Run Goldens

run-goldens — run the golden LLM tests against a real model, by choice

Golden tests (@GoldenLlmTest-style) exercise an LLM surface — routing, strict-JSON extraction, synthesis grounding — against a real local model, asserting structure, not exact text. They are opt-in (a GOLDEN_LLM gate), skipped in CI, and slow (a real model call per assertion), so you run the ones that touch what you changed — not the whole suite reflexively.

Procedure

1. Discover the goldens. Find the repo's golden classes — grep for the golden marker (this repo family uses @GoldenLlmTest): grep -rl "@GoldenLlmTest" --include=*.java (adapt the marker/glob to the repo's language). Group the hits by module (the Maven -pl path / package root) so each has a runnable address.
2. Present the inventory + get a scope. Show the discovered classes grouped by module and ask which to run: one class, several (comma-separated), a whole module, or all. You can narrow to a single method with Class#method. Prefer the smallest scope that covers the surface just changed.
3. Check prerequisites. The golden lane needs a real model. Use the repo's golden runner (scripts/golden.sh in this repo family) — it auto-starts the local inference engine + an LLM gateway pointed at it, sets GOLDEN_LLM=true, and handles model quirks (e.g. suppressing "thinking" so a call is seconds not minutes). The required model must already be pulled; the runner names the pull command rather than downloading multi-GB blobs unasked. Do NOT hand-roll the env — go through the runner so the setup stays identical to everyone else's.
4. Run the chosen scope through the runner, e.g.:
   • one class: scripts/golden.sh -pl <module> -Dtest=<GoldenClass>
   • several: scripts/golden.sh -pl <module> -Dtest='<GoldenA>,<GoldenB>'
   • one method: scripts/golden.sh -pl <module> -Dtest=<GoldenClass>#<method>
   • all: run per module (loop the -pl <module> -Dtest=<its Golden*> addresses). The runner leaves the engine + gateway up, so the next run is instant.
5. Read the result — real regression vs flaky case. On failure, extract only the failing assertion (e.g. «…» should route to 'finance' but went to 'tasks'), not the full log. Then apply the flakiness rule below before calling it a bug.
6. Report. A short per-class pass/fail; for each failure, the assertion + a verdict: real regression (stable, reproducible) or flaky borderline case (input too ambiguous for the model — tighten the test, not the product).

Flakiness — the rule that keeps goldens trustworthy

A small local model is non-deterministic on borderline inputs — a phrasing sitting between two domains can route one way now and another next run. So:

• A golden case must be crisp: unambiguous intent, one obvious answer. Vague or "clever" phrasings belong in exploratory notes, not a must-pass golden.
• On a failure, re-run that case once before concluding. If the verdict flips between identical runs, the case is too borderline — fix the test input (make it crisp) rather than chasing the model or loosening the assertion. Only a stable, reproducible failure is a real regression to investigate in the product (prompt / manifest / router).
• When you add a golden case, run it twice to confirm it is stable before committing — a flaky green is a debt that fails someone else later.

When NOT to reach for this

• You changed only non-LLM code (plumbing, DB, config with no prompt/router effect) — there is no LLM surface to re-validate; the normal unit/slice tests cover it.
• You just want the fast inner loop — goldens are the real-model gate, not the iterate loop. Use mocked slice tests while iterating; run the goldens before the PR for the surface you touched.

Triggering contract (examples)

SHOULD fire: "прогони голдены по роутингу", "run the finance golden", "проверь, не сломал ли я синтез на реальной модели", "run all the goldens before I open the PR". SHOULD NOT fire: "run the unit tests" (not the real-model gate), "why is my build failing" (compile, not goldens), "add a golden test" (that is authoring — new-skill / write the test; this skill only runs them).