System And Research Workflow Overview

This repository is designed for an experimental personal multi-agent research workstation, with an emphasis on combinatorics and graph theory workflows. It is not guaranteed to work as desired in every environment. Codex, Claude, and DeepSeek each keep their own local configuration directory, but the reusable research instructions live here as canonical skill bodies. The default auto install mode follows each agent’s policy: Codex uses copies, Claude uses symlinks, and DeepSeek uses reference adapters. Other target policies and explicit overrides are described in Installation. Absent agents are left alone.

The system has three layers:

Layer

Role

Agent frontends and targets

Codex, Claude, DeepSeek, CodeWhale, Copilot, OpenCode, Antigravity, Grok, Kimi, and chatgpt-local-coder receive user requests and load installed skill instructions; OpenClaw is a restricted fake-root target for normal installer flows, with reviewed v2 real-system skill-file writes only through openclaw-target-*, legacy dual-route /aas workspace copies may still exist, but new publishing is blocked and existing copies are not replaced or removed.

Shared skill repository

manifest/ selects skills and profiles; canonical/skills/ stores reusable workflows; targets/ holds agent-specific notes.

Runtime and software tools

Python, TeX, optional SageMath, local library tools, document parsers, public databases, and external retrieval helpers do the actual work when a skill needs them.

The installer links these layers without embedding private state. It does not store credentials, session logs, local library databases, downloaded papers, or machine-specific paths. Instead, precheck detects logical capabilities such as python-runtime, tex-runtime, sage-runtime, library access, and optional Python packages on the current system.

A typical research workflow looks like this:

  1. A request enters one installed agent, for example Codex, Claude, or DeepSeek.

  2. The agent loads a shared skill such as research-briefing, deep-research-workflow, zotero, docling, or tikz-draw.

  3. The skill routes to the right software capability: local libraries first, document parsing when files are involved, public databases for structured records, TeX for figures, and SageMath or Python for math checks.

  4. Research phases keep explicit guard outputs: scope boundaries, goal/backward success checks, delegated-agent result packets, source handoffs, and a bounded iteration ledger.

  5. The final answer passes through review or verification skills when the task needs stronger evidence control. Verification reports concrete issues and remaining gaps instead of an aggregate quality score.

  6. Reusable failures, corrections, or missing capabilities are logged with self-improving-agent. When they affect shared skills or settings, the learning includes a canonical repo integration plan before any target-home or repo mutation.

Examples:

  • Current literature brief: research-briefing scopes the question, deep-research-workflow preserves source IDs across search and synthesis, paper-lookup or database-lookup fills metadata gaps, then research-report-reviewer and research-verification-gate check the final report.

  • Paper review from a local library: zotero checks the paper library first, calibre is used for book-like review inputs, docling parses local documents when structure matters, and paper-review or annotated-review performs the review workflow.

  • Research figure or math-heavy answer: deep-research-workflow produces a figure brief, tikz-draw turns it into a structural diagram using TeX, and sagemath or graph-verifier handles graph or algebra checks when available.

  • Windows plus WSL-backed tools: Windows agents can receive the same skill bodies as Linux agents. Tools such as SageMath may be detected as WSL-backed capabilities, so the dependency graph records the substrate instead of hardcoding a personal path.

  • Paper formalization: use ordinary Lean/Lake and pinned Mathlib by default. Lax discovery is optional; independently verify any selected Lax dependency. Local or authorized remote execution keeps proof evidence separate from paper correspondence. For later Lax delivery, reuse the existing from-existing-lean workflow in Lax Formalization And Zenodo Archival. Archival and publication remain optional, separately authorized operations.

  • Reusable workflow improvement: self-improving-agent records local .learnings/ entries, then proposes repo-first changes across canonical/, manifest/, generated docs, runtime helpers, and tests with explicit target/OS coverage limits.

Related pages: Installation, Skills, Dependencies, Multi-Agent Examples.