kainakamura/understand

Your task is to thoroughly understand the existing codebase and how it pertains to the user's problem. Carefully analyze the prompt and the codebase to understand what they are asking. Your job is to gather as much context as possible about the problem and understand how to approach it.

Structure your output following template.md.

Steps

1. Determine the current iteration (highest iteration-NN/ directory)
2. Read these thoughts:
   • problem.md: Problem statement
   • iterations.md: Cross-iteration summary
   • iteration-NN/progress.md: Progress summary for the current iteration
   • Any thoughts explicitly specified by the user via $ARGUMENTS
   • If your task requires context beyond these files, consult log.md to identify other relevant artifacts by their descriptions and slugs
3. Do NOT spawn agents by default. Work directly unless the user explicitly requests it via natural language (e.g., 'agent team of 3', 'use an agent team'). When spawning agents, understanding tasks primarily benefit from code-searchers and web-searchers for broad research. Include critics when validating existing research.
4. If spawning agents, for each teammate provide:
   • Its specific, focused research question
   • Relevant context: problem summary, current iteration, what is already known
   • Include repo paths and context from the hub file body so agents know where source code lives
   • If the hub file links to related projects, consider whether those projects have relevant findings. Read their hub files if the research question would benefit from cross-project context. Use judgment: don't read unrelated projects.
5. Wait for all agents to complete and collect their text summaries (skip if working directly)
6. Synthesize all summaries into iteration-NN/understanding/understanding-NN-topic.md (include a short kebab-case topic in the filename)
7. Append a one-line entry to iteration-NN/log.md describing what was researched and linking to the artifact
8. Ask follow-up questions for clarification if anything remains unclear
9. Clean up the team when done

Success Criteria

• iteration-NN/understanding/understanding-NN-topic.md created in current iteration
• Produce a coherent, unified document, not a concatenation of agent outputs
• No code files have been changed
• Summary returned to user for verification
• Summary does not contain any potential solutions, only an understanding of the problem

Guidelines

• Ask follow-up questions to the user to clarify unclear parts
• Do not propose solutions, just give an overview of your understanding of the problem