Documentation
¶
Overview ¶
Package classify infers a tool's reversibility class from its name, description, and schema - the single biggest onboarding barrier this system had. Hand-classifying an entire tool catalog into A/B/C before you can run anything is real work nobody wants to do on day one, and it was the difference between "docker run and see something" and "read the docs for an hour first."
Two honest limits, stated up front because they shape the whole design:
This is a HEURISTIC, not an authority. It reads names, descriptions, and schemas - it cannot know that `archive_record` in your system actually hard-deletes. Every result carries the signals that fired and a confidence, and anything it calls Class C (or can't read confidently) is flagged for human review rather than silently trusted.
It fails CLOSED. An unrecognized verb doesn't become "probably safe Class B" - it becomes "unknown, review this," because the expensive mistake here is under-classifying something irreversible, not over-classifying something harmless. A false Class C costs someone an approval click; a false Class B costs them a sent email they can't recall.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type Input ¶
type Input struct {
Name string
Description string
Method string // optional: GET/POST/PATCH/DELETE, when known
ArgNames []string // optional: property names from the tool's schema
}
Input is what the classifier can actually see about a tool. MCP gives name/description/schema and no HTTP verb; a REST-style catalog gives a method too. Both paths feed the same engine, the method is just an extra signal when it happens to be there.
type Report ¶
type Report struct {
Source string `json:"source"`
TotalTools int `json:"total_tools"`
ByClass map[string]int `json:"by_class"`
Irreversible []string `json:"irreversible_tools,omitempty"`
NeedsReview []string `json:"needs_review_tools,omitempty"`
Unclassified []string `json:"unclassified_tools,omitempty"`
UndoableInName []string `json:"compensable_but_no_pre_image,omitempty"`
Results []Result `json:"results"`
}
Report is the aggregate view over a whole catalog - the thing that actually answers "how much damage can this agent do." The individual classifications matter less than these three numbers: how much of this catalog is irreversible, how much of it nobody has reviewed, and how much of it claims to be undoable without any way to actually undo it.
type Result ¶
type Result struct {
Tool string `json:"tool"`
Class registry.ReversibilityClass `json:"class,omitempty"`
Confidence float64 `json:"confidence"`
Rationale string `json:"rationale"`
Signals []Signal `json:"signals,omitempty"`
// NeedsReview is set whenever a human must look before this
// classification is trusted in production: anything irreversible,
// anything ambiguous, and anything unrecognized.
NeedsReview bool `json:"needs_review"`
// Unclassified means no rule matched with enough confidence to name a
// class at all. Deliberately distinct from "Class C" - "I don't know"
// and "I know this is dangerous" are different messages to a human.
Unclassified bool `json:"unclassified,omitempty"`
}
Source Files