dsh-rigor-4

dsh-rigor-4

English | 中文

Rigor-4 adds explicit requirements, plans, evidence, child-agent briefs and independent review records to DeepSeek Harness. DSH retains ownership of sessions, model calls, tools, approvals and subagents. This is an independent community plugin.

Release 0.1.3 targets the native contracts checked against DSH 0.1.5-rc.2. There are no external npm runtime dependencies. Other DSH releases require renewed contract testing; the plugin does not claim compatibility with every future release.

Install

Initialize a native DSH profile first and ensure Node, npm, pnpm and DSH are available. Clone or extract this repository, then run:

node scripts/install.mjs --dry-run
node scripts/install.mjs --apply

The default mode only previews. The installer packs the release with npm pack --ignore-scripts, forwards installation to dsh plugin --profile web add, asks native pnpm to refresh its lockfile, verifies package resolution, and copies four agent presets into /.agent-presets. Existing customized files are preserved unless --replace-presets is explicitly passed.

The installer uses offline mode. Other dependencies already present in the target profile must be available locally; if pnpm reports missing cache entries, restore that profile's dependencies with the normal official DSH command before retrying. Installation never embeds credentials or relies on a developer's private directory. It preserves versioned release archives for rollback and does not promise a transaction across native package-manager operations.

Useful options are --dsh-home PATH, --profile NAME, --dsh-package PATH (the installed official package directory), and --store-dir PATH. The default home is the official DSH_HOME value or ~/.dsh. install.ps1 provides the same thin wrapper on Windows. Run --help for details.

Restart DSH and select a Rigor preset in the agent selector. The package is activated by its agent preset rather than a global dsh.bundle; the official CLI warning that it is installed as a plain dependency is expected.

Uninstall

Finish active Rigor tasks and choose a built-in preset for new tasks. From the retained source or extracted release directory:

node scripts/uninstall.mjs --dry-run
node scripts/uninstall.mjs --apply

The uninstaller forwards package removal to official dsh plugin --profile NAME remove dsh-rigor-4. Shared presets remain if another profile still depends on Rigor. Otherwise, it removes only preset files that exactly match this release. Customized presets stop the operation before removal; export and explicitly remove those definitions before retrying. Session state, credentials and release archives are preserved. Restart DSH and refresh the browser afterward.

Tools and presets

Tool	Purpose
rigor_read	Record the request, explicit and inferred requirements, and falsifiable acceptance conditions.
rigor_plan	Associate steps and expected evidence with requirements.
rigor_brief	Define a child agent's question, role and relevant requirements.
rigor_evidence	Link observations, commands or artifacts already present in this session.
rigor_review	Read a structured review from the bound review child's own log.
rigor_report	Check a completion claim or record an honest partial/blocked result.

Preset	Native base	Child agents and user-question channel
rigor-4	standard	Available
rigor-4-ptc	PTC	Available
rigor-4-cordis	cordis	Available
rigor-4-minimal	minimal	Unavailable; registers the four usable record tools

State is isolated per session under /.rigor4/sessions/.json. Reading, plan and work revisions prevent older evidence or reviews from silently proving a newer delivery. Restarting or disposing an agent preserves these records.

Model experience and limitations

Rigor adds stable prompt guidance and tool schemas, plus variable tool results. These consume context and may affect cache reuse according to the host/provider. The plugin does not make direct model requests, but its workflow can cause more tool calls, model turns and review subagents.

Evidence and review records can still be wrong. Tool and shell classifications are finite heuristics, and role prompts or plugin checks are not operating-system read-only sandboxes. In particular, a reviewer allowed to run shell checks may execute commands with side effects. The native sandbox and approval policy remain responsible for execution permissions. A minimal preset cannot provide independent review and must report that limitation.

Development and validation

node --test
npm pack --ignore-scripts

Tests use synthetic state and native-command substitutes. They verify logic, event contracts and installer/uninstaller failure behavior without calling a live model. Passing them does not demonstrate every real LLM workflow or compatibility with untested DSH releases. Package archives exclude tests, user state and credentials; the Git repository includes the synthetic tests for contributors.

License and attribution

MIT. The four preset compositions are adapted from DeepSeek Harness 0.1.5-rc.2; its original MIT notice is retained in LICENSE-DeepSeek.