Summary
Define the first input format for omni cache-sim.
Candidate JSONL Format
{id:1,timestamp:2026-07-09T10:00:00Z,model:gpt-4o,prompt:...}
{id:2,timestamp:2026-07-09T10:00:03Z,model:gpt-4o,prompt:...}Alternative Format
Structured prompt parts:
{
id: 1,
model: gpt-4o,
parts: [
{name:system,stable:true,text:...},
{name:user,stable:false,text:...}
]
}Questions
- Should MVP support raw final prompts or structured parts?
- Should we support OpenAI/Anthropic request JSON directly?
- How much timestamp/session metadata is needed for TTL simulation?
- Should this be a stable public schema or experimental first?
Summary
Define the first input format for
omni cache-sim.Candidate JSONL Format
{id:1,timestamp:2026-07-09T10:00:00Z,model:gpt-4o,prompt:...} {id:2,timestamp:2026-07-09T10:00:03Z,model:gpt-4o,prompt:...}Alternative Format
Structured prompt parts:
{ id: 1, model: gpt-4o, parts: [ {name:system,stable:true,text:...}, {name:user,stable:false,text:...} ] }Questions