Skip to content

Architecture

This page condenses the architecture sections scattered across README.md and maintainer guidance into one stable overview.

Core Principle

Qiongli separates:

  • canonical contract truth
  • capability routing
  • functional ownership
  • reusable execution specs
  • workflow sequencing
  • runtime execution
  • client-facing distribution

That separation is what keeps the system extensible without turning every workflow tweak into a hidden contract change.

Layer Model

LayerPrimary locationResponsibility
Contractstandards/research-workflow-contract.yamlTask IDs, artifacts, quality gates
Capability Mapstandards/mcp-agent-capability-map.yamlRuntime routing, MCP and skill requirements
Functional Agentsroles/Ownership, quality thresholds, tone
Internal Skill Specsskills/Reusable execution behavior
Pipelines / Workflowspipelines/, .agent/workflows/Step sequencing and entry UX
Bridgesbridges/Runtime adapters and orchestration
Portable Skill Packageqiongli-workflow/Cross-client distributable entry skill

Dependency Direction

mermaid
flowchart TD
  Contract["Contract"]
  Capability["Capability Map"]
  Roles["Roles"]
  Skills["Internal Skills"]
  Pipelines["Pipelines / Workflows"]
  Bridges["Bridges"]
  Portable["Portable Skill Package"]

  Contract --> Capability
  Contract --> Roles
  Contract --> Skills
  Contract --> Pipelines
  Capability --> Roles
  Capability --> Skills
  Capability --> Pipelines
  Pipelines --> Bridges
  Roles --> Pipelines
  Skills --> Pipelines
  Bridges --> Portable

Stable User-Facing Entry Modes

Entry modeBest forEntry
Claude Code workflowsSlash-command UX inside a project.agent/workflows/*.md
Shell/Python installer CLIInstalling and upgrading assetsqiongli, ql, research-skills, rsk, rsw
Orchestrator CLITask planning, task execution, validationpython3 -m bridges.orchestrator ...
Portable skill packageCross-client distribution surfaceqiongli-workflow/

Subject Packages And Discipline Domains

The canonical source stays generic. Installable specialization is generated from subjects/catalog.yaml, subject overlays, selected profiles, and subject-specific skills. CLI/npm installs default to coverage=complete, which keeps the full core framework and adds the requested subject layer. coverage=focused is the slim selected package used by Desktop/Web ZIPs.

For the detailed user/developer distinction, see Subject Packaging Model.

Runtime domain flags and domain profiles still matter for a single task packet, but they no longer replace subject packaging. This keeps the source tree unified while still allowing:

  • domain-specific libraries
  • diagnostics and robustness checks
  • reporting standards
  • venue or methodology priors
  • official composite subjects such as economics-accounting

Multi-Model Runtime Collaboration

At execution time, the system can coordinate codex, claude, and gemini through the orchestrator.

Common modes:

  • parallel: same prompt, multiple runtimes, one synthesis
  • task-run: one canonical task with contract-driven review flow
  • team-run: one task, multiple work units, then merge/review

Two external projects are especially relevant to how this repository evolved:

  • fengshao1227/ccg-workflow
    • Important influence: strict separation between spec, planning, execution, and review.
    • Important difference: CCG is a general software-engineering collaboration system, while qiongli localizes that discipline into academic research workflows and Stage-I tasks I5 -> I6 -> I7 -> I8.
  • GuDaStudio/skills
    • Important influence: packaging cross-model collaboration capabilities as installable Claude-oriented skills.
    • Important difference: GuDaStudio/skills is a general-purpose skill collection, while qiongli is organized around one canonical contract, one task catalog, and one RESEARCH/[topic]/ artifact tree.

Where To Go Next

Last updated:

Qiongli documentation