deepseek-harness
deepseek-ai
DeepSeek Harness: Everything is a Plugin.
PROJECT TOPICS
PROJECT README
English | 中文
Delete one complete closed turn from a DeepSeek Harness conversation without deleting or replacing the Session. The action appears beside the final assistant response for each completed top-level turn.
From GitHub:
dsh plugin --profile web add github:hanshenmesen/dsh-turn-delete
From a prebuilt GitHub Release:
dsh plugin --profile web add https://github.com/hanshenmesen/dsh-turn-delete/releases/latest/download/dsh-turn-delete.tgz
Restart dsh web after installation. To remove it:
dsh plugin --profile web remove dsh-turn-delete
The plugin refuses deletion when the target turn is still open or has already been merged into a compaction replacement with other history. This prevents deleting only part of model-visible context.
The Host half registers POST /dsh-turn-delete, claims the target Agent's maintenance lease, validates the complete closed-turn surface span, appends a zero-content replacement, and waits for sessions.flush() before acknowledging success.
The browser half contributes to the public conversation.chat.assistant-actions and conversation.chat.turnTail slots. A Conversation definition projects the durable tombstone onto the target Turn. DeepSeek Harness does not yet expose a public whole-Turn visibility hook, so the tail contribution applies a narrow presentation compatibility layer to the Chat rows between adjacent Turn tails. Model context deletion does not depend on that compatibility layer.
0.1.0-rc.6 and 0.1.0-rc.7DeepSeek Harness is in developer preview. After upgrading Harness, verify the plugin once with a disposable Session containing three short turns and delete the middle one.
npm install
npm run check
dsh plugin --profile web add -w link:/absolute/path/to/dsh-turn-delete
npm run check type-checks and builds both plugin halves, runs the unit/UI tests, and verifies the npm package contents.
Like every DSH plugin, this package runs with the Harness process's permissions. Review the source before installing third-party plugins. The deletion changes future model context but deliberately retains the original append-only events for audit and replay.
MIT
CLASSIFICATION EVIDENCE
系统优先读取 GitHub Topics,再与站内分类词典和词根规则比对。当前命中: 无有效分类标签。