Behavior Inventory

Defines what a Behavior record contains. Does not define how to extract behaviors — that is agents/behavior-extractor.md's job, and identity resolution across runs is scripts/resolve-identity.py's job (REQ-31a), never an agent's.

Schema

schemas/behavior.schema.json. One JSON object per behavior; inventory/behaviors.json is a JSON array of these.

Required fields

id (BHV-NNNN), summary (≥10 characters, one atomic fact, not a module summary), source_services (≥1), evidence (≥1), confidence.

Content obligations

• kind classifies the behavior: validation, calculation, routing, state-transition, error-mapping, side-effect, emission, or other.
• failure_path: true for anything describing failure, timeout, or partial-completion — these are under-represented by default (01-spec.md DR1) and must be marked, not left to inference.
• confidence: "unknown" requires a reason. Do not default to "medium" to avoid writing a reason.
• status defaults to "confirmed". "unconfirmed" is set only by scripts/resolve-identity.py when a previously-seen Behavior is absent from the latest run — never set by hand.

Behavior-ID citation

A Behavior is the thing everything else cites, not the other way around. Its own spec_clauses and tests fields, once populated downstream, are how coverage becomes computable (REQ-12): % of Behavior IDs with a passing characterization test.

Unknowns

An undeterminable field is "confidence": "unknown" with a reason — never omitted, never guessed (REQ-09, Constitution C10).

Minimal worked example

{
  "id": "BHV-0042",
  "summary": "Rejects an order when requested quantity is zero or negative",
  "source_services": ["orders-svc"],
  "interfaces": ["IFC-SYNC-0001"],
  "kind": "validation",
  "failure_path": true,
  "evidence": [ { "source": "code", "locator": "orders/validate.py:42" } ],
  "confidence": "high"
}