agentflight agentflight

15 Deterministic Rules

AgentFlight combines 10 static MCP schema ergonomics rules (TS001–TS010) with 5 runtime circuit-breaker guards (CB001–CB005). Filter by group on the left rail or search below.

Rule Explorer

Rule ID Group Check Description & Threshold Severity / Penalty

Automated Schema Hardening

Running agentflight fix tools.json automatically resolves common ergonomics findings by adding missing property descriptions, adding explicit confirm guards on destructive functions, and injecting missing required fields. 

agentflight fix tools.json --output hardened_tools.json

Ergonomics Scoring & CI Grades

  • Score Calculation: Every MCP suite begins at 100. Each unique violation subtracts its deterministic penalty weight (-5 to -12) down to a floor of 0.
  • Letter Grades: 90–100 = A+ / A, 80–89 = B, 70–79 = C, 60–69 = D, <60 = F.
  • Strict CI Gate: agentflight ci --schema-file tools.json --fail-on-smell exits non-zero whenever any violation remains.

Technical FAQ

How does TS008 detect overlapping MCP tool names?

ToolSmellLinter computes normalized bag-of-words cosine similarity across tool names and descriptions in the same suite. Any pair meeting or exceeding 0.72 similarity is flagged so models do not route randomly between lookalike tools.

Does redaction affect SHA-256 chain verification on replay?

No. FlightRecorder scrubs secrets prior to computing the canonical-JSON SHA-256 step digest, and CassetteReplayer checks both raw and redacted input digests during key-rotated CI runs.

Can I enforce rules TS001–TS010 on remote MCP servers like OpenAny?

Yes. Run agentflight openany lint-tools --json or call agentflight_lint_tools directly over Streamable HTTP via https://agentflight-tau.vercel.app/mcp.