[Technical Report] GLM Coding Pro quota consumption anomaly - empirical analysis shows 5.6x faster burn rate than expected

Suggestion from a user: peanut_the_fox_terrier (Discord)

My hypothesis is the medium thinking level might not saves tokens in a multi-round implementer-reviewer workflow, in which lower thinking effort could lead to more rounds of reviews. On the other hand, lower thinking level in a single-pass workflow could, or very likely, to save tokens as Agnes suggested.

Unfortunately, it will be a little tricky to reproduce this data locally, since the experiment was done on my personal project, which is not ready to be open-sourced yet.

If someone from your team is interested to reproduce it, the workflow I used is https://github.com/BoTime/custom_toolkit/tree/main/plugins/autopilot/skills/autopilot.

After the plugin is installed:
/autopilot

Full Report Thread:
https://discord.com/channels/1346756824233148527/1534315858921259088

Start with the linked custom_toolkit autopilot workflow and the full Discord report thread; the issue provides no repository file or test to inspect. Install the plugin and use /autopilot with a comparable feature description, then compare quota usage across single-pass and multi-round implementer-reviewer workflows. Done means producing reproducible evidence that explains or confirms the reported 5.6x burn rate.