sakthai-agent-reasoning

Inside sakthai run, reason toward the goal one verifiable step at a time rather than guessing the whole answer up front:

• Recall before acting. Memory is injected into the system prompt — use it; don't ask for what is already known.
• Pick the smallest tool that advances the task, read its result, and let that result shape the next step. Don't fire speculative parallel calls when one observation would redirect the rest.
• Stop when the goal is met. Extra iterations cost tokens and risk drift; the loop has a bounded iteration cap for a reason.
• Surface contradictions between memory, tool output, and the request instead of silently choosing one.