Perf: session bootstrap/compaction becomes very slow after ~50 rounds of conversation

Summary

After roughly 50 rounds of conversation, resuming / bootstrapping an existing session becomes noticeably slow — session load, restore, and compaction latency all balloon, and the degradation feels super-linear in transcript size.

Environment

• opencode (anomalyco/opencode), local CLI, Windows 11
• Long-lived sessions (50+ rounds), resuming from disk

Expected vs Actual

• Expected: load/resume time stays roughly proportional to transcript size; compaction runs in the background without blocking interaction.
• Actual: at ~50 rounds, resuming the session takes a long time; larger transcripts make load/compaction markedly slower.

Suggested areas to investigate

• Full-file transcript parsing on every resume (no incremental/streamed load)
• Compaction running synchronously in the interactive/init path instead of backgrounded
• No lazy hydration of large snapshots -> every resume re-parses the whole session

Acceptance criteria for a fix

• Session resume time scales near-linearly (or better) with transcript size
• Compaction is backgrounded / non-blocking
• No functional regression in session fidelity after compaction

Additional context

Long agent sessions are the main use case for this tool; this is a day-one DX pain point for anyone running long-lived conversations. Happy to provide a profile or minimal repro transcript on request.

No files or tests are named in the report. Start by profiling session resume/load and compaction with a 50+ round transcript, focusing on full-file parsing, synchronous compaction, and snapshot hydration. Done means resume scales near-linearly, compaction does not block interaction, and session fidelity remains intact.