Overview
dsh-plugin-continue-on-limit
DeepSeek Harness plugin that auto-continues a turn the model ended by exhausting its output-token limit, with a per-session burst cap and a minimum interval between continuations.
README / EN
Package documentation
Registry summary
DeepSeek Harness plugin that auto-continues a turn the model ended by exhausting its output-token limit, with a per-session burst cap and a minimum interval between continuations.
dsh.pub verifies the pinned bundle contract, runtime facts, and distribution semantics. The complete README remains in the source repository.
Read the full README on GitHubLIMITATIONS
Known limitations
- **只认 `max-tokens`。** `aborted` / `error` / `blocked` / `interrupted` / `forked` 一律不动。它们要么是模型自己停的，要么是被人停的，发「继续」救不回来。 - **计数器是内存态。** 进程重启会清零，一条跨重启的链会被当成新链。 - **429 / 配额耗尽不在范围内。** 那是另一个问题域。 - **不判断模型是否真的写完了。** 它只是替你按键。如果模型陷入循环复读， `maxConsecutive` 会兜住轮数，但代价还是要你自己看日志确认。 - **不感知子代理。** `ctx.agents.get()` 对子代理会话同样返回 agent，所以子代理的 截断也会被续写，共享同一套 burst 上限。如果只想管主会话，把 `maxConsecutive` 调小或直接 `enabled: false`。