dsh-think-summary

dsh-think-summary

English | 简体中文

A DeepSeek Harness plugin that turns each reasoning step into 1–3 readable bullet points, so a 30-minute long-horizon run stays skimmable. The full chain of thought is still one click away.

Community plugin. Not affiliated with or maintained by DeepSeek AI.

The problem

Harness renders the model's raw chain of thought. In a long task that means:

• collapsed: a one-line excerpt that tells you nothing, plus a step list like "ran a command, read a file, searched code";
• expanded: several thousand words of stream-of-consciousness, in the order the model happened to think it.

Codex feels different because it shows a summary that the model itself produces (OpenAI's reasoning-summary channel). DeepSeek has no such channel — the only way to get that experience is to generate the summary locally, one extra cheap call per step.

What it does

Before:

› 思考   The user reports a UI issue after the latest changes: a message shows raw
         evidence (the fallback) at the front, trace panel shows 24 tools…

After:

› 思考 · 摘要
  • Root cause: the backend blocks the final event on memory extraction, so the client
    falls back to raw evidence
  • Plan: emit the final event before sealing, add a 12s timeout, fix the continuation
    helper to skip orphan user messages
  • Still needs tests for event ordering and timeout behaviour

Click the row to read the original reasoning, rendered as Markdown exactly as before.

Install

dsh plugin --profile <your-profile> add @drifter-yh/dsh-think-summary

Some builds also expose a plugin registry under Settings → Plugins; otherwise use the command above.

Restart Harness once after installing: the host half is loaded as a forked module and cannot be hot-swapped.

Configure

Configuration lives in your profile's cordis.patch.yml (or the plugin's config panel):

- id: think-summary
  name: '@drifter-yh/dsh-think-summary'
  config:
    provider: opencode-go          # any registered LLM route
    model: deepseek-v4.1-flash     # a cheap, fast model is the point
    enabled: true                  # master switch; false = no model calls at all
    maxOutputTokens: 2000          # budget for (reasoning + answer) tokens
    maxInputChars: 6000            # longer reasoning is clamped head+tail
    maxCallsPerSession: 300        # per-session cost ceiling
    maxCallsTotal: 600             # per-process cost ceiling

Emergency switch: localStorage.setItem('dsh-think-summary.enabled', 'false') in the browser disables requests for that page without touching the profile.

How it works

browser                                              host process
┌───────────────────────────┐   POST /summarize   ┌──────────────────────────────┐
│ shadows assistant-step    │ ──────────────────▶ │ ctx.llm.stream()             │
│ collapsed → bullets       │ ◀────────────────── │   provider/model from config │
│ expanded  → raw reasoning │     1–3 bullets     │ content-hash cache + budget  │
└───────────────────────────┘                     └──────────────────────────────┘

• The client registers conversation.chat.node at priority: -1 for assistant-step, filtering blocks exactly like the shipped renderer (groupPart rules, text / reasoning / image; tool calls stay separate nodes).
• A summary is requested only when a step settles — never mid-stream.
• Summaries are keyed by a SHA-1 of the (clamped) reasoning text, so identical text is never billed twice, and the cache survives restarts.
• Any failure (no route, timeout, truncation, empty output) degrades to the original one-line excerpt. The plugin can never make the transcript worse than stock.

Cost and latency

One cheap call per settled step. On deepseek-v4.1-flash a typical step takes 3–20 s and a few thousand input tokens; the answer itself is ~3 short lines. A 40-step run is well under one unit of currency on most routes — but this is a paid feature, so the budgets above exist on purpose.

Rows older than three hours are not backfilled; opening an old session does not trigger a wave of calls.

Privacy and security

• Reasoning text is sent only to the provider you configure — the same provider that already serves the conversation.
• Summaries are cached locally at ~/.dsh/storages/think-summary/cache.json.
• The HTTP endpoint (/dsh-think-summary/*) accepts loopback connections only, caps request bodies at 512 KB, requires POST for summarising, and runs at most two model calls at a time.
• No telemetry, no analytics, no network calls other than the model request.

Compatibility

• Harness >=0.2.2-rc.1 <0.2.3-0; verified on 0.2.2-rc.1 using the desktop profile (the same web client Harness serves).
• The plugin shadows the shipped assistant-step renderer. A Harness release that changes the node contract can break rendering; the renderer is wrapped in an error boundary that falls back to plain text.

Limitations

• File mentions (@path) inside assistant text render as plain text for now.
• Chinese bullets only — the prompt asks for Chinese; other languages are not wired up.
• Reasoning effort cannot be disabled on routes that only expose low/high/max, so maxOutputTokens must leave room for thinking. That is why the default is 2000.
• History is not fully backfilled (3-hour window) to keep costs predictable.
• Tool calls are still rendered by Harness itself; only assistant-step nodes are shadowed.

Troubleshooting

Development

node --check index.js
node --check client.js
npm pack --dry-run

Releases are published from CI on tag push (.github/workflows/release.yml) with npm provenance; it needs an NPM_TOKEN repository secret.

The host half is a Cordis plugin (apply(ctx, config)); the client half is an ESM module registered through window.__ModuleLoader__.load.

License

MIT — see LICENSE. Third-party notices: THIRD_PARTY_NOTICES.md.