dsh-ctx-visualizer

English | 简体中文

See exactly what context DeepSeek Harness assembles for every model call — system vs user content, tool schemas, parsed tool calls, tool executions and the follow-up calls — as a per-turn timeline and a dynamic architecture view.

A community plugin for DeepSeek Harness (tagged dsh-plugin), not an official package. It is for anyone who wants to study the agent loop in detail: inside one multi-turn exchange — call the model → parse tool calls → execute tools → feed results back → call the model again — what exactly was sent to the model at each step, what was added compared with the previous step, and which part is growing.

Features

After installing, the conversation header gains a Context Lens tab:

• Left · turns: one card per turn (user question, call count, tool-run count, context tokens, elapsed time, most-used tools). The column is a fixed-height window you can scroll to reach earlier turns.
• Top · horizontal timeline: every step inside the selected turn — Step 1 · model call → runs bash / results fed back → Step 2 …. Click any node to switch.
• Right · detail for the selected step:
  • Architecture (default): generated dynamically from the actual content — the system prompt is split into sections by its own headings, tools are grouped by origin (MCP server / built-in), and the conversation history is drawn as user → model call → tool result chains per user input, with messages new to this step highlighted and compaction summaries called out. On the right, this step's flow: request → model → each tool call's outcome/duration/result size → fed back into the next request.
  • Request context: shows only the messages new in this step by default; system content and the tool list collapse automatically when unchanged; a stacked bar of the context composition (system / user input / tool results / plugin injections / assistant history / tool schemas); input/output tokens and cache-hit rate.
  • Response and Tool runs.
  • Raw JSON: the full record plus sub-blocks for system content, user content (user input, tool results, injections) and the model response — all collapsed by default and individually copyable.
• Persistence: live captures are written to ~/.dsh/ctx-visualizer/captures/ and restored after a plugin reload or a Harness restart.
• Backfill: steps from before the plugin was enabled are rebuilt by replaying the session log's context changes (including compaction) and marked as rebuilt.
• Redaction: secrets shaped like Bearer …, sk-…, tos_…, ghp_…, xoxb-… are masked before storage.
• Language: follows the Harness locale automatically (Chinese/English), with a manual switcher in the toolbar; the choice is remembered.

Compatibility

Item	Version
Developed and verified on	DeepSeek Harness Desktop 0.2.0-rc.2, @deepseek-ai/dsh-* 0.2.2-rc.1 (Web GUI)
Node.js (to run the tests)	≥ 20

DeepSeek Harness is still in Developer Preview, so the events and interfaces this plugin relies on may change incompatibly. If something breaks, check the version first and open an issue.

Install

The plugin has no npm dependencies, so cloning is enough:

git clone https://github.com/cenbonew/dsh-ctx-visualizer.git

Then install that local bundle in DeepSeek Harness, either way:

• Ask the agent: say “install the local bundle at /absolute/path/dsh-ctx-visualizer with the Plugin Manager”.
• Settings page: Settings › Plugins (if your build supports installing a bundle from a local path).

dsh.bundle.patch in package.json is what makes it installable as a bundle; the frontend half is declared by dsh.client and joins the Web GUI automatically.

After installing, refresh the page and switch to “Context Lens” at the top of any conversation. Model calls made after the plugin was enabled are captured live; earlier steps are backfilled from the session log automatically.

To update: git pull, then toggle the bundle off and on in the Plugin Manager (or restart Harness) and refresh the page.

How it works

Read-only observation built on first-class Harness events; the agent loop is never modified:

Event / service	Purpose
llm/stream (waterfall)	Receives the full GenerateOptions of every model request (messages, tools, model, …) and collects text, reasoning, tool calls and usage while passing the response stream through unchanged
agent/request	Associates turn / step numbers
tools/execute, tools/result	Tool arguments, results, duration and outcome, correlated back to their step by callId
agent/error	Records failed steps
sessionQuery.readSession	Backfill: replays the surfaceOp (append / replace over a seq range) of system/message, user/message, assistant/message and tool/result to reconstruct the context each call received

The frontend is a conversation.view slot plugin that reads the host-side ctxVisualizer service over typert Remote. An Inspect Provider named CtxVisualizer is also registered so an agent can query capture status with cordis_inspect_query.

Every listener is wrapped in try/catch, so a capture failure never affects the agent itself.

Configuration

There is no settings form yet. Pass config to the ctx-visualizer row in the profile's cordis.patch.yml, or edit DEFAULTS at the top of lib/index.js:

Key	Default	Meaning
capture	true	Master switch
persist	true	Write captures to disk
persistDir	~/.dsh/ctx-visualizer/captures	Capture directory (0700, files 0600)
backfill	true	Rebuild earlier steps from the session log
backfillIntervalMs	30000	Minimum gap between two backfills of one session
maxStepsPerSession	80	Live steps kept per session (rebuilt steps do not count)
maxSessions	40	Sessions kept
maxCharsPerText / maxCharsPerToolResult	400000 / 120000	Truncation bound per stored text block
redactSecrets	true	Redact before storage

Privacy and security

• Captures are the full model context and may contain code, file contents and business data returned by tools. Everything stays on this machine; nothing is uploaded.
• The default redaction only recognises well-known secret prefixes and cannot mask every sensitive value. Check before sharing a screenshot or a copied JSON block.
• Model replies and reasoning text are not redacted yet.
• “Clear captures” in the toolbar also deletes the files on disk.

Limitations

• Auxiliary calls such as session-title generation and context compaction are not part of the session log, so they cannot be backfilled (tick “Show aux calls” to see them when captured live).
• File attachments are projected to text in the actual request, so a rebuilt step shows the original attachment blocks.
• Character counts are estimates; for exact token counts, trust the usage reported by the provider.

Development

npm test   # self-contained; needs no Harness install and never writes to ~/.dsh

• lib/index.js — host half (zero npm dependencies, Node built-ins only)
• lib/client.js — frontend, hand-written in Harness's ModuleLoader format; no build step
• cordis.patch.yml — bundle patch that inserts the ctx-visualizer row

License

MIT