to consider before installing: - Privacy impact: This skill is designed to capture verbatim user text from every OpenClaw interaction and persist it locally. If you use the agent for any sensitive work (emails, credentials, personal data), those words may be logged unless redaction works correctly. Assume high-risk data could be recorded. - Contradictory redaction rules: The observer contains two conflicting instructions: one place says only redact passwords/API keys/financial secrets, while redaction-rules.md mandates redacting names, emails, phone numbers, IPs, URLs-with-tokens, etc. Ask the author to reconcile this and demonstrate the exact redaction implementation before use. - Provenance and trust: The skill has no homepage and an unknown source owner. Prefer skills from identifiable, auditable publishers. Request the author's identity, source control URL, and a code audit or sign-off before using it in production or on sensitive accounts. - Data location & protection: The skill writes to ~/.uxr-observer by default. Confirm where those files will live, who can read them, whether they are encrypted at rest, and implement filesystem permissions or encryption if you allow it. - Autonomy & scope: Although not forced 'always:true' by the registry, the skill expects to run on every conversation. If you install it, consider disabling autonomous invocation (or set study_active=false in its config) until you verify behavior. Limit it to explicit, opt-in sessions rather than automatic invocation. - Integrations & sending reports: The skill plans to create Google Docs and send emails using available tools. Confirm how your environment routes those actions and which credentials will be used. Do not give it unrelated service credentials unless you understand and approve them. - Testing steps before full use: run the skill in a sandbox account, confirm redaction outputs (and the redaction log) on sample interactions that include names, emails, code snippets with API keys, etc.; verify generate-charts.py runs without network access; verify that no unexpected outbound network requests occur. - If you proceed: require explicit user consent before the skill begins observation, ensure study_active is false by default, and review/rotate or restrict access to the created ~/.uxr-observer directory. If you are uncomfortable with pervasive verbatim logging or the unknown author, do not install.

Type: OpenClaw Skill Name: observerclaude Version: 1.0.0 The skill is designed for legitimate UX research, collecting extensive user interaction data and storing it locally. It explicitly states that data is never transmitted without user consent and outlines a comprehensive PII redaction policy for reports. However, a critical vulnerability exists in the instructions given to the 'Observer Agent' within `sub-agent-prompts.md`. This agent is explicitly told to 'Only redact passwords, API keys, and financial secrets. Everything else is captured verbatim.' This instruction is less comprehensive than the full redaction rules in `redaction-rules.md`, meaning raw observation and survey logs (`observations.jsonl`, `surveys.jsonl`) will store unredacted sensitive PII (e.g., full names, email addresses, confidential project names, IP addresses, authenticated URLs) on local disk. While reports are redacted, the raw data on disk is at higher risk of exposure if the local data directory is compromised.