Code Flow Studio

An independently written successor to a group AI-in-SDLC capstone: a local-first development workbench for isolated polyglot checks, configurable AI proposals, and replayable revision evidence.

Version 0.2: real private local accounts. Create an account, unlock your encrypted source/history workspace, and use the same interface in the desktop app or your browser. No public backend, shared login, analytics, ads or cloud AI calls are enabled by default. By Jason Stys; contact via LinkedIn.

Download OS-specific installers/portable packages from Releases. Read private-account operation, current validation and security boundaries. Packages are unsigned. This is a local development tool, not a public hostile-code SaaS sandbox or an autonomous correctness oracle.

The development loop

1. Upload flat source/data files or edit code in the offline Monaco editor. Select a compiler adapter and entry point. Write requirements and approve independent stdin/stdout cases.
2. Optionally ask a configured model to generate code, tests, improvements or live suggestions. Source transmission is opt-in. Responses are schema-validated unverified proposals, not executed results.
3. Save an immutable revision and run it through the container broker. Compilation, approved cases, supplied pytest harnesses, Python coverage/mutation, bounded fuzzing and exploratory benchmarks produce separate evidence.
4. Review findings and add a candidate revision. Baseline expectations and existing tests cannot be removed. Compare results under the same environment and adjustable runtime/memory/readability/footprint preferences.
5. Download escaped HTML, machine-readable JSON and a source/test/harness/report ZIP. Reopen saved revisions and replay them.

Features and limits

Feature	Implemented contract
Private accounts	Real alias/password registration, hashed sessions and one-use login recovery. Independent browser-derived vault key, authenticated AES-256-GCM locked source/history/provider archives and per-session unlock proof. No developer plaintext read API. Trusted local service and OS can access active plaintext; not browser-only E2EE.
Desktop application	Electron shell and bundled Python service; one-click Windows installer plus OS-specific Linux/macOS builds. No IntelliJ or separately installed Python required for the UI. Unsigned development packages.
Web application	Same responsive editor/API served on loopback. Narrow-screen browser support; not a native Android/iOS installer or public hosted service.
Polyglot execution	Fixed CLI adapters for Python, JavaScript, TypeScript, C/C++, Java, C#, Rust, PHP, Bash, SQLite SQL and GNU x86-64 assembly. Actual toolchains live inside the worker image.
Testing	Approved CLI oracles and reviewed pytest harnesses, including subprocess harnesses for non-Python targets. 48-category applicability catalog; human, hardware, cloud/load and specialized tool checks need independent integrations.
Coverage/mutation	Python entry line/branch coverage; up to eight operator mutants. No fabricated condition/MC/DC, function, edge or universal mutation scores.
Fuzzing	Seeded black-box generation, mutation, dictionary, boundaries and CSV/SQL grammar; instrumented libFuzzer/AFL++ for reviewed C/C++ harnesses and Atheris for Python harnesses. Differential/metamorphic/stateful relations require custom oracles.
AI routing	Account-owned OpenAI-compatible, Anthropic and Ollama endpoints/keys, four task routes and destination-version-bound consent. Keys stay server-side after explicit settings submission, outside histories/exports/workers. Protocol tests use mocks, not a claim of live model quality.
Metrics/ranking	Repeated process wall times, peak child RSS, source bytes and transparent long-line proxy. Passing independent cases is mandatory. No syntax-based Big-O claims or automatic “best” model claims.
Training	Rights-checked, project-disjoint JSONL preparation, actual small causal-model creation or approved local-model fine-tuning, safe-tensor output and authenticated local inference API. Tiny synthetic smoke training is not a useful general coding model.
CI	SHA-pinned, read-only Actions validate controller/browser behavior, real worker scenarios, training smoke and native packages. Push analysis retains generated tests/candidate evidence when a provider is explicitly configured; never auto-commits model changes.

Binary formats can be retained as fixtures. There is no promise to understand every document, image, encrypted archive or binary. Supply text requirements and explicit harnesses for formats without a parser; unsupported extraction must not silently succeed.

Run and install

For end users, use the checksummed OS-specific packages in Releases. The UI runtime is bundled. Open the app's Code Flow → Open local website menu for browser use; no login token appears in the URL. Compiler jobs require an isolated local Unix/named-pipe Docker engine and the trusted worker image. Missing isolation is blocked, never passed; there is no host-execution fallback. Save/edit/static review/history/help remain usable without Docker.

Source-development instructions, provider setup, first-run diagnostics and packaging are in setup. All documentation is indexed in docs. Original private capstone code, team files, API credentials and vendor trees are not distributed here.

Source map

Location	Responsibility
studio/accounts.py, studio/permissions.py	scrypt logins, session/recovery digests, bounded attempts and private per-OS data permissions.
studio/private_api.py	Account-only HTTP/CSRF boundary, own provider settings, per-session key proof, backup/recovery/deletion and scoped controller proxy.
studio/private_workspace.py, studio/vault.py	Single-instance ownership, account-specific engine lifecycle, authenticated streaming encryption, SQLite online snapshots and safe restoration.
studio/private_cli.py	Token-free loopback account service, dedicated data root and exact-parent graceful shutdown monitor.
ui/Account.tsx	Sign-in/unlock/recovery, contextual help, encrypted backup, private AI configuration and creator contact.
studio/models.py	Validated flat-file, project, case, preference and AI proposal schemas; size/path limits.
studio/api.py	Authenticated loopback API, bounded jobs, history, reports, comparisons and AI tasks.
studio/store.py	SQLite transactions, append-only revisions, immutable baseline oracles and interrupted-job recovery.
studio/execution.py	Fixed container flags, readiness checks, output/deadline bounds and exact owned-container cleanup.
studio/analysis.py	Non-executing source heuristics and correctness-gated weighted comparisons.
studio/providers.py	Private provider configuration, task routes, consent, bounded requests and proposal schema validation.
studio/catalog.py	All 48 supplied test categories with applicability/integration requirements.
studio/reports.py	Escaped, inert HTML/JSON/ZIP evidence exports.
studio/cli.py	Shared web launcher and manifest-to-report CI command.
runner/Dockerfile, runner/worker.py	Trusted compiler image and isolated compile/test/coverage/mutation/fuzz/benchmark execution.
ui/App.tsx	Workspace, cases, priorities, live suggestions, proposal approval, evidence and history views.
ui/main.tsx, ui/api.ts	Offline editor worker registry and authenticated same-origin client/downloads.
ui/demo.ts, ui/types.ts	Synthetic CSV example and renderer contracts.
ui/index.html, ui/style.css	Local HTML entry and responsive accessible layout.
desktop/main.cjs	Sandboxed Electron renderer and private bundled-service lifecycle.
training/dataset.py, training/train.py, training/serve.py	Provenance/split validation, parameterized real training, local model API.
examples/projects.py, examples/web-demo.project.json	Compiler-contract fixtures and the exact synthetic CSV source/manifest used by the UI, checked in CI.
examples/training.json, examples/training.jsonl	Tiny rights-owned training smoke parameters and data; no quality claims.
scripts/frozen_entry.py	Frozen-service entry point for installers.
scripts/worker_checks.py	Actual isolated compiler/fuzz/security-boundary integration scenarios.
scripts/private_worker_checks.py	Real twelve-adapter execution through authenticated account API and encrypted-history reopen.
scripts/license_bundle.py, scripts/release_archive.py	Generated third-party notices, portable app packages and SHA-256 checksums; never package user accounts.
scripts/push_analysis.py	Push generation/test/candidate workflow without automatic source promotion.
tests/	Controller/security/history/provider/metric tests, browser flows and native launch smoke.
.github/workflows/	Validation, real isolated workers, training and native OS-package workflows.
pyproject.toml, package.json, lockfile, TypeScript/Vite/Playwright configs	Dependency and reproducible build/test definitions.

Every source header points to the checked symbol/variable line map. Architecture explains trade-offs; testing separates test claims from evidence.

Historical 0.1 bearer-token CLI mode remains explicitly opt-in for existing automation; it is not account mode. No automatic import of legacy plaintext projects or global AI keys occurs.