testbed-generator

Testbed generator — mine + validate + enrich + compile + generate

Drives the deterministic scai testbed mine, validate, enrich, compile, and generate phases for the whole workload. The data lives in the CLI's opaque state; this skill only sequences subcommands and surfaces their JSON. It replaces production data, not source access — a source connection is still required downstream to capture the baseline.

The core phases are independent and resumable: mine (init → list-unsolved → whole-workload branch drill-down) inventories the constraint/branch coverage; validate reads the mined state and reports readiness (fk gaps / type conflicts / unsatisfied constraints) — ready: false is a report, not a failure; enrich is the reasoning + propose/validate loop that turns unsolved constraints into enrichment JSON (see the Orchestrate phase below); compile materializes the data-coupling clusters + coordination spec that data generation consumes; generate (readiness-gated) writes the CSV pool + manifest deliverable. Each phase is deterministic — re-running it is safe and cheap.

On Entry

Tell the user:

"Running the testbed mine → validate → enrich → compile → generate phases: I'll inventory the branch/constraint coverage your converted workload needs and show the categorized unsolved-constraint view, report readiness (fk gaps / type conflicts / unsatisfied constraints), run the enrichment loop that turns the unsolved constraints into enrichment JSON (assemble → propose → validate, with a bounded retry/iterate budget), compile the data-coupling / coordination spec and show its cluster summary, then — readiness permitting — generate the CSV pool + manifest and show the row counts. This reads the conversion artifacts and the testbed state; the enrich phase mutates that state (through propose-enrichments) and only the generate phase writes data files."

Scope

• In scope: init → list-unsolved (mine) → surface the categorized view; validate → surface the readiness view; compile → surface the cluster summary; the two-pass ENRICH reasoning (structural then value prompts) that turns unsolved constraints into enrichment JSON (§enrich below); generate (readiness-gated) → surface row counts + CSV/manifest deliverable; resume each phase; classify failures.
• In scope (orchestrate): assemble the per-type fragments into the 8-array envelope, drive propose-enrichments/validate with a bounded retry/iterate budget, emit observability counts (SNOW-3782435).
• Out of scope: generate's CSV internals; the per-object test-YAML bridge.
• Phase-scoped runs: each driver subcommand (mine, validate, enrich, compile, generate) can be run and stopped independently — e.g. run validate to inspect readiness without generating data.