Skip to content

SACM agent operating plan

Agent workflow

Use specialized agents rather than asking one agent to do everything. The goal is to reduce drift, force traceability, and prevent implementation from outrunning the standard.

sacm-implementation-lead
  -> sacm-researcher        (specification analysis)
  -> sacm-researcher        (metamodel cartography)
  -> sacm-library-architect
  -> sacm-xmi-test-engineer
  -> sacm-implementer       (library scope)
  -> sacm-conformance-verifier
  -> sacm-implementer       (adapter scope)
  -> sacm-researcher        (interoperability research, when needed)

Slice rules

A slice should be small enough to complete through tests and verification.

Good slices:

  • Minimal editable document: create AssuranceCasePackage, ArgumentPackage, Claim, preview/apply deletion, save/load, semantic round trip.
  • Base LangString and multilingual strings.
  • IDs and reference resolution.
  • One asserted relationship family.
  • One artifact provenance cluster.
  • One adapter projection from library Claim to Assurance Forge Goal display node.

Bad slices:

  • Implement all of SACM argumentation at once.
  • Rewrite the entire parser before tests exist.
  • Add UI features while the library edit/model/XMI behavior is untested.
  • Import a third-party example and call success compliance without matrix coverage.

Required artifacts per slice

  • Slice brief under docs/sacm/slices/ or in the issue/PR description.
  • Conformance matrix entries.
  • Metamodel inventory for the classes/associations in scope.
  • Positive and negative fixtures.
  • Library-level tests.
  • Adapter tests if Assurance Forge behavior changes.
  • Verification report.

Suggested branch naming

sacm23/<area>-<short-slice-name>

Examples:

sacm23/editable-minimal-package-claim
sacm23/base-lang-string
sacm23/argument-claim
sacm23/af-adapter-source-of-truth

Review checklist

Before merging a slice:

  • Does the library remain independent of Assurance Forge?
  • Is the SACM library model the source of truth?
  • Do tests fail without implementation?
  • Do tests verify edit behavior and semantics, not just parser success?
  • Are unsupported standard features visible in the matrix?
  • Are compatibility behaviors separated from strict SACM 2.3 behavior?
  • Are destructive edits previewable and explicit?
  • Is strict save free of Assurance Forge layout metadata?
  • Are external fixtures license-compatible or minimized?

Coordination prompts

Use the prompts in docs/sacm/prompts/.

The most important one is sacm-library-master-prompt.md. It gives the lead agent the library-first constraints and tells it not to write production code before tests exist.