Company Workflow

Company Workflow

Use this as the shared operating layer for the project-local company harness. This is a Codex-native file workflow, not an OMX team runtime.

Core Rule

Do not invoke omx, omx team, tmux worker panes, .codex/agents/*.toml, or .omx/state as the execution mechanism unless the user explicitly asks to operate OMX. Use:

• project AGENTS.md for baseline behavior
• .agents/skills/*/SKILL.md for task-specific workflows
• durable Markdown files under company/ when task state or handoff matters
• Codex native subagents only when the user explicitly asks for multi-agent work or when the main task clearly requested delegation

Company Layout

Expected mutable state, when initialized in a consuming repo:

company/
  supervisor/{inbox,active,waiting,done,status}/
  develop/
    research/{inbox,active,waiting,done,status}/
    model_dev/{inbox,active,waiting,done,status}/
    paper_reenact/{inbox,active,waiting,done,status}/
    debug/{inbox,active,waiting,done,status}/
  audit/{inbox,active,waiting,done,status}/
  shared/{templates,policies,registry}/

If company/ is absent, do not invent hidden state. Either create only the specific artifact the user requested, or ask before initializing a full company tree.

Workflow

1. Supervisor intake: normalize the human request into TASK, route to the right team, and record branch/task decisions when they matter.
2. Research / design / develop: produce the smallest durable packet needed for the next owner. Research does not default to code.
3. Debug gate: verify the smallest realistic executable path before audit.
4. Audit gate: independently review pushed branch/provenance/evidence before promotion.
5. Supervisor closure: summarize operational status and, for experiments or paper reproduction, scientific closure.

Routing

• Pre-work project/material analysis: use analyze-project.
• Full paper-to-experiment research loop: use research-to-experiment.
• Literature survey, paper cards, field Q&A: use literature-field-surveyor.
• Paper-to-experiment packet: use paper-intake.
• New idea novelty / baseline comparison screen: use research-idea-screen.
• Repo mapping before design: use framework-mapping or repo-audit.
• Implementation planning: use implementation-plan.
• Plan execution: use execute-plan.
• Experiment proposal: use experiment-design.
• Feasibility decision before implementation: use experiment-feasibility.
• Run execution from a ready spec: use run-experiment.
• General test verification: use run-test.
• Result interpretation: use analyze-results, then skeptic-review for final claims.
• Document strategy, draft, or refinement: use document-workflow.
• Artifact/document lint before handoff: use artifact-audit.
• Gate verification: use debug-review, then audit-review.
• Skill repository maintenance: use skillset-sync.

Required Evidence

Every durable artifact should include:

• objective or claim
• scope / out-of-scope
• evidence paths, commands, logs, papers, specs, or runs
• blockers and next action
• git state when code or experiment execution is involved
• explicit caveats when scientific interpretation is partial

See workflow.md for templates, policy vocabulary, and closure rules.