Problem
The product is the markdown in commands/vine/ and agents/ — but those get only structural linting (/trellis, .vine/scripts/trellis-check.sh). The only behavioral test harness, .vine/scripts/run-tests.sh, covers the shell hooks (journal-check, trellis-gate, main-guard) and nothing else. There is no scenario/golden test that the actual workflows behave: that vine-coder resumes correctly from a partial NAVIGATION.md, that vine-reviewer flags a scope breach, that verify catches a knowledge-record-vs-live-code mismatch, that the gitignored/tracked artifact rules hold.
On top of that, nothing runs on PR. The only GitHub Action is .github/workflows/publish.yml (manual workflow_dispatch). So even the shell tests that do exist aren't gating anything — dogfooding is the entire regression suite. For a framework whose value proposition is reliability under delegation, that's the weak link.
Suggestion
Two parts, independently shippable:
- CI gate (cheap, do first). A GitHub Action on PR that runs
.vine/scripts/run-tests.sh and .vine/scripts/trellis-check.sh. This alone makes the existing teeth enforceable.
- Behavioral/scenario layer for the product. Fixture repos under a test dir + expected-behavior assertions for the command/agent workflows. Start with the highest-stakes, most-mechanical claims (coder resume-from-journal, reviewer scope-breach detection, verify knowledge/code mismatch). Golden-transcript style where a full run is too heavy.
Keep it contributor-only (never shipped by create-vine), same as the existing scripts.
Problem
The product is the markdown in
commands/vine/andagents/— but those get only structural linting (/trellis,.vine/scripts/trellis-check.sh). The only behavioral test harness,.vine/scripts/run-tests.sh, covers the shell hooks (journal-check, trellis-gate, main-guard) and nothing else. There is no scenario/golden test that the actual workflows behave: that vine-coder resumes correctly from a partial NAVIGATION.md, that vine-reviewer flags a scope breach, that verify catches a knowledge-record-vs-live-code mismatch, that the gitignored/tracked artifact rules hold.On top of that, nothing runs on PR. The only GitHub Action is
.github/workflows/publish.yml(manualworkflow_dispatch). So even the shell tests that do exist aren't gating anything — dogfooding is the entire regression suite. For a framework whose value proposition is reliability under delegation, that's the weak link.Suggestion
Two parts, independently shippable:
.vine/scripts/run-tests.shand.vine/scripts/trellis-check.sh. This alone makes the existing teeth enforceable.Keep it contributor-only (never shipped by create-vine), same as the existing scripts.