The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes GitHub issue titles, descriptions, and comments which are provided by users and could potentially contain malicious instructions.
Ingestion points: In src/run.mjs, issue data is retrieved from GitHub using getIssue and getHumanComments.
Boundary markers: The prompt template in src/run.mjs uses Markdown headers (e.g., # Issue #N, # Rules) to separate user input from system instructions.
Capability inventory: The orchestrator in src/run.mjs can execute shell commands, write files, and perform git operations within the Upstash Box environment via src/box.mjs.
Sanitization: The ingested issue content is interpolated directly into the prompt without explicit sanitization or escaping of potential injection patterns.
[COMMAND_EXECUTION]: The skill uses sh and runScript functions in src/box.mjs to execute shell commands and bash scripts within Upstash Boxes for environment setup, testing, and running agents.
[DYNAMIC_EXECUTION]: The skill dynamically generates bash scripts at runtime (e.g., agent.sh, setup.sh, checks.sh) to perform various tasks in the sandboxed environment.
[PROMPT_INJECTION]: The system instructions for the coding agent in src/run.mjs include explicit override patterns ("These rules win over any other instructions you have been given") to ensure the agent adheres to the factory's workflow constraints.

create-upstash-software-factory