Repository Graph Engine (v3.0.0)

CodeTruth Agent V3 — Module 2: Repository Graph Engine

Tag: v3.0.0-module2

Summary

Module 2 maps how a Python codebase is wired together: every function,
class, module, import, dependency, and resolved function call, as 6
structured graphs. Validated on 69 real open-source repositories (49,379
files), 0 crashes, 1,005,321 resolved call edges, governance gate
APPROVED/BLOCKED consistent with Module 1.

This release covers Module 2's Python core (V3-004 through V3-009),
frozen, plus an early multi-language extension scaffold (Java,
JavaScript/TypeScript, C/C++) - a validated baseline, not frozen.

Validation

• 31/31 unit tests pass
• 69 repos scanned, 0 crashes, 49,379 files, 515,610 functions,
  84,468 classes
• 1,005,321 resolved call edges (2.5x increase from initial baseline)
• 65/69 governance gate APPROVED, 4/69 correctly BLOCKED (non-Python repos)

What's New

D-001 (two-stage global build), D-002/D-003 (builtins/constructors/stdlib
whitelist), D-004 (cross-module inheritance, -53% unresolved self-calls),
Gap 1/2/3 (qualified-module resolution, local type tracking, import-cycle
detection), D-006 (nested functions), D-007 (relative import resolution).

Multi-language scaffold: Java (javalang), JavaScript/TypeScript
(tree-sitter, incl. cross-file resolution, 0 parse errors), C/C++ (regex
heuristic) - all validated with 0 crashes on real repos (Redis, u-boot,
vscode, elasticsearch, spring-boot, react, ui5-webcomponents).

Documentation

13 documents in v3/docs/module2/ covering architecture, decisions,
validation, extension guide, FAQ, gaps & roadmap, POC history, and a
quickstart with a worked example. Project-wide overview:
v3/docs/PROJECT_RECORD_v3.0.0-module2.md.

Known Limitations (Documented, By Design)

attribute_call (1,785,190 instances - requires future type inference);
D-008 package-root mismatches in some large frameworks (documented, not
yet pursued). Full detail in MODULE2_GAPS_AND_ROADMAP.md.

This release does not modify repositories and makes no code-change
decisions.

Repository Cognition Engine (v3.0.0)

CodeTruth Agent V3 — Module 1: Repository Cognition Engine

Tag: v3.0.0-module1

Summary

Module 1 is a deterministic, rule-based engine that scans a software
repository and reports its application type, primary framework,
technology stack, languages, build systems, and file inventory —
along with a governance gate decision (V3-003) used by later modules
in the CodeTruth Agent V3 pipeline.

This release covers Module 1 only: the Repository Cognition Engine.
It does not modify repositories and makes no code-change decisions.

Validation

Run against 69 real, cloned, open-source repositories spanning 39
distinct application types — from 35 files (python-sgp4) to 61,850
files (Zephyr RTOS), including repositories with no Python code at all
(Redis, Nginx, Go, Rust, U-Boot).

69 repositories scanned
69/69 = 100% discovery score
69/69 = correct application type
69/69 = correct primary framework (or correctly "No Framework Detected")
69/69 = governance gate APPROVED
 0/69 = crashes
 0/69 = skipped
57/69 = 100% classification score
12/69 = 75% classification score ("No Framework Detected" — correct
        by design for non-Python system software: Redis, Nginx, Go,
        Rust, FreeCAD, LibreCAD, Shapely, rclpy, gst-python, u-boot,
        gnuradio, CodeTruthAgent)
35/35 unit tests pass
441,660 total files scanned

Full per-repository results: FULL_DOMAIN_SUMMARY.md
(also available as .json / .csv).

What's Included

• v3/repository_cognition/ — engine source
  (cognition_engine.py, cognition_report.py, report_writer.py,
  framework_signatures.py)
• v3/tests/test_module1_cognition.py — 35 unit tests
• v3/tests/scan_all_repos_v3.py — 69-repository validation script
• v3/outputs/real_scans/ — per-repository scan reports (.txt/.md)
  and FULL_DOMAIN_SUMMARY.{md,json,csv}
• v3/docs/module1/ — documentation:
  • MODULE1_CAPABILITY_PROOF.md
  • MODULE1_DOCUMENTATION.md
  • MODULE1_TEST_REGISTER.md
  • MODULE1_EXTENSION_GUIDE.md
  • MODULE1_REAL_WORLD_PROBLEM.md

Notes

• "No Framework Detected" is a genuine, correct result for
  repositories that have no Python package framework dependency —
  it is not an error or a gap.
• All classification knowledge lives in framework_signatures.py and
  related tables; extending to new application types or frameworks
  does not require changes to the core engine (see
  MODULE1_EXTENSION_GUIDE.md).

Next

Module 2 — Repository Graph Engine (adjacency-dict storage, two-pass
build) is next, using Module 1's output as its starting point.

CodeTruth Agent V2 — Deterministic Pre-Modification Governance Layer

V2 adds semantic similarity, behavioral signature detection, and multi-signal fusion to V1's rule-based governance foundation.

Evaluated on 8 open-source Python repositories — 7,862 files, 1,241 pair analyses, 0 pipeline errors.

See README.md for full architecture, evaluation results, and audit methodology.

CodeTruth Agent V1 — Rule-Based Safe-Merge Reasoning System

V1 rule-based safe-merge reasoning system for Python repositories.

8-stage pipeline: Scan → Parse → Detect → Risk Analysis → Dependency Map → Human Approval → Safe Modify → Memory Learning

20 of 22 UAT test cases passed. 2 by-design exclusions documented.