dshbase

Blog · Review

A real job, end to end: 66 minutes and a few yuan for a RAID reassembly

August 25, 2026 · dshbase · hands-on test

DeepSeek Harness hands-on test hero: deepseek-v4-flash model plus the harness completes a RAID reassembly in 66 minutes, 200-400万 token, ¥3-6, 99% cache hit

On August 13 DeepSeek open-sourced its first agent framework, DeepSeek Harness (v0.1 developer preview, MIT). The formula is one line: Model + Harness = Agent — the model thinks, the harness actually executes. This note skips the concepts and runs one real workload through it, with numbers.

Install and first run

Installation is unremarkable: pull the official npm package and launch the web UI locally with npx — no database, no server to configure, out of the box. You need an API key from the DeepSeek open platform to start, billed per token. Cache hits are the money lever here: re-using the same input segments hits cache and cuts cost, which is exactly why repeated analyses get cheap.

Prompting an assistant tool to install DeepSeek Harness from the official npm package, model set to deepseek-v4-flash

The four modes

Different tasks load different default plugin sets:

  • Standard — the full tool combination. Best for everyday dev and general tasks; the most feature-complete.
  • PTC (Programmatic Tool Calling) — the model composes multiple tool calls as a single block of code. Best for long call chains; a real token-saver.
  • Minimal — just one shell tool and one file editor. For minimal-environment model benchmarks and well-defined small jobs.
  • Create — inspect the running runtime, experiment with Cordis plugins in memory, and compose new presets. For advanced, high-trust tinkering.

Practical rule: for long, cost-sensitive tasks prefer PTC, and keep an eye on the cache-hit rate in the corner — the higher it is, the cheaper it gets.

The job: RAID reassembly of a forensic image

Using Standard mode on a forensic evidence folder, the agent was pointed at a RAID reassembly task. The status bar streams inference rounds, per-step timing, and input/output tokens in real time — the whole reasoning process is transparent, which is the most visible difference from an ordinary chatbot. When the same evidence gets re-analysed or similar questions repeat, cache hits make those runs nearly free.

The outcome on deepseek-v4-flash: about 66 minutes total, roughly 2–4 million tokens, and around ¥3–6 spent. A complete RAID-reassembly analysis for pocket change. Two caveats worth flagging: that token volume is normal for deep reasoning plus long tool chains, and after the August 17 price adjustment the same job is expected to land around ¥10–20 — so the cheap window was exactly that.

Trajectory: the part that matters later

Every step — system prompt, tool call, reasoning — lands in a replayable trace. That Trajectory can be replayed for review, branched, or folded straight into a Skill so a successful run becomes a reusable standard operating procedure instead of a one-off. For anyone who will redo the same kind of analysis, that's the quiet killer feature: the working procedure, not just the result.

The takeaway for installers

This is a single real workload, not a general benchmark. What it demonstrates is that the harness is comfortably usable for a non-trivial forensic/computation task at a cost that makes it a daily tool — while the transparent status and the trajectory-to-skill loop are where the unlike-chatbot value actually lives. Run your own representative task in Standard or PTC mode, watch the token counter, and reuse the trace.

All articles →