Metabolic Maintenance for DSH

A native Cordis skill-provider plugin for DeepSeek Harness (DSH). It teaches the agent to maintain a workflow according to error consequences, preserve consumer evidence, and verify downstream work after changes and recovery.

The plugin supplies instructions on demand. It registers no cleanup hooks, runs no background jobs, and grants no permissions. Use WM's MCP integration when executable maintenance tools are wanted.

Where this repository sits

This is the plugin's distribution and discovery home. Development, the canonical instructions, and packaging live in metabolism-tools/workspace-metabolism:

• plugin source — plugins/dsh-metabolic-maintenance
• canonical skill text — skills/metabolic-maintenance

Packaging generates an identical distributable copy of that skill text. Edit the upstream sources, not this mirror. Releases are copied here by .github/workflows/sync-release.yml, which mirrors the newest dsh-metabolic-maintenance-v* release with its assets and checksums; if this repository and the source repository ever disagree, the source is authoritative.

Install

Download the plugin .tgz from the latest release and extract it into a directory you control; npm archives contain a package/ directory. Keep that directory intact. No npm account, build step, or plugin dependency installation is required to load the extracted plugin.

From the workspace you want to work on, generate a patch with the plugin's absolute location, then launch your existing DSH installation:

node /absolute/path/to/package/configure.mjs > maintenance.patch.json
npx @deepseek-ai/dsh web --patch ./maintenance.patch.json

Alternatively, add a plugin row to your existing DSH composition using the file URL of the extracted index.js:

- insert:
    - id: metabolic-maintenance
      name: 'file:///absolute/path/to/package/index.js'

On Windows use a file URL, for example file:///D:/tools/metabolic-maintenance/package/index.js. The generator handles spaces and non-ASCII paths automatically. The standard DSH composition must provide its skills registry and skill tool. Remove the overlay or row to disable the plugin.

Ask the agent: Use the metabolic-maintenance skill to assess this workspace's maintenance needs. Execute only authorized actions and verify the affected downstream results.

The model can load it through skill({name: "metabolic-maintenance"}). The evaluation reference is available on demand; its contents are not added to every conversation step. Same-name project skills outrank this bundled provider, so check for a local override if the body differs.

Which part should I use?

Part	Responsibility	Use it when	What it cannot establish
Metabolic Maintenance skill	Guides the agent's maintenance judgment and evidence collection	An agent must decide whether maintenance is needed and how to verify its result	Enforced permissions, model obedience, or business correctness
WM CLI / MCP tools	Audit, policy-based previews, reversible cleanup, integrity checks, and rollback	You need executable file-lifecycle operations	Whether a downstream research report, build, or other business result is correct
Your downstream consumer and acceptance check	Runs the affected workflow and checks its actual output	You need to accept maintenance or confirm recovery	Authorization for unrelated changes

Use WM alone when an existing script already defines the maintenance and acceptance procedure. Use the skill alone with other adequate tools when the agent needs maintenance guidance. Combine them when the agent needs both guidance and WM operations. The skill is an advisory layer; it is not a decision authority or an enforcement boundary.

The four rules in practice

Rule	Meaning	Example
Maintain for an actual reason	Identify a due policy, capacity limit, broken dependency, or explicit objective; otherwise take no action	A 30-day retention rule applies to an unused 45-day-old scratch report. A fresh report does not become a cleanup target merely because an agent is idle.
Match autonomy to error consequences	Assess affected scope, dependencies, isolation, and recovery. More serious consequences require stronger conditions before autonomous action, not just a lower confidence score	A regenerable cache can use bounded checks. Shared input evidence needs a known consumer, preserved provenance, and a tested recovery route. Unresolved high-impact choices go to the authorized human.
Preserve consumer evidence	Retain the information needed to reproduce and judge downstream results, not merely copies of the files being removed	Keep input identity, relevant policy/tool/consumer versions, acceptance criteria, actual output, and recovery material outside the cleanup target.
Verify downstream work after maintenance and recovery	Run the relevant consumer and compare its output with the agreed acceptance condition. Recheck after rollback	A report must still total 18 after cleanup. If cleanup breaks it, restoring files is only the first step: the report must produce 18 again. Missing verification leaves the outcome unresolved.

Checks should be proportional. These rules do not require repeated approval for already authorized reversible work, or a research experiment for every cache cleanup.

End-to-end example

The packaged plugin ships an isolated maintenance cycle: assess a due scratch report, preview and execute WM cleanup, check a real file-reading consumer, and recover a deliberately broken fixture. It also demonstrates holding a cycle when its consumer is unknown. The example separates scripted execution from the DSH agent's proposed role; it is not evidence of autonomous model performance.

Version 0.1.3 adds maintenance observation examples — startup failure, stale evidence, notification retry, incomplete recovery, and legitimate no-change/not-due outcomes. Detailed observation instructions and a case record are loaded only when needed. No scheduler, notification service, or new WM CLI command is installed.

Pair with WM

Install workspace-metabolism into your chosen Python environment and generate a combined patch from your target workspace:

node /absolute/path/to/package/configure.mjs --with-wm > maintenance.patch.json

This adds the same official MCP bridge used by the existing WM integration and pins its working directory to where you generated the patch. WM's Python executable must be on the DSH process's PATH. If WM is already configured, use the skill-only patch to avoid duplicate plugin entries. Generated patches contain local paths; regenerate after moving directories and do not include them in public reports.

The skill can also use existing scripts. It does not require experimental wm change commands. WM integrity checks do not by themselves prove consumer correctness, and policy receipts do not authenticate human approval.

Build and verify from source

With Node 22+ and Python 3.11+, from plugins/dsh-metabolic-maintenance in the WM repository:

npm ci --ignore-scripts
npm run build
npm test
npm pack

Packaging generates the distributable copy of the canonical skill; edit the canonical source, not generated assets.

Verification boundary

Compatibility checks use the real Cordis 4.0.2 and DSH skill registry 0.0.1-rc.1, covering discovery, loading, reference access, project override, and disposal. The provider follows the official DSH skill API. These checks do not establish full DSH UI compatibility, model obedience, lower token costs, or production maintenance safety. DSH is evolving; keep the version boundary visible when diagnosing compatibility issues.

License

MIT — see LICENSE.