Problem

waza suggest already bootstraps test cases from a SKILL.md, but in practice authors:

1. Get a small fixed batch and have to re-run for more coverage.
2. Can't ask for cases targeting a specific behavior (negative triggers, edge fixtures, "DO NOT USE FOR" cases).
3. Have no signal for which generated cases are high vs. low confidence.
4. Risk overwriting hand-curated eval.yaml entries on regeneration.

So skills still ship with 3–5 hand-written cases and discover edge behavior in production.

Proposal

Enhance the existing waza suggest (and its internal/suggest/ pipeline) rather than introducing a new command:

• --count N — how many cases to propose.
• --focus <category> — triggers | negative-triggers | edge-fixtures | do-not-use-for | parameters.
• Emit per-case confidence and rationale (which SKILL.md span it came from) so authors can triage.
• --dry-run prints proposals; --apply merges into eval.yaml. Never overwrite an existing task id without --force.
• Generated tasks must validate against the existing eval schema before they're written.

Why this matters for agentic-first

Coverage of agentic intent (when to invoke, when to refuse, which tools, which parameters) is the long tail of skill quality. The current suggest is great for a first pass; this issue is about giving authors a steerable second/third pass.

Acceptance criteria

• waza suggest accepts --count, --focus, --dry-run, --apply, --force.
• Each proposed case carries confidence (0–1) and a rationale field referencing the SKILL.md span.
• --apply refuses to overwrite an existing task id unless --force is set; surface a clear diff.
• Generated cases pass internal/validation/schema before being written.
• Tests cover focus categories, overwrite safety, and schema validation.
• Docs updated in site/ with a worked example.

Non-goals (filed separately)

• Generating tests from captured traces — depends on feat: Trace replay and deterministic snapshot for agent runs #367 snapshot work.
• Coverage analysis against SKILL.md requirements — see feat: Natural-language requirements → executable evals (spec-to-test) #361.

Related

• Existing module: internal/suggest/
• Roadmap: feat: refactor waza new to use shared FileWriter #58 #66
• Coverage verification: feat: Natural-language requirements → executable evals (spec-to-test) #361
• Trace-derived generation: depends on feat: Trace replay and deterministic snapshot for agent runs #367