Experiment with improved openagents spec writing flow

[DawidWraga]

• dont want claude to write full spec (slow,expensive)
• but without full spec, subagent will lack context
• pass in full contxt-> evey single tool call is costing us for ALL input tokens
• better appraoch:
  • claude writes spec TITLE
  • -> pass spec TITLE + FULL CONTEXT (eg 100k) to cursor cli
  • -> instruct it to write a detailed spec (eg 1k) based on title/context
  • -> pass 1k spec to the full cli agent to execute
• outcome:
  • claude only writes title -> fast & cheap
  • cursor agent 1 reads full contxt only ONCE -> faster than getting claude to write it; cheaper than handing full context to execution agent; smarter than giving only title
  • cursor agent 2 gets detailed spec -> effective