Agent Preflight

• 

  One command installs Agent Preflight. Zero runtime dependencies.

• 

  Compiled in Rust — single static binary, every platform.

• 

  Scan silently parses every tool decorator without executing code.

• 

  Review surfaces every pending capability rule that needs approval.

• 

  Approve signs the capability contract into the repository.

• 

  Contract locked. The approved rules are now version-controlled.

• 

  Source drifted from contract. CI exits 1. Pipeline blocked.

• 

  The full scan → approve → verify lifecycle in one terminal.

• 

  One line added. needs_approval=True. The fix is in source.

Inspiration

No agent framework proves, before deployment, that every high-stakes tool has the safety control its owner intended. Agents ship to production with @function_tool and no human-approval gate. The gap isn't tooling — it's evidence.

What it does

A static-analysis CLI that scans agent codebases and verifies every tool decorator has the approval control its owner approved — without ever executing code.

Scan → Approve → Verify. Three commands. If source code drifts from the approved contract, CI fails. Uncertainty never becomes a green result. Supports OpenAI Agents SDK, Google ADK, and Anthropic Claude.

How we built it

• Rust — zero-dependency, single static binary, cross-platform.
• Tree-sitter — parses Python ASTs directly from source text. No imports, no execution.
• YAML contracts — human-readable, diffable, version-controlled.
• GitHub Actions — 15 test suites across Windows, macOS, and Linux.
• Codex (GPT-5.6) — architecture partner for designing the scan/approve/verify lifecycle.

Challenges we ran into

Making the analyzer deliberately conservative. If it can't prove a control exists, it fails — never guesses. Cross-SDK support was equally hard: OpenAI, Google ADK, and Anthropic each express approval controls in completely different idioms.

Accomplishments that we're proud of

Zero false positives. Sub-second scans. One binary, every platform. 15 CI suites on 3 operating systems — we practiced what we preach.

What we learned

The hardest part of AI safety tooling isn't detection — it's deciding what to do when you're uncertain. We chose to fail loudly. That one decision shaped the entire architecture.

What's next for Agent Preflight

MCP server support, pre-commit hooks, a VS Code extension, and TypeScript/Go framework adapters.

Built With

• actions
• adk
• agents
• codex
• github
• google
• openai
• python
• rust
• sdk
• tree-sitter
• yaml