The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes data from project files (STATE.md and loop-run-log.md) to generate budget negotiation requests, which introduces an indirect prompt injection surface. External data like bug reports or CI logs reflected in these files could theoretically influence the justification drafted by the agent.
Ingestion points: Reads outcomes, action counts, and task priorities from loop-run-log.md and STATE.md in the Negotiation Protocol steps.
Boundary markers: No explicit delimiters or instructions are provided to the agent to distinguish between its own logic and potentially untrusted content within these ingested logs.
Capability inventory: The skill is restricted to reading and appending to specific project configuration and state files; it has no network access or shell execution capabilities mentioned.
Sanitization: No explicit sanitization or validation of the log content is described before interpolation into the drafted request.
[SAFE]: The skill specifically implements safety boundaries that prohibit the agent from self-modifying the loop-budget.md configuration. A human must manually review and edit the file to grant an extension, serving as a robust mitigation against autonomous financial or resource escalation.

budget-negotiator