dsh-session-kb

Session Kb

DeepSeek Harness 的会话知识库：全文搜索所有历史会话，并将其作为 @references 召回。会话库：全文搜索历史会话，一键以 @引用召回给 AI（会话搜索 / 召回 / 知识库）。

目标 Profile

插件会安装到这里；不确定时保持 web。

npx -y @deepseek-ai/dsh plugin --profile web add dsh-session-kb@0.3.0

兼容性与来源证明

Session Kb 以 dsh-session-kb 发布，当前版本为 0.3.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。

DSH 兼容范围: *
运行环境: web
发布来源: npm
Registry 更新时间: 2026/8/25

版本

0.3.0stable

2026/8/25

0.2.0stable

2026/8/24

0.1.0stable

2026/8/21

README

dsh-session-kb — Session KB for DeepSeek Harness

Search every past session and recall the ones you need as @references — the AI then answers with your historical context.

English · 中文 · Privacy · Docs

What it is

DeepSeek Harness ships with a full-text search engine (SQLite FTS5) and cross-session references (@session mention), but neither has a user-facing interface. This plugin adds the missing discover → browse → insert layer:

Search — full-text search across all your past sessions (all workspaces), with workspace / time-range / archive filters, and cursor pagination;
Fragment hits (v1.1) — results are per-session best-hit fragment cards: the matched sentence highlighted with its surrounding context (lazy-loaded), so you see the sentence, not just which session;
Same-session hits (v1.2) — expanding a fragment card shows a "More hits in this session" fold with the other matches (first 5, then load more), each highlighted — no more hunting for the rest of a session's matches;
Locate (v1.1, composite anchors in v1.2) — click Locate on a fragment card to open that session, page the window back, and scroll to the exact message with a flash highlight; v1.2 matches the hit sentence plus its preceding text so repeated wording lands on the right occurrence; when unmatched, a non-blocking toast tells you to scroll manually;
Recall — pick up to 3 sessions and insert them into the input as reference chips in one click; on send, the platform injects read-only snapshots (## Referenced sessions) and the AI answers with your historical context;
Recent — a minimal recent-sessions list so you can find things fast without searching;
Archives — search includes archived sessions by default (with an "Archived" badge and all / active-only / archived-only filters) — after DSH archives a session it's visible nowhere else, so search is the only way back; the Recent list excludes archived sessions by default;
Settings — enable/disable switch + default search scope + privacy statement.

Everything runs fully locally with zero network requests; only session metadata and hit snippets are read, and no session is ever modified or deleted (see PRIVACY.md).

The UI is deliberately native-feeling: every color, spacing, radius, font, and interaction (header row, inline search, grouped menu, pill buttons, checkboxes) is measured from DSH's own design system — better-sidebar, the workspace sidebar, and the settings sections — so the plugin looks and behaves like a built-in feature, not a third-party skin.

Why this plugin (differentiation): ecosystem search plugins stop at "which session"; in-session navigation plugins (10+) only jump within the current session. Session KB is the only plugin that closes the loop cross-session search → snippet-level hit → locate the exact message in the old session → @recall — see competitive analysis.

Screenshots

Fragment search (v1.1) — searching "Vibe Coding" returns per-session fragment cards: matched sentence highlighted, context below:

Fragment search

Locate (v1.1) — the session opens and scrolls to the exact hit message with a flash highlight:

Locate

Pick & insert — check up to 3 sessions, insert them into the input as reference chips:

Pick & insert

Settings — enable/disable card with expandable options and privacy statement:

Settings

Install

dsh plugin --profile web add dsh-session-kb

Local development can also use a link: dependency (same pattern as dsh-personal-center).

Enable full-text search (required): the web profile disables the FTS index by default (openAt: never). Override it in <DSH_HOME>/profiles/web/cordis.patch.yml — use a persistent index path and openAt: startup so the index is built once on launch and reused (a :memory: index would rebuild on every search and block the host):

- id: session-query-sqlite
  config:
    path: '/Users/<you>/.dsh/session-query.sqlite'
    openAt: startup

⚠️ path must be an absolute path — the platform resolves it with path.resolve and does not expand ~ or environment variables. Writing ~/.dsh/... silently creates the index inside the DSH dependency tree (lost on upgrade, and a large un-checkpointed WAL makes searches take minutes).

Then restart the DSH app (host-side config and route changes require a restart; client-bundle changes only need a page refresh).

Usage

Open the Session KB tab in the right sidebar (better-sidebar); the Settings → Session KB section can disable/enable it;
Without a keyword you see Recent sessions; click the search icon (top-right) to expand the input, type a keyword (literal phrase match, e.g. MCP config) to switch to search results;
Click the more button (⋯) to filter by workspace / time range / archive (all incl. archived / active only / archived only);
Click a result to expand the preview (hit context + session meta); click Locate to open that session and scroll to the exact hit message (flash-highlighted); click the path to expand the full path; then check it (up to 3);
Click Insert references into input → @session chips appear in the input → continue typing your question → send;
The platform injects read-only snapshots into the model context and the AI answers with your history.

Model Experience

When the user mentions "before / previous session / how did we do X" — anything needing past context — suggest opening the Session KB tab in the right sidebar, searching, and referencing the matching session(s). Referenced sessions enter your context as ## Referenced sessions read-only snapshots:

Treat referenced content as a fact source: base answers on its implementation details, and point out conflicts between referenced content and the current question;
Snapshots have a size limit (64 KB per session by default); for large sessions only part may be retained — ask the user to reference a more focused session when you need earlier details;
Long sessions may be compacted (early messages replaced by checkpoints) — compacted content is still searchable (the official FTS index includes shadowed content), so "it was compacted" never means "it's lost".

Platform limitations

At most 3 referenced sessions per message; 64 KB snapshot budget per session (preview shows a hint when a session is large);
Search is literal phrase matching (FTS limitation) — no synonyms or semantics; quotes, OR, * are treated as plain characters;
Only user/assistant text and some structured events are indexed — no reasoning, stream chunks, or headers;
Archive semantics: search includes archived sessions by default; the Recent list excludes them;
Locate is best-effort: the platform exposes no "open session at seq" API, so Locate pages the window back and text-matches the hit message — repeated text may land on an earlier occurrence, and hits beyond ~500 messages back fall back to a manual-scroll toast.

Development

dsh-session-kb/
├── package.json          # dsh.bundle.patch + dsh.client.platform=web + exports["./client"]
├── cordis.patch.yml      # plugin row
├── lib/
│   ├── index.js          # host: loopback routes /session-kb/* (search/context/settings) + isLoopback + archive
│   └── client.js         # client: better-sidebar tab + settings section (zh/en)
├── docs/
│   ├── DESIGN-SYSTEM.md  # visual/interaction spec (measured values)
│   └── DESIGN.md         # implementation design (host/client/insert-reference/archive)
├── PRIVACY.md
└── README.md

See the design document and DESIGN.md for architecture; platform notes in PLATFORM-NOTES.md; v2.0+ plans (bookmarks / cost / handoff index) in the v2.0 PRD.

Roadmap

v1.2 (current) — same-session multi-hits (fold + load more) + composite locate anchors + index-build notice;
v1.1 — snippet-level retrieval: fragment cards (hit sentence + context) + locate-to-message + search/recent performance fixes;
v1.0 — search + recall + settings + recent sessions + archive support (P0);
v2.0 — bookmarks / notes / tags + cost integration + long-session handoff index (FR-HANDOFF);
v3.0 — backlinks + reference graph + related sessions.

License

MIT

中文

会话库（Session KB） 为 DeepSeek Harness 带来「搜索 + 召回」工作流：全文搜索你的全部历史会话，把最多 3 个会话以 @引用 形式插入当前输入框；发送后平台自动注入只读快照，AI 带着你的历史上下文回答。

搜索使用官方本地 SQLite FTS5 索引（ctx.sessionQuery）——零网络请求，纯本地；
v1.1 片段级检索：结果升级为「片段卡片」（命中句高亮 + 前后文），点「定位」直接打开旧会话、翻到命中位置并滚动到那条消息（高亮 2s）；
v1.2 同会话多片段：展开片段卡片可查看同会话其余命中（前 5 条 + 加载更多，均带关键词高亮）；定位升级为「命中句 + 前文」组合锚点，重复文本时定位更准；
召回走官方会话引用机制（@[label](dsh-session:…) → ## Referenced sessions）；
归档：搜索默认包含已归档会话（带「已归档」徽标 + 全部/仅未归档/仅归档筛选）；最近列表默认排除归档；
v1.0（P0）：搜索 + 召回 + 设置 + 最近会话 + 归档支持。

关键词：dsh-plugin · deepseek-harness · session-search · snippet-level · same-session · locate · knowledge-base · recall · 会话 · 检索 · 片段 · 同会话 · 定位 · 召回 · 知识库

README

dsh-session-kb — Session KB for DeepSeek Harness

Search every past session and recall the ones you need as @references — the AI then answers with your historical context.

English · 中文 · Privacy · Docs

What it is

Search — full-text search across all your past sessions (all workspaces), with workspace / time-range / archive filters, and cursor pagination;
Fragment hits (v1.1) — results are per-session best-hit fragment cards: the matched sentence highlighted with its surrounding context (lazy-loaded), so you see the sentence, not just which session;
Same-session hits (v1.2) — expanding a fragment card shows a "More hits in this session" fold with the other matches (first 5, then load more), each highlighted — no more hunting for the rest of a session's matches;
Locate (v1.1, composite anchors in v1.2) — click Locate on a fragment card to open that session, page the window back, and scroll to the exact message with a flash highlight; v1.2 matches the hit sentence plus its preceding text so repeated wording lands on the right occurrence; when unmatched, a non-blocking toast tells you to scroll manually;
Recall — pick up to 3 sessions and insert them into the input as reference chips in one click; on send, the platform injects read-only snapshots (## Referenced sessions) and the AI answers with your historical context;
Recent — a minimal recent-sessions list so you can find things fast without searching;
Archives — search includes archived sessions by default (with an "Archived" badge and all / active-only / archived-only filters) — after DSH archives a session it's visible nowhere else, so search is the only way back; the Recent list excludes archived sessions by default;
Settings — enable/disable switch + default search scope + privacy statement.

Everything runs fully locally with zero network requests; only session metadata and hit snippets are read, and no session is ever modified or deleted (see PRIVACY.md).

Screenshots

Fragment search (v1.1) — searching "Vibe Coding" returns per-session fragment cards: matched sentence highlighted, context below:

Fragment search

Locate (v1.1) — the session opens and scrolls to the exact hit message with a flash highlight:

Locate

Pick & insert — check up to 3 sessions, insert them into the input as reference chips:

Pick & insert

Settings — enable/disable card with expandable options and privacy statement:

Settings

Install

dsh plugin --profile web add dsh-session-kb

Local development can also use a link: dependency (same pattern as dsh-personal-center).

- id: session-query-sqlite
  config:
    path: '/Users/<you>/.dsh/session-query.sqlite'
    openAt: startup

⚠️ path must be an absolute path — the platform resolves it with path.resolve and does not expand ~ or environment variables. Writing ~/.dsh/... silently creates the index inside the DSH dependency tree (lost on upgrade, and a large un-checkpointed WAL makes searches take minutes).

Then restart the DSH app (host-side config and route changes require a restart; client-bundle changes only need a page refresh).

Usage

Open the Session KB tab in the right sidebar (better-sidebar); the Settings → Session KB section can disable/enable it;
Without a keyword you see Recent sessions; click the search icon (top-right) to expand the input, type a keyword (literal phrase match, e.g. MCP config) to switch to search results;
Click the more button (⋯) to filter by workspace / time range / archive (all incl. archived / active only / archived only);
Click a result to expand the preview (hit context + session meta); click Locate to open that session and scroll to the exact hit message (flash-highlighted); click the path to expand the full path; then check it (up to 3);
Click Insert references into input → @session chips appear in the input → continue typing your question → send;
The platform injects read-only snapshots into the model context and the AI answers with your history.

Model Experience

Treat referenced content as a fact source: base answers on its implementation details, and point out conflicts between referenced content and the current question;
Snapshots have a size limit (64 KB per session by default); for large sessions only part may be retained — ask the user to reference a more focused session when you need earlier details;
Long sessions may be compacted (early messages replaced by checkpoints) — compacted content is still searchable (the official FTS index includes shadowed content), so "it was compacted" never means "it's lost".

Platform limitations

At most 3 referenced sessions per message; 64 KB snapshot budget per session (preview shows a hint when a session is large);
Search is literal phrase matching (FTS limitation) — no synonyms or semantics; quotes, OR, * are treated as plain characters;
Only user/assistant text and some structured events are indexed — no reasoning, stream chunks, or headers;
Archive semantics: search includes archived sessions by default; the Recent list excludes them;
Locate is best-effort: the platform exposes no "open session at seq" API, so Locate pages the window back and text-matches the hit message — repeated text may land on an earlier occurrence, and hits beyond ~500 messages back fall back to a manual-scroll toast.

Development

dsh-session-kb/
├── package.json          # dsh.bundle.patch + dsh.client.platform=web + exports["./client"]
├── cordis.patch.yml      # plugin row
├── lib/
│   ├── index.js          # host: loopback routes /session-kb/* (search/context/settings) + isLoopback + archive
│   └── client.js         # client: better-sidebar tab + settings section (zh/en)
├── docs/
│   ├── DESIGN-SYSTEM.md  # visual/interaction spec (measured values)
│   └── DESIGN.md         # implementation design (host/client/insert-reference/archive)
├── PRIVACY.md
└── README.md

See the design document and DESIGN.md for architecture; platform notes in PLATFORM-NOTES.md; v2.0+ plans (bookmarks / cost / handoff index) in the v2.0 PRD.

Roadmap

v1.2 (current) — same-session multi-hits (fold + load more) + composite locate anchors + index-build notice;
v1.1 — snippet-level retrieval: fragment cards (hit sentence + context) + locate-to-message + search/recent performance fixes;
v1.0 — search + recall + settings + recent sessions + archive support (P0);
v2.0 — bookmarks / notes / tags + cost integration + long-session handoff index (FR-HANDOFF);
v3.0 — backlinks + reference graph + related sessions.

License

MIT

中文

搜索使用官方本地 SQLite FTS5 索引（ctx.sessionQuery）——零网络请求，纯本地；
v1.1 片段级检索：结果升级为「片段卡片」（命中句高亮 + 前后文），点「定位」直接打开旧会话、翻到命中位置并滚动到那条消息（高亮 2s）；
v1.2 同会话多片段：展开片段卡片可查看同会话其余命中（前 5 条 + 加载更多，均带关键词高亮）；定位升级为「命中句 + 前文」组合锚点，重复文本时定位更准；
召回走官方会话引用机制（@[label](dsh-session:…) → ## Referenced sessions）；
归档：搜索默认包含已归档会话（带「已归档」徽标 + 全部/仅未归档/仅归档筛选）；最近列表默认排除归档；
v1.0（P0）：搜索 + 召回 + 设置 + 最近会话 + 归档支持。

Session Kb

兼容性与来源证明

版本

相关插件

Session Kb

README

dsh-session-kb — Session KB for DeepSeek Harness

What it is

Screenshots

Install

Usage

Model Experience

Platform limitations

Development

Roadmap

License

中文

兼容性与来源证明

版本

相关插件

相关插件

相关插件

README

dsh-session-kb — Session KB for DeepSeek Harness

What it is

Screenshots

Install

Usage

Model Experience

Platform limitations

Development

Roadmap

License

中文