deepseek-harness-flowx

FlowX for DeepSeek Harness

The web-profile interface is an owner-only, read-only workflow viewer. The FlowX sidebar footer action opens a left-side overlay without replacing the conversation; it overlays part of the window instead of adding a native, resizable sidebar column. The overlay lists workflows, renders the current plan graph with node descriptions, and browses the workflow directory. It is a UI and host integration, not an agent tool. Use the separately installed FlowX workflow skill to create, evolve, and run workflows.

The viewer intentionally has no Start a build, repair, or evolve controls. The authenticated host API retains build endpoints for existing host compatibility, but the browser interface does not surface them.

Prerequisites

• DeepSeek Harness dsh 0.2.2-rc.1, Node.js 24.9.0 or newer, and pnpm on PATH for the CLI plugin installer. This package declares an exact Harness version, not compatibility with other releases. Only the web profile is supported; installing it does not install a Desktop plugin.
• Python 3.10 or newer with the FlowX SDK and its dependencies installed in the interpreter used by the plugin. Installing this npm package does not install Python, the SDK, or the FlowX skill.
• The viewer and its settings need only the Python SDK. The FlowX skill is needed only by the retained build compatibility endpoints, which invoke its build.py.
• A provider and model with authorized API access. Export FLOWX_LLM_API_KEY, FLOWX_LLM_PROVIDER, and FLOWX_LLM_MODEL into the Harness host process environment before launch. Credentials are not entered in the viewer or saved in its settings; they may be inherited by spawned skill processes. The SDK and skill scripts do not automatically load a FlowX .env file; even inspection can require a provider key. Harness model settings do not automatically configure FlowX.

Install and configure

The following macOS/Linux example starts at a FlowX checkout root. Replace the placeholder path, use a trusted checkout, and obtain the API key securely in your environment; do not put it in a command, shell history, chat, or a committed file. Create the virtual environment only if the chosen path is unused, or use an existing Python 3.10+ environment with FlowX installed. Git is needed to install FlowX's runtime dependency from its Git repository.

cd /absolute/path/to/FlowX
python3.10 -m venv .venv
.venv/bin/python -m pip install -e .
source .venv/bin/activate

# Install the complete skill where the Harness filesystem skill provider can find it.
# Review or back up an existing destination before copying over it.
mkdir -p .dsh/skills/flowx-workflows .flowx_workspaces
cp -R skills/flowx-workflows/. .dsh/skills/flowx-workflows/

export FLOWX_PYTHON="$PWD/.venv/bin/python"
export FLOWX_WORKSPACE="$PWD/.flowx_workspaces"
export FLOWX_SKILL_DIR="$PWD/.dsh/skills/flowx-workflows"
export FLOWX_LLM_PROVIDER=deepseek
export FLOWX_LLM_MODEL=deepseek-flash
# Set FLOWX_LLM_API_KEY securely in this environment before continuing.
: "${FLOWX_LLM_API_KEY:?Set FLOWX_LLM_API_KEY securely before launching Harness}"
export FLOWX_LLM_API_KEY

dsh plugin --profile web add ./plugins/deepseek-harness-flowx
dsh web

The installer resolves the local path from the invoking directory and adds the bundle to the web profile. Restart an already-running profile after installation if necessary. The last command starts the web profile; open its reported URL. dsh plugin manages a profile: it does not start Harness or globally launch a process. For other installation methods, see the Harness CLI guide.

Viewer settings

Select Settings in the overlay to edit two existing absolute paths and the API key. No skill directory is needed to save settings or view workflows; build endpoints locate the skill on demand (FLOWX_SKILL_DIR, the source checkout, then $DSH_HOME/skills/flowx-workflows or ~/.dsh/skills/flowx-workflows, then ./.dsh/skills/flowx-workflows). The dialog reloads the workflow list immediately after a successful save; changing saved values does not require a host restart.

The separately installed skill's scripts read this same settings file, so they use the saved python, workspace and API key by default (explicit --workspace still wins).

The host persists these paths in $DSH_HOME/plugins/deepseek-harness-flowx/settings.json, or ~/.dsh/plugins/deepseek-harness-flowx/settings.json when DSH_HOME is unset. The file is owner-only, mode 0600; the plugin settings directory is mode 0700. DSH_HOME, when set, must be an absolute path; a symlinked DSH_HOME base is resolved once, while links inside the plugin-managed paths remain rejected. Saved settings override the corresponding environment fallbacks and persist across restarts. Invalid or unsafe paths and storage are rejected without overwriting the previous file, and the dialog shows the failure so it can be corrected.

The host snapshots environment variables at initialization, so restart it after changing environment fallbacks or provider credentials. The source-checkout layout supplies a default skill path when available; for linked, installed, copied, or packed plugins, explicitly configure the separately installed skill path. Neither local add nor link: bundles the skill. Reinstall or restart as appropriate after updating the plugin or skill. Activating the environment also puts its Python on PATH for the separate skill's runtime and smoke tests; check that dsh and pnpm remain available after activation.

The example's skill copy is separate from the plugin package. For another project or user skill root, follow the FlowX skill installation instructions and the Harness skill provider's discovery rules, then point skillDir or FLOWX_SKILL_DIR at that installed directory. Configure the skill separately to use the same absolute FlowX workspace; viewer settings do not configure the agent's skill process. When redistributing the separate skill, retain its attribution and applicable FlowX LICENSE and NOTICE.

Use

1. Select the FlowX sidebar footer action to open the left overlay. In narrow windows this dismisses Harness's navigation drawer first; the current conversation and its selection are preserved. Press Escape, select Close, or use the footer action again to close the overlay. Select Settings to configure skillDir, python, and workspace.
2. Select a workflow to inspect its graph, dependencies, requirements, main entrypoint, and available node artifacts. The graph is the current plan, not necessarily the published revision: published workflow and node revisions or artifacts can differ after a graph-only update. Rebuild affected stages before relying on a consistent published workflow.
3. Expand the Workflow folder tree to browse the actual directory under workspace/workflows/<workflow_id> and open eligible files. Select Refresh after a separate skill operation finishes; there is no automatic synchronization with other processes.

The viewer has no build, repair, evolve, or runtime controls. Use the separate skill for those operations, review generated code before running it, and follow the skill's runtime instructions. A service started by the skill is owned by its runtime process, not by the viewer.

Folder browsing limits

Folder browsing omits hidden and credential-named paths, symlinks, and special files; requests traversing symlinks are denied. Only regular, single-link .json, .md, .py, .txt, and .log files of at most 1 MiB that pass UTF-8 and safety checks are readable; other files may appear but cannot be opened. Invalid JSON, JSON-shaped text that cannot be safely parsed, and disallowed control characters are rejected. JSON is structurally sanitized, and manifest error details and absolute entrypoint paths are withheld.

Each folder listing returns at most 200 entries, scans at most 1,000 entries, and is limited to 256 KiB of entry metadata and 4 MiB of text probes. A listing can therefore be truncated or contain files that are not marked readable when the probe budget is exhausted; it is not an exhaustive filesystem listing.

Host coordination of builds, inspections, and settings is local to this Node process. It does not lock other Harness, SDK, MCP, skill, or external processes, or prevent external workspace edits. Avoid concurrent writers to the same workspace and inspect only after the other operation finishes.

Security and privacy

Read-only describes the viewer UI, not the host's full capabilities: for compatibility, its authenticated owner-only API still exposes build endpoints even though the viewer has no build form. Do not expose the host route outside Harness's authenticated owner-only boundary. Inspection can also create the workspace's workflows storage directory if it is absent.

Builds through those endpoints or the separate skill send prompts and potentially repair context to the configured model provider. Generated tests and code run without a sandbox under the host user's privileges and may access secrets, local files, and the network. Obtain explicit consent for each build and its side effects; use only authorized inputs and trusted environments. Auditing and tests do not establish safety. The host filters some inherited environment variables and performs best-effort redaction of displayed configuration and output; neither is a security boundary or a guarantee that secrets are removed. Prompts, repair details, logs, and artifacts can be sensitive even after redaction. Restrict workspace access and review content before sharing it.

Tests and source

From the FlowX checkout root, with Node.js 24.9.0 or newer:

node --test plugins/deepseek-harness-flowx/tests/*.test.js

These offline unit tests use stubs and do not validate a real Harness or model integration; the tests are not included in the packed plugin. The plugin entry is index.js, its browser interface is client.js, and the host implementation is host/service.js. The separate skill's scripts and instructions are maintained upstream.

License and attribution

Copyright 2026 FlowX contributors, for their respective contributions. Original upstream: AIpRoBuilder/FlowX. FlowX origin: urn:uuid:ea62b8c8-902a-4eb4-a42e-e5412ed08466. Distributed under the Apache License 2.0; see NOTICE.