Context Explorer

A panel for DeepSeek Harness that shows what a Session actually sends to its model: the rendered system prompt, the tool schemas, and every message, drawn as a grid of token squares. It appears as a tab in the Web UI's right sidebar, titled 上下文 / Context.

This is an unofficial third-party plugin, not affiliated with or endorsed by DeepSeek. MIT licensed — see LICENSE; THIRD_PARTY_NOTICES.md records the two files reproduced from DeepSeek projects (estimate.js, serve.js).

Install

dsh plugin --profile <name> add dsh-context-explorer

dsh plugin forwards its arguments to pnpm in that profile, so the flag names the profile to install into. The package declares dsh.bundle.patch → ./cordis.patch.yml, so the profile loader applies its one host row the next time the profile loads. The runtime files are plain checked-in JavaScript with no build step and no install script, so a github: install needs no build permission. Node ^22.19.0 or >=24.0.0. locale/*.json carry the display name and description per language (上下文浏览器 / Context Explorer). Verified against @deepseek-ai/dsh 0.2.0-rc.2; the plugin declares no version gate, so a Harness release that moves these extension points would show up as a missing or empty panel rather than a refused install.

What the panel shows

A headline. 已用 {used} / {total}: used is the prompt size the provider reported for the last request, and until a call reports one it is the panel's own total. total is the context window, or — when the host has no window yet. The tooltip says which figure is which.

Composition chips. One per kind of content present, heaviest first: an 8×8 swatch in that kind's colour, the kind's name, its token total and its share of the panel total. They wrap, and they double as the grid's legend. Clicking a chip filters the list to that kind; clicking it again clears the filter.

The grid. The whole context window, not just the part in use, as squares each covering an equal slice of the window. Occupied squares carry the colour of the kind holding most of that slice; a slice with at least 35% held by a second kind is drawn as a diagonal blend; the space past the end of the context is drawn flat as free space. Squares are 7 device pixels on a side with a 1-pixel gap, so every square paints the same size on any display scale. Hovering names the kind or kinds. Clicking a square opens the region that holds most of it.

The list. One row per region, in the order the model sees them: colour swatch, a one-line preview, a 已截断 / clipped badge on a closed row whose fragment was cut, and the region's token figure with its share. A figure the provider counted is printed bare; this plugin's estimate carries ≈. Clicking a row opens the region's own fragment of the prompt the model receives — role markers and all — rather than a summary of it; the badge gives way to detail.clippedNote in the body, so the fact is stated once. A count and a toggle switch the list between model order and size order.

With nothing to draw — no published view yet, or a view with no regions — the panel says so instead of an empty shell. A fold failure prints above the headline. A rendering error replaces the tab rather than leaving a blank panel.

Where the numbers come from

The headline is the provider's own count of the last request. Each region's figure is the provider's count where one covers that region, and this plugin's estimate otherwise — marked ≈. Two things are measured: an answer's own size, when the report states the output it was generated with, and a message that is the only one in a request window, which is what the prompt grew by less the previous answer. The system prompt, the tool envelope, a compaction summary, a window holding several messages, and anything appended after the last report have normally no provider count of their own, so they stay estimated.

The estimate is the Harness's own fixed heuristic, the same one its context meter uses: one token per four UTF-16 code units, plus four per content block and four per message. It prices content and not wire framing, and it drifts with the language of the text, so read an estimated figure as composition, not as a bill.

Known limitations

• The prompt text is a reconstruction, not a capture. The Harness talks to the hosted API, which renders the prompt server-side, and the session log records no request body. serve.js transcribes the published serving implementation so the panel can show a region in the form the model receives; tools/prompt-golden.mjs holds that transcription byte for byte against the official open-weights encoder. Treat it as what that implementation renders.
• The panel shows fragments, not one continuous prompt. Each region carries its own piece. Two pieces travel whole because each exists once per request: the hoisted system prompt (the earliest one, not the newest), whose section boundaries exist nowhere else, and the ## Tools block. A message's fragment is capped at 4000 UTF-16 code units and spent in the order the prompt spends it; a region whose fragment was cut is marked clipped and its expansion says only the beginning was kept. The row's preview is bounded separately at 400 code units.
• Host-half changes need a server restart. index.js, serve.js and estimate.js are read when the DSH server starts, so restart dsh web after editing them. client.js is served through the client module system, which reloads the browser half without a restart.
• No automated visual check. The suites cover the fold, the wire, the browser helpers, the packaging and the prompt text; nothing renders the panel in a browser, so its layout is checked by hand.

Layout

One npm package, one plugin, both halves in it:

Path	Role
index.js	Host half: the sessionProjections unit that folds a session into the view
client.js	Browser half: the tab, the grid, the region list
serve.js	The served-prompt renderer, transcribed from the DeepSeek serving implementation
estimate.js	The token estimator, a local copy of the shipped one
package.json, cordis.patch.yml	Manifest and the bundle patch (one host row)
icon.svg, locale/*.json	The mark, and the name and description per language
docs/prompt-injection-format.md	How a Harness conversation becomes prompt text
tools/	Checks and diagnostics

The host half imports nothing but its sibling estimate.js and serve.js, which is what lets a profile link this directory directly instead of installing a published copy.

Checks

Run from the repository root; each exits non-zero on failure.

• node tools/check-bundle.mjs — packaging and bindings: the manifest ships every runtime file, both halves resolve through it, every class the stylesheet styles is applied by an element, and no declaration shadows a builtin.
• node tools/context-explorer.test.mjs — the host fold: region order, the accounting behind every share, the prompt surviving a rewrite, the truncation flag, and agreement with the installed estimator.
• node tools/context-explorer.contract.mjs — the wire the two halves share: locale keys, the fields each half names, the grid's square budget, the palette.
• node tools/context-explorer.helpers.test.mjs — the browser half's pure helpers, loaded through its real module registration.
• node tools/prompt-golden.mjs — serve.js against the official encoder, byte for byte.

tools/replay-session.mjs <session.v4.jsonl.zstd> replays a real log, and tools/activation-probe.mjs mounts the plugin as the Loader does and asserts the projection key reaches a snapshot.