Persistent reviewed research and engineering runtime

version
0.1.8
GitHub stars
336

Install

/plugin install argus@argus

Run it inside Claude Code after adding the marketplace with /plugin marketplace add lbx154/Argus.

Contents

Discovered from source

Skills 3

Run With Argus

Use the installed Argus MCP tools. Do not imitate the Argus roles in the host session and do not write Argus state files directly.

  1. Call argus_project_list with the exact current work directory.
  2. Reuse the most recently active exact-workdir project when one exists; otherwise call argus_project_create with that directory and a concise name.
  3. Call argus_message with the user's complete objective. Manager owns chat versus task routing, lifetime, workflow, domain, and execution handoff.
  4. Report the project ID and whether the turn was handled inline or dispatched.
  5. For dispatched work, stop the current turn. Do not create an unchanged polling loop; use the argus-status Skill when status is requested.

If the MCP server is unavailable, run argus --doctor and report the exact installation or backend failure. Do not silently replace Argus with an ad hoc single-agent workflow.

Never claim completion from dispatch, daemon activity, loss, or intermediate artifacts. Argus completion comes from its persisted Reviewer/Manager state.

Inspect And Control Argus

  1. Resolve the project ID from the user's explicit ID or call argus_project_list for the exact work directory.
  2. Call argus_status. Report daemon state, active role, completed and pending work, pending operator questions, and continuous objective separately.
  3. Call argus_artifacts when the user asks for outputs. Report only existing allowlisted artifacts returned by Argus.
  4. Call argus_doctor only when status exposes a runtime/backend fault or the user requests diagnostics.
  5. Send ordinary steering or answers through argus_message; Manager remains the operator front door.
  6. Call argus_stop only on an explicit stop request. Keep force=false unless the user explicitly requests a force stop after understanding interruption risk.

Do not count provider, transport, interrupted, or resource failures as task failures. Do not translate training loss or activity into a successful result.

Target-Disease Research

This workflow supports pharmaceutical research and portfolio decisions. It does not diagnose a patient, select treatment, or assert that a new drug was created.

  1. Require a target and disease. Capture aliases, population or subtype, date bounds, intervention class, exclusions, decision question, and output language when supplied.
  2. Stop if the request contains patient-identifying information and ask for a de-identified research question.
  3. Reuse only the project-resolution part of argus-run: call argus_project_list for the exact work directory, then reuse that project or call argus_project_create. Do not dispatch while resolving the project.
  4. Call argus_message exactly once with an objective that asks Manager to select the medical vertical (shipped by the argus-verticals package) and to produce: medical/evidence.jsonl, medical/evidence_matrix.csv, medical/target_disease_memo.md, and medical/review.json.
  5. Require source IDs and URLs, exact retrieval/query provenance, conflicts, failures, missing fields, and a separate infrastructure-failure count.
  6. Treat PubMed metadata as metadata, not full-text verification. Treat trial registration as trial existence/design, not evidence of efficacy. Qualify cross-trial comparisons unless population, endpoint, comparator, follow-up, treatment line, and data cutoff align.
  7. Return the project ID. Use argus-status for later progress or artifact inspection instead of polling in the same turn.

Only describe a dossier as reviewed when Argus returns a Reviewer-certified state. Preserve negative and contradictory evidence.

MCP servers 1

Sponsors

Details

Marketplace
lbx154/Argus
Source
plugins/argus
Category
Productivity