Dashboard
Cost dashboards tell you what you spent. This tells you how you prompt — it reads your Claude Code session logs and surfaces the prompts that unlock autonomous work and the topics where you end up fighting the agent.
user turns you've sent (system messages excluded)
distinct Claude Code sessions
% of turns that didn't trigger a correction
prompts followed by ≥10 uninterrupted autonomous turns
Clusters of similar prompts with high correction rates.
Things you said right before the agent ran autonomously for a long stretch.
ultrathink through how to decomplect and then make the complete plan and then execute it if gemini approves
make sure readme is strong and accurate and clear
re: pass 3 the qui thing should be fixed the way other secrets work in .env, the Cap repo has some custom build stuff check docker compose to understand. i think you actually can do everything, do it
<blob>
get wrong.") ⎿ 2026-04-04T20:03:04.191470Z ERROR codex_core::codex: failed to load skill