dsh-explain

Turn everyday DSH work into a private, continuous learning loop.

dsh-explain is a learning-mode plugin for DeepSeek Harness. It turns useful concepts from completed work sessions into structured explanations, keeps them in one global learning thread, and adapts future explanations to what the user already knows.

The primary agent stays untouched: Explain uses its own model calls, scheduler, context, and local SQLite database.

See it in action

Select text or choose Learn from this answer, review the editable /explain draft, generate a learning card, then mark it understood. This demonstration used real DSH Web sessions and real DeepSeek main-agent and Explain model rounds; the exact commits and recording conditions are preserved in PR #16.

Quick start

Explain currently targets DSH 0.1.0-rc.7.

npx @deepseek-ai/dsh@0.1.0-rc.7 plugin --profile web add github:yuezengwu/dsh-explain
npx @deepseek-ai/dsh@0.1.0-rc.7 web

Then open Settings → Learning, select an auxiliary provider and model, enable learning mode, and save. Explain observes only future completed top-level turns; it does not scan existing history.

Git-hosted plugins build during installation. If pnpm requests build approval, add the printed dsh-explain entry to the profile's pnpm-workspace.yaml, then repeat the install command.

Ways to learn

Use /explain on, /explain off, and /explain status to control or inspect the runtime without leaving the composer.

Each learning card answers three questions:

• What is it? A concise explanation of the concept.
• Why does it matter? The practical reason it matters in the source work.
• What is the common pitfall? A mistake or misconception to avoid.

Choose Got it to close the card, or Not yet to request a different explanation. Rephrasing remains available even if the source session is later deleted.

One learning thread, many work sessions

Every $DSH_HOME owns exactly one Explain learning thread. Individual work sessions contribute material, but resumes and forks never copy the learning state.

• Each source session has at most one explanation awaiting feedback.
• All work sessions display the same global history in the first-party Learning tab.
• One global scheduler serializes manual explanations, autonomous evaluation, rephrases, and compaction.
• The default autonomous budget is 50 requests per rolling 24 hours and survives restarts.
• A private ExplainContext tracks explanation preferences, knowledge level, and learning progress.
• When structured observations or closed explanations are pending, auxiliary history compacts after 30 minutes without an Explain action, or before a request would exceed 50% of the selected model's context window.

Local-first by design

Explain uses first-party DSH conversation.view, composer, assistant-action, and settings extension points. It does not require better-sidebar or patches to other plugins.

Compatibility and verification

• Current compatibility line: DSH 0.1.0-rc.7.
• Unit suite: 64 tests.
• Assembled DSH Web acceptance: 4 scenarios.
• Explain-owned shortcut acceptance: 3 M6 scenarios.
• Real-model workflow evidence: PR #16.
• Detailed acceptance matrix: docs/ACCEPTANCE.md.

DSH is still a developer preview. Explain follows the current public API line and does not retain compatibility layers for earlier private-preview packages.

Local development

The default install uses the published rc.7 API packages. Assembled-Web tests also need a built DSH rc.7 source checkout:

pnpm install
DSH_SOURCE_DIR=/absolute/path/to/dsh pnpm run dsh:link
DSH_SOURCE_DIR=/absolute/path/to/dsh pnpm run dsh:link:check
pnpm run typecheck
pnpm test
DSH_SOURCE_DIR=/absolute/path/to/dsh pnpm run test:web
DSH_SOURCE_DIR=/absolute/path/to/dsh pnpm run test:m6
pnpm run build

Install this checkout directly for manual development:

dsh plugin --profile web add /absolute/path/to/dsh-explain
dsh --profile web --dump-config
dsh --profile web

test:web starts a fresh keyless DSH Web composition and verifies the Learning view, settings, source navigation, and missing-source fallback. test:m6 installs only Explain and verifies both editable-draft shortcuts plus clean unload and reinstall behavior.

Documentation

The detailed design documents are currently written in Chinese.

License

MIT