Conformance describes individual capabilities and the evidence for each one. It is not a single certification of a client, model, deployment, or platform. Use the runtime contract, validation record, and operating manual together. ACK remains read, not task completion (REQ-30). The server cannot determine which reader issued an ACK.
For each run record source revision, harness/client version, OS/architecture, execution context, isolated transport/origin, scenario, expected and observed results, and reproducible command or sanitized trace reference. Include poll interval, order, observer and deadline where the scenario uses them. Never include keys, claim tokens, private configuration, or unrelated conversations.
Use PASS, FAIL, NOT_TESTED, or NOT_APPLICABLE per item. Missing facilities
are NOT_TESTED; NOT_APPLICABLE needs a reason. Evidence categories are:
| Evidence | What it establishes |
|---|---|
| static/config | Source-derived call paths, selected policy or configuration; no execution claim |
| fixture | Executed synthetic requests/state and their asserted behavior |
| subprocess/protocol | A real executable or MCP protocol exchange in the recorded environment |
| actual-context | Content was presented in an actual agent context and followed by its explicit read decision; no proof of understanding |
| scheduled-peer | Actual scheduled run and independent peer observation in the recorded deployment |
Isolated fixture transports reject origins outside their synthetic test entry point. In-process HTTP fixtures and temporary databases do not establish real scheduler behavior, native sandbox enforcement, or cross-host compatibility. A durable-ingestion receipt must be labeled separately from explicit reading in the model context; neither label changes service ACK semantics or silently reclassifies an existing deployment's authorization.
| Profile | Required observations |
|---|---|
| transport-invariants | Key and exact source IP; exclusive claims and deadlines; each ACK outcome; independent fan-out; coupled retention and current-comparison deduplication |
| model-explicit-read | Actual claimed body enters the agent context before a subsequent explicit ACK; fixtures only establish call ordering |
| legacy-durable-ingestion observation | Storage before ACK and pending-work wake behavior, with provenance; not automatically model-explicit-read |
| durable-work | Authorized retained work, source deduplication, recovery after ACK/crash, unchanged blocked work not repeatedly runnable |
| outbox-recovery | Immutable per-target intent and outcome, unknown sends, bounded fresh retained observations, restart recovery without blind re-POST |
| harness-lifecycle | Actual tools, quiet runs, overlap rejection before effects, whole-turn deadline and owned-child cleanup; only the verified harness/version/OS combination |
Apply these profiles only to delivered capabilities. A stateless transport need not implement a local work store. Message lookup, own-read GET, epoch/cache validation, reference local components, and persistent threads require their own fixtures before their capabilities can be advertised.
The runnable agent-context, route-reply and ACK-read server fixtures cover computed agent context while preserving original message representations and administrator history; reply routing and permissions including commit-time changes; strict input and multipart rejection before file staging; legacy attachments; unchanged ACK fencing with nullable account read; and persistence remaining schema 1. Adapter tests exercise open JSON response parsing and the separate single-object reply result contract. These tests do not certify arbitrary third-party closed response schemas.
Run the focused fixtures with:
go test ./internal/server -run '^Test(AgentContext|RouteReply|AckRead|LegacyRepresentation)' -count=1
go test ./internal/client ./internal/cli ./internal/mcp -count=1TestLegacyWorkflowCallCounts freezes selected old shared-client workflows.
TestRouteReplyCallCounts exercises the corresponding reply path:
go test ./internal/client -run '^Test(LegacyWorkflow|RouteReply)CallCounts$' -count=1 -vThese count scripted HTTP calls against synthetic transports. Counts exclude bootstrap, authentication discovery, model turns and scheduler work. They are not universal minima or latency/efficiency measurements. A known private recipient can already use one raw direct send. No production helper is modified or polled, and no actual new/old collaboration benefit is inferred.
This section specifies a future diagnostic operation; no doctor command is
implemented or advertised. It must report only nonsecret configuration and
versions, authenticated identity, advertised features, statically inspectable
permissions and any authorized local-state binding. It must distinguish static
permissions from actual enforcement and verified binding from a mere HTTP 200.
Doctor must never send, claim, ACK, join, start models, change schedules, or
write task/message state. Normal authenticated GETs can still update service
last_access and storage; read-only means no messaging/task business writes.
File/database write capability checks belong in explicitly isolated fixtures,
not the default diagnostic path. An older server's missing features are
unsupported rather than a failed overall diagnosis. Credentials and token
material must be redacted even on failures.