Imported from Vinayak0987/Orion (AGENTS.md). Install upstream with npx skills add Vinayak0987/Orion. Copyright stays with the author.

AGENTS.md

Read docs/HANDOVER.md before doing anything else. It contains the full state of this project: what is built, what is pending, the next steps, and the environment gotchas. You do not need the conversation history.

One-line summary

Orion (codebase identifier mabgp) is an AI Business Operating System. A generation layer turns one sentence into a founding business plan using 21 agents across 5 departments; an operating layer then runs that plan as a living company on department cadences. B.E. final-year major project, University of Mumbai.

Verify state first

$env:PYTHONPATH="C:\Users\Vinayak\Desktop\kiro\backend;C:\Users\Vinayak\Desktop\kiro"
cd backend
..\.venv\Scripts\python.exe -m pytest -q

Expect 989 passed, fully offline. Trust the test output over any document.

The invariant you must not break

Anchors are computed, never reported. No department, planner, worker or reflection process may write an anchor metric.

Enforced structurally: anchors are not stored, so there is nowhere to write one. backend/tests/test_anchors.py asserts it. If a task would require relaxing this, say so rather than working around it.

Binding rules

• Stack is LOCKED: LangGraph, Groq/Gemini/Ollama via app/llm/ only, FastAPI, SQLite, ChromaDB, React 18 + Vite. Free tiers only ? flag anything costing money.
• Every LLM call goes through app.llm.llm_client.
• Every generation-layer state write goes through apply_department_patch.
• Field names in app/models/departments.py are FROZEN. Add optional fields, never rename ? the metrics depend on them.
• The test suite must stay fully offline. conftest.py fails any outbound socket.
• Add a regression test for every change. Runnable code only. Pinned versions.
• Frontend depends on react and react-dom ONLY, 162 kB bundle. No router, no chart library. Charts are inline SVG.

Environment

• Two Pythons: .venv\Scripts\python.exe for the app; system python for playwright (used only to render docs/orion/ PDFs and figures).
• PowerShell has no && ? use ;.
• Set $env:PYTHONIOENCODING="utf-8" before running the demo scripts. They print ?, ? and box-drawing characters; the default Windows console codepage raises UnicodeEncodeError and the script exits 1 for that reason alone, not because anything is broken.
• PYTHONPATH must include both backend and the repo root.
• Money is integer paise in the operating layer. Divide by 100 for display only.
• No git repo. Nothing to revert to. Be careful with destructive edits.

Where to start

docs/HANDOVER.md ?8 lists pending work in priority order. Short version: three non-code items gate the paper (35-prompt sweep, sourced benchmarks, competitor ground truth); the largest code task is the ablation harness in docs/WORK_REMAINING.md ?5.1; the frontend spec is ready in docs/FRONTEND_PROMPT.md.

Honesty requirement

The generation layer is measured. The operating layer runs unattended but its planners are rule-based. The ablation harness is built and has been run at n=10: the direction favours the full graph but nothing reaches p<0.05, so RQ4?RQ7 have no answer. Documents in docs/orion/ mark every figure Measured or Prospective. Do not let any document, slide or UI imply the system is experimentally validated when it is not.

Read docs/OPERATING_LAYER.md ?4b before touching the ablation. It records five defects the first sweep exposed ? including that the documented "load-bearing" learning branch could never fire in the live path ? and the modelling choices the result is conditional on.