foggy-projects/foggy-deepseek-harness-plugin

This package is the lightweight DeepSeek Harness Bundle for Foggy's Java data analysis engine. Installing the Bundle adds a native Foggy tab under Settings → Built-in plugins. The private Python runtime, Launcher, and isolated CLI environment are downloaded only when the user selects Initialize and start (or the equivalent component action). Existing installations are not silently changed when the plugin package is upgraded.

The 0.4.3-rc.1 compatibility candidate targets DeepSeek Harness 0.2.0-rc.2 and requires a system Node.js >=24.0.0. Foggy's Java Launcher uses a system Java 17+. Foggy manages its own pinned Python 3.12 runtime inside the per-user component directory; it does not require a system Python, modify PATH, or register Python globally. Advanced users may explicitly set FOGGY_PYTHON to a compatible Python 3.11+ executable.

Local release-candidate installation

dsh plugin --profile web add --workspace-root ./foggy-projects-deepseek-harness-plugin-0.4.3-rc.1.tgz

Restart dsh web, use the browser it opens (or the complete printed URL, including ?token=...), open Settings → Built-in plugins → Foggy Data Analysis, and initialize the components from there. Do not share the launch-token URL.

After the version is available on npm, install the exact DSH 0.2.0-rc.2 compatibility candidate so the stable beta channel is not changed:

dsh plugin --profile web add --workspace-root @foggy-projects/deepseek-harness-plugin@0.4.3-rc.1

DSH 0.2.0 candidates use the dsh020 npm dist-tag. Check that the tag resolves to the version you intend to install before using it. GitHub's published tarball can be installed directly when registry publication is still pending:

dsh plugin --profile web add --workspace-root @foggy-projects/deepseek-harness-plugin@dsh020

The earlier DSH 0.1.7-rc.2 and 0.1.5-rc.2 candidates remain available as @foggy-projects/deepseek-harness-plugin@dsh017 and @foggy-projects/deepseek-harness-plugin@dsh015. The dsh020 candidate does not move the existing beta or latest channels.

For the existing stable beta channel, the corresponding one-line install is:

dsh plugin --profile web add --workspace-root @foggy-projects/deepseek-harness-plugin@beta

After updating the plugin package, open Settings → Built-in plugins → Foggy Data Analysis and use Update components (or Update and start) to download the pinned CLI, Launcher, and Skills. The settings page shows current and target versions. Updates are blocked while Runtime is running so the active Java process and the next-launcher state cannot diverge. Component-specific repair actions remain available under Advanced repair.

DeepSeek Harness profiles can apply pnpm's minimum-release-age policy to their lockfiles. An upgrade performed shortly after publication can therefore fail while naming either the new plugin or a previously installed version. This is a DSH profile-policy failure, not a damaged Foggy package. Use the exact profile directory printed in the error and rebuild its lockfile, then rerun the same dsh plugin add command:

Set-Location "<the DSH profile directory printed in the error>"
npx --yes pnpm@11.7.0 clean --lockfile

The failed add normally records the requested exact version under minimumReleaseAgeExclude before returning. If the error repeats, confirm that both the installed and requested exact Foggy versions named by the error appear under that key in the profile's pnpm-workspace.yaml; do not disable the policy globally.

For development, FOGGY_ASSET_CACHE_DIRS can contain platform-delimited verified asset-cache directories. It is not required for a normal install.

The bundled onboarding Skill is intentionally development-first. Users may provide a local datasource password in the conversation, a user-managed connection file, an Agent environment variable, or a compatible Runtime Console. Direct passwords are submitted to the already-running Runtime without being copied into onboarding state or evidence. The Skill then drives schema discovery, semantic drafting, local Bundle registration, and bounded query verification without restarting Runtime.

Native Skill and workspace contract

The Bundle registers foggy-deepseek-onboarding and the downloaded foggy-ai-analysis through DeepSeek Harness's native Skill provider API. Skills are available in every DSH workspace without copying or symlinking .agents. The current session cwd remains the workspace boundary for semantic drafts and evidence. An existing Harness session keeps its original workspace even if another workspace becomes the default. Create a new session in the selected workspace when testing model-file isolation; an isolated DSH_HOME alone does not change a session's cwd. The --workspace-root option in the plugin install command is a pnpm installation option, not a Harness workspace selector.

Opaque CLI profiles remain available for users who prefer them and default to the private persistent <dataRoot>/cli-profiles directory, but they are not a prerequisite for ordinary development. Composite onboarding commands are idempotent: unchanged completed phases are resumed rather than re-adding a datasource or re-registering a local Bundle. A resumed datasource checkpoint is also reconciled against the live Runtime; if the datasource was removed outside the wrapper, only datasource configuration, verification, and schema discovery are resumed. Settings detects legacy temporary profiles and offers an explicit, validated migration into the persistent store.

Private Python, CLI, Launcher, the analysis Skill, install state, and Runtime state live in the user-level Foggy component directories. The managed CLI is intentionally isolated and does not need to be on PATH. Re-download / Repair verifies the global analysis Skill, backs up modified or outdated managed content, restores it, and invalidates DSH's Skill catalog. The onboarding Skill is bundled with the plugin and is restored by reinstalling or upgrading the plugin package.

The Foggy settings tab shows the persisted database/semantic onboarding stages, offers pinned checks and repair for CLI, Launcher, and the managed analysis Skill, private Python, and exports a private redacted diagnostics report. The report includes bounded, sanitized tails of Runtime logs while refusing to read paths outside the managed Runtime directory. Runtime start monitors the Launcher PID during wait-ready, fails promptly if Java exits, and remains idempotent: an already-recorded process is verified with wait-ready and capabilities instead of being treated as a failed second start.

The same settings tab owns a persistent local Runtime port. New installations default to 18166; users can choose another port while Runtime is stopped. Startup checks the same wildcard binding used by the Java server, so a conflicting application or Windows port proxy produces an immediate, high-visibility error instead of a readiness timeout. The CLI and Skills resolve the resulting stable Runtime URL from managed state.

For a successful standard QueryModel DSL call made through the native foggy_query tool, Harness shows a compact result card below the completed reply, outside the folded tool trace. A visible "Generated by Foggy plugin" label distinguishes this programmatic record of executed results from the AI reply. The card shows the model, namespace, row count, an optional five-row result preview with DSL/SQL details, and View data. If a turn has more than six successful queries, the first six are shown until the user expands the list. The detailed DSL and execution record remains in the tool trace. Clicking View data calls the configured Foggy MCP connector's dataset.open_in_viewer tool through Harness's tool policy pipeline. Configure exactly one connector that exposes this tool. Its transport retains the configured authorization headers; the plugin does not put credentials in the browser contract and does not fall back to anonymous HTTP link creation. Asking the assistant to open a data view uses the same MCP tool. Both entry points follow the connector's service authorization scope and open without an additional username/password form. The page re-executes the DSL against the current model and data rather than freezing the original rows. Launcher 0.1.24 stores viewer links in SQLite by default. Links default to no expiry and can be given a TTL or revoked; each browser session defaults to one day and has a separately configurable TTL. The historical card can request a new link while the transcript, connector, Runtime, and model are available. Explicit aggregate having conditions are retained alongside detail-level slice conditions. CTE preview is not included. The Harness card retains the complete executed result; the web view displays fields supported by the Runtime's model metadata. The DataViewer page can export the current filtered and sorted model-backed result to CSV (up to 10,000 rows). It refuses larger exports rather than silently truncating them.

The native foggy_query tool uses the managed CLI's separately configured Runtime authorization (for example FOGGY_RUNTIME_AUTHORIZATION). It does not extract the MCP connector's headers. Verify the query tool and its identity when demonstrating restricted data; the shared MCP scope described above applies to the two viewer entry points.

This beta remains a local dev/test integration. It does not automatically extend local credentials or approvals into a formal environment. Production model publication should use a separate manual or dedicated deployment workflow with an explicit target, model Git commit/tag, credentials, verification, and rollback plan.

See docs/PUBLIC-BETA-READINESS.md for the tested public Beta scope, release gates, and stable-release blockers. The DSH 0.2.0-rc.2 compatibility candidate is documented in docs/RELEASE-CANDIDATE-0.4.3-DSH-0.2.0-RC.1.md. The previous DSH 0.1.7-rc.2 candidate is documented in docs/RELEASE-CANDIDATE-0.4.2-DSH-0.1.7-RC.2.md. The earlier DSH 0.1.5-rc.2 candidate is documented in docs/RELEASE-CANDIDATE-0.4.1-DSH-0.1.5-RC.1.md. Native Windows acceptance instructions are in docs/WINDOWS-BETA-ACCEPTANCE.md; database credentials intentionally remain outside the public repository.

Community discovery

DeepSeek Harness does not currently publish a first-party plugin marketplace. For discoverability, this repository is intended to carry the GitHub dsh-plugin topic. Independent community directories and their submission rules are tracked in docs/COMMUNITY-MARKETPLACES.md; inclusion in any directory is community curation, not an endorsement by DeepSeek or Foggy.

Linux and WSL2 experience

Ubuntu and WSL2 users can use the checked-in preflighted installer under experience/linux. It keeps DSH, its profile, the project workspace, and Foggy data on the Linux-native filesystem.