dsh-openai-server-compaction

Long conversations on OpenAI Responses routes hit their context ceiling quickly, and naive client-side truncation throws away the very context you need. dsh-openai-server-compaction brings Codex Remote Compaction V2 into DSH so the server performs the compression while the client keeps only the most recent real user messages plus a single opaque compaction item—saving tokens without losing meaning.

The plugin supports manual, token-pressure, and overflow trigger modes, retaining up to 64,000 tokens of recent messages in a single pass. Opaque compaction state is persisted as an encrypted file outside the session JSONL and restored automatically after a DSH restart. A portable text checkpoint capped at 4,096 output tokens is also written into standard DSH history so the session can continue on a different provider. Missing credentials, ambiguous configuration, malformed Responses streams, and unsupported route facts all fail explicitly—there is no silent Chat Completions fallback.

Built for DSH developers who run openai-responses routes and need precise long-context management. If you are migrating from the legacy llm-openai-server-compaction namespace, the plugin refuses to start while a non-empty legacy section remains, steering you through a clean migration and preventing conflicts between the two configuration schemas.