deepseek-harness
deepseek-ai
DeepSeek Harness: Everything is a Plugin.
PROJECT TOPICS
PROJECT README
Point checkpoints and ancestry-aware session continuation for DeepSeek Harness.
dsh-point gives a DSH workspace an immutable Point history without loading all prior sessions into every new context. Point is the only history unit: save a Point, continue from a Point, and let the agent read an authorized ancestor Point only when needed.
/point-stash [title] saves complete participant-visible events since the previous Point.point_read(point_id) lets the model read original events from an ancestor Point; unrelated Points are rejected.point CLI available on PATHInstall Point first if needed:
python3 -m pip install "git+https://github.com/guix4ever/point.git"
Add the plugin to the DSH Web profile:
dsh plugin --profile web add dsh-point
Restart DSH after installation. The package contributes both the Host plugin and its Web client bundle through cordis.patch.yml.
Save the current completed work:
/point-stash explain the change briefly
Open the Points tab in the DSH conversation view to inspect the current workspace DAG. Selecting a node is read-only; press Continue from this Point to open its deterministic continuation Session.
The continuation receives only the selected Point ancestry. If historical detail is relevant, the model can call:
point_read(point_id)
Only Point ids in that continuation's ancestry are accepted.
The save boundary waits for agent idle, flushes DSH persistence, and advances its cursor only after point stash succeeds. Projection fails closed on non-contiguous events, open turns, incomplete assistant chunks, duplicate or missing tool identities, dangling calls, and empty visible Points.
The browser cannot choose a filesystem path, Point binary, or shell argument. Workspace path and agent preset are derived by the Host from the addressed DSH Session.
npm ci
npm run typecheck
npm test
npm pack --dry-run
Maintained sources are TypeScript in src/ and client-src/. Runtime entry points are built into lib/ and committed so Git installation does not require a local build step.
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。