AllDSH
中文

DSH Context

bowenliang123 · bowenliang123/dsh-context

Context dashboard, browser and command that show how the context window is composed and how it evolves.

DSH Context is a DSH plugin by bowenliang123. Context dashboard, browser and command that show how the context window is composed and how it evolves. It is filed under Memory & Context and Development & Operations, for Memory & Context. Verification level L3 (Install spec), an automated security scan has not run yet, and it targets DeepSeek Harness 0.1.2-rc.1 || 0.1.3-alpha.2 || 0.1.5-rc.1; the listing was last checked on 2026-09-18 and is published under the Apache-2.0 licence. Install it with dsh plugin --profile web add dsh-context

Listing checked by Chiyu , data updated · How to cite

L3active
DSH Context verification panel: level L3 of 5 (Install spec) — Install command parsed, DSH target declared. Security scan: unverified — no scan has run. Compatibility target: DeepSeek Harness 0.1.2-rc.1 || 0.1.3-alpha.2 || 0.1.5-rc.1. Licence: Apache-2.0. GitHub stars: 1,430. Listing last checked 2026-09-18.
Verification panel for DSH Context. The listing holds level L3 of 5 — Install spec: Install command parsed, DSH target declared. Security scan: unverified — no scan has run. Compatibility target: DeepSeek Harness 0.1.2-rc.1 || 0.1.3-alpha.2 || 0.1.5-rc.1. Licence: Apache-2.0. Listing last checked 2026-09-18. How verification works

In use

Screenshots taken by this directory of the plugin running in a throwaway DeepSeek Harness on a generated demo project — not of your own machine, and not supplied by the author. Click any capture for the full-size version.

DSH Context — Installed and enabled, as the expanded row in Settings → Plugins → Plugin list: full package name, configuration state and runtime state.
Installed and enabled, as the expanded row in Settings → Plugins → Plugin list: full package name, configuration state and runtime state.
DSH Context — A real session that ran one short task on a generated demo project.
A real session that ran one short task on a generated demo project.
DSH Context — A session tab this plugin adds: “Context”.
A session tab this plugin adds: “Context”.
DSH Context — A right-sidebar panel this plugin adds: “Context”.
A right-sidebar panel this plugin adds: “Context”.

Captured by AllDSH on 2026-09-19 against DSH 0.1.5-rc.1 at a 1920×1080 viewport, package dsh-context@^0.53.3.

Install command

dsh plugin --profile web add dsh-context

Requires a running harness: start it withnpx @deepseek-ai/dsh webthen run the command above.New to this? See the installation guide.

What it does

DSH Context answers the question a long session eventually raises: what is actually in the window right now, where did it come from, and how did it get there. It adds a Context tab to every session, a matching right-sidebar panel, and a /context command, all reading the same data.

One distinction matters more than any single card: it observes, it does not manage. Its own README calls the plugin “full context lifecycle management”, and its package.json description says “context insight and management” — but nothing in either source says the plugin modifies, trims, compacts or prunes the window, rewrites prompts, or changes model behaviour. Read “management” there as loose README wording, not as a mechanism the sources describe.

The three surfaces

  • Context tab — the full dashboard, available in every session: stats, composition, per-request trend, events, file activity and the agent network.
  • Context panel — the same dashboard as a right-sidebar tab on harness 0.1.5-rc.1 and later. You pick Context on the sidebar’s guide page and the panel opens beside the chat.
  • /context command — a centered dialog with the current composition card and the context browser, so you can look without leaving the chat.
  • Settings → Plugins → Plugin configuration holds per-user defaults: trend granularity (Step/Turn), trend mode (Total/Delta) and the File Activity sort order. In-chart toggles stay per-view and do not overwrite the stored preference.

The cards, and what each one shows

  • Context Stats — turns, steps, human inputs, live tool calls, the session’s cache-hit rate, plus a cost estimated from models.dev list prices, with DeepSeek peak/off-peak awareness.
  • Token Stats — where the billed tokens went, split by composition (system, tools, messages and so on), closing with the provider-exact output total.
  • Timing — how active time split across model calls, tool runs and overhead.
  • Current Context — what is in the window right now, as a six-colour stacked bar against the model’s full window, with hatching for free headroom.
  • Context Trend — one stacked bar per model request, with a plain-language step brief, compaction and prune marks pinned to the bar where they happened, and hover or click for a pinned breakdown.
  • Context Browser — what any retained request was assembled from: six collapsible categories drilled down to the actual prompt text, tool JSON schemas, message content, reasoning, tool arguments and tool outputs.
  • Context Events — every injection, compaction, prune, model switch and plan-mode toggle, labelled with its producer, net token delta, turn/step and time.
  • File Activity — one row per touched file, with read/written/searched/image counts, estimated line deltas, and click-through to the exact tool result.
  • Agent Network — the current agent, its parents and every subagent on one map, each ring scaled to its window occupancy.

Estimates, not measurements

This is the heart of an honest reading of the plugin, and the card UI is built around it:

  • Category breakdowns use dsh’s own fixed-density heuristic — the same one behind the built-in token meter.
  • Headline occupancy comes from the official token-meter projections (contextPressure / contextBreakdown), so the figures match the context ring above the composer.
  • The Trend card shows provider-reported actuals — Actual Prompt / Output / Cache — next to the estimates, not instead of them.
  • Line deltas in File Activity are estimated, and tool-source chips are best-effort. Steps older than a compaction are reconstructed from the removed-message archive, and the card itself says when a step’s makeup is only approximate.

Treat every number as a well-grounded approximation. Only the provider-reported fields are exact.

Requirements

  • Node ^22.19.0 || >=24.0.0, per the package’s engines field.
  • Harness — the README says it works on @deepseek-ai/dsh 0.1.2-rc.1 and later, across several session-log generations. The manifest is stricter: it maps exactly three releases as compatible — 0.1.2-rc.1, 0.1.3-alpha.2 and 0.1.5-rc.1. Do not assume compatibility with a release outside that list.
  • The README points to a docs/compatibility.md per-release matrix; that file is not among the sources here, so its contents cannot be summarised.
  • Runtime dependencies are thin: zod is the only one. React and the UI primitives are declared as optional peer dependencies.
  • The package ships a dsh block declaring a bundle patch (a Cordis composition patch) plus a web client that injects six client packages. The client platform is declared as web — a headless or CLI-only mode is not described.
  • The README states no permission model, file-read/write behaviour, network access or telemetry. Assume these are undocumented rather than assume they are safe.

Known limits

  • It cannot change the window, so it cannot fix a session that is filling up. Reading the cards is the whole of its capability.
  • Cost is estimated from list prices and is not a bill. The sources do not claim the plugin reduces cost.
  • Exact token or context measurement is not on offer; the estimates can diverge from provider actuals.
  • Post-compaction reconstruction is explicitly approximate.
  • The right-sidebar panel, and the file preview it opens, require harness 0.1.5-rc.1 and later; on an older harness the file name opens on your system instead.
  • The README carries a star badge and asks for stars, but no star figure is verifiable from the sources here, so none is quoted.

When to use it — and when not to

Use it when a session has grown long and you need to know what is taking up the space, or when you are debugging why a prompt looks the way it does. The Context Browser is the part with no obvious substitute: it lets you read the actual system prompt, tool schemas and tool outputs that a specific request was assembled from. File Activity and Agent Network are useful on their own for auditing what an agent did and how a subagent tree is shaped.

Do not install it expecting it to manage, trim or shrink anything. If you want a smaller window, this plugin will show you the problem and stop there — changing the window is outside everything the sources describe. Be cautious about reading its numbers as measurements where an exact figure matters; use the provider-reported actuals for that and treat the rest as guidance.

FAQ

Does DSH Context manage or optimise my context window?

No, not in any way the sources describe. It reads and displays context data. Nothing in the README or the package manifest says it modifies, trims, compacts or prunes the window, changes prompts, or alters model behaviour. The word “management” appears in the README’s tagline and the package description, but no mechanism behind it is documented.

Are the token numbers exact?

Mostly not. Category breakdowns use dsh’s own fixed-density heuristic, occupancy comes from the official token-meter projections, file line deltas are estimated, and tool-source chips are best-effort. Provider-reported Actual Prompt / Output / Cache figures are shown alongside the estimates on the pinned trend details, and the Token Stats card pairs its estimated composition shares with the provider-exact billed total. Steps older than a compaction are reconstructed from the removed-message archive and marked as approximate.

Which harness versions are supported?

The README says 0.1.2-rc.1 and later. The package manifest is more specific and lists exactly three compatible releases: 0.1.2-rc.1, 0.1.3-alpha.2 and 0.1.5-rc.1. Compatibility beyond that list is not established by either source, and the compatibility document the README links to is not available here.

Does it need React and the harness UI packages?

For the web UI, yes — the client injects six client packages — five @deepseek-ai/dsh-client-* bundles plus @deepseek-ai/dsh-api-remotes, and React plus the UI primitives are declared as peer dependencies. Both of those peers are marked optional, which is worth understanding before you rely on the client half rendering. The only ordinary runtime dependency is zod.

Can I use it in a headless or CLI-only setup?

The sources do not describe one. The manifest declares the client platform as web, and every surface it documents — the Context tab, the right-sidebar panel, the /context dialog and the settings card — is browser UI.

Sources & references

The technical claims on this page rest on the primary documents below. Each link is followed so it can be checked directly.

  1. [1]bowenliang123/dsh-context — GitHubThe plugin’s own repository and README — the primary document every claim in this listing is checked against.
  2. [2]dsh-context on npm — npmThe published package the install command on this page resolves.
  3. [3]deepseek-ai/deepseek-harness — GitHubUpstream repository for DeepSeek Harness, the Cordis runtime and the plugin loading model.
  4. [4]@deepseek-ai/dsh on npm — npmPublished harness package, including the release tags our dshTarget field tracks.
  5. [5]dsh-plugin GitHub topic — GitHubThe discovery mechanism used to find plugins listed in this directory.

All links last verified 2026-09-11

Questions about this listing

Is DSH Context safe to install?
DSH Context holds verification level L3 of 5 — Install command parsed, DSH target declared — and its security status is unverified — no scan has run. A verification level records how far the install path was exercised; it is not a safety guarantee. Safety depends on what the plugin can actually reach, so read its source and install script and check its file, shell, network and credential access before you run it.
Which DeepSeek Harness version does DSH Context target?
DSH Context declares dshTarget 0.1.2-rc.1 || 0.1.3-alpha.2 || 0.1.5-rc.1, the DeepSeek Harness release this listing was checked against. dshTarget is the compatibility target the plugin author declares: it records that a release was checked, not that the plugin keeps working after upstream ships something new. Test in an isolated profile before upgrading your harness.
How do I install DSH Context?
Start the harness with npx @deepseek-ai/dsh web, then copy the install command shown at the top of this page. That command is generated from the package name the repository publishes — or from an install spec the listing declares — so it cannot drift from what the repository actually ships. This listing was last checked on 2026-09-18.
Does DSH Context manage or optimise my context window?
No, not in any way the sources describe. It reads and displays context data. Nothing in the README or the package manifest says it modifies, trims, compacts or prunes the window, changes prompts, or alters model behaviour. The word "management" appears in the README's tagline and the package description, but no mechanism behind it is documented.
Are the token numbers exact?
Mostly not. Category breakdowns use dsh's own fixed-density heuristic, occupancy comes from the official token-meter projections, file line deltas are estimated, and tool-source chips are best-effort. Provider-reported Actual Prompt / Output / Cache figures are shown alongside the estimates on the pinned trend details, and the Token Stats card pairs its estimated composition shares with the provider-exact billed total. Steps older than a compaction are reconstructed from the removed-message archive and marked as approximate.
Which harness versions are supported?
The README says 0.1.2-rc.1 and later. The package manifest is more specific and lists exactly three compatible releases: 0.1.2-rc.1, 0.1.3-alpha.2 and 0.1.5-rc.1. Compatibility beyond that list is not established by either source, and the compatibility document the README links to is not available here.
Does it need React and the harness UI packages?
For the web UI, yes — the client injects six client packages — five @deepseek-ai/dsh-client-* bundles plus @deepseek-ai/dsh-api-remotes, and React plus the UI primitives are declared as peer dependencies. Both of those peers are marked optional, which is worth understanding before you rely on the client half rendering. The only ordinary runtime dependency is zod.
Can I use it in a headless or CLI-only setup?
The sources do not describe one. The manifest declares the client platform as web, and every surface it documents — the Context tab, the right-sidebar panel, the /context dialog and the settings card — is browser UI.

Similar DSH plugins

Browse all

DSH-IM

xmanrui

Bridges ten IM platforms to DSH sessions over persistent outbound connections, with per-channel access policies.

L3
1.4k JavaScript

deepseek-design

Devin-AXIS

A native design suite for DSH — AI generation, visual editing, curated templates and a PPT studio.

L1
1.3k JavaScript

dsh-mnemon

omdsh-dev

Composable three-tier memory for DSH — runtime context, searchable documents, and long-term Memory Spaces on pluggable providers.

L2
384 TypeScript

deja-vu

vshulcz

Local cross-agent session memory — it indexes the coding-agent histories already on disk and recalls them into DSH, with no LLM.

L3
829 Go

Graph Memory

adoresever

A DSH memory plugin that keeps the newest turns, archives the rest, and recalls source-backed answers from a local SQLite graph.

L2
626 TypeScript

Mnemon

mnemon-dev

Editor's pick

LLM-supervised persistent memory with graph-based recall, cross-session stores and a single native binary.

L2
587 Go
ESC

Type to search the index. Built at deploy time by Pagefind.