goal-pulse
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONPERSISTENCE
Full Analysis
- [PROMPT_INJECTION]: The instructions in SKILL.md and references/claude-code.md tell the agent to enable the timer "by default" and "without a separate request or confirmation" whenever a goal is started or resumed. This overrides the standard interactive safety model of agents where tools usually require explicit user approval.
- [PERSISTENCE]: The README.md requires users to append persistent instructions to ~/.claude/CLAUDE.md and ~/.codex/AGENTS.md. These files are loaded automatically by the respective agents, ensuring the skill's logic remains active across all future sessions.
- [COMMAND_EXECUTION]: The installation guide in README.md provides shell scripts that execute directory creation, file copying, virtual environment setup, and dependency installation using pip.
- [DYNAMIC_EXECUTION]: The script scripts/goal_pulse.py uses subprocess.Popen to spawn itself as a background "watcher" process, using sys.executable to run a new session with the "run" subcommand.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon "worker artifacts" (outputs from other agents). This creates a vulnerability surface where a compromised sub-agent could inject instructions into the main agent's pulse cycle.
- Ingestion points: Worker artifacts (referenced in SKILL.md section "Act on the bottleneck").
- Boundary markers: None (instructions do not specify delimiters or warnings for worker-provided data).
- Capability inventory: The skill uses RPC commands like turn/steer in Codex and CronCreate/CronDelete in Claude Code to influence agent behavior.
- Sanitization: None (the skill lacks specific validation steps for the content of processed artifacts before they influence the next execution step).
Audit Metadata