dsh-learning-bundle

Unofficial community bundle for DeepSeek Harness.

A proof-carrying correction loop for DSH. Record a correction as a candidate, adopt it deliberately, recall it only in the right profile/workspace/session, and reconstruct every model-visible delivery from the session log.

This is a deliberately narrow Stage 0+1 release. It does not auto-generate memories, silently activate them, or claim to be a general RAG system.

The 60-second proof

The same headless task was run through the real plugin twice on DSH v0.1.1-rc.2:

Arm	Learned item	Model-visible injection	Final response
L−	candidate	0	npm install
L+	explicitly adopted	1	pnpm install

The other profile received zero injections, five explicit npm/Yarn overrides won, and the L+ delivery was reconstructed from the durable session log. Read the current machine-readable transcript or run the fast public contract yourself:

git clone --depth 1 --branch v0.3.0 https://github.com/B1lli/dsh-learning-bundle.git
cd dsh-learning-bundle
npm test && npm run benchmark

Who it is for

This bundle is for people who repeatedly correct a coding agent and need the correction to survive the next session without becoming an invisible global rule: heavy DSH users, independent developers, professional engineers, and plugin/platform authors validating agent behavior.

The pain is cumulative rather than theatrical: repeated teaching wastes turns, an over-broad memory silently changes unrelated work, and an opaque recall makes the next mistake hard to diagnose. For frequent agent users, that is a high-frequency trust and productivity failure.

Why this instead of “more memory”

Agent corrections are useful only if they are safe to reuse. A global, invisible memory can leak one project's convention into another; an opaque prompt injection is difficult to audit; automatic activation can turn a bad guess into a durable rule.

This bundle uses four concrete boundaries:

• new items remain candidate until an explicit adopt;
• profile, workspace, and session scopes fail closed;
• a current explicit instruction overrides a learned default;
• recalled items are logged as user/message records with source.kind = dsh-learning-recall.

The checked rc.2 evidence also demonstrates an observable difference: the same task produces npm install before adoption and pnpm install after the adopted rule is delivered.

Approach	What it optimizes	Where this bundle differs
General memory / RAG	broad storage and retrieval	explicit adoption, fail-closed scope, current-instruction override
AGENTS.md / static instructions	human-authored repository policy	a record/adopt/recall lifecycle plus per-delivery receipts
dsh-mnemon	a full supervised memory control plane with Web UI and multiple providers	this bundle is dependency-free and deliberately limited to correction delivery
dsh-continual-evolve	versioned self-evolution, approval, rollback, and benchmark acceptance	this bundle publishes a smaller candidate→adopt contract with per-delivery receipts and a same-task L−/L+ proof
This bundle	a narrow correction-delivery reference	profile/workspace/session fail-closed recall, causal L−/L+ proof, and log reconstruction

The agent-memory category is not empty, and the projects above are more capable choices for broad memory or self-evolution. This project's narrower market hypothesis is that plugin authors and evaluation-minded users also value a small, inspectable reference that proves a correction reached the model only where allowed and changed the result. Adoption—not this README—must determine whether that niche is actually under-served.

Behavior status

Behavior	Evidence
candidate → explicit adoption → active recall	implemented and tested
profile/workspace/session isolation	implemented and tested
current explicit instruction wins	implemented and tested
adjacent-negative prompts do not recall	implemented and tested
logged, reconstructable delivery	implemented and tested
real headless plugin assembly on current dsh rc.2	checked transcript
historical rc.8 assembly and live DeepSeek 3× L− / 3× L+ sample	checked snapshots
Web UI, team sharing, automatic extraction, sealed holdout	not claimed

The live sample is a bounded demonstration, not a population success rate.

Real-world scenario evidence

The public scenario benchmark covers eight common pains: package-manager drift, Python test entry points, formatters, monorepo builds, deployment defaults, session-only debugging, bilingual release notes, and .env bootstrap. All 72/72 deterministic product-contract probes pass across candidate blocking, adopted recall, paraphrases, adjacent negatives, identity isolation, current overrides, exact delivered content, and delivery reconstruction.

This is public regression evidence, not a claim about model quality. It stays separate from the checked 3× L− / 3× L+ live DeepSeek sample and the current rc.2 assembled consumer. See the scenario report and machine-readable result.

Install and try it

Requirements: Node 22.19+/24+, pnpm on PATH, and the dsh CLI.

Recommended: download the prebuilt release tarball, inspect it if desired, and install it without running a package build script:

curl -LO https://github.com/B1lli/dsh-learning-bundle/releases/download/v0.3.0/dsh-learning-bundle-0.3.0.tgz
dsh plugin --profile headless add -w ./dsh-learning-bundle-0.3.0.tgz

The package CLI is installed inside that profile. From the workspace whose convention you want to preserve:

cd /path/to/your-project
DSH_LEARNING_BIN="${DSH_HOME:-$HOME/.dsh}/profiles/headless/node_modules/.bin/dsh-learning"

"$DSH_LEARNING_BIN" record \
  --profile headless \
  --statement "Use pnpm instead of npm for dependency installs." \
  --scope workspace --workspace-id "$(pwd -P)" \
  --keyword install,installs,installed,installing --keyword dependencies \
  --exclude "explicitly use npm" \
  --override-term npm --override-term yarn --override-term bun

"$DSH_LEARNING_BIN" adopt <item-id>
dsh --profile headless "Install the JavaScript dependencies for this workspace."

For development, link a checkout instead:

git clone https://github.com/B1lli/dsh-learning-bundle.git
cd dsh-learning-bundle

# Link this checkout into the headless profile and activate its bundle layer.
dsh plugin --profile headless add -w "$(pwd -P)"

# Record a correction. It is a non-recallable candidate at this point.
node scripts/learning.mjs record \
  --profile headless \
  --statement "Use pnpm instead of npm for dependency installs." \
  --scope workspace --workspace-id "$(pwd -P)" \
  --keyword install,installs,installed,installing --keyword dependencies \
  --exclude "explicitly use npm" \
  --override-term npm --override-term yarn --override-term bun

# Adopt the returned item id, then run a relevant task.
node scripts/learning.mjs adopt <item-id>
dsh --profile headless "Install the JavaScript dependencies for this workspace."

Each --keyword is one required trigger group. Comma-separated values inside the same flag are literal alternatives: plain ASCII words use word boundaries, while developer tokens such as C++, @scope/pkg, CLI flags, and dotfiles are matched by their literal spelling. Each --override-term names a choice in the current task that suppresses the learned default. The CLI resolves workspace paths to their physical identity so symlink checkouts match dsh's recorded cwd.

The CLI and plugin resolve the same store in this order: DSH_LEARNING_STORE, $DSH_HOME/dsh-learning-store.json, then ~/.dsh/dsh-learning-store.json.

Core API

The Cordis plugin is the package's default export. Dependency-free learning primitives are available through the explicit dsh-learning-bundle/core subpath:

import {
  createLearningStore,
  recallLearning,
  buildLearningMessage,
  reconstructLearningDelivery,
} from 'dsh-learning-bundle/core'

The plugin itself is read-only: it recalls adopted items during agent/pre-step, but never records or adopts them.

Verify

The fast checks do not need credentials and do not modify checked-in evidence:

npm test
npm run benchmark
npm run benchmark:scenarios

To rebuild the current compatibility transcript, clone and build the exact official rc.2 commit below. zstdcat from the zstd package is required to inspect the durable session logs.

git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
git checkout b150a551b8d465e31e418e1b2eaf5e79bbb7d28e
corepack pnpm@11.7.0 install --frozen-lockfile
corepack pnpm@11.7.0 run build

cd /path/to/dsh-learning-bundle
DSH_SOURCE_ROOT=/path/to/deepseek-harness \
DSH_CLI_PATH=/path/to/deepseek-harness/apps/cli/lib/bin.js \
npm run benchmark:assemble:current
npm run benchmark:refresh

To refresh the preregistered live semantic snapshot, supply the API key only to the process; it is never written to results:

DEEPSEEK_API_KEY=... \
DSH_SOURCE_ROOT=/path/to/deepseek-harness \
DSH_CLI_PATH=/path/to/deepseek-harness/apps/cli/lib/bin.js \
npm run benchmark:live
npm run benchmark:refresh

See BENCHMARK.md for the gates, retry policy, evidence layers, and explicit non-claims.

Project layout

• src/index.js — dependency-free learning lifecycle and recall core;
• lib/index.js — Cordis agent/pre-step consumer;
• scripts/learning.mjs — manual record/adopt/list CLI;
• scripts/run-benchmark.mjs — fast acceptance decision;
• scripts/run-scenario-benchmark.mjs — eight-scenario, 72-probe regression;
• scripts/assemble-transcript.mjs — real pinned-version headless assembly;
• scripts/run-live-semantic.mjs — bounded live DeepSeek semantic gate;
• benchmark/results/ — sanitized, inspectable evidence snapshots.

Limits and status

• Current DSH rc.2 compatibility: PASS on the pinned official commit.
• Deterministic public contract: PASS, 8 scenarios / 72 probes.
• Historical bounded live DeepSeek sample: PASS, 3 L− + 3 L+ decisions.
• Automatic correction extraction, team sync, UI, adoption/retention, aggregate canary, and sealed holdout: UNPROVEN / not implemented.
• This project is unofficial and does not imply DeepSeek endorsement.

License

MIT. See NOTICE for provenance.