Toollery for DSH v0.3.1

Toollery for DeepSeek Harness v0.3.1

v0.3.1 completes the first native skill-manual workflow and refines the target-aware controls without changing DSH's ownership of registration, permissions or execution.

Native skill manuals

• Load eligible skill bodies through DSH's scoped native loader only for manual preparation.
• Generate bounded body-derived intent queries and store them in a profile-local skill-cache with provenance-linked identities.
• Support spec, manual and spec_manual skill representations. spec remains metadata-only; manual modes reuse valid entries and invalidate them when body, catalog metadata or generator identity changes.
• Keep full skill bodies and generated manuals out of model-visible catalog history. DSH's native skill provider still loads and executes selected bodies.

Pre-submission authorization

• Check the current scoped skill catalog and local cache before a prompt is sent.
• Ask for explicit approval only when skill manuals are missing. The cache check itself makes no model call.
• Preserve the composer draft and block submission while generation is running; unlock it when preparation completes. /toollery prepare-skills remains available for deliberate pre-generation.
• Report preparation progress and approximate ETA based on completed uncached manual work.

UI and boundaries

• Preserve standalone SDK presets that do not provide a native skill service.

• Use Tool rep and Skill rep labels in the target-aware panel; skills shows only Skill rep and both shows both.

• Remove the redundant Indexed tool pool line.

• Keep tool k and skill k independent; in both, one Update k action applies both displayed values together. /toollery k N remains the command-line compatibility form.

• Update OFF messaging to explain that Toollery retrieval is bypassed while DSH's host-eligible tools and native skill catalog remain available under host restrictions.

• Toollery does not yet make every native DSH tool permanently resident while ON. The native skill declaration is retained, and complete host-eligible tool visibility is preserved by OFF and the skills target.

Verification

Validation passed: 83 Python tests (80 passed, 3 optional ToolTalk skips), 29 Node CI/UI/service tests, 5 packaged-plugin tests, and 12 checks in the real native DSH fixture (paid_api_calls: 0). The totals include 22 focused Python skill-manual tests and 6 Node UI/route tests. The wheel and source archive passed twine check; the npm tarball embeds the matching worker wheel. See the verification summary for the detailed scope and historical v0.3.0 checks.

Install the matching @jkl-agents/dsh-toollery npm version or the v0.3.1 .tgz asset from the GitHub release.

v0.3.0 — Native DSH skill retrieval

Toollery for DeepSeek Harness v0.3.0

v0.3.0 adds native DSH skill adaptation alongside Toollery's existing tool retrieval.

Native skill retrieval

• Set the retrieval target to tools, skills, or both.
• In skills mode, Toollery builds a local BM25 index from the host's current skill names, descriptions and optional metadata. Explicitly named available skills receive priority and the result is deduplicated and bounded by k.
• DSH remains responsible for the native skill declaration, full skill body loading, permissions and execution. Toollery does not read or copy skill bodies and does not turn skill ranking into an execution grant.
• Skill retrieval keeps the complete host-eligible tool pool available. both runs the tool and skill paths independently.
• Settings and selections remain frozen for the user turn, including later model steps. Host restrictions remain effective.

The package includes 95 collision-resistant sample-* prompt skills for trying metadata routing and native body loading. They add no runtime packages, although individual instructions can refer to project files or optional services. See plugin/sample-skills/README.md for attribution and limits.

Tool and worker updates

• Skill-aware cards show the target, skill candidate count, applied/latest k and metadata retrieval time. Tool representation controls are disabled when the selected target does not use them.
• Top-k filling now handles explicit-name matches without duplicate entries or k-1 underfilling when enough positive matches exist.
• Startup safely installs the embedded worker offline when DSH's bundled or an existing Python is available. No model call or runtime download happens automatically; the setup control remains available for retry and explicit private-runtime consent.
• Existing tool retrieval, cache-based manual generation, provider-owned discovery ranking and the standalone demo remain available.

Compatibility and limits

The tested compatibility target is Windows with DSH Desktop 0.12.1 and DSH runtime 0.2.1-rc.2. Native root conversations are covered; PTC and subagent retrieval are outside this release's scope.

Skill retrieval is metadata-only in this release; generated skill manuals, history-aware rewriting, dependency expansion and recovery search are not included. Hosts without a native skill catalog retain their normal behavior. Tool retrieval still uses the latest user text, so increase k or switch OFF for tasks that need broad follow-up access.

Install from DSH Desktop with the package name @jkl-agents/dsh-toollery, or use the matching .tgz asset from the GitHub release. The npm package and GitHub bundle contain the same versioned worker and UI.

v0.2.3

Highlights

• Fix upgrades of existing Desktop pip-free profile environments by clearing only stale Toollery files and metadata before installing the bundled worker.
• Report the setup stage when worker installation or validation fails.
• Document recovery for local npm metadata/cache-blob errors during plugin reinstall.

The npm package is published as @jkl-agents/dsh-toollery@0.2.3.

v0.2.2

Toollery for DeepSeek Harness v0.2.2

This patch makes the conversation panel less intrusive and simplifies getting started. Retrieval behavior and the supported DSH runtime are unchanged.

Changes

• Collapse Toollery controls and turn cards to a compact ON/OFF button. Expand them in place with settings, unsaved inputs and result details retained. Folding affects display only; its preference persists in browser storage.
• Simplify the root README around Desktop installation and everyday use. Move manual demo setup and developer commands into dedicated guides.
• Record successful Desktop installation reports from the maintainer and a colleague on another computer, separately from automated verification.
• Include an optional candidate-batch discovery example and integration instructions in the source distribution. The example is installed separately; it ranks candidates without registering or executing them.
• Align the runtime-reported package version with the distribution version.

Install or upgrade

In DSH Desktop, use Plugins → Add plugin → Install a third-party plugin with the URL or absolute path of this release's jkl-agents-dsh-toollery-0.2.2.tgz asset. Restart if requested, then select Set up local worker to update the embedded worker. Existing configuration and caches are retained.

The npm package is @jkl-agents/dsh-toollery. GitHub and npm publication are separate steps; use the release asset for this exact version until npm lists 0.2.2. The optional discovery example is available in the source archive and is not automatically installed by the main plugin.

Validation and scope

The panel passed an isolated React/Edge interaction check covering collapse/restore, preserved drafts/results and ON/OFF state, reload persistence, disabled browser storage and zero configuration mutations from folding. The discovery example has focused tests and an isolated DSH host check. Release CI covers offline regressions, packaged-worker setup and installed-wheel smoke checks. No paid API requests are needed.

The local compatibility target remains DSH Desktop 0.12.1 with DSH runtime 0.2.1-rc.2 on Windows. Reports from another computer confirm installation, not every feature or platform. Native root conversations are supported; PTC and subagent retrieval remain outside scope. See verification.

Toollery for DeepSeek Harness v0.2.1

This patch release contains the latest Windows provider-bridge fixes and publishes the DSH bundle as @jkl-agents/dsh-toollery for direct installation from DSH Desktop.

Install

In DSH Desktop, open Plugins → Add plugin → Install a third-party plugin and enter:

@jkl-agents/dsh-toollery

For a pinned or offline install, download the .tgz asset from this release and enter its URL or absolute local path. The standalone Windows demo remains available from the repository source ZIP.

Changes

• Detects the built-in provider from DSH profile configuration rather than Windows executable path equality.
• Keeps the provider bridge isolated from checkout paths and unrelated Python environments.
• Publishes the same built bundle through npm and GitHub release assets.
• Clarifies the Desktop, CLI, pinned archive and standalone demo installation paths.

Scope

This release retains the v0.2.0 compatibility boundary: Windows and the pinned DSH Desktop/runtime path are verified; PTC, subagent retrieval, Linux/macOS and arbitrary plugin/provider combinations remain outside the verification claim. See the verification report.

v0.2.0

Toollery for DeepSeek Harness v0.2.0

Released 2026-10-23. Install Toollery into an existing DSH profile. The standalone Windows demo remains available; the retrieval algorithm and earlier benchmark conclusions are unchanged.

Install the plugin

In DSH Desktop 0.12.1 (DSH runtime 0.2.1-rc.2), open Plugins → Add plugin → Install a third-party plugin and enter the .tgz asset URL or downloaded file's absolute path. Enable and restart if requested. Desktop supplies its Node/package manager; no second DSH installation is needed.

Open Toollery's details, set up the local worker, and save your manual-generation model, OpenAI-compatible endpoint and key in the form. Windows x64 users without Python can explicitly allow a private Python download; other platforms require Python 3.12+ with venv/pip. Setup and saving settings make no model calls.

Start a new native conversation, review the pool and missing-manual count, confirm API charges, and generate manuals. Valid caches are reused. DSH keeps its chat model and credentials; Toollery does not copy them. See the plugin guide, including the current Desktop verification boundary in verification.

CLI alternative (requires its own Node/pnpm/Python environment):

dsh plugin --profile web add C:/absolute/path/jkl-agents-dsh-toollery-0.2.0.tgz
dsh --profile web

Replace web with your CLI profile, not Desktop's application-owned profile. /toollery setup and /toollery prepare remain available; use the UI for runtime-download consent and model configuration. The key is stored in a local profile file, not an OS keychain, and is never returned to the form.

Changes

• Official-format DSH bundle with profile-local worker, cache and configuration.
• Existing host models, personas, providers and tool registrations are preserved; retrieval operates within the tools the host allows.
• Web controls and /toollery commands for setup, preparation and next-turn settings.
• Current-turn schemas replace earlier subsets on DSH's addition-only model projection, while conversation history stays intact.
• Added per-turn pool reconciliation and bounded manual hot refresh: additions and schema changes are detected on the next user turn, valid manuals are reused, and small cache misses can be generated once without restarting DSH. Refreshes use revision checks and never overwrite a newer host snapshot.
• Added an optional provider-owned discovery hook: providers may submit a remote candidate batch for local manual/BM25 ranking. Toollery persists candidate schemas and manuals, returns snapshot-bound IDs, opaque provider references and schema hashes, and leaves remote fetching, authorization, executor mapping, registration and execution to the provider. This is an API, not autonomous agent discovery.
• Manual preparation/refresh shows progress, elapsed time and an approximate ETA. Refresh waits default to 120 seconds; timeout preserves background generation for reuse. Above-policy updates require approval and explicit message resubmission.
• OFF works even when the worker is unavailable. Unloading cleans up plugin-owned resources and retains local configuration/cache.
• Shared retrieval/UI code with the standalone demo. Python SDK/runtime dependencies move to the optional [demo] extra; manual standalone installations must install .[demo] or the wheel with [demo].
• The standalone demo waiting page now shows startup stages, elapsed time and failure feedback.

Scope

Verified locally on Windows with pinned DSH runtime 0.2.1-rc.2, including Desktop 0.12.1 bundled Python reuse in an isolated profile. Desktop GUI onboarding remains unverified. Native root conversations are supported; PTC and subagent retrieval are outside this version's scope. Settings/selection are fixed per turn; registered pool changes reconcile on a later turn without restart. Retrieval still uses the latest user text, without history rewriting or recovery search.

Targeted offline checks cover official bundle installation, embedded-worker setup, cache reuse, host tool restrictions, changing subsets, retained history, dynamic tool registration with hot refresh, and the retained standalone demo. This is bounded integration coverage, not a claim of compatibility with every plugin. See verification for results.

For the standalone double-click demo, download Source code (zip) and run Start-Toollery.cmd; the .tgz is for an existing DSH installation. Distribution remains GitHub-only, with no npm/PyPI publication. Existing releases are unchanged.

Toollery for DeepSeek Harness v0.1.2

Windows double-click installation and a browser-based demo setup. Retrieval behavior and benchmark conclusions are unchanged.

Try it

1. Download the v0.1.2 repository ZIP, or choose Source code (zip) below, and extract it to a writable folder.
2. Double-click Start-Toollery.cmd and keep the terminal open. The launcher installs private Python, Node.js, pnpm and DSH dependencies; no global installations or administrator rights are required.
3. On the browser setup page, choose 12 utility tools, enable Scripted preview, and select Start demo. Try Use add_numbers to add 19 and 23. No API key or paid API requests are needed for this preview.

The wheel is for manual Python installation. Use the repository ZIP for the double-click entry point. See the README and launcher guide for real-model setup and configuration.

Changes

• Add Windows x64 private dependency installation without relying on preconfigured Python/Node/pnpm PATH entries.
• Add browser settings for tool pools, ON/OFF, retrieval representation, top-k, independent chat/manual credentials and explicit manual-generation permission.
• Add a free scripted utility preview with isolated mock manuals and real DSH tool execution.
• Open an authenticated DSH workspace automatically, reuse valid manuals across launches, select an available localhost port and stop owned child processes.
• Reorganize the README into free preview, real-model setup, restart/troubleshooting and optional manual installation.

Scope and known limitations

This is a standalone demo with its own DSH environment, not yet a bundle installed into an existing DSH profile. Windows x64 is the verified launcher platform; DSH remains pinned to 0.2.1rc2. Scripted preview supports example requests, not arbitrary chat. Real chat and missing/stale manual generation use the user's configured APIs and may incur charges.

First use requires internet access and can spend several minutes installing Web dependencies after Start demo. The new tab displays a static preparation message during this stage; detailed installation progress is not yet shown there. Keep the setup page and terminal open. Later starts reuse installed dependencies. Each launch creates fresh session state; manuals and playground files persist.

The release includes a wheel, source archive and SHA-256 checksums. Existing releases remain unchanged. This is a GitHub release, not a PyPI/npm publication.

Offline release checks passed on the tagged commit.

Toollery for DeepSeek Harness v0.1.1

Compatibility and startup diagnostics update. Retrieval behavior and benchmark conclusions are unchanged.

• Check Node.js and pnpm compatibility before Web setup/startup, with bounded probes and actionable PATH diagnostics.
• Resolve configured generator and host commands from the configuration directory; handle Windows npm command shims consistently.
• Validate Web runtime metadata, port availability, MCP executables/directories and referenced credentials before startup.
• Treat pnpm as a setup-only dependency and support custom generators without assuming API credentials.
• Fix the Windows CI short-path assertion and update checkout/setup actions to Node.js 24.
• Clarify prerequisites and authenticated browser URLs; add two real DSH Web UI demo recordings.

Windows with Python 3.12 and DSH 0.2.1rc2 is the verified platform. Node.js 22.19+ on 22.x or 24+, and pnpm 11.x for Web setup, are required. See the README for source-based installation and the environment audit for the checks behind this update.

The release includes a Python wheel, source archive and SHA-256 checksums. This is a GitHub release, not a PyPI/npm publication. One-click installation and demo launchers are not included in v0.1.1.

Offline release checks passed on the tagged commit.

Toollery for DeepSeek Harness v0.1.0

Initial independent DSH implementation of Toollery's proxy-query retrieval method.

• Installable Python CLI with an isolated retrieval/cache core and pinned DSH 0.2.1rc2 integration.
• Configurable, independent chat and manual-generation providers, with local environment-file credentials.
• Three cached proxy queries per tool; BM25 spec/manual/combined retrieval; ON/OFF and all/top-k controls.
• Official DSH Web UI controls, selected/called tool cards, utility demo and filesystem MCP preset.
• Complete configured-source discovery, per-turn frozen schemas and incremental cache invalidation on restart.
• English installation/integration guides, MIT license and explicit upstream attribution.
• ToolRet development results and an executable ToolTalk OFF/ON supplement, including negative results and cost/latency boundaries.

GitHub source and release artifacts are the distribution channels. This is not yet an npm, PyPI or DSH bundle release. Windows/Python 3.12 is the verified platform. Linux/macOS and arbitrary existing-host presets are not certified.

The default is ON / spec_manual / top_k / k=3. Latest-query retrieval may omit tools required by conversation history or dependencies. The pool is frozen until the agent restarts. Manual preparation fails on missing entries unless generation is explicitly allowed. Read the README and experiments before choosing settings for your workload.

This project is not affiliated with or endorsed by the original Toollery authors or DeepSeek. The original method and repository are acknowledged; the Appendix A.1 prompt is reproduced with attribution. No original Toollery source code is incorporated.

Release checks passed on the tagged commit: Offline release checks.