dsh-jspace-trigger

Configurable, near-field trigger for J-Space in DeepSeek Harness.
No blanket injection. No forced system prompt. Only prompt the model to load j-space when rules say it is worth it.

简体中文 · Design · Contributing · Issues

---

What is this?

dsh-jspace-trigger is a lightweight DeepSeek Harness (DSH) plugin that decides when to remind the agent to use the J-Space Cognition Suite.

It does not inject J-Space every turn. Instead:

• It watches real user messages.
• It evaluates configurable rules: regex / keywords / length / score / explicit commands.
• Only when a rule fires does it append a small near-field nudge to the agent inbox.
• When nothing fires, it stays completely silent.

Compatibility

DSH runtime	Status
DSH 0.1.0-rc.7 (@deepseek-ai/cordis 4.x)	✅ verified against installed type contracts
later 0.1.0-rc.x	expected compatible; report issues otherwise
< 0.1.0-rc.7	unsupported (event/data shapes differ)

This plugin is runtime dependency-free (Node built-ins only). It pegs peerDependencies to @deepseek-ai/cordis >=4.0.0-rc <5 and the DSH service packages so npm can warn on major mismatches without bundling anything.

Why?

Plain J-Space as a DSH skill is available but not automatic. A fixed J-Space system-prompt injection is too noisy for simple tasks.

This plugin is the middle ground:

Mode	Behavior
near-field (default)	Matched tasks receive one short nudge in the inbox; everything else is silent
none	Observe-only: rules still run and metrics still update, but no message is ever injected

Features

• No blanket injection — no J-Space section forced into every system prompt.
• Configurable rule engine — regex/keyword patterns, any / all / score matching, rule-local thresholds, and opt-out exclusions.
• Length fallback — long non-chat messages can automatically route to full or loop.
• Session-safe delivery — a hit in one session is never delivered to an agent from another session.
• Deduplication — one event triggers at most one nudge.
• Observe-only mode — injectMode: none gives pure telemetry without touching the conversation.
• Skill detection & guided install — jspace_trigger_status reports whether j-space is installed; triggered nudges include a missing-skill hint, and jspace_install_skill installs it explicitly.
• Runtime metrics — jspace_trigger_status reports events, triggers, injections, and observe-only hits.
• Invocation funnel analysis — jspace_trigger_analytics connects a rule hit to delivery and following tool calls, including whether the j-space skill was actually loaded.
• Dry-run tool — jspace_trigger_test shows exactly what the current config would do.

Quick start

Prerequisites

• DSH web profile installed and working.
• J-Space Cognition Suite installed as a DSH skill, e.g. at ~/.agents/skills/j-space/.

Install

From GitHub (recommended for end users):

dsh plugin --profile web add "github:EricWang1358/dsh-jspace-trigger"

Or, install the packed tarball from a local clone or a GitHub release:

dsh plugin --profile web add "<path-or-url-to>/dsh-jspace-trigger-<version>.tgz"

Local development link (replace <repo> with your own clone path):

dsh plugin --profile web add "link:<repo>"

Then restart:

dsh --profile web

Verify

dsh --profile web --dump-config | Select-String dsh-jspace-trigger

After restart, the tools jspace_trigger_status, jspace_trigger_test, jspace_trigger_analytics, and jspace_install_skill are available to the agent.

Marketplace discoverability

This is a host bundle plugin: package.json declares dsh.bundle.patch = ./cordis.patch.yml, which is what DSH plugin markets use to classify it as an plugin (auto-enable-able) rather than a client or nonplugin. For the GitHub-topic-driven markets, the repository must carry the dsh-plugin GitHub topic — with a non-empty repository description, so search and one-click install can find it.

GitHub topic: dsh-plugin  (plus: dsh, dsh-bundle, j-space)

Keepers of a fork: add the topic via gh repo edit <owner>/<repo> --add-topic dsh-plugin.

How triggering works today

Priority order (first match wins, short-circuit):

built-in opt-out (always first) > explicit > ignore/chat > workspace-research > loop > research > complex > length fallback > none

The opt-out rule (jspace-optout) is built in and always evaluated first, even when you supply a custom rules array, so an explicit "不要使用 j-space / do not use j-space" can never be overridden by a content keyword. The chat rule is configurable too: if you redefine chat, your patterns replace the hardcoded greeting list; if you drop it, a minimal built-in greeting guard still keeps one-word pleasantries silent.

Default rules

Rule	Examples	Decision
explicit	/j-space, use j-space, 启用 j-space, 加载 j-space	loop + capacity, broadcast
jspace-optout	不需要使用 j-space, do not use j-space	ignore (silent; overrides complexity signals)
chat	你好, hello, thanks, ok, 嗯	ignore (silent)
workspace-research	文件夹/目录/仓库 + 调研/盘点/画像等综合意图	loop + capacity, broadcast, markers, self-monitoring
loop	仓库级、跨文件、多阶段、多轮、long-horizon、multi-file	loop + capacity, broadcast, markers, self-monitoring
research	调研、盘点、梳理、尽调、research、survey	full + deep-reasoning, self-monitoring
complex	重构、架构、全面、详细、调试、审查、refactor、architecture	full + deep-reasoning, self-monitoring

Real trigger examples

# explicit -> loop
/j-space 请审计这个仓库并跨文件保持一致

# loop keywords -> loop
做一个仓库级跨文件重构，并保持全局一致

# workspace-research -> loop (two independent signals)
深度调研此文件夹下的内容、TODO 和 DDL，梳理现状与潜在风险

# research intent -> full
调研一下这个技术方案的可行性

# complex keywords -> full
详细分析一下这个项目的架构，并检查潜在风险

# long fallback -> full
超过 120 个字符的非寒暄任务……

# long fallback -> loop
超过 1800 个字符的多阶段长任务……

# chat -> silent
你好

Dry-run a message

jspace_trigger_test "仓库级跨文件重构"

Expected output shape:

action=trigger
pass=loop
modules=capacity,broadcast,markers,self-monitoring
matched=loop
reason=rule:loop
matchMode=any
hits=2
threshold=1
signals=多阶段|多个文件|多轮|长程|长期|仓库级|跨文件|系统化|完整项目|长时|agentic|long-horizon|multi-stage|multi-file|multi-turn|repository-wide|workflow|loop
skillInstalled=true
---
[jspace-trigger] J-space pass: loop. Suggested modules: capacity, broadcast, markers, self-monitoring. If this task needs structured workspace control, load the `j-space` skill and follow its gate.

Observe-only mode

Set in cordis.patch.yml:

config:
  injectMode: none

In this mode matched events are recorded in jspace_trigger_status but never injected into any session.

Analyze real invocations

jspace_trigger_analytics is a bounded in-memory funnel for checking whether a nudge was useful:

jspace_trigger_analytics scope=current limit=20

It reports rule -> delivery -> toolCalls -> jspaceSkillLoaded. It deliberately stores no user-message text and no tool arguments—only timestamps, session/event IDs, rule/pass metadata, delivery outcome, and tool names. Set analytics.enabled: false to disable it, or use maxRecords (1–500) to bound the retained recent records.

Skill detection & installation

The plugin never downloads or installs J-Space automatically. It only checks and helps when you ask.

• jspace_trigger_status reports skillInstalled and installed paths.
• When the skill is missing, triggered nudges append:

  J-Space skill is not installed. Run `jspace_install_skill` to install it.
  

• jspace_install_skill explicitly clones the upstream repository and copies j-space/ into your configured skill root.

Install the skill manually

jspace_install_skill

To force a reinstall:

jspace_install_skill force=true

To install into a custom root:

jspace_install_skill root="C:\path\to\skills"

Coexistence with routing presets

This plugin does not change a preset's persona or tool surface. It only appends a near-field message after a matched real user message. The following combinations therefore have no tool-name or system-prompt collision, but may affect whether a guide is visible to the model.

Router Standard

When Router Standard selects its weak band, it appends its own routing guide after each real user message. If a J-Space rule also matches, the session can receive two near-field guides: Router Standard's build/fix guidance and this plugin's J-Space suggestion.

dsh-jspace-trigger deliberately does not detect or mute Router Standard. If you want Router Standard to be the only near-field guide, use observe-only mode:

config:
  injectMode: none

Otherwise, the two messages are compatible but add prompt noise. This is most likely for a weak-band task containing J-Space complexity keywords such as 详细 or 分析.

梁神模式 (Liangshen / anchored standard)

梁神模式的首轮锚定阶段只允许真实用户消息到达模型。本插件的提示使用 source.kind: plugin，因此即使规则命中，首轮提示也会被梁神模式过滤；这 保护了它的 Minimal 锚定，不是错误。模式晋升后，本插件的后续提示可以正常 参与会话。

如果你要保持全程最纯净的梁神模式轨迹，同样建议使用 injectMode: none； 如果希望在完成首轮锚定后获得 J-Space 建议，则保持默认 near-field 即可。

Configuration

Configuration lives in the plugin config, normally edited in cordis.patch.yml.

- insert:
    - id: jspace-trigger
      name: dsh-jspace-trigger
      config:
        enabled: true
        injectMode: near-field      # near-field | none
        analytics:
          enabled: true             # metadata only; no prompts or tool arguments
          maxRecords: 50            # bounded to 1..500
        trigger:
          minScore: 1               # threshold only for matchMode: score
          loopChars: 1800           # text longer than this -> loop fallback
          fullChars: 120            # text longer than this -> full fallback
          rules:
            - id: explicit
              action: trigger
              pass: loop
              excludePatterns: ["不要使用 j-space", "do not use j-space"]
              modules: [capacity, broadcast]
              patterns: ["/j-space", "use j-space"]

            - id: chat
              action: ignore
              patterns: ["^你好[!。.!？?~～]*$", "^(hello|hi|thanks|ok)[!。.!？?~～]*$"]

            - id: workspace-research
              action: trigger
              pass: loop
              matchMode: all
              modules: [capacity, broadcast, markers, self-monitoring]
              patterns:
                - "文件夹|目录|仓库|代码库|工作区|(?:todo|ddl).*(?:文件|列表|状态)|(?:文件|列表).*(?:todo|ddl)|folder|directory|repository|repo|workspace"
                - "调研|盘点|梳理|画像|审计|研究|分析|了解|research|survey|audit"

            - id: loop
              action: trigger
              pass: loop
              modules: [capacity, broadcast, markers, self-monitoring]
              patterns: ["多阶段|多文件|跨文件|长程|仓库级", "long-horizon|multi-file|repository-wide"]

            - id: research
              action: trigger
              pass: full
              modules: [deep-reasoning, self-monitoring]
              patterns: ["调研|盘点|梳理|尽调|研究|调查|research|investigate|survey"]

            - id: complex
              action: trigger
              pass: full
              modules: [deep-reasoning, self-monitoring]
              patterns: ["重构|架构|全面|调试|审查", "refactor|architecture|debug|review"]

        # Optional: skill install/check settings
        skillRoots:
          - ~/.agents/skills
          - ~/.dsh/skills
        repoUrl: https://github.com/Tiger3807861189/J-Space-Cognition-Suite-V3.6.git
        branch: main

Rule fields

Field	Meaning
action	trigger, ignore, or none
pass	fast, full, loop
modules	J-Space modules to suggest on a hit
patterns	Regex source strings or RegExp objects
matchMode	any (default), all, or score
minScore	Per-rule score threshold; overrides trigger.minScore for matchMode: score
excludePatterns	Regexes that veto this rule, useful for explicit opt-outs

Tools

Tool	Purpose
jspace_trigger_status	Show config, counters, skill installation state, and recent hit count
jspace_trigger_test <text>	Dry-run a message through the current rules
jspace_trigger_analytics	Inspect the bounded, privacy-safe trigger → delivery → tool-call funnel
jspace_install_skill	Explicitly install/repair the J-Space skill from upstream

Project structure

dsh-jspace-trigger/
├── docs/design.md                # research + rule design rationale
├── src/trigger-core.mjs          # pure rule engine (zero dependencies)
├── src/skill-utils.mjs           # skill detection + explicit installer
├── src/call-analysis.mjs         # bounded privacy-safe funnel
├── src/index.js                  # DSH plugin entry (Cordis lifecycle)
├── scripts/check.mjs             # cross-platform pre-pack validator
├── scripts/build.sh              # CI/pipeline build entry (bash)
├── index.js                      # package entry (re-exports plugin)
├── index.d.ts                    # TypeScript declarations
├── cordis.patch.yml              # DSH bundle mount
├── .github/                      # CI, release, dependabot, CODEOWNERS, issue/PR templates
├── SECURITY.md                   # security policy
└── tests/                        # rule engine + simulated Cordis event tests

Development

node scripts/check.mjs   # validate entry + manifest (no child-process capture)
npm test                 # full rule + simulated Cordis event suite

The suite runs on Node built-ins only — there is nothing to npm install, and CI deliberately skips install to prove it (see .github/workflows/ci.yml).

Not yet done: real-session validation after a DSH restart.

See docs/design.md for the complete design rationale.

Supply-chain & repository hygiene

• No runtime npm dependencies — the shipped code imports only node:* built-ins; peerDependencies are declared for warning purposes only.
• .npmrc sets ignore-scripts=true so no postinstall can run during any install.
• SECURITY.md documents supported versions and private reporting.
• CODEOWNERS routes review ownership; Dependabot keeps the (nearly empty) dependency graph and GitHub Actions up to date.
• CI (ci.yml) runs the suite on Node 20/22 plus an npm audit gate; release.yml publishes a signed-by-tag GitHub release with the packed tarball.

Contributing

Bug reports, feature requests, docs improvements, and PRs are welcome.

Please read CONTRIBUTING.md first.

License

MIT