How gca-spatial-reasoning/gca works
GCA turns a benchmark sample and its images into a structured spatial answer. The principal flow formalizes the task, plans tool calls, executes geometric and code-based computations, and returns a numeric result with a natural-language summary. The README also identifies five supported evaluation benchmarks; their individual adapters are grouped because their dispatch wiring was not sampled. The graph reflects documented and excerpted relationships only; unsampled components and links are represented at directory level where necessary.
Main entry points: workflow.py Workflow runtime · executor.py Tool execution · reconstructor.py Geometric reconstruction · planner.py Tool planning
Components 25
Evaluation
- Benchmark adapters
evals - Evaluation base
evals/base.py - Agent entrypoint
entrypoints/agent.py - Spatial answer
Agent workflow
- Workflow runtime
workflow/workflow.py - Task formalization
workflow/nodes/analyst.py - Task routing
workflow/nodes/meta_planner.py - Tool planning
workflow/nodes/solver/planner.py - Tool execution
workflow/nodes/solver/executor.py
Spatial tools
- Geometric reconstruction
tools/apis/reconstructor.py - 3D reconstruction
tools/apis/vggt_model.py - Scene alignment
tools/apis/scene_aligner.py - Vision tools
tools/apis - Answer generation
tools/apis/final_answer.py
Models and execution
- Reasoning model
tools/apis/cot_reasoner.py - Code computation
tools/apis/python_tool.py - LLM clients
tools/llm_client - LLM retry helper
tools/utils/llm_invoke.py
State and reporting
- Runtime configuration
workflow/config.py - Agent state
workflow/state.py - Tool and context types
tools/apis/base.py - Tool dispatch
tools/tool_call.py - Image loading
tools/apis/io.py - Session logging
workflow/logging.py
Other components
- Benchmark sample
Connections 32
- Benchmark sample → Agent entrypoint: submits sample
README.md - Agent entrypoint → Benchmark adapters: selects benchmark
- Benchmark adapters → Evaluation base: implements
evals/__init__.py - Agent entrypoint → Workflow runtime: starts agent
- Workflow runtime → Runtime configuration: reads settings
workflow/workflow.py - Workflow runtime → Task formalization: routes workflow
workflow/workflow.py - Workflow runtime → Task routing: optionally routes
workflow/workflow.py - Workflow runtime → Tool planning: routes workflow
workflow/workflow.py - Workflow runtime → Tool execution: routes workflow
workflow/workflow.py - Workflow runtime → Agent state: manages state
workflow/workflow.py - Task formalization → Agent state: updates context
workflow/nodes/analyst.py - Task formalization → Reasoning model: analyzes task
- Tool planning → Agent state: reads workspace
workflow/nodes/solver/planner.py - Tool planning → LLM retry helper: invokes with retry
workflow/nodes/solver/planner.py - LLM retry helper → LLM clients: calls invoker
- Tool planning → Tool execution: produces tool plan
workflow/nodes/solver/planner.py - Tool execution → Agent state: records results
workflow/state.py - Tool execution → Tool dispatch: dispatches calls
workflow/nodes/solver/executor.py - Tool dispatch → Tool and context types: uses tool interface
tools/tool_call.py - Tool execution → Code computation: executes code calls
workflow/nodes/solver/executor.py - Tool execution → Geometric reconstruction: executes geometry calls
- Geometric reconstruction → Reasoning model: asks for strategy
tools/apis/reconstructor.py - Geometric reconstruction → 3D reconstruction: reconstructs scenes
tools/apis/reconstructor.py - Geometric reconstruction → Scene alignment: aligns scenes
tools/apis/reconstructor.py - Geometric reconstruction → Vision tools: requests detections
tools/apis/reconstructor.py - 3D reconstruction → Image loading: loads images
tools/apis/vggt_model.py - Code computation → LLM retry helper: generates code
tools/apis/python_tool.py - Code computation → Reasoning model: uses code model
- Workflow runtime → Answer generation: extracts final result
- Workflow runtime → Session logging: logs session
workflow/workflow.py - Answer generation → Spatial answer: returns result
- Workflow runtime → Spatial answer: returns summary
workflow/workflow.py