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-<fingerprint12> 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:
[](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
- 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
| ID | Architecture defect | Tier | Severity | Primary guidance |
|---|
| AA001 | Unbounded agent loop | Static | critical | Agents SDK runner |
| AA002 | Missing output/run budget controls | Static | critical | OpenAI practical guide |
| AA003 | Ungated destructive action | Static + judgment | critical | Agents SDK approvals, OpenAI safety |
| AA004 | Unvalidated tool arguments | Static | critical | OpenAI function calling, Anthropic tool definitions |
| AA005 | Raw untrusted content in prompts | Judgment | critical | OpenAI safety, Anthropic guardrails |
| AA006 | Secret or narrow PII exposure | Static | critical | OpenAI safety |
| AA007 | Missing timeout, retry, or retry backoff | Static | warning | OpenAI practical guide |
| AA008 | Unjustified orchestration complexity | Judgment | warning | Anthropic effective agents, OpenAI practical guide |
| AA009 | Poor tool schema | Judgment | warning | OpenAI function calling, Anthropic tool definitions |
| AA010 | Unvalidated model output before side effect | Static + judgment | warning | OpenAI function calling, OpenAI safety |
| AA011 | No agent eval coverage | Static + judgment | warning | OpenAI evals, Google ADK evaluation |
| AA012 | No agent observability | Static | info | Agents SDK tracing |
| AA013 | Unbounded context growth | Static | warning | Anthropic effective agents, OpenAI practical guide |
| AA014 | Network MCP without authentication | Static | critical | MCP authorization, FastMCP token verification |
| AA015 | MCP tools without annotations | Static | warning | MCP tools |
| AA016 | Non-strict MCP tool arguments | Static | warning | MCP tools, Pydantic configuration |
| AA017 | Undifferentiated tool errors | Static + judgment (experimental) | warning | MCP tool errors, Claude tool-call errors |
| AA018 | Prompt-only policy enforcement | Static + judgment (experimental) | warning | NIST AI 600-1, OWASP Excessive Agency |
Each static verdict requires recognized syntax and source-backed evidence. Unsupported or dynamic constructs produce coverage gaps or analysis warnings; they are never silently declared clean.
VS Code
The standalone extension in editor/vscode-uptocode/ scans Python documents on save and
turns valid UpToCode JSON findings into file diagnostics. It runs no daemon or background
repository scan and provides no auto-fix. Build and install a local package with:
cd editor/vscode-uptocode
npm ci
npm run verify
code --install-extension uptocode-vscode.vsix
Configure uptocode.executable, uptocode.extraArgs, and uptocode.failOn in VS Code.
Marketplace publication remains operator-only.
Supported and unsupported constructs
Recognized evidence includes:
- OpenAI
Runner.run/run_sync limits: omitted max_turns is clean because the SDK default is bounded; explicit numeric limits are clean; max_turns=None is flagged.
- LangGraph
recursion_limit when statically resolvable.
while True loops with detectable exits, constant bounds, and simple recursive base cases.
- OpenAI Responses model calls, output caps, client/call timeouts, bounded client retries, and simple source-level run budgets.
@function_tool approvals, its model-controlled parameters reaching common SQL/shell/file/network sinks, parameter-linked validation, agent eval markers/tests, and nearby logging/tracing evidence.
- Anthropic Messages/Agent SDK, CrewAI, PydanticAI, and LlamaIndex framework contracts documented under
docs/frameworks/.
- Tree-sitter analysis for
.ts, .tsx, and .mts: custom infinite loops, OpenAI Agents JS maxTurns, LangGraph.js recursionLimit, token/timeout/retry options, Zod tool parameters, and tracing imports.
Known limitations:
- TypeScript coverage is intentionally limited to AA001, AA002, AA004, AA007, and AA012;
every other rule is reported per-language as not applicable.
- Dynamic imports, metaprogramming, dispatch beyond the supported one-hop project call graph, and runtime-only behavior remain inconclusive and produce coverage warnings when recognized.
- Static side-effect and schema analysis is conservative and can produce false positives; findings should be reviewed before planning.
- Secret/PII recognition covers a small explicit pattern set, not arbitrary credentials or personal data.
- No instruction-file linting, runtime tracing service, in-process source-changing codemod,
browser repository upload, or account system. Explicit
fix --apply remediation is
delegated to an external runner in an isolated Git worktree.
Configuration
Create .uptocode.yml at the scan root:
exclude:
- generated/
- vendor/
max_file_size: 1048576
Paths in reports are normalized relative to the scan root. Report 2.2 coverage records aggregate and per-language discovered/analyzed files, detected frameworks, evaluated/not-applicable rules, suppressions, warnings, judgment status, and redaction counts.
Standalone HTML reports include keyboard-accessible Approve/Reject decisions and download
a report-bound ReviewManifest 2.0 locally through a Blob. The report performs no network
requests; uptocode plan remains the only manifest consumer and never edits source.
Generate matching HTML and JSON from one scan before reviewing:
uptocode scan . --format html --output report.html --json-output report.json
# Review report.html and download uptocode-review-manifest.json
uptocode plan report.json --manifest uptocode-review-manifest.json --output FIXPLAN.md
MCP registration
Run the stdio server directly with uptocode serve, or register the module with an MCP client:
{
"mcpServers": {
"uptocode": {
"command": "python",
"args": ["-m", "uptocode.mcp_server"],
"cwd": "/absolute/path/to/this/repository"
}
}
}
Local mode exposes repository/file/source/diff audits, strict schema and loop checks, the rule catalog, review, and FIXPLAN generation. It is bounded to the configured canonical workspace root, including symlink defense, and may load only explicitly trusted local rulepacks.
Hosted Streamable HTTP mode registers only submitted-content tools. It has no repository or
filesystem-path tools, never persists submitted code or results, and requires private-beta
bearer keys configured as SHA-256 digests in UPTOCODE_API_KEY_HASHES. Authenticated
requests use an in-memory per-key token bucket (UPTOCODE_RATE_LIMIT_PER_MINUTE, default
30); excess requests return 429 with Retry-After. Logs attribute traffic only to an
eight-character digest prefix. Hosted judgment is rejected unless
UPTOCODE_HOSTED_JUDGMENT=true; the default is false. The double-consent
judgment=true, send_code=true rule still applies when the global gate is enabled.
Local mode resolves its canonical workspace root once when the server is created. Relative
and absolute file, repository, and diff-base paths are all contained against that same root,
including symlink resolution. The pinned MCP SDK compatibility guard fails server startup
with the installed SDK version if the strict-schema private contract changes; the upgrade
procedure is documented in docs/operations.md.
The functional landing and connection generator live in web/. /connect generates local stdio or hosted Codex MCP configuration; bearer keys remain in the user's local environment and are never entered into or transmitted by the page.
UpToCode audits itself
The release gate runs python scripts/compliance.py against UpToCode's production Python
source. The current verified result is zero findings, zero suppressions, and zero unexplained
analysis warnings. Deliberately bad fixtures and synthetic redaction sentinels remain test
evidence and are excluded from that production claim. This dogfood check is part of every
release candidate, alongside the full offline suite and branch-coverage floor.
Built with Codex and GPT-5.6
Codex was the development collaborator throughout UpToCode: it translated the product
specification into vertical slices, implemented and reviewed the scanner, CLI, reports,
MCP servers, site, and tests, fixed clean-clone defects, and ran the Windows/POSIX
submission gates. UpToCode also produces a report-bound FIXPLAN.md with a copy-paste
Codex hand-off, so approval—not automatic source mutation—connects analysis to coding.
The competition-first work order, decisions, progress log, and validation report remain
in the repository as an auditable record of that collaboration.
GPT-5.6 powers the optional judgment tier through OpenAI Responses Structured Outputs.
It evaluates only bounded, redacted evidence for candidate rules that need contextual
judgment. The feature is opt-in twice (--judgment --send-code), uses store=false, and
never replaces or deletes deterministic static findings when a request fails or is
refused. Ordinary tests, the judge test build, and the demo dry-run are entirely offline;
any live paid call requires separate operator approval.
Project evidence
No repository, package, video, deployment, or submission is published by this project workflow without explicit approval.