MCP Error Suggestions

[withtwoemms]

As an AI agent using ucon-MCP tools,
I want error responses that localize the problem, name the fix, and distinguish high-confidence corrections from exploratory hints,
So that I can self-correct in a single retry without human intervention.

---

Acceptance Criteria

Response model

• ConversionError model with fields: error, error_type, parameter, got, expected, likely_fix, hints
• error_type is one of: "unknown_unit", "dimension_mismatch", "no_conversion_path"

Unknown unit errors

• convert() catches UnknownUnitError for both from_unit and to_unit
• check_dimensions() catches UnknownUnitError for both inputs
• list_units() catches invalid dimension filter
• parameter field identifies which input was bad
• Single high-confidence fuzzy match (≥ 0.7) populates likely_fix
• Multiple or lower-confidence matches populate hints only
• Unrecognizable inputs produce no spurious matches — only generic hints

Dimension mismatch errors

• convert() catches DimensionMismatch
• Error message uses readable dimension names (never Vector(...))
• got and expected fields carry dimension names
• Compatible units listed from graph edges (not theoretical dimension matches)

No conversion path errors

• convert() catches ConversionNotFound
• Pseudo-dimension isolation is explained (not just stated)
• Hints include alternative computation approach (e.g., "compute angle/2π explicitly")
• Missing-edge case suggests intermediate conversion path when dimensions match

Structural

• Suggestion logic lives in ucon/mcp/suggestions.py, not in server.py
• No error response error field contains the string "Vector("
• All error responses are valid ConversionError instances (no raw exceptions)

[withtwoemms]

closing as these positive eval results (attached) were against the ucon.mcp:convert tool. the #152 changes were merged and published as ucon v0.7.0: