Skip to main content

DSH-pipeline-kernel

28Stars0Forks0Issues0Watchers

DSH Multi-Agent Pipeline Management Kernel: defines pipelines, submits tasks, auto-routing, ledger three-strike mechanism, inspection self-healing, panel archiving and package reuse. Business-agnostic.

Evidence5/5methodologySourceInstallMaintenanceDSH versionSecurity scan
Machine-auditedInstall commandRepo verifieddsh-plugin topicLicenseREADMEAI wiki
Language
TypeScript
License
MIT
Branch
master
agentcordisdshdsh-pluginmulti-agentpipeline

Install

cmdweb profile
$ dsh plugin --profile web add github:not-big-dog/DSH-pipeline-kernel

Run the command above in your terminal to install this plugin via the dsh CLI. You can switch Profile in the top-right corner. New to dsh? Read the beginner tutorial

Install via your agent

Install the DeepSeek Harness plugin not-big-dog/DSH-pipeline-kernel for me: review the repository at https://github.com/not-big-dog/DSH-pipeline-kernel first, then run the install command and verify the plugin loads successfully.

Paste this instruction to the DSH Web GUI assistant — it will install and verify for you.

One-Line Positioning

DSH multi-Agent collaboration "pipeline management kernel" - abstracts unidirectional task chains like supervisor → designer → prompter → reviewer into definable, deployable, inspectable "pipelines" (one row in the pipelines table), providing task board, ledger, registry, auto-routing, inspection self-healing, control panel, and packaging reuse capabilities. The kernel itself does not bind to any specific business.

Core Capabilities

  • Pipeline first-class citizen and archiving: define/undeploy/undefine pipelines (pipeline_define/pipeline_deploy/pipeline_undefine), archive (pipeline_archive/pipeline_unarchive/pipeline_purge) only marks, recoverable
  • Task board and auto-routing: push/claim/done/cancel/list (pipeline_push/pipeline_claim/pipeline_done/pipeline_cancel/pipeline_list), auto-route to next step on completion via route:<role|UID> tag
  • Role registration and seeding: pipeline_seed creates sessions for unregistered roles and writes to registry, node directory auto-numbers by chain (node 0 = supervisor, node 1..N = business roles)
  • Ledger and three-strikes: pipeline_ledger_write/pipeline_ledger_list records deliverable changes; pipeline_strike cumulatively counts rework times, auto-marks task as blocked when threshold is reached
  • Inspection self-healing and Web panel: watchdog scans every 5 minutes by default to wake idle targets and self-heal zombie tasks; Web bottom-right capsule expands to full-height right sidebar (Activity/Pipeline/New/Archive four tabs, 1-second polling snapshot)
  • Packaging reuse and slash commands: pipeline_pack packages entire pipeline as JSON+MD, paired with pipeline-create fromPack to restore definition and seed in one click; /pipeline gate|status|pack three sub-commands

Technical Implementation

  • Language: TypeScript (src/) + Node.js ESM (compiled output in lib/)
  • Key dependencies: @deepseek-ai/cordis (4.x plugin framework), @deepseek-ai/dsh-storage-domain (domain table JSON backend), @deepseek-ai/dsh-tools (defineTool tool definition), @deepseek-ai/dsh-agent + @deepseek-ai/dsh-llm (session runtime, encapsulated in lib/agent-runtime.js thin adapter layer)
  • Architecture pattern: Cordis dual-side plugin; node side follows standard Cordis spec name + Config (zod Schemastery) + inject + apply, registers 20 pipeline_* tools + 1 /pipeline slash command + 6 Web endpoints; client side uses ctx.locale.register to register Chinese-English dictionary and mount React control panel to body portal
  • Entry files: lib/index.js (node-side Cordis apply) + src/client/index.tsx (browser-side mount entry)

Use Cases

Scenarios requiring orchestration of multi-Agent unidirectional task chains on DSH (typical: supervisor collects requirements → designer creates solution → prompter writes prompts → reviewer reviews → generates deliverables), unified task progress viewing, deliverable清单 management, cross-session auto-routing and reminders. Ordinary "single-Agent + single Q&A" won't need this plugin; projects with more than two sessions that need role-based handoffs are best suited for this plugin.

Prerequisites & Compatibility

DependencyMinimum VersionDescription
DSH (deepseek-harness)0.1.0-rc.8package.json peerDependencies locks ^0.1.0-rc.8, before rc.8 will fail to install due to @deepseek-ai/cordis 4.x, service name changes like webServer/workspaceRegistry/agentDefaultModel
Node.js^22.19.0 || >=24.0.0package.json#engines.node; below 22.19 will cause tool validation failures due to zod 4 and built-in fetch behavior differences
PlatformmacOS / Windows / Linuxlib/web.js:317-334 openFolder calls open / explorer / xdg-open on three platforms respectively; rest is pure ESM + node:fs/node:path, no platform-specific code
Native modulesNoneNo node-pty / better-sqlite3 etc.; only uses built-in modules like node:fs node:path node:os node:child_process node:crypto node:url

Installation

dsh plugin --profile web add github:not-big-dog/DSH-pipeline-kernel

Configuration Options

All config goes through cordis.patch.yml under pipeline-kernel node (default written to ~/.dsh/profiles/web/cordis.patch.yml), restart dsh web to take effect.

ConfigTypeDescriptionDefault
projectRootStringGlobal workspace root (pipeline projectRoot falls back here), all inboxDir/artifacts/pack paths must be within itprocess.cwd()
inboxDirStringRole inbox root directory (relative to workspace), pipeline_push writes task slips hereexchange/收件箱
defaultPipelineStringDefault pipeline id for tasks without pipeline:<id> tag (note: just default, no auto-definition creation)default
watchdogEnabledBooleanEnable unattended inspection (self-healing + wake + stall reporting)true
watchdogIntervalMsNumber (ms)Inspection tick interval300000 (5 minutes)
watchdogStallMsNumber (ms)How long full chain is idle before considered "stuck" and reports to supervisor1800000 (30 minutes)
watchdogWakeCooldownMsNumber (ms)Minimum cooldown for same task wake-up600000 (10 minutes)
watchdogReportCooldownMsNumber (ms)Minimum cooldown for same pipeline supervisor report1800000 (30 minutes)
strikeOutNumberRework count threshold; tasks exceeding this get blocked and suggest human intervention3

FAQ

Q: After installation, why don't existing sessions have pipeline_* tools?

A: Old session tool directory is frozen at session creation time; restarting dsh web only takes effect for new sessions. Need to close existing sessions and open a new one to get the updated tool list from the host.

Q: I got BAD_INPUT for invalid pipeline id with pipeline_define, what's the rule?

A: Pipeline id must match ^[A-Za-z0-9_-]{1,64}$ (see lib/deploy.js:38). Cannot contain /, \, spaces, Chinese characters, or empty strings; dots or .. are also rejected because id is used as both registry key and path segment, path traversal is prohibited.

Q: I see "unregistered" roles in the control panel, can I run tasks directly?

A: No. "Unregistered" means the pipeline/role → sessionId row is missing from the registry table; tasks will be delivered but no one processes the task slips in the inbox, and watchdog cannot wake them. Must first pipeline_seed (with preset in role definition) or manually pipeline_register.

Q: Are role sessions supervisor-resident or created on-demand?

A: Per definition. Each role (including supervisor) is created by pipeline_seed calling host agents.create + agentPresets.mount; not resident, not reused. Seed validates that registered session's cwd matches current pipeline workspace - if not, it's treated as collision residue and cleared/re-seeded.

Q: How does the three-strike mechanism count one "redo"?

A: Based on attempt:<n> in task tags, higher n means more times the task was sent back for rework. pipeline_strike {taskId} reads n ≥ strikeOut (overridable in pipeline definition, falls back to global config) and changes status to blocked with "suggest escalate to human or change direction" message. attempt doesn't need manual maintenance, just add it in route fallback or rework flow.

Q: What are the packaged *.pack.json / *.pack.md for?

A: Output of pipeline_pack - *.pack.json is machine-readable full-node pipeline definition (includes roles/chain/gates/skill list/document template list), *.pack.md is human-readable "green reusable full-node pipeline documentation". Take these two files to another project, use Web route POST /plugins/pipeline-kernel/pipeline-create with body fromPack: "<path>" to restore definition, seed sessions, and create node directories in one click (see lib/web.js:106-173).

Q: "Open Folder" in panel doesn't work?

A: Path whitelist validation. Check if the path given by panel is within the workspace - lib/workspace.js:77-87 uses isWithin to strictly require path within any known pipeline workspace, absolute paths or .. get BAD_INPUT rejected. If folder path is clearly valid but still fails, the explorer/open/xdg-open command may not exist on the system (common on headless Linux servers).

Q: Where is kernel data stored? Can it sync across devices?

A: Data goes through host ctx.storageDomain (JSON backend, 4 tables: pipelines / registry / tasks / ledger). Write path is determined by DSH host; DSH rc.8 typically at ~/.dsh/profiles/web/domains/pipeline_kernel/. Can copy directory to sync across devices, but don't have two devices running the same profile simultaneously - kernel has no cross-process lock.

Learning Curve

Advanced — Users must understand DSH's agent/session model, preset system, and multi-session collaboration semantics, and write their first pipeline definition (roles/chain/gates) using pipeline_define; otherwise, after installation the panel will just show "No pipelines yet".

Known Issues & Limitations

  • Path hard constraint: artifactRoot and inboxDir must be relative paths under workspace and cannot escape projectRoot (lib/workspace.js:18-30), otherwise pipeline_define throws BAD_INPUT
  • Pipeline id and role name hard constraint: must match ^[A-Za-z0-9_-]{1,64}$ (lib/deploy.js:38, lib/pipeline.js:17-22), ids with Chinese/dots/spaces are rejected
  • Domain version number fixed: kernelDomain hardcodes version: 1 (lib/domain.js:90), adding new fields doesn't bump version; if version is bumped in the future, old data will be rejected on open due to JSON backend's unit header validation, migration needed
  • Web control panel only shows archived pipelines with ?archived=1: lib/web.js:48 and lib/snapshot.js:19, main view doesn't include archive box
  • Inspection doesn't感知 host restart: watchdog runs within DSH process, after process crashes and restarts watchdog state is lost, pipeline_reconcile can manually fallback (lib/tools.js:561-584)
  • Node directories are created all at once during pack reuse, new pipelines only create supervisor directory (nodes 1..N created when supervisor finishes documents and calls pipeline_mkdirs or triggers on first pipeline_push), won't actively prompt supervisor to fill in

Read the usage guide →

Install steps, key points, FAQ and compatibility for this plugin — auto-derived from indexed fields.

Listing badge

Listed on deepseek-plugin.org
[![Listed on deepseek-plugin.org](https://img.shields.io/badge/listed_on-deepseek--plugin.org-007EC6)](https://deepseek-plugin.org/plugins/not-big-dog/DSH-pipeline-kernel)

Paste this markdown into your GitHub README to link back to this listing. The badge only states the listing — not a security endorsement.

← Back to plugin directory