为什么
The current public host test reports prompt_tokens=10 and disables automatic compaction. It verifies serialization and manual summarization but does not cover the user's requirement that dynamic pruning work during long autonomous runs without interrupting tools. The audit reproduced reference-marker suppression and a separate host Native LLM slow-tool abort under context pressure.
范围
Real OpenCode plugin loading, public session API, outgoing model requests, automatic compaction/continuation and tool completion. Use an isolated pinned host checkout, configuration and database with a deterministic local model endpoint. Preserve the existing model-switch and summary-fidelity checks.
做法
Add or adapt public-host scenarios using reported usage that grows with request payloads, keep automatic compaction enabled, exercise references and slow tools, assert tool results and exit codes as well as agent continuation, and document the bounded budget behavior. Track the Native LLM runtime correction in the host repository; pin its reviewed commit before enabling its release regression. Do not change session lifecycle from inside DCP.
验收
- Long tool loops demonstrate actual request-local pruning and retain persisted outputs.
- File references are present in history while pruning still operates before the first native summary.
- Automatic summaries receive original tool content and continuation finishes the requested work.
- High-usage slow tools complete successfully on supported runtime paths; explicit cancellation still works.
- Small-window protection constraints are tested and documented without claiming guaranteed avoidance of native compaction.
- Required CI and SpecGit acceptance pass before merge and publication.
为什么
The current public host test reports prompt_tokens=10 and disables automatic compaction. It verifies serialization and manual summarization but does not cover the user's requirement that dynamic pruning work during long autonomous runs without interrupting tools. The audit reproduced reference-marker suppression and a separate host Native LLM slow-tool abort under context pressure.
范围
Real OpenCode plugin loading, public session API, outgoing model requests, automatic compaction/continuation and tool completion. Use an isolated pinned host checkout, configuration and database with a deterministic local model endpoint. Preserve the existing model-switch and summary-fidelity checks.
做法
Add or adapt public-host scenarios using reported usage that grows with request payloads, keep automatic compaction enabled, exercise references and slow tools, assert tool results and exit codes as well as agent continuation, and document the bounded budget behavior. Track the Native LLM runtime correction in the host repository; pin its reviewed commit before enabling its release regression. Do not change session lifecycle from inside DCP.
验收