dsh-dashboard

DSH Dashboard

Per-session statistics for DeepSeek Harness: token usage, timing, tool and skill activity.

Adds a Statistics tab to the conversation view. The numbers update live as the session progresses.

中文 | English

Compatibility

Supports DeepSeek Harness v0.1.2-alpha.1 through v0.1.3-alpha.2 (npm alpha dist-tag): the client half targets the conversation.view slot, the tokenUsage / sessionStats projections, the chat snapshot's legacy.nodes timeline and the turn-tail turn data — verified present and field-identical across the dsh-v0.1.2-alpha.1 through dsh-v0.1.2-alpha.5 tags, re-verified at dsh-v0.1.2-rc.1 (the rc.1 tree diff against alpha.5 is a repo-wide package.json version bump only: zero non-package.json files and zero non-version lines changed, so the plugin-consumed contract surface is byte-identical), and re-verified at dsh-v0.1.3-alpha.2. The re-verification covers: tool-result node fields (time, call.name/argsRaw, callTime, isError), the turn-tail business data keys (tokenUsage, ttftMs, tokensPerSecond), the sessionStats / tokenUsage projection keys, and the built-in view entries (chat at order 0, trajectory at order 10 — the statistics entry at order 20 does not collide). 0.1.3 changed the streaming event generation (assistant/attempt + assistant/live-chunk replace assistant/chunk and the chunkrow/* rows) and the projection sampling point, but every field this plugin reads is unchanged. Peer dependencies use the ^0.1.2-alpha.1 || ^0.1.3-alpha.1 range: a bare ^0.1.2-alpha.1 does not match 0.1.3-alpha.2, because npm's prerelease rule only matches a prerelease whose major.minor.patch has a comparator inside the range.

Not supported: 0.1.1-rc.2 and earlier — that generation runs the pre-useChat client-runtime contract, so the tool and skill tables would silently stay empty.

Compatibility risk to track: the chat snapshot's legacy slice is the host's compatibility projection; the shipped StatsLine still consumes it through 0.1.3-alpha.2 (re-verified), but re-check it on every harness upgrade.

Install

Releases are version-aligned with the harness: each version is built for, and named after, the matching @deepseek-ai/dsh release.

From npm

Pick the dist-tag that matches your harness:

1. npm alpha (0.1.3-alpha.2) — compatible with dsh v0.1.3-alpha.1 ~ alpha.2, and with the whole v0.1.2-alpha.1 ~ v0.1.2-rc.1 range (the contract surface it reads is unchanged):

   dsh plugin --profile web add @wenqi_bian/dsh-dashboard@alpha
   

2. npm latest (0.1.2-rc.1, kept in sync with next) — conservative channel for dsh v0.1.2-alpha.1 ~ rc.1:

   dsh plugin --profile web add @wenqi_bian/dsh-dashboard
   # equivalent: dsh plugin --profile web add @wenqi_bian/dsh-dashboard@next
   

The previous generation 0.1.1-rc.2 no longer backs any dist-tag; install the exact version if you still need it, but it is not compatible with current harnesses (see Compatibility).

The npm package ships the prebuilt host and client bundles, so no build step is needed on install.

From source

The source version — 0.1.3-alpha.2, compatible with the harness source at the dsh-v0.1.2-alpha.1 through dsh-v0.1.3-alpha.2 tags (rc.1 equals alpha.5 apart from version numbers; 0.1.3-alpha.2 keeps every field this plugin reads) — is what the GitHub Release v0.1.3-alpha.2 ships. Install it from source for local development:

dsh plugin --profile web add .

Run it from the repo root; it links the plugin into the web profile that $DSH_HOME points at. Rebuild after any code change with pnpm run bundle (see Quickstart).

Quickstart

pnpm run bundle                    # emit lib/index.js (host) + lib/client.js (browser)
dsh plugin --profile web add .
# restart dsh web, open a conversation, pick the Statistics view

The tab appears beside Chat / Trajectory and reads everything on the client from the session's tokenUsage / sessionStats projections and the conversation snapshot.

What it shows

• Token layer: input (fresh + cache read), output tokens, cache hit rate.
• Timing: LLM time, tool time, first-token latency, decode rate. The per-turn TTFT curve is drawn only while a session streams; after a reload that card falls back to the durable session average (harness 0.1.3 keeps the first-token timestamp in a client-only event).
• Tool activity: per-tool calls, fails, total and average duration, drawn as proportional segments with a four-color cycle.
• Skills: which skill tools were invoked, as chips.
• Insights: tool success rate, think/tool time split, tokens per turn, input:output ratio.
• Locale follows the DSH host (zh / en).

Data source

Stats are derived on the client from the conversation snapshot's tool-result nodes and the built-in sessionStats / tokenUsage projections, the same source the shipped composer StatsLine uses. Every harness-owned name and type the client half touches — the slot name, the projection keys, the turn-tail data key, node kinds, and the theme/DOM variables its styles borrow — lives in src/client/contract.ts, so a harness upgrade has one file to check. There is no host @Remote call, because @Remote is not generated for a runtime-loaded out-of-repo plugin. The workspace cross-session aggregate is not implemented.

License

MIT.