Imported from MaCoredroid/Lumo_FlyWheel-qwen-diffusion (
benchmark_blueprints/families/responses-tool-schema-cutover/workspace_bundle/v1-clean-baseline/AGENTS.md). Install upstream withnpx skills add MaCoredroid/Lumo_FlyWheel-qwen-diffusion --skill v1-clean-baseline. Copyright stays with the author.
Responses Replay Repair
Variant: v1-clean-baseline
Visible baseline: two same-name tool calls, out-of-order tool results, no extra distractors.
Required solver behavior:
- Inspect at least one replay fixture before patching.
- Repair both
gateway/adapter.pyandgateway/reducer.pyif needed. - Keep repeated same-name tool calls distinct by
call_idall the way through render output. - Keep the public CLI summary format stable for successful runs.
- Update
codex/config.tomlanddocs/migrations/responses-cutover.mdto the Responses-native contract.
Anti-shortcuts:
- Do not dedupe repeated tool calls by tool name.
- Do not synthesize
call_idfrom fixture position or visible ordinals. - Do not delete fixtures or relax tests to get green.