clearai-dsh

What it solves

Most agent loops track one thing — whether the task is done — and nothing about what makes a conclusion trustworthy. ClearAI implements an Epistemic Loop inside DeepSeek Harness: frame → hypothesize → plan → observe → verify → evaluate → record. Every edge has to be earned through evidence and independent evaluation, instead of being piled up by extraction and assertion.

The output is a domain ontology that grows as you research:

• Vocabulary — the concepts, predicates, value forms and units your project speaks (conventions carry no truth value; sentences written with them do);
• Established entries — knowledge that passed the loop, each with its boundary, support level and evidence chain, so it can be cited safely;
• Ontology and entity graphs — what the domain looks like, and what you have actually verified;
• Conflict readings — contradictory conclusions surface automatically; the system reports them, and retracting or keeping is the human’s call.

At runtime the seven stages compress into four beats — plan, execute, observe, reflect. State is derived from the session record with no second store, and the tools the model holds contain no field in which it could declare a step complete.

The authors explicitly do not claim recursive self-improvement: ClearAI is the epistemic substrate such a system would need (see docs/positioning.md and docs/research-openrsi.md).

Who it’s for

• Research and analysis workflows that need traceable, safely citable conclusions;
• Teams that want agent output to accumulate into a reusable knowledge structure rather than a chat transcript;
• Evaluation, review and scientific-discovery settings where “who judges, and on what grounds?” matters.

Install and use

# Recommended: pinned version, prebuilt from npm
dsh plugin --profile web add clearai-dsh@0.2.7

In the app, use Plugins → Add plugin and enter clearai-dsh@0.2.7. Nothing is compiled on your machine, so there is no allowBuilds prompt. Restart dsh web, create a session, and pick ClearAI in the preset picker at the top.

Other paths:

• npx clearai-dsh install — resolves the current release and pins it for you (--lang zh|en to force output language);
• github:Clearailhc/clearai-dsh — development only: git fetches source, pnpm ≥10 refuses the prepare script until you grant allowBuilds, and pinning a commit is advised — that grant means this package’s code may execute on your machine at install time.

Caveat

pnpm ≥ 11’s minimumReleaseAge (default 1440 minutes) makes a bare package name or @latest silently fall back to the newest version older than a day, and DSH’s plugin manager reports that downgrade as success (its pnpm view preview won’t show the difference either). To be exact, pin the version as above, or exempt the package once in the profile’s pnpm-workspace.yaml:

minimumReleaseAgeExclude:
  - clearai-dsh

Licensed Apache-2.0; the only runtime dependency is zod, with the graph stack bundled into the client half at build time.