中文

The best DeepSeek Harness plugin for Agent's context insights and management.

Recommended and starred by DeepSeek Harness staff @tianyicui and @turtle1999 — see this post on X.

dsh-context provides full context lifecycle management features.

• Context Insights page — the cross-session overview as a first-level sidebar panel (beside Plugins and Automation tasks): KPI band with a 7-day token/cost usage chart, activity heatmap, aggregate composition ring, and filterable session cards that jump straight into any session.
• Context tab — a UI context dashboard for DeepSeek Harness's context stats, composition, trend, events, messages, and commands.
• Fleet tab — the agent family's own page: the current agent, its parents, and every subagent as one live card each.
• Context panel — the same dashboard as a right-sidebar tab: pick Context on the sidebar's guide page and the panel opens beside the chat, agent network included.
• /context command — the slash command shows the context model for current context composition and recent context evolution.

Fill dsh-context in the Add plugin wizard's search box, and click Install:

Install dsh-context plugin from DeepSeek Harness:

dsh plugin --profile web add dsh-context

Or update the dsh-context plugin:

dsh plugin --profile web update dsh-context@latest

Then start the web UI with dsh web. No build step, no restart.

Five surfaces, one story — what your agent is carrying, how it got there, and what it did with it:

Click Context Insights / 上下文洞察 in the sidebar's panel list — the same row level as Plugins and Automation tasks — and the page opens as a full center-column panel:

Open any session and click the Context / 上下文 tab:

The cost estimate follows one documented rule set: How model pricing is mapped.

The headline occupancy and composition read the same official token-meter projections as the chat composer's context ring (contextPressure / contextBreakdown), so the figures always match what the ring tells you — the card's percentage is that same occupancy, printed with one decimal.

The Context Stats card's Agent Network cell (the billed tokens and estimated cost of the whole family) jumps to the Fleet tab. Its Input / Output pills are jump links too: Human Inputs opens the Context Browser on your messages, and Read / Written / Searched / Images land on the File Activity card with that kind picked.

A six-color stacked bar against the model's full window (the two hatched masks at the end = auto-compaction's reserve: its headroom, then the request's reserved output window): system prompt, tool schemas, user messages, injected context, assistant replies, tool results — each with its ≈token figure and share. When a conversation starts degrading, this is where you see which part is responsible.

One stacked bar per model request — finer than per-message — so you watch the window grow turn by turn:

• ✨ Step brief — three plain-language rows under the chart: User recalls the message that opened the turn, In lists what newly entered (usually the previous tool results), Response shows the reply and/or tools called. Click a row to open that exact message in the Context browser.
• ✂ marks the events — compactions and prunes are pinned to the bar where they happened, so the drops explain themselves.
• Read it your way — Step / Turn granularity, Total (cumulative makeup) or Delta (each request's signed change), and sideways scrolling through the whole session. In Delta mode, growth piles up above the baseline and a compaction dives below it:

• Overview — one toggle squashes the whole session (steps or turns) into the pane's width as a continuous stacked ribbon: every compaction drop and composition shift at a glance, no sideways scrolling, nothing extra to load. Flags mark the quartiles and the full span; hover to scrub, hover a band to light that category across the whole ribbon; click pins the step and opens that band's category in the browser (exclusive with DNA mode and Adaptive).
• Hover & pin — scrub for an instant tooltip; click to pin the full breakdown, with provider-reported Actual Prompt / Output / Cache next to the estimates.
• Live linkage — hovering a bar previews that step's assembled context in the Context browser beside the chart; leaving the chart returns to your own pick.

Pick Live (next request) or any retained step, and browse what that request was assembled from: seven collapsible categories expand into one row per element with its token price, and every element expands again into its actual content — the system prompt, each tool's JSON schema, message text, reasoning, tool arguments, and tool outputs. Skill content (the available-skills catalog, /skill invocations, and skill-tool loads) has its own Skill Injections category, so a stealthy skill's context footprint can't hide inside the injected-context and tool-result buckets.

• Who provides each tool — every tool-schema row carries a best-effort source chip: tool-* first-party packages, dsh-* capability packages, mcp:<server> proxies, or the exact plugin watched live from tools.register(). Sort by size / name, and filter every category by its own searchable fields:

• Filter assistant replies by kind — the assistant category's search box carries Thinking / Tools / Answer chips, each with the shown step's message count (one reply can carry several); click one to keep only the matching replies, click again to clear:

• Tool results open into the full call — the tool name and arguments with its OK/error status, the result body with line count and a Raw / Markdown toggle:

• Attachments render as ordered cards — a message's image and file blocks sit together in the order the message carries them, one card each: images with a thumbnail plus their dimensions, stored size, and the official DeepSeek image-token estimate; other files with their type glyph, extension, and stored size. One click opens the attachment in dsh's own preview column (markdown, code, PDF, office, and image viewers), under the file's real name:

• Diff against the previous turn — signed delta badges per category (+N items, ±Nk tokens) tell you what a turn added or reclaimed at one glance. Steps older than a compaction are reconstructed from the removed-message archive — and the card says so when a step's makeup is only approximate.

• DNA mode lays the whole context out flat — the bar draws one band per item in the model's reading order, and below it the same items become one flat list with the newest message on top (the model reads oldest-first, so the list runs that order bottom-up): every row carries its prompt position (#12), category, introducing step, and token price, and expands into the same full content; the tool schemas and the system prompt close the list as two group rows. The card ends flush with the composition + trend cards beside it — a fully flattened session scrolls inside instead of flooding the page.

Every injection, compaction, prune, model switch, and plan-mode toggle — labeled with its producer (instruction file, plugin id, skill name), its net token delta (compactions show what they reclaimed), turn/step, and time. The Inject / Compact / Prune / Switch / Mode chips filter the log by kind, each carrying its whole-session event tally.

One row per touched file — read, written, or searched — aggregated up to whichever step you pick on the trend chart:

• Per-purpose counts with header chips doubling as filters (Read / Written / Searched / Images) and a path search box.
• Line deltas — every edit/write contributes its estimated +added / −removed footprint, per file and summed.
• Every mode counts — native tools, the Minimal preset's str_replace_editor, and the nested calls inside PTC run_code programs are all folded into per-tool rows.
• Searches land on real files — matched files get their own ops rows with hit counts.
• Click a row to expand its full operation log — every op jumps straight to the exact tool result in the Context browser.
• Click a file name to open its preview in the right Sidebar, exactly as the built-in Files sidebar does — the same viewer, the same tab-per-file behavior. On a harness without that column the name opens on your system as before.

Every shell and code run with its status, wall time, and step stamp, plus a session summary, attention shortcuts, and a program ranking. The ranking stays global while the search and status chips narrow the list; clicking a program drills into it and marks the row — the ranking never collapses under its own filter.

Open any session and click the Fleet / 编队 tab (right of Context) — the agent family on a page of its own, with room for the collaboration surfaces that follow:

A one-line summary opens the page — agents and their status, live context, estimated cost, and task progress — then hands the same family to four linked surfaces: Agent Network, Fleet Members, Tasks, and Agent Comms. Picking an agent on the graph or the member board narrows the tasks and channels to that agent, so the page answers for one member at a time.

The current agent, its parents, and every subagent — one card per agent: its title, live context size and window occupancy, a composition bar, and how long it has run for and how many steps it has taken. Edges fan out from each parent in the lineage's color — hover a card to light its whole chain and read the full breakdown in the detail strip; click to jump into that session's own Context tab. Running agents breathe with a green pulse on their incoming edge and turn over the harness's own rolling spinner beside their title, which settles into the same green dot the header's subagent list uses once the work is done; narrow panes fold the family into tidy rows — never a horizontal scrollbar.

Each roster member's card carries its role (lead or teammate), model, latest reply, run time and steps, and active task count; the family's other agents sit below in the same card shape. A card's buttons open that agent's chat in the sidebar or the whole view, search and the status chips narrow both sections, and the sort menu orders the other agents by spawn order, tokens, duration, steps, or recency. The Tasks board puts team tasks and subagent delegations in the same place — owner groups, status filters, and blocked-by links that light the whole prerequisite chain when you focus a task. Click a task to open its full description in place, raw or rendered, and the board scrolls it to the front under a sticky owner header.

The comms card reads the family's session logs automatically — no scan step. Team relays, mailbox deliveries, and the runtime's settlement notices merge into channels: each card pairs the lead or current agent with one counterpart, labels both sides, and shows how many messages each side sent. The timeline names both parties and their roles. Clicking any message opens the full record inline below it — the session log it came from and its sequence, a notice when the runtime wrote the settlement rather than the child, the complete closing report (with its structured result or diagnostic), and participant cards that open that agent's chat in the sidebar or the whole view. Filter by message kind or text, flip the order, and the shared agent pick from the graph or member board narrows the whole hub; older history and clipped excerpts are disclosed instead of silently truncated.

The right Sidebar's Context panel keeps the same card inline, so the family stays one click away while you read the chat. Don't want the tab at all? Turn Fleet tab off in the preferences card — the card then lives on in that panel, and the Context tab's family cell opens it there whenever a Sidebar is served.

Type /context (or pick it from the / menu) and press Enter:

A centered dialog opens with the Current Composition card and the Context browser — the same composition bar, per-step picker, and vs previous turn diff badges as the tab:

The Context preferences card holds this plugin's per-user settings — the Context panel entry (the Context conversation tab, the right Sidebar panel, both, or Hide for neither — the reply-row jump goes with it), the Fleet tab's own Show/Hide switch (the panel entry never gates it), the Context Insights default range (Today / Last 7 days / Last 30 days / All, Last 30 days unless changed; picking a preset on the page writes it back), trend granularity (Step/Turn), trend mode (Total/Delta), the trend duration curve (shown by default), the trend auto-overview step threshold, the Context Browser diff baseline, tool and File Activity sort, and the Context Insights Show/Hide switch. It lives on the Plugins page: the sidebar Plugins entry → the dsh-context bundle's page → its Configuration section (the card is served by the entry's live Config form).

The System Prompt · Prefix / Suffix rows hold free text joined at the very start / end of every session's system prompt, for all agents (subagents included) — a per-deployment persona or behavioral rule that rides the harness's own prompt assembly. Edits apply from the next turn; leave a row empty to disable it.

The Context Insights preset buttons write the pick back as the stored default; every other in-chart and in-card toggle stays per-view and never overwrites the stored preference.

• Estimates vs actuals — category figures use dsh's own fixed-density heuristic (the same one as its built-in token meter); the pinned trend details show provider-reported actuals next to them, and the Token card pairs its ≈-estimated composition shares with the provider-exact billed total.
• Compatibility — works on @deepseek-ai/dsh from 0.2.0-rc.2 (the 0.2.0 and 0.2.1 lines; verified on their rc.2 releases), the V4 session-log generation, from one shape-driven code path. The per-release matrix and how it is verified: docs/compatibility.md.
• I18n — UI in English and 简体中文.

If dsh-context helped you understand what your agent is carrying around, a ⭐ on GitHub is much appreciated — and issues/PRs are welcome!

Apache-2.0