Home HOOTL agent runs on DeepSeek API after Logfare cut

The author's setup splits the work across three pieces: a PowerShell worker that polls GitHub and launches or resumes the agent, Claude Code acting as the agent, and a Python checker running on GitHub Actions through the GLM model. The worker also counts spending every 30 seconds from token usage in the output log and stops the agent when the budget is reached, so the loop has an external brake rather than relying on the model to stop itself.

The security design separates a trusted worker account from an untrusted working account. The worker can read API key files and write state, budget and logs; the working account can freely write its own folders but only read state and cannot log in directly. The author acknowledges the limit: API keys are passed as environment variables to the running agent, so a determined agent could call the APIs directly and bypass the worker's budget accounting, which is why prepaid APIs and a remaining balance act as the final ceiling.

A notable design choice was abandoning a separate orchestrator agent. The author found that splitting research and decision roles lost nuance and fixed the boundaries the model should decide, and Claude Code already has sub-agent capability. Instead, one agent owns one goal end to end, and the author concludes that the quality of the Issue text — purpose, audience, success conditions, prohibitions — matters more than the plumbing.