RFC0015 DISPATCH SEQUENCE PROPOSAL
Versioned draft snapshot
RFCs, reference catalogs, and evidence reflect Spec 3b2e3e7a114fb826b1445d39a48756b8fff629cd. Counts and pending/status notes belong to that snapshot. Read the dated project update for later progress.
Source repositories are private during staging. Links marked “repository access required” need access to the linked repository.
This is a trace and gap record from the pinned draft specification. It does not establish proof or implementation conformance.
Status: owner decision pending. Owners: Runtime, AI.
Conflicting requirements
- §3.4 universally requires a successful health check before dispatch, while
§7 permits bindings without
healthCheckand Lean b17 makes probes conditional. - §7 says a missing probe proceeds to Step 4, skipping Step 3 prompt construction and its authorization checks.
- §5 says every failure halts the sequence, while Step 4 restarts at Step 2 after a transient dispatch failure.
These contradictions are reflected in the authored RDF claim summaries. They must be resolved before an executable agent runtime can claim conformance.
Recommended first-release contract
Keep health checks optional, as the grammar, §7, and Lean b17 already specify. If configured, every inference attempt must be preceded by a successful probe. An unhealthy probe ends the invocation with M01501 and is not retried by the inference retry policy. An omitted probe emits M01513 and proceeds to Step 3, never directly to dispatch.
Prompt construction and authorization are mandatory for every inference attempt. A transient inference failure may retry from Step 2 according to the configured retry policy; Step 3 then reconstructs and reauthorizes the request against the current effective read authority. No unauthorized context may be serialized into the outbound payload. Previously staged outputs are not committed between attempts. Other failed steps terminate the invocation.
This proposal does not decide whether the read-scope authoring syntax is a
separate readScope block or derived from permissions: the companion QA and
§13 disagree on that ownership. Whichever scope contract is selected, unknown
or denied effective read authority excludes data before network dispatch.
Alternative: require a configured probe for every binding. That introduces a new mandatory field and changes M01513 from an advisory to a rejection; it also requires a health endpoint for every supported transport.
Concrete source edits after approval
- Qualify §3.4 with “when a healthCheck is configured.”
- Change §7’s absent-probe destination from Step 4 to Step 3.
- State the transient inference retry exception in §5’s opening rule and require prompt reauthorization for each retry.
- Align the two health claim summaries, ordered-execution and retry claims, and corresponding Lean statements; preserve existing claim IRIs.
- Regenerate reference and RDF artifacts, refresh reviews, and check drift.
Runtime acceptance
| Scenario | Required outcome |
|---|---|
| No healthCheck; authorized context | M01513; construct and authorize prompt; dispatch |
| No healthCheck; unauthorized context | M01500; no inference request containing denied data |
| Configured probe fails | M01501; no inference call or inference retry |
| Transient inference failure then retry | Probe if configured, rebuild and authorize prompt, then dispatch |
| Authority revoked between attempts | Exclude denied data or fail with M01500 before redispatch |
| Response or write validation fails | No inference retry under this exception; no partial graph commit |
Verify these through the actual runtime with a request-capturing inference endpoint and durable graph reads after failures. A standalone sequence helper or mocked callback order does not prove prompt confidentiality or rollback.