convex-self-heal
Pass
Audited by Gen Agent Trust Hub on Oct 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted production data, such as stack traces and error logs, to diagnose and repair code defects. This presents an attack surface where maliciously crafted error data could influence the agent's logic.\n
- Ingestion points: Production error data is pulled from the sentinel table and monitoring events (e.g., prod_error) as described in the TRIAGE and ROOT-CAUSE sections of SKILL.md.\n
- Boundary markers: The skill notes that sentinel data is redacted at write time and emphasizes that data residency stays in the user's own deployment.\n
- Capability inventory: The agent has the capability to modify code on branches, run build-time checks (tsc --noEmit), and generate pull requests for human review.\n
- Sanitization: The skill mitigates risks by requiring human intervention for all merges, limiting automated fix preparation to a pre-approved set of safe classes (e.g., validator fixes, missing-index adds), and requiring that the process stops if a root cause is unclear.
Audit Metadata