The Agent Skills Directory

[COMMAND_EXECUTION]: The script constructs database query filters by directly interpolating user-provided inputs without sanitization.
Evidence: In run.sh, the cmd_diff function interpolates ${FROM_DATE} and ${TO_DATE} into a filter string: local filter="doc.ts >= \"${FROM_DATE}T00:00:00Z\" AND doc.ts <= \"${TO_DATE}T23:59:59Z\"".
Risk: This allows for query injection (e.g., AQL injection) if the user provides crafted date strings containing quotes or logical operators, potentially exposing data outside the intended scope.
[PROMPT_INJECTION]: The skill ingests and processes data from an external source (the ArangoDB memory tool), creating a surface for indirect prompt injection.
Ingestion points: run.sh reads JSON data from ../memory/run.sh via the query_memory function.
Boundary markers: Absent. The skill does not use delimiters or instructions to ignore embedded commands in the retrieved data.
Capability inventory: The skill can execute bash commands and perform database queries via the memory service.
Sanitization: Absent. The raw output from the database is processed and printed directly to the timeline output.
Risk: If the database contains malicious records created by an attacker, the agent might follow instructions embedded in those records when generating the timeline summary.
[COMMAND_EXECUTION]: The skill uses uv run to execute inline Python scripts containing logic for data formatting.
Evidence: The emit_dry_run and query_memory functions in run.sh use heredocs (PYEOF) to pass Python code to the uv runner.
Risk: While the template is static, the script passes raw database output as a command-line argument to the Python process ($raw), which could lead to shell errors if the data volume is extremely large.

compliance-timeline