dsh-context-taxonomy

1. 安装并启动 DeepSeek Harness：npx @deepseek-ai/dsh web
2. 在终端执行上面的安装命令（CLI 会解析插件并核验来源）
3. 用 dsh plugins list 确认已安装，必要时重启 Harness 生效

插件以当前 dsh 进程的权限运行，安装时可能执行代码。请先通读仓库源码与许可证，确认无破坏性命令与越权访问；本站只做索引，不对第三方插件安全性作担保。

事实信息来自公开插件目录快照（2026-10-01），介绍文案由本站再加工。

以下为插件仓库 README 全文（原始内容，由公开目录抓取整理）。

# DSH Context Taxonomy

English | [简体中文](README.zh-CN.md)

> **See how DeepSeek Harness builds the context for every ordinary agent call.** Inspect the complete system prompt, conversation history, current prompt, tool definitions, model options, token composition, cache usage, and reasoning evidence in one explorable view.

Context Taxonomy is a learning and debugging companion for [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness). It turns the provider-neutral logical request at Harness's public `llm/stream` dispatch layer into an inspectable taxonomy, then keeps the result beside the conversation in a dedicated **Context Taxonomy** tab.

![A real DeepSeek Harness session opening Context Taxonomy and inspecting the assembled system prompt, messages, tools, options, and token composition](https://raw.githubusercontent.com/ArtificialNotImbecile/dsh-context-taxonomy/codex-initial-plugin-assets/context-taxonomy-demo.gif)

_Recorded from a real DeepSeek-V4-Flash run on the official Harness `0.1.0-rc.6` Web profile—not a mock or fixture._

## How this differs from Harness Trajectory

Context Taxonomy complements the official [**Trajectory**](https://github.com/deepseek-ai/deepseek-harness/tree/master/packages/client/ui-trajectory) view; it does not replace it. The two views deliberately overlap on system prompts, tool definitions, request options, and reported token usage, but organize that evidence around different questions.

| | Harness Trajectory | Context Taxonomy |
| --- | --- | --- |
| Primary question | What happened during the agent run, and in what order? | What context made up this particular logical model call? |
| Organization | A Turn/Step event ledger covering User, Assistant, Tool, Subtool, retries, compaction, and timing. | A per-call snapshot divided into System, Conversation, Current prompt, Tools, Options, and Unclassified sections. |
| Best at | Following execution flow, inspecting tool inputs/results, diagnosing latency, and locating failures or retries. | Explaining prompt composition, context growth, exposed tool schemas, message provenance, and unexpected logical-request fields. |
| Context analysis | Shows the recorded prompt, tools, options, usage, and prompt changes alongside the execution ledger. | Adds estimated composition by category, `MessageSource.kind` / `ContextForm` provenance, unclassified-field surfacing, a logical reasoning-retention check, and sanitized canonical JSON. |
| Coverage and history | Reconstructs the Session timeline, including compaction, from official Session data. | Observes ordinary agent-loop calls that reach this plugin's `llm/stream` listener; it excludes auxiliary calls and cannot reconstruct calls made before installation. |
| Storage | Uses the official Session data already retained by Harness. | Keeps a separately retained, sanitized sidecar so each observed logical call remains independently inspectable. |

Use **Trajectory** when you need the execution story: model responses, tool calls, nested work, timing, retries, and compaction. Use **Context Taxonomy** when you need the request anatomy: why a call had this prompt, which context and tools were exposed, where its estimated input composition came from, or which fields do not fit the expected taxonomy. Using both gives the most complete debugging and learning view.

## Quick start for existing Harness users

If `dsh --version` already works and reports `0.1.0-rc.6`, install the plugin into your Web profile and start Harness with the short commands:

```sh
dsh plugin --profile web add @artificialnotimbecile/dsh-context-taxonomy@0.1.0
dsh web
```

If `dsh web` is already running, stop and restart it after installation so the new bundle is loaded. Then create or open a Session, send an ordinary agent request, and select the **Context Taxonomy** tab. The plugin records calls made after it is installed; it cannot reconstruct older calls.

If the `dsh` binary is not on your `PATH`, use `npx` as a reproducible fallback:

```sh
npx --yes @deepseek-ai/dsh@0.1.0-rc.6 plugin --profile web add @artificialnotimbecile/dsh-context-taxonomy@0.1.0
npx --yes @deepseek-ai/dsh@0.1.0-rc.6 web
```

The longer form does not install Harness globally: `npx` runs the pinned Harness CLI from its cache. In both forms, `plugin --profile web add` modifies only the local Web profile, normally stored under `$DSH_HOME/profiles/web` or `~/.dsh/profiles/web` when `DSH_HOME` is unset.

Use it when you want to answer questions such as:

- What system prompt did Harness assemble for this call?
- Which messages were conversation history, current context, or plugin injections?
- Which tools and JSON schemas were exposed to the model?
- Which provider, model, reasoning effort, sampling options, and token limit were selected?
- How much of the input came from the system prompt, messages, tools, and options?
- Did the adapter report cache reads/writes or reasoning tokens?
- What changed between retries or later calls in the same Session?

That makes the plugin useful for learning Harness architecture, teaching plugin development, debugging prompt growth, comparing agent presets, auditing tool exposure, and investigating model behavior without reading a raw Session log by hand.

The plugin is intentionally precise about what it observes: it does **not** capture provider HTTP payloads, headers, endpoints, serializer output, transport attempts, or delivery status. Calls that fail before LLM dispatch, or that are intercepted before this listener, are not visible. In the UI, provider-reported usage is labeled actual and locally derived composition is labeled estimated.

## What you can inspect

- System, conversation, current-prompt, tool, option, and unclassified sections.
- DSH `MessageSource.kind` and `ContextForm` provenance when messages carry it.
- Reported prompt/cache/output/reasoning usage without treating missing fields as zero.
- Estimated per-section composition, clearly separated from provider-reported usage.
- A logical-only DeepSeek reasoning-retention check; it is never presented as wire evidence.
- Sanitized canonical logical JSON with lazy, bounded paging.
- Durable per-Session captures across Harness restarts.

### Read the call at a glance

![Context Taxonomy overview showing actual input usage, estimated composition, cache evidence, reasoning evidence, and the categorized context tree](https://raw.githubusercontent.com/ArtificialNotImbecile/dsh-context-taxonomy/codex-initial-plugin-assets/context-taxonomy-overview.png)

### Inspect the sanitized logical request

![The lazy raw-request explorer showing the sanitized Harness logical request and its explicit non-wire disclaimer](https://raw.githubusercontent.com/ArtificialNotImbecile/dsh-context-taxonomy/codex-initial-plugin-assets/context-taxonomy-logical-request.png)

## Install from GitHub source

The first release targets DeepSeek Harness `0.1.0-rc.6` exactly and supports its `web` profile only. Installation is the explicit opt-in to local capture.

For an auditable source install, pin the release and select this repository's publishable subdirectory:

```sh
npx @deepseek-ai/dsh@0.1.0-rc.6 plugin --profile web add \
  'github:ArtificialNotImbecile/dsh-context-taxonomy#v0.1.0&path:packages/dsh-context-taxonomy'
```

Git installs execute the package's `prepare` build. On the first attempt, pnpm 11 stops safely and prints the exact codeload URL + commit + workspace-path key required under `allowBuilds` in `$DSH_HOME/profiles/web/pnpm-workspace.yaml`. Review the pinned source, add that exact key, and repeat the command. A package-name-only entry is intentionally insufficient for Git dependencies. The precompiled npm release is the recommended install path.

## Data handling

The compact index uses Harness `storageDomain`; sanitized logical JSON is stored as private gzip blobs under `$DSH_HOME/context-taxonomy`. Built-in redaction covers secret-like keys, bearer credentials, credential query parameters, data URLs, and large base64 values, and cannot be disabled. Hashes are computed only after sanitization.

Defaults retain 30 days, 200 captures per Session lifecycle, 512 MiB globally, and 16 MiB per capture. Directories use mode `0700` and blobs `0600`. The plugin does not provide at-rest encryption or a mutation/deletion Remote in v1.

## Development

```sh
pnpm install --ignore-scripts
pnpm run build
pnpm run test
pnpm run pack:plugin
```

The repository root is a private workspace. The publishable package lives at [`packages/dsh-context-taxonomy`](packages/dsh-context-taxonomy), while Kimi K3's UI specification and framework-free prototype live under [`design`](design).

## License

MIT