UpToCode

UpToCode

UpToCode is a design-time architecture-quality scanner for Python and TypeScript agent applications. It finds missing execution bounds, run budgets, approvals, validation, resilience controls, evals, and observability; explains the evidence; and produces an approval-driven plan for Codex without rewriting source code.

The product is deliberately narrower than a general agent-security scanner. Version 1.7 recognizes OpenAI, Anthropic, CrewAI, PydanticAI, LlamaIndex, LangGraph, conservative custom Python/TypeScript loops, and positive evidence of unbounded context growth. Static scans are local and offline. Optional GPT‑5.6 judgment remains explicit, bounded, redacted, and code-sharing gated.

The Python distribution, import package, and command are all uptocode. Version 1.0.0

is the published baseline; this source tree is prepared as 1.10.0 for the operator-gated Batch 3 release train.

Install locally

UpToCode requires Python 3.11 or newer. Install the verified 1.0 release from PyPI:

python -m pip install "uptocode==1.0.0"

uptocode --help

For repository development, use the editable test environment:

python -m pip install -e ".[test]"

uptocode --help

All ordinary tests are mocked/offline:

python -m pytest -q

python scripts/acceptance.py

Judge test build

The repository and package are public. No hosted credential or paid model call is required for the complete offline test build.

git clone https://github.com/DDYRich72/UpToCode.git

cd UpToCode python -m venv .venv

Activate the environment with .venv\Scripts\activate on Windows PowerShell or source .venv/bin/activate on POSIX, then run:

python -m pip install --upgrade pip

python -m pip install -e ".[test]" uptocode scan fixtures/bad_python --fail-on critical uptocode scan fixtures/bad_python --format json --output report.json uptocode review report.json --approve AA001,AA003 --reject AA012 --non-interactive uptocode plan report.json --manifest .uptocode/manifest.json --output FIXPLAN.md python scripts/acceptance.py

The first scan intentionally exits 1 because the fixture contains critical findings. The remaining commands produce a report-bound review manifest and an approved-only FIXPLAN.md; they do not edit the fixture. To inspect the local MCP surface, run uptocode serve --transport stdio --root . from an MCP client or use the JSON registration under MCP registration. python scripts/acceptance.py also launches a real stdio server, lists all ten tools, and calls check_loop offline.

The submitted-content hosted transport is deployed at https://uptocode-mcp-1015314816960.us-central1.run.app/mcp. Its bearer credential is issued separately and read from UPTOCODE_API_KEY; never commit it or enter it into the website. The immutable production revision passed readiness, unauthorized-access, AA001, tool-discovery, judgment-gate, rate-limit, and payload-free-log checks. Hosted judgment is off, no OpenAI key is attached, and verification made zero paid model calls.

Scan, review, plan

uptocode scan PATH [--format terminal|json|html|github|sarif] [--output FILE]

[--judgment --send-code] [--fail-on critical|warning|info] [--include-experimental] [--share-safe] [--baseline FILE|--update-baseline FILE] [--changed-since REF] [--select RULES] [--ignore RULES] [--exclude GLOB] [--severity RULE=LEVEL] [--fail-on-analysis-warning] [--github-summary PATH] [--verbose] uptocode review REPORT [--approve IDS|--approve-all] [--reject IDS] [--reuse MANIFEST] [--non-interactive] uptocode plan REPORT --manifest MANIFEST [--output FIXPLAN.md] uptocode fix REPORT --manifest MANIFEST --runner codex|command [--command TEMPLATE] [--verify-command COMMAND] [--apply] uptocode serve --transport stdio --root PATH uptocode serve --transport streamable-http --mode hosted

Scanning, review, planning, and the default fix preview are non-mutating:

uptocode scan . --format json --output report.json

uptocode review report.json --approve AA001,AA003 --reject AA012 --non-interactive uptocode plan report.json --manifest .uptocode/manifest.json --output FIXPLAN.md uptocode fix report.json --manifest .uptocode/manifest.json --runner codex

review binds decisions to the exact report fingerprint. plan refuses a mismatched manifest and includes only approved findings. Neither command edits the scanned repository. fix is also a no-write dry-run unless --apply is supplied. The applying path requires a clean Git checkout, creates one retained uptocode/fix- branch/worktree per approved finding, invokes only the selected external runner there, verifies fingerprint absence, and never commits, merges, deletes work, or changes the invoking checkout.

Exit codes are 0 for no configured threshold breach, 1 for a finding at or above --fail-on (or a requested analysis-warning gate), and 2 for invalid configuration or an unrecoverable scan error. Experimental findings remain visible but require --include-experimental to affect --fail-on. --share-safe sanitizes JSON, HTML, or SARIF for distribution while retaining the repository revision. Baseline 2.1 reports new, aging, and resolved debt and preserves first_seen when updated. Suppressions may include owner, quoted reason, and an inclusive expires date. HTML requires --output. GitHub workflow commands escape untrusted command data and properties. SARIF 2.1.0 omits absent fields, declares default rule levels, and carries stable partial fingerprints. --github-summary PATH appends a bounded Markdown summary; when GITHUB_STEP_SUMMARY is set, the summary is appended there automatically.

Pre-commit

repos:

• repo: https://github.com/DDYRich72/UpToCode rev: v1.1.0 hooks:
• id: uptocode

Run pre-commit install, then use pre-commit run uptocode --all-files. To copy the project badge into your own README:

[![Scanned with UpToCode](https://img.shields.io/badge/scanned%20with-UpToCode-175cd3)](https://github.com/DDYRich72/UpToCode)

GitHub Action

The repository ships a composite Action that installs UpToCode from the Action checkout, runs the scan, uploads SARIF with always(), and only then returns the preserved scanner exit code:

permissions:

contents: read security-events: write

steps:

• uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
• uses: DDYRich72/UpToCode@v1.0.0 with: path: . fail-on: critical version: source upload-sarif: "true"

version: source installs from the checked-out Action source. An exact semantic version, such as 1.0.0, installs the matching uptocode PyPI release. Set upload-sarif: "false" when code-scanning upload is not desired; otherwise the calling workflow needs security-events: write.

Example terminal output:

UpToCode scan

Coverage: 1/1 files analyzed Findings: 11 | Warnings: 0 | Redactions: 3 [CRITICAL] AA001 agent.py:21 - Unbounded agent loop [CRITICAL] AA003 agent.py:11 - Ungated destructive action

Privacy model

• Static scanning is the default and makes no network calls.
• --judgment is rejected unless --send-code is also present.
• Judgment sends only normalized evidence and bounded, redacted excerpts, grouped once per candidate rule. Requests use gpt-5.6, Pydantic Structured Outputs, store=false, a 2,000-token output ceiling, a 30-second timeout, and a six-rule call budget.
• Recognized OpenAI keys, AWS access keys, bearer tokens, email addresses, and US Social Security numbers are replaced before report or judgment output. Detection is intentionally narrow and is not a substitute for a dedicated secret/PII scanner.
• Nested .gitignore files, default build/dependency/generated exclusions, .uptocode.yml, 1 MiB per-file and 4 MiB aggregate limits, binary/non-Python exclusion, and explicit suppressions are honored. Production UpToCode source is gated at zero suppressions.
• Refusal, timeout, authentication failure, or API failure never deletes static results; the report records partial or failed judgment status and an analysis warning.

No live API call is part of the offline test or acceptance suite. The submission smoke test requires a key and separate approval for paid usage.

Rules

| AA014 | Network MCP without authentication | Static | critical | MCP authorization, [FastMCP