Skip to content

ADR 0006: Slice Workflow Skill CLI and Stack Rubric

Status: Accepted Date: 2026-04-29 Deciders: Alex Furrier Generated from: GitHub issues #6 and #7 Plan: .ai/plans/2026-04-29-154802-followup-contract-stack-rubric/

Context

The deterministic slice work left two follow-up concerns.

First, the slice workflow had grown into several repo-local scripts. The planning, prompt rendering, and handoff checks were still exposed through stable mise tasks, but their implementation lived partly in scripts/. That split made the workflow feel less like a portable skill and more like scaffold-only plumbing.

Second, the Rust stack made the project's implicit stack quality bar visible. Python, Go, and Rust all cover a broad happy path, but the project did not yet state what future stacks must provide before they count as supported.

Decision

The mise tasks remain the stable agent-facing interface. Their implementation now delegates to the skill-local CLI at templates/.agent/skills/slice-workflow/cli in scaffold source and .agent/skills/slice-workflow/cli in generated repositories.

The CLI uses a tiny uv project and the slice_workflow_cli package:

  • plan.py creates plan directories from templates.
  • workflow.py renders planner, implementer, and reviewer prompts.
  • checks.py runs the plan/spec/evidence/review/sync contract checks.
  • The contract package contains plan, git, markdown, artifact, and docs helpers.

The project now also has docs/reference/stacks/acceptance-rubric.md, which defines the required future-stack bar and reviewer checklist.

Consequences

Positive:

  • Workflow code moves with the slice-workflow skill into generated repos.
  • The deterministic command surface remains mise run ....
  • Future plan-contract changes should land in smaller files with clearer names.
  • Future stack PRs have an explicit review checklist instead of relying on the Rust stack as an implicit example.

Negative / Trade-offs:

  • The skill now carries a small CLI project and lockfile.
  • The task wrappers duplicate a small launcher to keep .mise/tasks/* directly executable and provider-neutral.

Alternatives Considered

Alternative Reason not chosen
Keep one plan_contract.py file Too much unrelated logic in one review unit
Keep repo-local modules under scripts/ It preserves old imports but conflicts with the skill-local mini CLI pattern
Make the new CLI user-facing More disruption than needed; task names are already the stable interface
Add a third-party YAML or CLI library Generated repos should keep contract checks dependency-light