SketchExtract Codex Control Pack v0.1

This control pack defines the first agentic workflow for converting an arbitrary floor-plan image into a reviewable CAD reconstruction.

The user does not have to follow SketchMark or any special symbol language.

Primary input classes:

• hand-drawn floor plans;
• BTI / cadastral / inventory-style plans;
• scanned or photographed architectural plans;
• PDF or screenshot floor plans.

Primary target:

• reconstruct the area of interest (AOI);
• preserve auxiliary enclosed spaces inside or adjacent to the AOI;
• recognize walls and wall geometry;
• recognize openings;
• bind each opening to its host wall;
• distinguish doors, including double-leaf doors;
• identify windows / viewing-window candidates;
• interpret written dimensions and wall-thickness information;
• reconstruct metric geometry;
• produce an editable CAD-ready layout;
• record every uncertainty, assumption, default and conflict in a reconstruction legend.

Core operating model

The model performs the interpretation. Each major stage writes durable artifacts to disk.

Reasoning between stages is disposable. Stage artifacts are durable.

A later stage should read the smallest set of earlier artifacts it needs. The final aggregation stage rereads the summaries and authoritative artifacts before producing the final model.

Pipeline

source image
  ↓
01 global scene scan
  ↓
02 AOI selection
  ↓
03 focused AOI interpretation
  ↓
04 opening + annotation interpretation
  ↓
05 constraint model
  ↓
06 metric reconstruction
  ↓
07 validation / focused reinspection / reconciliation
  ↓
08 final aggregation + CAD preparation

Important separation

Three kinds of information must remain distinguishable:

1. Observed — directly visible in the source.
2. Inferred — interpreted from visible evidence and context.
3. Defaulted / assumed — introduced by a domain rule because the source does not state the value.

A default or assumption must never be silently presented as an observation.

Output philosophy

The goal is a useful editable drawing, not perfect autonomous certainty.

When the model cannot resolve an ambiguity reliably, it should:

• choose no value if the ambiguity is material;
• preserve the competing hypotheses;
• write the issue into the reconstruction legend;
• optionally produce a best-effort CAD preview only when the uncertain choice is clearly marked.

See:

• control/methodologies/floor_plan_reconstruction_v1.md
• control/profiles/ct_room_profile_v1.md
• control/templates/run_layout.md

Repository interaction modes

Development work follows AGENTS.md and remains in DEVELOPMENT MODE. Pipeline execution is explicit and follows this chain:

AGENTS.md → control/AGENT_EXECUTION.md → playbook → methodology → skills
           → contracts → tools → runs

control/ is the canonical pipeline control plane. tools/ contains deterministic support utilities. tests/ is development/test infrastructure. experiments/ contains current non-canonical experiments, while archive/ contains historical evidence only. fixtures/ contains reusable example inputs. runs/ contains generated execution artifacts and is not control-plane source.

Development checks

Install dependencies with .venv/bin/python -m pip install -r requirements-dev.txt, then run pytest -q. Before verified materialization, run .venv/bin/python tools/control_preflight.py <run-dir> --mode reconstruct-and-materialize.