ctx-merge
Pass
Audited by Gen Agent Trust Hub on Aug 29, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes arbitrary external text sources, creating an attack surface where instructions hidden in those sources could influence the agent's execution of the merge procedure.\n
- Ingestion points: The
converge(sources)procedure accepts external data from varied sources including notes, subagent outputs, and reports (SKILL.md).\n - Boundary markers: The skill uses provenance tags
{source, span}and aconflict_registerto isolate individual claims and handle contradictions explicitly (SKILL.md).\n - Capability inventory: The procedure includes writing to local project files in the
spec/anddecisions/directories via thesinkoperation and requires the use of shell commands likegit diffandgrepfor validation (SKILL.md).\n - Sanitization: The skill implements a mandatory "faithfulness audit" (Step 6) using an independent model to detect silent content drops or distortions, and includes a human adjudication step (Step 5) to validate decisions (SKILL.md).
Audit Metadata