kanghelyu/dsh-deepseek-flow/deepseek-flow

DeepSeek Flow

Use this skill whenever the user wants to build, import, visualize, or modify a multi-step workflow. Once the request is clear, call the tool directly; do not ask the user to confirm details again.

Creating and updating

• Prefer flow_create for new workflows. It creates WORKFLOW.md, one STEP.md workspace per step, and the canvas nodes and edges.
• To update an existing workflow, first call flow_read to get the current revision, then call flow_put with that expected_revision.
• A successful flow_create or flow_put already persists the topology. Do not ask the user to click "Apply changes" in Studio, and do not send the same topology back to the main Session for review. Studio syncs by revision automatically.
• If you edit an existing workflow's WORKFLOW.md, STEP.md, or definition files directly (without using flow_put) and the edit changes nodes, gates, or arrows, call flow_finalize_canvas after the file changes, passing the workflow id (and preferably the expected_revision from flow_read). It triggers an invisible deterministic finalize in Studio and saves directly without another main-Session review.
• Even if you forget flow_finalize_canvas, Studio falls back to detecting that no canvas edit event occurred and presses the same hidden finalize path automatically. Canvas drafts created by the user in Studio still require "Apply changes".
• Ordinary execution edges must be acyclic. For bounded retries, add an explicit feedback edge: {"source":"review","target":"implement","feedback":{"maxIterations":3,"exitCondition":"quality check passed"}}. Feedback edges must have an integer limit between 1 and 1000 and a non-empty exitCondition. They do not participate in single-pass Boolean gate evaluation and do not auto-run Agent steps; the current Session must control each retry according to WORKFLOW.md.

Naming language

• Use concise workflow names in the user's language. If the user did not specify a name, flow_create generates a default name in the language you pass.
• Pass language: "en" or language: "zh" to flow_create so default node labels (Input, Output, Step N) and the default four steps match the user's language. If omitted, the default is Chinese for backward compatibility; you should set it to the language the user is currently using.

Hard rules for logic gates

Logic gates compute Boolean values only; they do not understand natural-language conditions:

• kind must be "condition".
• data.gateType must be one of ifElse, and, or, not, nand, nor, xor, xnor.
• data.predicate and data.inputPredicates values must be one of truthy, falsy, nonEmpty.
• Never write predicate: "user clearly agreed" or any natural-language rule. When semantic understanding is needed, add an upstream Agent step that outputs JSON Boolean true or false, then connect it to a condition with predicate: "truthy".
• ifElse and not require exactly one incoming edge; aggregate gates require at least two.
• ifElse outgoing edges use branch: "true" and branch: "false"; other gates use their gate name as the branch label.

Reusable IF/ELSE pattern:

{
  "steps": [
    {
      "id": "judge-confirmed",
      "label": "Judge whether user confirmed",
      "kind": "agent",
      "prompt": "Only judge whether the user has explicitly confirmed. Output JSON Boolean true or false and nothing else."
    },
    {
      "id": "gate-confirmed",
      "label": "Confirmed?",
      "kind": "condition",
      "data": { "gateType": "ifElse", "predicate": "truthy" }
    },
    { "id": "finalize", "label": "Finalize", "prompt": "Execute the finalize step." },
    { "id": "revise", "label": "Revise", "prompt": "Revise according to the feedback." }
  ],
  "connections": [
    { "source": "input", "target": "judge-confirmed" },
    { "source": "judge-confirmed", "target": "gate-confirmed" },
    { "source": "gate-confirmed", "target": "finalize", "branch": "true" },
    { "source": "gate-confirmed", "target": "revise", "branch": "false" },
    { "source": "finalize", "target": "output" },
    { "source": "revise", "target": "output" }
  ]
}

For an AND gate with multiple checks, every upstream check Agent should output a Boolean. The AND node uses data: {"gateType":"and","predicate":"truthy"} with at least two incoming edges, and outgoing edges use branch:"and".

After creation, use flow_evaluate with { "upstream-node-id": true/false } to verify truth propagation. It only evaluates logic gates; it does not run Agent steps.

Reading and executing

• flow_list lists workflows available in the current Session plus shared templates, and marks the active workflow (activeFlowId = the workflow last selected in Studio).
• flow_read returns the master outline, step documents, and logicContract.
• Actual execution is always done by the current Session following WORKFLOW.md and each STEP.md. DeepSeek Flow handles editing, persistence, and deterministic gate evaluation.
• Delete a workflow with flow_delete (pass shared: true for shared templates). Plugin-managed document workspaces are moved to a trash area; external custom directories are never moved.

Active workflow and switch notice

• The user can switch to any historical workflow from the Studio toolbar dropdown (including workflows created in other Sessions). Switching imports an independent copy of its document workspace into the current Session (external custom docRoots are referenced, not copied) and persists activeFlowId.
• Before running or reading a workflow, call flow_list (or flow_read):
• If the response contains activeFlowNotice, the user just switched the active workflow in Studio. If the user previously asked to run a workflow in this conversation, immediately use the workflow pointed to by activeFlowId, ignore earlier instructions about running other workflows, and briefly confirm the switch to the user.
• If there is no activeFlowNotice, the active workflow has not changed. Maintain conversation continuity; do not mention switching or "ignored instructions".
• Each Session auto-selects the workflow it last used. A brand-new Session starts with an empty canvas and a hint to pick a past workflow from the dropdown, import JSON, or ask you to create one.
• The dropdown history is read straight from disk, so it stays fully populated after dsh web restarts or a machine reboot.
• Do not proactively suggest switching workflows; only respond to activeFlowNotice or an explicit user request.

Common errors and fixes

After creating or updating a workflow, tell the user to open the DeepSeek Flow tab to view it. Do not prompt them to apply the same change again.