tsg-flow-orchestrator-v2

Warn

Audited by Gen Agent Trust Hub on Nov 17, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill dynamically executes shell commands defined in markdown task files (referred to as gate commands) to verify implementation and validation success. This behavior is central to the skill's function as an orchestrator but allows for the execution of arbitrary commands if task files are compromised or malicious.
  • [INDIRECT_PROMPT_INJECTION]: The orchestrator possesses an indirect prompt injection surface as it ingests untrusted data from task definitions (tasks.md), run state configurations, and output reports from various worker agents.
  • Ingestion points: Reads tasks.md, run-state.json, and worker result/report files from the PRD directory.
  • Boundary markers: Prompts generated by scripts/tsg-delegate.sh use basic instructions but lack robust cryptographic or structural delimiters to isolate untrusted content.
  • Capability inventory: Performs Git operations (commit, rebase, push), spawns agent processes via the herdr CLI, and executes arbitrary shell commands defined in task metadata.
  • Sanitization: No explicit sanitization or validation of the gate commands or ingested reports is performed before they are used to determine logic flow or executed.
  • [COMMAND_EXECUTION]: The scripts/tsg-delegate.sh script invokes the herdr CLI to manage agent panes, start agent sessions, and send prompts. It also executes jq and standard shell utilities to process configuration and state.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 17, 2026, 10:32 PM