dsh-token-telemetry

A DeepSeek Harness (dsh) plugin — part of the dsh-plugin-* family — that surfaces token throughput telemetry from real harness data, in three layers:

• Per-turn inline card (conversation.chat.turnTail) — under each completed turn, a compact row: thinking / output tokens, throughput tok/s, wall time, a mini dual-line sparkline, and a "明细" toggle that expands a per-step table (step · think · out · duration · tok/s · tool · state).
• Aggregate view (conversation.view tab "Token 遥测") — full thinking-vs-output TPS curve, settlement cards (decode TPS, mean TTFT, total tokens), and a cross-turn overview table.
• Live counters — the curve and numbers recompute every 250 ms from the streaming events, so they animate during a turn (the durable sessionStats/tokenUsage projections only tick at step/end).

All figures come from the harness's own durable surfaces — no simulated data:

Panel field	Source
decode TPS / model TPS / TTFT / turns / steps / toolMs	sessionStats projection (@deepseek-ai/dsh-session-stats)
total / output / cache-read tokens	tokenUsage projection (@deepseek-ai/dsh-token-meter)
live TPS curve	SessionBinding.eventSource → settled assistant/message stream (timed AssistantStreamRecords) + transient assistant/live-chunk deltas

The curve reconstructs from the durable timed stream carried on each settled assistant/message event, so it renders correctly even after a turn completes (not only mid-stream).

Architecture

dsh is "everything is a plugin" on Cordis. A plugin is one package with two halves:

• Host half — exports["."] → apply(ctx) (Node). This plugin has no host-side behavior (it only surfaces data the host already computes), so apply is a no-op that makes the row appear in the composed profile.
• Client half — exports["./client"] → lib/client.js, a browser module registered via window.__ModuleLoader__.load(...). It declares inject = ["slots", "sessions"] and in apply(ctx) mounts a React component into the conversation.view slot.

The client bundle is hand-authored in the exact shape the loader expects (React is an external resolved by require("react")), so no bundler is required to build it.

Install into a dsh profile

1. Make the package resolvable from the profile (here: ~/.dsh/profiles/web, boot dir's node_modules), e.g. copy it to node_modules/dsh-token-telemetry.

2. Add one insert row to the profile's cordis.patch.yml:

   - insert:
       - id: token-telemetry
         name: dsh-token-telemetry
   

3. Restart dsh web. The "Token 遥测" tab appears in a session's conversation view.

Verified

Tested against a clean dsh web instance (@deepseek-ai/dsh@0.1.7-alpha.1) with real StepFun step-5-preview turns:

• Aggregate + live counters: confirmed against real data. Settlement matched dsh's native stats pill exactly (74 tok/s, 11.9K tok), and a multi-step agent task (a Gomoku build) produced a correct per-step cross-turn table (13,427 output tok / 132,814 input+cache / ~92 tok/s avg).
• Per-turn inline cards and the trajectory ribbon render with real per-step curves, colored by phase (thinking / output / tool) and sized by wall time — confirmed via headless screenshots. Phantom turn-0 rows and empty step blocks are filtered out.

Notes / limitations

• Token counts on the live curve use a CJK-aware heuristic (estTokens); durable settlement uses the provider-reported sessionStats/tokenUsage values.
• The plugin is per-session (a conversation-view tab). A multi-session "queue" overview would add a main + sidebar.panellist panel reading useSessions.
• dsh is in developer preview; plugin surfaces can change.

MIT