dsh-context/README.md

dsh-context

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

dsh-context provides full context lifecycle management features.

• Context tab — an UI context dashboard for DeepSeek Harness’s context stats, composition, trend, events, and messages.
• /context command — the slash command shows the context model for current context composition and recent context evolution.

Install / Update

To Install from any DeepSeek Harness installation:

dsh plugin --profile web add dsh-context

Or to 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.

Use it

Context tab

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

⌨️ /context command — In-session Context Insight modal

Type /context (or pick it from the / menu) and press Enter: a centered dialog shows the Current Composition card and the Context browser — the same composition bar, legend, and per-step browsing as the tab, so you can inspect what any request was assembled from without leaving the chat.

⚙️ Settings — per-user preferences

In Settings → Plugins → Plugin configuration, the Context / 上下文 card holds this plugin's per-user preferences — the default trend granularity (Step/Turn) and the default trend mode (Total/Delta) the Context tab opens with. In-chart toggling stays per-view and never overwrites the stored preference. The card appears only when the Host half is installed and the settings document is writable (a remote browser keeps settings process-local and shows no card).

What you'll see

📊 Context stats — the session at a glance

Turns, steps, how many injections, compactions, and prunes have happened.

🧱 Current composition — what's in the window right now

A six-color stacked bar scaled against the model's full context window (the gray track is your remaining headroom): system prompt, tool schemas, your messages, injected context, assistant replies, and tool results — plus the top-5 most expensive tool schemas. When a conversation starts degrading, this is where you find out which part ate the budget.

The headline occupancy and the composition counts read the same official token-meter projections the chat composer's context ring reads (contextPressure / contextBreakdown), so the legend's ≈ figures and proportions match the ring's click-open panel exactly; the message bucket is subdivided into the four surface categories by the fold's per-category ratios.

📈 Context Trend — every request's size and its story

One stacked bar per model request — finer than per-message — so you watch the window grow turn by turn, and drop in one ✂ when compaction hits:

• ✨ Step brief — what a step was, not just how big. Three plain-language rows under the chart: User recalls the message that opened the turn (on any of its steps), In lists what newly entered the context — usually the previous tool calls' results, failures flagged — and Response shows what the model returned: its reply text and/or the tools it called. Hover a row's tag to learn what the row means; click any row to open that exact message in the Context browser.
• Read it your way — Step or Turn granularity, Total (cumulative makeup) or Delta (each request's signed change), and sideways scroll through the whole session.
• Hover & pin — scrub for an instant tooltip (turn/step, time, tokens, a one-line reply preview); click to pin the full category breakdown, with provider-reported actual prompt/output/cache figures next to the estimates.
• ✂ marks the events — compactions and prunes land exactly where they happened, so the bars' drops explain themselves.
• Live linkage — hovering a bar previews that step's assembled context in the Context browser beside the chart; leave the chart and it returns to your own pick.

Above: Turn 1 · Step 15 of a real session — the brief recalls the turn's opening message, the files just read in, and the reply that called read next.

A longer session tells the dramatic version — ~563k tokens across 48 turns, then compaction (✂) recycled −535.5k in one step, and the conversation continued from a fresh, small window:

Switch the chart from Total to Delta and each bar becomes the change that request made to the window instead of its cumulative size: diverging stacks pile up from the solid zero baseline when the window grew and hang below it when it shrank, tooltips read Δ ±Nk, and the pinned detail card re-prices every category as a signed delta — so you can tell exactly which part of a request added (or reclaimed) tokens. Below, Turn 5's first step grew the window by +1.6k: injected context +803, the user message +649, the assistant reply +178 — and nothing else moved:

⚡ Context events — when and why the window changed

Every compaction, tool-output prune, skill or plugin context injection, model switch, and plan-mode toggle — each labeled with its producer source (instruction file paths, plugin id, skill name), its token delta (compactions/prunes show the net reclaimed amount, matching the chart's drop), turn/step attribution, and timestamp. Filter by category (Inject / Compact / Prune / Switch / Mode) to see exactly when each kind of event happened and its impact — e.g. when a skill was injected, when instructions were added, or how much a compaction reclaimed:

📁 File Activity — what the agent did to your files

Not what the context is made of, but what it was good for: one row per touched file — read, written, or searched — aggregated up to the step you pick on the trend chart (the card scrubs live as you hover the bars; the default is the whole session to date):

• Per-purpose counts — how many times each file was read, written, and searched, with the header chips doubling as purpose filters (Read / Written / Searched), an Images chip for the multimodal view (read_image calls and image extensions), and a path search box.
• Line deltas — every edit/write contributes its estimated footprint (+added / −removed, read off the call arguments), per file and summed in the header.
• Click to inspect — a row expands into the file's own operation log (tool, time, per-op delta, failures flagged); each operation jumps straight to the exact tool result in the Context browser. Operations whose paths aged out of the retained window are honestly counted in a footnote.

🧭 Context browser — open the box of any request

Pick Live (next request) or any retained step from the picker, and browse what that request was actually assembled from:

Six collapsible category sections (system prompt, tool schemas, user messages, injected context, assistant replies, tool results) expand into one row per element — each with its token price — and every element expands again into its actual content: the full system prompt, each tool's description and JSON schema, message text, reasoning, tool-call arguments, and tool outputs.

• Linked with the trend chart — hover any bar in the Context trend card and the browser previews that step instantly; leave the chart and it returns to your own pick. Keep a category open while scrubbing to compare one category across steps. Clicking a step-brief row (User / In / Response) opens that exact message here, expanded and scrolled into view.
• Honest about coverage — steps before a compaction are reconstructed from the removed-message archive, and the card says so when a step's makeup is only approximate. Elements older than the loaded chat window page older history in automatically when you expand them, and live injections (AGENTS.md, session-start context, …) are always listed — never a token sum without its items.
• Diff against the previous turn — switch the picker to vs previous turn and every category gets signed delta badges (+N items, +Nk tokens), so one glance tells you what the conversation added since the end of the last turn.

Tool results open into the full call and response: the tool's name and arguments with its OK/error status on top, the result body with its line count and a Raw / Markdown display toggle, and any image payload (e.g. read_image output) rendered as a thumbnail card with its name, dimensions, stored size, and estimated token cost — instead of a flattened blob of text:

🖼 Multimodal — image attachments in full view (DeepSeek Harness 0.1.1+)

Fully adapted to DeepSeek Harness 0.1.1's multimodal pipeline and the vision capability of DeepSeek-V4-Flash-Vision-Exp. A user message carrying images expands into a card layout — prose in the text card (with the usual raw/Markdown toggle), each image attachment as a thumbnail card in an equal-width two-column grid with its name, normalized dimensions (plus the pre-normalization size when 0.1.1's image pipeline downscaled it), stored size, and estimated token cost — priced by DeepSeek's official image-size→token conversion (the docs' image token calculator; 117–384 tokens per image under the provider's per-image cap), the same estimate the message/token breakdowns carry — and anything unrecognized as raw content:

Images load through the harness's own session-authorized loader — the same one the chat history uses — and degrade to metadata-only cards when it is unavailable. Image blocks in assistant messages and tool results (e.g. read_image output) now render too, instead of being silently dropped.

Like it?

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

License

Apache-2.0