[adewale]

This PR now carries two earlier PRs, folded together at the owner's request. #4 is merged in with a merge commit, so its commits keep their original SHAs in this history.

From #3 (this PR's original change): strict known-bug semantics, uniform startup, lint in CI

• Strict known-bug semantics.
  • Resolved issue 2 is now a plain regression assertion.
  • Active issues 3, 4a and 4b use @pytest.mark.xfail(strict=True, raises=PlatformBugReproduced). If an upstream fix lands, the run fails with XPASS(strict) and prints NOT REPRODUCED ... move it to Resolved Issues.
• tests/conftest.py.
  • Startup budgets stay at main's values: 300 s for 2- under CI, 30 s otherwise. PYWRANGLER_DEV_TIMEOUT overrides them when set by hand.
  • Server output is drained on a background thread.
  • An early pywrangler exit fails immediately.
  • Teardown kills the whole process group, so wrangler/workerd processes no longer leak.
• CI.
  • The install step warms every repro directory.
  • pre-commit run --all-files runs as a step in main's existing test job, not as a new lint job. It uses SKIP=ruff, because main already runs ruff check.
  • Ruff is main's 0.16.0 (Upgrade Ruff to 0.16.0 and make Workers CI reliable #2), which supersedes this PR's 0.15.2 pin.
• ruff format cleanup of files that already failed the hook on main.
• README "Run the tests" section, and LESSONS_LEARNED entries 9 and 10 (renumbered after main's new entries 7 and 8).

From #4: unbreak main, byte-level issue 2 oracle, wire-level issue 5 oracle

• Product fix. Pins fastapi<0.142 in 4-. 2- no longer needs the pin, because main (Upgrade Ruff to 0.16.0 and make Workers CI reliable #2) now imports its FastAPI app inside the request handler.
  • FastAPI 0.142 imports opentelemetry-api, which calls uuid4() at import time. That raises Errno 29 Randomness is not allowed during Worker startup.
  • pywrangler vendors from pyproject.toml, not the lockfile, so the new release broke main without any commit.
• test_2. /seed now writes 32768 big-endian words, where word i holds i. /read is checked word by word, so a reordered, repeated or zeroed chunk is caught.
• test_5. Superseded by main (Upgrade Ruff to 0.16.0 and make Workers CI reliable #2). Its ECHO_URL Worker variable and header_echo_server fixture assert on what the local echo server received. This PR's ?echo= override, its loopback validation and bypass check, tests/echo_server.py and the live httpbin.org call were dropped in the merge.

Deferred patch applied

• 7a68976: test_2 checks the /stream body word by word. Test audit: unbreak main (pin fastapi below 0.142), byte-level issue 2 oracle, wire-level issue 5 oracle #4 wrote this as an exact patch and deferred it only because Strict known-bug tests, issue 2 byte oracle, issue 4 fastapi<0.142 pin, pre-commit hooks in CI #3 rewrote those lines. It is applied verbatim and ruff formatted.

Follow-ups still open

• Delete the unused tests/conftest.py::_worker_name_from_config. No longer blocked: Upgrade Ruff to 0.16.0 and make Workers CI reliable #2 is merged into main, and main is merged into this branch. Not done here.
• Leads from Test audit: unbreak main (pin fastapi below 0.142), byte-level issue 2 oracle, wire-level issue 5 oracle #4 that were never implemented. These are not patches, so they were not applied:
  • A wire-level echo-server oracle for test_3_httpx_headers. Done by main (Upgrade Ruff to 0.16.0 and make Workers CI reliable #2): test_3 now asserts on header_echo_server and keeps this PR's strict marker.
  • Issue 4 still seeds with bytes(range(256)) in three places.
  • Report "fastapi>=0.142 cannot be imported in Python Workers" upstream, or add it as a repro directory.
  • CI liveness: whether to add a scheduled run. Ruled out: the owner's verification-cost standard allows no new schedules.
  • A LESSONS_LEARNED.md entry for Test audit: unbreak main (pin fastapi below 0.142), byte-level issue 2 oracle, wire-level issue 5 oracle #4's findings.

Local checks on the combined tree (7a68976)

These results predate the merge of main and the pruning. The current checks are under Cost and limits.

Commands are taken from .github/workflows/ci.yml and run with CI=true. Results are decided by exit status.

Command	Result
uv sync --dev && for dir in [0-9]*-*/; do (cd "$dir" && uv run pywrangler dev --help); done (under set -e)	exit 0
uv run pre-commit run --all-files --show-diff-on-failure (lint job)	exit 0, all hooks passed
uv run pytest -vv -rxXs (test job)	exit 0: test_2 PASSED, test_5 PASSED, test_3 XFAIL, 4a/4b/4c SKIPPED (deploy-only). 2 passed, 3 skipped, 1 xfailed in 36 s
git merge-tree --write-tree HEAD origin/main	clean; origin/main is already an ancestor, so no main merge was needed

• The merge of origin/claude/test-audit-campaign was clean: tests/test_examples.py auto-merged.
• test_5 contacts httpbin.org, a public echo service, as the test itself does.
• Skipped:
  • semgrep.yml, which needs SEMGREP_APP_TOKEN;
  • deployed tests 4a, 4b and 4c, which need --deploy;
  • GitHub Actions itself, which was not run.

Needs the owner

• Deployed tests 4a, 4b and 4c. Run --deploy. 4a and 4b now carry strict markers, so their outcome is unverified, and a 4b XPASS(strict) is worth investigating. The fastapi<0.142 pin's effect on a deployed issue 4 Worker is also unverified.
• PR Upgrade Ruff to 0.16.0 and make Workers CI reliable #2. Already merged into main. main is now merged into this branch (merge commit 24bf211), so there is no merge order left to decide.
• Scheduled CI run. Not added: the owner's verification-cost standard allows no new schedules.

Cost and limits

Pruned to the owner's verification-cost standard on 2026-10-10.

main merged in first. Merge commit 24bf211 brings in #2 (Ruff 0.16.0, reliable Workers CI) and #5 (CI lessons). Where main already did what this PR did, main's version won:

• Ruff 0.16.0 replaces this PR's 0.15.2 pin.
• In 2-, main imports the FastAPI app inside the request handler, so 2- no longer needs the fastapi<0.142 pin.
• For issues 3 and 5, main has the ECHO_URL Worker variable and the header_echo_server fixture. They replace this PR's ?echo= override, its loopback validation and bypass check, tests/echo_server.py, and the live httpbin.org call.
• Dev-server teardown is main's process-group stop. No wrangler/workerd process from this repo was left after a local run.

Lockfiles were regenerated with uv lock (uv 0.12.3), not hand-merged.

What CI runs for this PR. It is still one workflow (ci.yml) with one job (test). There is no new workflow, job, matrix entry or schedule.

• Install step. It warms every repro directory with uv run pywrangler dev --help, not only 2-. The first local warm took 30, 18, 16 and 14 s for 2-, 3-, 4- and 5-. For 3- and 5-, that time would otherwise be spent inside their timed dev-server start, so the net addition is mostly the 16 s for 4-. No CI test starts 4-; warming it only checks that its dependencies install and vendor.
• New step after main's uv run ruff check .. It runs SKIP=ruff uv run pre-commit run --all-files --show-diff-on-failure: ruff format, codespell and file hygiene. It replaces this PR's separate lint job, so ruff check runs once. It took 17-19 s locally with an empty hook cache.
• Tests. Same uv run pytest -vv and the same three local tests. Locally: 41 s, with 2 passed, 3 skipped and 1 xfailed. Each dev server was ready in 10-13 s.
• Added per run: roughly 35-60 s. These are local timings; GitHub runner time was not measured.
• Timeouts and retries. None were raised. The dev-server startup budgets are main's: 300 s for 2- under CI, 30 s otherwise. This PR had raised every directory to 300 s under CI, and that was reverted. PYWRANGLER_DEV_TIMEOUT is a manual override that CI never sets. There are no retries.

Removed from CI

• The lint job, folded into the test job as described above.
• The live httpbin.org call in test_5. main took httpbin out of CI on purpose.

Manual only (exact commands)

• Live HTTPS from requests/urllib3 to a real host (issue 5): cd 5-sync-http-libraries && uv run pywrangler dev, then curl http://localhost:8787/test. Without ECHO_URL, the Worker calls httpbin.org.

• Deployed tests 4a, 4b and 4c (unchanged, always manual): uv run pytest -vv -rxXs -k test_4 --deploy or --deployed-url <url>. This deploys to Cloudflare, so it was not run here.

• Issue 4 Worker startup: cd 4-r2-large-binary-roundtrip && uv run pywrangler dev, then curl http://localhost:8787/. Checked locally:

  • with fastapi<0.142, the Worker vendors FastAPI 0.141.1 and / returns 200;
  • without the pin, it vendors 0.143.0 and fails at startup with Errno 29 Randomness is not allowed.

  CI does not start issue 4.

• Cold local test run: PYWRANGLER_DEV_TIMEOUT=120 uv run pytest -vv, or warm each directory first with uv run pywrangler dev --help.

PBT, fuzzing and mutation testing. This PR has none, so there are no budgets to set and no mutation lanes to remove.

Checks on 11ba67c. Run with CI=true UV_PYTHON=3.13, uv 0.12.3 and Python 3.13.15. Results are decided by exit status.

CI command	Result
uv sync --dev + for dir in [0-9]*-*/; do (cd "$dir" && uv run pywrangler dev --help); done	exit 0
uv run pytest -vv	exit 0: 2 passed, 3 skipped, 1 xfailed, 41 s
uv run ruff check .	exit 0
SKIP=ruff uv run pre-commit run --all-files --show-diff-on-failure	exit 0: all hooks passed

Not verified

• GitHub Actions itself (not run).
• Deployed tests 4a, 4b and 4c, and the effect of the issue 4 pin on a deployed Worker.
• semgrep.yml, which needs SEMGREP_APP_TOKEN; it is unchanged.
• Live HTTPS to a real host for issues 3 and 5. This is manual only, as on main.
• Timings on a GitHub runner.

Permission denials: none.

🤖 Generated with Claude Code

https://claude.ai/code/session_012E6xQEfTtV2ws3dCSMnoCo

---

Generated by Claude Code