Measure compaction cost and retrieval-backed alternatives #2471

Junior's compaction approach is modeled on Codex, but its frequency and cost have not been measured. Explore the tradeoffs of retaining full context in storage and letting the agent retrieve relevant history on demand, rather than relying exclusively on summarized context. Inspired by this discussion.

• Quantify current compaction frequency, duration, and token/cost impact.
• Compare retrieval-backed context for conversation history and large tool outputs, including latency, cache cost, continuity, and extra tool calls.
• Related work: #431 and #872.

via @thomasjiangcy.

--

View Session [Sentry]