Whole-specification RFC 2119 conformance audit of this SDK against the stable ACP v1 specification. This directory is outside the Maven reactor and outside the shipped artifacts; nothing in acp-* depends on it.

Results: the audit report, with one page per run under results/: run 001 (the code released as 0.80.0) and run 002 (with the fixes released in 0.81.0).

The roster is labelled derived ACP v1 testing requirements. It is not a replacement specification: every imported clause is the original wording, located verbatim in the retained bytes, with its original keyword strength (only the RFC 2119 synonyms REQUIRED→MUST and OPTIONAL→MAY are normalized, and the original is kept beside the normalized form).

python3 -I protocol-conformance/scripts/import-spec.py check   # hashes, manifest, coverage, rendered files current
python3 -I protocol-conformance/scripts/import-spec.py write   # re-render after editing the manifest

check fails if any retained byte changed, any clause is not found verbatim in its declared line range, any keyword occurrence in a selected page is unclaimed, any schema.mdx keyword line is not mirrored from a schema.json description, or the rendered files are stale. Re-retaining from a local clone of the protocol repository:

python3 -I protocol-conformance/scripts/import-spec.py retain \
  --protocol-repo /path/to/agent-client-protocol --jsonrpc-html /path/to/specification.html

• Scope: the 21 pages in the docs/docs.json navigation group "v1" (tag "Latest"), the stable schema/v1/schema.json, and the JSON-RPC 2.0 clauses ACP incorporates by reference. Draft pages, v2, the unstable schema and RFDs are excluded (reasons in spec/manifest/scope.toml). The stable transports page still calls Streamable HTTP a draft proposal, so this SDK's HTTP and WebSocket transports are assessed only under the stable custom transport clauses.
• Keyword occurrences: every MUST, MUST NOT, SHOULD, SHOULD NOT, MAY, REQUIRED, OPTIONAL, SHALL, RECOMMENDED occurrence in the prose pages (188) is either inside an imported clause, a declared duplicate of a canonical requirement, or an explicit non-normative disposition.
• Schema: the 146 keyword occurrences in schema.json descriptions are dispositioned by JSON Pointer; schema.mdx (181 occurrences) is proven line by line to be a mirror of those descriptions. Structural constraints (required members, discriminated unions, enum values) are imported as labelled derivations whose member lists are generated from the schema bytes.
• Prose: lowercase "must"/"should" sentences were reviewed one by one; five are imported at a declared strength with the reason recorded, the rest are marked covered or non-normative.
• Roles and responsibility: each requirement declares agent, client or both, and whether the SDK alone can satisfy it (sdk), the SDK and the embedding application together (shared), or the application alone, with the SDK assessed on what it exposes (application). Capability- and lifecycle-conditional clauses carry an explicit applicability condition; only those may be judged not applicable.

protocol-conformance/scripts/prepare.sh                      # roster check, dependencies, identity, offline tests
protocol-conformance/scripts/run.sh --replay <run-id>        # deterministic replay; refused (exit 2) without a real recording
protocol-conformance/scripts/run.sh --replay <run-id> --assert-satisfied
protocol-conformance/scripts/run.sh --live --capture <run-id>   # one explicitly approved investigative run
./mvnw -f protocol-conformance/pom.xml -o test               # offline harness tests over synthetic answers
./mvnw -f protocol-conformance/pom.xml -o test -Dacp.conformance.run=<run-id>                     # replay regression
./mvnw -f protocol-conformance/pom.xml -o test -Dacp.conformance.run=<run-id> -Dacp.conformance.assert=true  # acceptance

Response semantics the offline tests pin: every requirement reaches one backend execution unchanged; CANNOT_DETERMINE is an abstention; NOT_APPLICABLE is accepted only on a requirement with a declared condition; malformed, duplicate, missing or undeclared answers are instrument errors, never SDK findings; a refused replay or a crashed backend is retained as a failed run.

Retention under the producer's bound. The producer bounds one V6 document at 1 MiB and refuses rather than truncates; a 240-item full-text verdict is about 1.7 MB. RetainedVerdicts therefore writes the verdict as V6 parts under verdict-v6-parts/: root.json carries the collective judgment and each part-NN.json is a complete V6 roster verdict over a contiguous slice of the items, every file under the bound, with manifest.json recording slices and digests. read reopens every part through the producer's codec and reassembles a Verdict that equals the one voted (asserted by the offline tests). Small verdicts stay one verdict-v6.json.

Two live runs have been made (2026-10-08); see the report. Their raw recordings are kept locally and not committed: the session transcripts carry the environment of the machine the judge ran on. The per-item results are published under results/. runs/ therefore holds no committed recording, and a fresh checkout can replay only a run made on that machine. The roster has not been reviewed by the protocol authors and makes no claim of official ACP certification.