DeepSeek Harness Interactive Learning

中文

An independently installable DeepSeek Harness plugin that adds an explicitly selected Learning mode. It separates explanation from interaction: a native visual is an exploratory illustration inside an ordinary response, not a form that owns the learner's turn or replaces the normal composer.

Architecture

• The package root provides the learningActivities Host broker and registers the required

learning/state session event; it registers no model-visible tools.

• ./agent is mounted only by preset/learning and exposes immediate learning_visual,

internal learning_state_update, and optional learning_checkpoint tools.

• ./client renders visuals and checkpoints inline; state updates have an explicitly empty view.
• ./bootstrap exports registerInteractiveLearningSessionCompatibility so the event type is

registered before persisted sessions are resumed.

• V1/V2 activities and V3 visuals remain available only for read-only historical replay and are no

longer exposed by the current Learning preset.

Non-blocking learning flow

1. The assistant first explains the genuinely missing idea in ordinary prose. 2. When a visual or manipulation materially improves understanding, it calls learning_visual once. 3. The closed protocol validates and immediately returns visual-result@4 { status: "ready" }; it creates no pending question, submit button, or Reveal turn. 4. The visual renders at the tool call and remains interactive on refresh and historical replay. 5. The assistant continues with the interpretation and can ask one natural question through ordinary conversation.

The old learning_activity, learning_question, and learning_reveal tools remain only as read-only V1/V2 replay support. V3 parameter charts are likewise replay-only and are no longer exposed by the Learning preset.

Session-scoped LearnerState

Learning stores only a small, tentative teaching state for the current session: the current goal, shown prerequisites, misconceptions or gaps, scaffolding needs, evaluation context, and evidence of independent completion or transfer. It is not a cross-session profile, personality model, or long-term mastery record.

Updates must come from concrete observable evidence and are not mechanical per-turn bookkeeping. Each accepted update appends an identity-free full learning/state snapshot; later model steps fold the durable events and render a bounded 100–300-token dynamic context. Refresh/resume folds the same event log, fork diverges under a new identity, and reset advances the revision so stale asynchronous results cannot resurrect old state.

The public root export includes LearnerState types, folding, reset, serialization, and session-event registration APIs:

import {
  createInitialLearnerState,
  foldLearnerStateSession,
  serializeLearnerStateSnapshot,
} from '@dsh-portable/interactive-learning'

Optional checkpoint protocol v1

dsh-learning/checkpoint@1 is reserved for a prediction, explanation, contrast, design choice, debugging diagnosis, boundary case, or transfer application that will materially change the next teaching move. It is not a per-turn Continue ceremony.

• At most one checkpoint may be pending in a session and at most one distinct checkpoint may be emitted in a model step.
• The five closed kinds are free_text, single_choice, numeric, prediction, and code_slot; single-choice results use stable option ids.
• Payloads contain only the current prompt, context, expected evidence, answer-free options, and a self-sufficient fallback.
• Terminal statuses are submitted, skipped, and cancelled; call and receipt replays are idempotent, while conflicting reuse fails closed.
• Skip, cancel, timeout, renderer failure, or an unavailable rich Client restores ordinary conversation without Reveal, animation, Continue, or a second wait.

Semantic Visual Protocol v4

dsh-learning/visual@4 routes concepts to eight trusted native renderer families:

• plot for functions, data, probability, points, lines, bars, and quantitative relationships;
• node_link for neural-network layers, trees, processes, causality, and topology;
• scene_2d for geometry, vectors, forces, and annotated spatial schematics;
• relation for comparisons, matrices, classification, and set membership;
• timeline for historical events, discoveries, phases, and eras;
• formula_steps for derivations, algebraic transformations, and proof chains;
• study_map for anchored sections, prerequisites, and concept roles in reference material;
• recall_deck for hinted active-recall cards with local review state.

Every family can include a sequence that focuses only declared objects. Renderers provide a visible title, keyboard operation, responsive layouts, structured text alternatives, and a local error boundary. Interactions include bounded sliders, plot probes, series toggles, node and edge selection, progressive focus, and local review state.

A request to recall a derivative formula is answered with the formula instead of an arbitrary exponent slider. A requested fully connected neural network is rendered with layers, nodes, and every real edge instead of a sigmoid plot or Markdown art.

When a learner supplies a document, PDF, handout, or several sources, the system preserves observed section and page/title anchors, uses study_map for an overview when useful, then routes each concept to a more specific visual. It does not flatten a source into one mega-graph or mechanically turn every attachment into flashcards.

All model-visible payloads use closed schemas. Curves accept only a bounded mathematical AST. Unknown fields, undeclared variables, non-finite values, invalid references, cyclic prerequisites, excessive payloads, and invalid ranges are rejected. Model-provided HTML, SVG, Markdown diagrams, and JavaScript never execute.

Install

The repository is private, so configure GitHub credentials first:

gh auth setup-git
dsh plugin --profile web add git+https://github.com/wsnxxxs/deepseek-harness-interactive-learning.git

Install the Learning preset:

& "$env:USERPROFILE\.dsh\profiles\web\node_modules\.bin\dsh-learning-preset.cmd" install

Before constructing the Loader, agent loop, or restoring any configured session, import the bootstrap:

import '@dsh-portable/interactive-learning/bootstrap'

Restart DeepSeek Harness and select Learning mode in a new session. If DSH_HOME is set, use its profiles\web\node_modules\.bin directory instead.

Update and uninstall

After updating the plugin, run the preset installer again. It preserves preset files that you changed yourself.

To uninstall, switch active sessions to another mode, then run:

& "$env:USERPROFILE\.dsh\profiles\web\node_modules\.bin\dsh-learning-preset.cmd" uninstall
dsh plugin --profile web remove @dsh-portable/interactive-learning

Restart DeepSeek Harness afterward.

Development

The installable lib/ artifacts are committed. For source changes, use Node.js 22+ and pnpm:

pnpm install
pnpm run build
pnpm test
pnpm run check
pnpm run test:package:purity
pnpm run test:package
pnpm run pack:check
git diff --check

pnpm run check also runs the credential-free deterministic teaching evaluation. The real desktop/web runtime reads package exports from lib, so rebuild and fully restart after source changes.

The browser component fixture can be started from the repository root with:

pnpm exec vite --config tests/browser/vite.config.mjs

Run the offline evaluation alone with:

pnpm run eval

Package tests check declaration closure, absolute-path purity, tarball installation, public exports, and preset install/upgrade/uninstall lifecycle.

The standalone package targets the current DeepSeek Harness kernel 0.1.0-rc.7, using the same release family as the portable integration. Its Host composition, Client bundler, and cordis.patch.yml remain adapted to the standalone package layout.

License

MIT