The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a plan-convergence loop that ingests data from external AI reviewers and local model servers to drive subsequent planning phases.
Ingestion points: External AI CLI outputs (e.g., Claude, Codex, Qwen) and review summaries generated by the gsd-review skill.
Boundary markers: No specific boundary markers or "ignore embedded instructions" warnings are defined in this orchestrator to wrap the feedback ingested from reviewers.
Capability inventory: The skill has access to Bash for command execution, Write for file modifications, and the ability to trigger other Skills and Agents (SKILL.md).
Sanitization: There is no evidence of sanitization or filtering applied to the feedback received from external AI tools before it is interpolated into the planning process.
[COMMAND_EXECUTION]: The skill is designed to invoke multiple external CLI tools and connect to local model servers via shell commands.
The logic parses flags to execute commands for tools like claude, codex, coderabbit, opencode, and antigravity.
It targets local network services for local AI model hosting including Ollama (http://localhost:11434), LM Studio (http://localhost:1234), and llama.cpp (http://localhost:8080). These are standard local development configurations.
[SAFE]: The skill accesses configuration and reference files located within the agent's core directory structures.
It references files such as ~/.claude/gsd-core/workflows/plan-review-convergence.md and ~/.claude/gsd-core/references/gates.md for execution context, which is standard behavior for a workflow-orchestration skill.

gsd-plan-review-convergence