dingyi580/dsh-conversation-rail
dsh-conversation-rail
dsh-conversation-rail is a community DeepSeek Harness plugin. Read the repository README before installing.
Install
npx @deepseek-ai/dsh plugin --profile web add github:dingyi580/dsh-conversation-railRestart `dsh web` after install. Bundle APIs can change during the developer preview.
README badge
[](https://dshhub.dev/plugins/dsh-conversation-rail)Paste this into your README. The star count updates with every catalog sync.
From the README
Excerpt from dingyi580/dsh-conversation-rail, cleaned of badges and images.
dsh-conversation-rail · session minimap
English | 中文
A locator for long conversations in DSH Web. A vertical rail sits along the left edge of the conversation, one bar per turn; hover a bar for a preview card, click it to jump to that turn.
▏───── ┌──────────────────────────────┐
▏─────────── │ Can you patch this without… │ ← question (one line, dark)
▏━━━━━━━━━━━━━━ ◀━━━━━┤ CRLF is already ruled out… │
▏────── │ only the legal re-encodes… │ ← answer (up to four lines, dimmed)
▏──────── └──────────────────────────────┘
- Bar length = how much text that turn holds (square-root compressed, 10–34px).
- The bold bar = the turn currently at the top of the scroll area.
- The preview card carries prose only: your question plus the assistant's plain text. Fenced code blocks, tool calls, shell commands, reasoning traces and injected context never reach the card — you cannot skim JSON.
It draws the whole session, not the loaded window
Conversation history is paginated: the front-end snapshot usually holds only the
most recent stretch, so an 8-turn session may have 1 turn in the window. The
rail is not bound by that:
- The skeleton comes from the host's outline endpoint
(
GET /dsh-conversation-rail/api/outline?sessionId=). The host reads the full log withctx.sessionQuery.readSession(), folds it into a table of contents of roughly a hundred bytes per turn, and returns that — the history itself never crosses the network. - The live snapshot overrides the part that is loaded: it carries jumpable DOM
anchors, and for a turn that is still streaming it is the only current source.
The two are aligned by the
user/messageevent seq, live winning over outline. - Clicking a bar that isn't loaded yet calls
session.loadOlder()repeatedly until that turn enters the window, then jumps. The bar pulses while paging, with a ceiling of 40 pages — better to not reach a turn than to page through an entire log because of one click.
The outline uses readSession rather than the lighter filterEvents because the
latter does not return source. The user/message event type carries both real
questions and context injected by agent.inject() (file-change notices,
AGENTS.md, skill bodies…); the official docs state all three project content
verbatim and are told apart by source. Without that field, injected context
would be drawn as turns.
How it attaches
It takes one seat: conversation.session.header.utilities. Not to draw anything in
the header — that slot is session-scoped, so the component receives the framework's
useSession snapshot hook and sessionId, and its lifetime follows the session.
The component renders null in the header; the actual rail is a fixed overlay on
body, aligned to the measured rectangle of [data-conversation-scroll]. It stays
out of the host layout, so switching skins, collapsing the sidebar or changing the
grid cannot squeeze it.
Turn data comes from ConversationSnapshot.chat: walk order, open a turn on a
user node, attribute the following assistant-step text blocks to it. Jumping
uses the host's own anchors ([data-chat-anchor-key]), the same path its internal
scroll positioning takes.
Known boundaries
…

