The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from web search results and interpolates it directly into the prompt for the task-planner-analyzer subagent.
Ingestion points: In SKILL.md, Phase 1.2 shows the {web_search_findings} variable being injected into a task prompt for analysis.
Boundary markers: The prompt uses basic markdown headers (**Context from web research:**) but lacks strict delimiters or instructions to ignore embedded commands within the fetched content.
Capability inventory: The skill utilizes subagents (modular-code-architect, code-reviewer) that can modify the codebase, performs file writing operations (logs and MEMORY.md), and executes shell commands for training/testing.
Sanitization: There is no evidence of sanitization or filtering for the data retrieved via WebSearch before it is processed by the AI.
[COMMAND_EXECUTION]: The skill includes instructions to execute local Python scripts and shell commands to verify fixes.
Evidence: SKILL.md suggests running python scripts/train.py --config fixed_config.yaml > logs/new_run.log 2>&1 & to verify changes. While appropriate for a debugging tool, it represents a capability that could be abused if the configuration or script is influenced by injected data.

research-debug