Code Factory

Review your code, find missing checks, and verify repairs using the tools you already know. Start with one command:

python -m pip install factoryline-code-factory
factory scan --root .

Your workflow: scan → review the findings → repair → run your tests and the affected checks again. The basic scan reads eligible Python source; it does not run your project. Missing evidence is INCOMPLETE, not a clean bill of health.

Current source preview — core 0.48.2, VS Code 1.2.1, JetBrains 1.2.0

The bounded security review follows values through supported local helpers to HTTP, file, and SQL sinks, and checks an upload directory separately from its sanitized filename. Studio surfaces a shared next action; optional JEV integration remains advisory; unavailable evaluations stay incomplete. These changes are not a claim of exhaustive analysis or measured detection accuracy. Review findings, run the affected checks, and decide what to accept before using a repair.

Want a visual workspace? Run factory studio --root .. Need options? factory --help shows the short path; factory --help-all shows all commands. Start with the quick guide.

Keep pytest, mutation testing, CodeQL, Semgrep and your review checklist. CF adds test-quality findings and source-bound local evidence alongside them. A basic scan needs no adoption ceremony, signing keys, tenant setup or Docker. Native worker execution has separate setup and safety requirements.

Code Factory is the product. Its internal engines and advanced governance features are optional workflows, not prerequisites for reviewing your code.

Starting from an idea?

Describe your idea to your coding agent and create PRD.md. The optional beginner guide helps turn it into concrete acceptance checks. For an existing repository, start with the scan above.

Graph Ops telemetry sample — October 8, 2026

Fixed: bounded evaluator output validation across all six runtime lanes; malformed results stay incomplete. Shared imported assertion helpers and inherited test mixins are resolved by bounded static analysis. Changed: audit findings carry candidate-bound repair packets with reproduction, negative-control and separate specialty AI review requirements. Added: Graph Ops agent repair telemetry, evidence requirements, stop conditions and searchable individual test outcomes. Repair packets are consumed by the host agent; this view does not execute or approve repairs.

Actual local Graph Ops capture, October 8, 2026: 415 targeted tests passed. JUnit results are local and unbound; missing runtime receipts and coverage remain explicitly incomplete. Agent repair cards show candidate bindings, required evidence and separate specialty AI review requirements.

Code Factory + ForgeLine (CF/FL) is a robust, local-first code-audit factory. It starts by collecting local software review evidence for the candidate change. It connects requirements, architecture checks, Python AST security analysis, behavioral test evidence, workflow integrity, specialty AI review, and actionable repair. Its receipts support review; the tool does not certify software, guarantee that defects are absent, or approve a release.

Product naming: Code Factory is the product and ForgeLine is its audit and architecture engine. SpecLine, HSF, and Prestige are advanced internal lanes that appear only when a project declares the corresponding scope; they are not separate products a new user must learn before running a scan.

Graph Ops in action

This full-page local capture shows Graph Ops tracing a change from sealed lineage through a forensic finding to a proposed recovery path. The guarded action remains locked; the dashboard is a read-only inspection surface.

The Counterfactual Arena compares repair candidates and shows their evidence and risk. Winner controls and the Evidence Frontier show how a reviewer can inspect the rationale and choose the next proof.

The current Graph Ops UI also separates native deep-scan progress, evaluated deep-audit receipts, six runtime lanes, and individual JUnit test cases. Each panel labels its evidence state and next repair. Filter or search the recorded tests, then load more cases as needed; an inventory or a missing report never appears as a passing run. The JUnit reader accepts up to 10,000 cases and marks larger or malformed reports incomplete. A JUnit report is local observation without a current-candidate binding.

This local UI capture shows each reported test case with its status. The audit cards keep checks without a bound result explicitly marked not run.

Add deeper audit lanes

factory scan --root . --deep

Deep mode orchestrates architecture, Python security and test-oracle checks, workflow integrity, requirements, runtime coverage, and dependency evidence. Languages including TypeScript/JSX, Go, Rust, Java, C/C++, C#, Ruby, PHP and Swift are inventoried explicitly; missing native analysis remains unmeasured.

For authorized native execution, use:

factory scan --root . --deep --worker-config .factory/worker-config.json --json

The configuration contains workspace-relative manifest, authorization, and trust_root paths plus manifest_sha256 and trust_root_sha256 pins. This invokes the existing signed Docker worker engine for configured CodeQL, Semgrep, secrets, dependency, container, runtime and fuzz lanes. Unsigned plans are rejected. Completed workers still require independent agent review; missing evidence and review cannot produce a passing verdict. Native execution requires its configured images and tooling and can take longer than a minute. Each unresolved lane includes its next action. Receipt presence alone is not verification, and no scan claims to detect every defect.

Package metadata now uses Production/Stable, starting with this source update. That release-status label is separate from detection accuracy, independent benchmark results, and provider publication.

In an interactive terminal, factory optionally makes one plain PyPI version check and caches the result for up to 24 hours, then prints a notice when a newer version is available. It never downloads or installs the update. The check is quiet in CI, JSON output, server/MCP, help, version, and non-interactive runs; set FACTORY_DISABLE_UPDATE_CHECK=1 to turn it off. The optional request does not include a project path, account identifier, or usage data. All audit and proof commands remain local unless an operator explicitly uses a separate provider integration.

Run repository commands from the project being reviewed. factory guide is a read-only orientation; it does not run tests or agents.

Inspect this repository

These commands show the current architecture policy, the bounded static security scan, and the runtime-audit setup:

factory architecture health --root . --json
factory audit security --root . --json
factory runtime-audit status --root . --json

factory audit security checks a limited set of source patterns. It is not a penetration test. The runtime audit needs a separately reviewed plan and its own environment evidence. A status of NOT_RUN or a clean static scan is not a complete project audit.

To display individual tests in Graph Ops, run your suite with a JUnit report at .factory/test-reports/pytest.xml (for example, python -m pytest --junitxml=.factory/test-reports/pytest.xml). The local Studio reads the report and labels its candidate binding UNBOUND; it does not infer that those results still apply after source changes.

The prior snapshot's repository self-audit receipt records A (97.9/100) and 1,055/1,085 functions attributed to test intent (97.24%). That static measurement applies to its recorded source digest and is superseded for this candidate by the clean tracked-snapshot result above; it is not runtime coverage or certification.

ForgeLine 0.10.8 computes Python complexity for public module functions and public class methods in its recorded metrics.scope.code_files; names starting with _ are excluded. Its AST metric counts branches, boolean alternatives, exception handlers, with, and assertions. Ruff C901 uses a different McCabe metric and also checks private functions. CI therefore runs python -m ruff check --select C901 factoryline tests separately at limit 10. The ForgeLine maximum must not be read as the maximum of every Python function. CI also checks action and reusable workflow references for immutable commit pins; the generic workflow evidence validator alone does not inspect those refs.

Release controls

Candidate preflight requires release-cadence admission and strict architecture health. Protected main requires CI and a separate specialty AI source review; the coordinator records that assessment, separately from test evidence. It is not a second human approval. Provider publication and marketplace approval are separate outcomes. See release channels.

Architecture health checks growth budgets. ForgeLine's repository inventory and feature QA are separate checks. The current gap review records historical parser behavior; the current candidate's pinned-parser result is listed above. CF requires ForgeLine 0.10.8 and checks real MJS, TS, and TSX feature QA during factory doctor --strict. Static signals are neither executed coverage nor a security certification.

More detail

• Engineering review workflow
• GitHub Proof Review
• Commercial availability and limits
• Audit condition inventory
• Runtime assurance limits
• Release channels and publication evidence
• Changelog
• Documentation index

Updated video preview

60-second CF/FL audit and repair walkthrough — October 8, 2026. Silent captions, current Graph Ops captures, bounded scanner improvements and agent repair evidence requirements. The source render is in videos/code-factory-proof-of-survival/remotion.

Native agent-harness preview — current source candidate

Status: Current source preview only; published Code Factory 0.48.1 remains unchanged. Local-wheel inclusion is not claimed.

The native harness keeps incident checkpoints for recovery, binds review context shards to source, and compares candidate repairs without applying them. Budgeted Jev triage keeps unprocessed incidents visible as pending, not assessed. Parallel audits are opt-in and require a signed policy declaring work non-mutating and resource-safe; serial remains the default, and the declaration does not prove isolation. Optional activation experiments call only hooks explicitly registered by the caller. Their observations are experimental, not model certification or accuracy evidence. The source taxonomy and manifest define the same routes for Junie while keeping execution and identity limits visible.

Native scanner expansion — core 0.48.2 / editor 1.2.0 source preview

Fixed: native scanner controls require vulnerable and safe results bound to image, profile and source hashes. Missing evidence remains incomplete. Test-intent auditing recognizes argument/await assertions, expected-exception failure checks and invoked local assertion helpers. Direct local serialization round-trips in tests retain informational findings and custom-reducer warnings. Changed: agents use one factory scan --deep overview and explicitly configured native workers selected for the project's languages and risks. JavaScript/Actions workers verify bundled compiled queries rather than rebuilding them, and receipt discovery skips explicit file checks for unrelated logs. Added: CodeQL JavaScript/TypeScript interprocedural analysis, GitHub Actions security-extended analysis, and ten Python/JavaScript/TypeScript Semgrep rules including local taint tracking. CodeQL controls require native flow evidence; Semgrep community output does not establish cross-file coverage. The external-source regression corpus now includes 31 upstream cases from requests, Rich and Click; it is separate from the eight seeded benchmark cases and does not establish held-out accuracy.

Code Factory is a robust code audit factory with Production/Stable package metadata. This source preview does not establish new marketplace publication, runtime verification of every repository, or measured detection accuracy.

Smaller models, stronger quality workflows

CF/FL is designed to guide smaller LLMs toward frontier-like coding performance through deterministic workflows, independent audits and evidence-driven repair. Sealed task packets keep requirements, scope and acceptance checks explicit. This is the design objective; model parity and cost savings require a paired benchmark and are not established by the current attribution metric.