`feature_audit`¶

Verifies that features documented in Feature Inventory still function, via the machine-readable registry at references/feature-registry.csv.

Procedure: references/feature-audit.procedure.md · Driver: scripts/run_feature_audit.sh · Workflow: .github/workflows/feature-audit.yml (daily, 04:37 UTC)

Why the registry is derived, not authored¶

The inventory body is already machine-parseable — each feature is <n>. **Name** under a ## Category. The registry is generated from it, so the prose stays the single authored source. A second hand-maintained list would rot exactly as the summary table did: it read 125 features across 12 areas against a body of 146 across 14, omitting two whole capability areas, because the only check compared the total to its own column.

Three rules that keep it honest¶

proven is self-enforcing — a row needs both a proof_test and a distinct negative_control; load_registry downgrades anything else. A proof with no negative control shows a test runs, not that it can fail.
A tier that executes zero proofs is a finding — otherwise an all-UNPROVEN registry passes every structural check and exits clean.
Unreachable is a value, not an exception — agentteams/research is degrade-don't-raise, so a live probe asserting only "no error" passes on an empty response. Live proofs assert positive content; an outage reports UNREACHABLE and never gates.

Outcomes¶

Code	Meaning
0	clean — every executed proof passed; registry matches the inventory
1	findings — a proof failed, the registry drifted, or a tier proved nothing
2	harness broken — malformed/missing registry, engine death, unclassified exit

Indeterminate is not a pass; code 2 outranks code 1.

Usage¶

python -m agentteams.feature_audit --tiers unit          # structural + unit proofs
python -m agentteams.feature_audit --structural-only     # parity and resolution only
bash scripts/run_feature_audit.sh                        # driver, with outcome classification
FEATURE_REGISTRY=/path/to/fixture.csv python -m agentteams.feature_audit   # test override

Current coverage¶

6 of 151 proven. UNPROVEN means nothing specifically demonstrates the feature — not that it is broken. See the procedure doc for how to close a row correctly.

feature_audit¶