subagent-driven workflow bypasses `mship test`; finish loses test-evidence trail

[atomikpanda]

Observed

Implementing #104 via subagent-driven-development, each subagent ran .venv/bin/pytest directly to verify its task. The full feature ran cleanly (1235 tests passing) but mship finish warned:

Test-evidence warnings: Tests not run in: mothership

Because no subagent invoked mship test, the .mothership/test-runs/<task>/ directory stayed empty, the test-iteration counter never incremented, the per-repo evidence chain didn't fill in, and finish had no signal that the tests had actually passed.

Why this matters

mship test is the substrate's evidence-recording entry point. When an agent workflow bypasses it (running raw pytest instead), the substrate value disappears: no iteration tagging, no "new failure / fix / regression" diff between runs, no audit trail for finish. The framing "mship is the brain, go-task is the muscle" depends on the muscle being mship's muscle — but the actual implementation tasks were calling pytest, not mship test.

This is a workflow-adoption gap. If mship test isn't the path of least resistance for agents, they'll skip it — exactly what happened here.

Why it happened on #104

• The plan (docs/superpowers/plans/2026-05-14-task-dependency-graph.md) wrote each TDD verification step as pytest tests/... rather than mship test. Agents followed the plan literally.
• The implementer-prompt template in subagent-driven-development doesn't mention mship test.
• The vendored TDD skill doesn't reference mship test.

Suggested direction

A few overlapping options:

1. Plans generated in a mship workspace default to mship test in TDD step commands. Today writing-plans emits pytest tests/...; in a mship workspace, emit mship test (or mship test -- <pytest-args>). Most impactful change.
2. Implementer-prompt template (subagent-driven-development/implementer-prompt.md) instructs the subagent to prefer mship test when a mothership.yaml is present.
3. mship test accepts a pytest passthrough cleanly so the substitution is trivial in plans (mship test -- -k myfilter -v). If this works today, document it.
4. mship finish could detect commit-journal entries with test-pass signals ("all passing", "N passed") and treat them as a softer form of evidence, with a --source=journal annotation. Narrower / opportunistic — but doesn't require workflow changes.

The bigger lever is (1)+(2). (4) is the auto-detect fallback for sessions that still drop through.

Acceptance

• writing-plans skill emits mship test in TDD steps when generating plans in a mship workspace.
• subagent-driven-development's implementer-prompt mentions mship test preference.
• mship test argument-passthrough syntax is documented (or fixed if it's awkward today).

Related

• Discovered during First-class task dependency graph (blocks / depends-on) — design issue, marinating #104.
• src/mship/core/test_evidence.py (where the warning fires).
• The vendored skills in src/mship/skills/writing-plans/ and src/mship/skills/subagent-driven-development/.