mirror of
https://github.com/razzant/ouroboros.git
synced 2026-10-15 17:36:54 +00:00
915f93604d
The documented local command (`make test`, CONTRIBUTING §4, README) was `pytest tests/ -q --tb=short`: the whole default suite (23k tests) in a single process, behind a stale "smoke tests" comment. An Ouroboros task followed it, hit its 1200 s timeout and withheld its PR. The correct recipe — a parallel `not serial` xdist pass, then the serial pass — lived only inside the hermetic commit gate. scripts/run_tests.py is the single entry point for every actor. A bare call always runs the FULL battery: the node lane (a missing or too-old node is a red NOT_RUN, never a silent skip), then every default-lane test in ONE xdist run under `--dist loadgroup --serial-shards=N`. tests/conftest.py gains that option: each serial FILE is pinned to one of N file-sharded groups (one per four workers, at most four; four cores keep a single group), so serial tests overlap the parallel ones while a serial file still never splits across workers or runs beside another file of its shard; other tests keep the scope loadscope uses. The option defaults to 0 and is inert for CI and the commit gate, which keep the two-pass split and stay the authority. `--sequential` keeps the gate's marker split and xdist flags in two passes; extra arguments forward a focused run. Marker exclusions and xdist flags are imported from ouroboros.preflight_runner, never restated. Measured on an 18-core Mac shared with other agents: the two-pass battery took 649 s (parallel 203 s + serial 446 s); the one-run battery took 217–287 s with identical verdicts (22 950 passed). Also: `make test` now calls the script; CONTRIBUTING, README, the Build & CI chapter, the architecture map and one sentence of prompts/SYSTEM.md point at it and say which opt-in lanes it leaves out and what to record when a full run could not finish; tests/test_plan_author_presentation.py resolves node through ouroboros.preflight_node like the rest of the suite instead of a bare PATH `node`; tests/test_run_tests_script.py pins the recipe parity, the bounded shard count, bare call = full battery, node NOT_RUN = red, and the grouping semantics. Not changed: the commit gate and .github/workflows/ci.yml; the three gate self-tests that wait out real 30/20 s timeouts (their slack exists for nested-xdist startup on slow runners); no version carrier. Co-authored-by: Ouroboros <311266734+ouroboros-agent@users.noreply.github.com> |
||
|---|---|---|
| .. | ||
| 01-role-and-authority.md | ||
| 02-naming-and-boundaries.md | ||
| 03-module-size-and-complexity.md | ||
| 04-core-governance-artifacts.md | ||
| 05-review-and-commit-protocol.md | ||
| 06-rules-by-change-class.md | ||
| 07-managed-update-rule.md | ||
| 08-mutation-attribution-rule.md | ||
| 09-process-custody-rule.md | ||
| 10-platform-abstraction-rule.md | ||
| 11-design-system.md | ||
| 12-mcp-client-integration.md | ||
| 13-gateway-boundary-pattern.md | ||
| 14-build-and-ci.md | ||