history-compression

After every agent step, compression triggers when used tokens exceed min(compactionThreshold times context window, 200k). The default threshold is 0.65, with the chat agent overriding to 0.5, and a hard cap of 200k tokens applies regardless of model context size.

Run npx tsx scripts/experiments/extract-compression-test-data.ts with a channel flag to read the local stagewise SQLite database. It writes per-compression bundles containing the system prompt, user message, serialized history, and actual output for direct playground comparison.

Check whether usedTokens exceeded the trigger formula min(compactionThreshold times contextWindow, 200k), verify compactionThreshold is at least 0, and confirm the _isCompressingHistory flag is not stuck from a prior run blocking new compressions.

The cascade tries gemini-3.1-flash-lite, then gpt-5.4-nano, then claude-haiku-4.5, each with a 30-second timeout and temperature 0.1. If all fail, it falls back to the active chat model, and if that fails the agent continues uncompressed.

The serializer stops at the first prior compressedHistory it finds and inlines it as a previous-chat-history block, so older raw messages are never re-serialized. The prompt then injects a ratio-based budget hint telling the model how aggressively to condense the prior briefing.

Raise COMPRESSION_TARGET_CHARS above the 30k default to allow longer briefings, or lower compactionThreshold so compression triggers earlier on smaller inputs. Keep the kept budget below the trigger threshold or compression never runs.