Pipeline Orchestration + Retry Cure
Goal
run-pipeline <program> composes build->run->verify->benchmark->record end-to-end deterministically; a mechanical failure returns a structured needs-fix result the skill routes to /add-bug + /fix-bug + re-run (verify-first / zero-halt).
Implementation Sketch
Scripts own every mechanical action; the LLM authors programs/tests and analyzes only. All mechanical work is a single python -m scripts.rosetta_test_runtime <subcommand> invocation directed by a work_item body line.
Work Items
- Implement
pipeline.pycomposing the §00-§03 mechanical actions in order; emit ONE structured JSON result (per-stage pass/fail + needs_fix classification + gate verdicts + recorded artifact paths). - Define the needs-fix cure contract: a mechanical failure (compile error, wrong output, leak, RC imbalance, checksum mismatch, perf loss) emits
needs_fix{stage, subsystem, repro}the skill surfaces for LLM RCA -> /add-bug + /fix-bug -> re-run; NO workaround, NO #skip, never weaken a baseline. - Author
scripts/rosetta_test_runtime/next_action.schema.jsonfor the structured exit; pytest covers the happy path + each needs_fix class;--self-testgreen.