How gaurshalendra2412/axi-runtime works
AXI is a runtime that keeps an AI agent’s editable state consistent across multi-step work. The principal flow is: a user request and current state go to a model; AXI parses the proposed delta, checks and completes it under graph rules, commits accepted changes, and returns the exact state for the next step. The Python engine also includes constrained decoding, scope and grounding logic, diagnostics, and a ledger. Experiment runners cover local and hosted models, long sessions, and ARC-style tasks. Some links below reflect README descriptions because the orchestration source is only partially excerpted.
Main entry points: agent_loop.py Agent loop · api_llm.py Model adapters · gemini_arc.py ARC experiments · cico_parser.py Delta parser
Components 19
Agent interaction
- Agent loop
python/axi/experiments/agent_loop.py - Model adapters
python/axi/experiments/api_llm.py
State runtime
- Delta parser
python/axi/engine/cico_parser.py - Constrained decoder
python/axi/engine/cico_decoder.py - Graph gate
python/axi/engine/gate.py - Rule completion
python/axi/engine/grounding.py - Scope filtering
python/axi/engine/scope.py - State response
python/axi/engine/answer.py
Runtime support
- State ledger
python/axi/engine/ledger.py - Diagnostics
python/axi/engine/diagnostics.py - State paging
python/axi/engine/pager.py - GPU token mask
python/axi/engine/gpu_mask.py
Experiments
- Graph tasks
python/axi/experiments/scope_tasks.py - Long-run experiments
python/axi/experiments/million.py - ARC experiments
python/axi/experiments/gemini_arc.py - Replay results
python/axi/experiments/million_hard_result.py - Resumable runs
python/axi/experiments/resumable.py
Other components
- User request
- AI model
Connections 18
- User request → Agent loop: submits request
README.md - Agent loop → AI model: requests proposal
README.md - AI model → Agent loop: returns delta
README.md - Agent loop → Delta parser: parses proposal
python/axi/experiments/agent_loop.py - Agent loop → Graph gate: checks delta
README.md - Agent loop → Rule completion: completes rules
README.md - Agent loop → Scope filtering: filters changes
python/axi/experiments/cascade.py - Agent loop → State ledger: records state
- Agent loop → Diagnostics: reports outcomes
python/axi/experiments/agent_loop.py - Agent loop → State response: returns exact state
README.md - Model adapters → AI model: calls hosted model
python/axi/experiments/api_llm.py - Model adapters → Agent loop: serves generation
python/axi/experiments/agent_loop.py - Constrained decoder → GPU token mask: builds token mask
python/axi/engine/gpu_mask.py - Graph tasks → Delta parser: parses ideal deltas
python/axi/experiments/scope_tasks.py - Long-run experiments → Agent loop: runs agent steps
- ARC experiments → Model adapters: uses model adapter
python/axi/experiments/gemini_arc.py - ARC experiments → Resumable runs: uses resumable runs
python/axi/experiments/gemini_arc.py - Replay results → Agent loop: reads run results
python/axi/experiments/million_hard_result.py