[dmccoystephenson]

Summary

• _send now opens each report through a module-level opener (_OPENER) whose HTTPRedirectHandler declines every redirect, in place of urllib.request.urlopen. A 301/302/303 no longer re-sends Authorization: Bearer <key> to the Location host as an empty GET; every 3xx reaches the existing non-201 DEBUG line with its own status (answered 302) and the report is dropped.
• install_opener was deliberately not used: it would change urlopen for the whole host program.
• README "The wire format" gains a paragraph saying redirects are not followed, which matches the trace-client-java README's "Redirects are not followed" quoted in A 301, 302 or 303 answer sends the reporting key to the redirect target as an empty GET #31.

Closes #31

Test plan

• New test_a_redirect_is_logged_and_dropped_never_followed_with_the_key covers 301, 302, 303, 307 and 308. The origin redirects to a second loopback server reached as localhost on another port. The test asserts that the target receives nothing and that the log names the redirect status at DEBUG.
• Regression evidence (stash-and-run on Python 3.8.10): with the trace_client.py change stashed, the new test fails because GET /elsewhere reached the other host with the key. With the change restored, it passes.
• test_an_error_answer_whose_body_cannot_be_read_still_logs_its_status now patches _OPENER.open instead of urllib.request.urlopen, because urlopen is no longer the call path. Its assertions are unchanged.
• python3 -m unittest (Python 3.8.10): Ran 73 tests, OK.
• CI matrix 3.8 / 3.10 / 3.12. Python 3.11 added 308 handling to HTTPRedirectHandler, so the 3.12 job is the check for that row.

Java parity

This brings Python in line with trace-client-java, which by its README already does not follow redirects. No Java change is needed. The Java README could not be re-read from this session's sandbox, so the parity claim relies on the quotation verified when #31 was filed.

Merge gate

This PR changes _send's behaviour on the wire, which is on this loop's do-not-auto-merge list (wire format must stay in lockstep with trace-client-java and the trace server). It is left open for owner review rather than merged autonomously, as #31 itself requested.

Deferred

• The version tag is added even when it pushes an event past MAX_TAGS #16 (version tag past MAX_TAGS) was not picked up. It still needs an owner decision on the README promise that tags "always holds at least version", and a change made in step with trace-client-java (see the comment on The version tag is added even when it pushes an event past MAX_TAGS #16).

This PR description was drafted during a Gardener session (https://github.com/Stephenson-Software/gardener).

🤖 Generated with Claude Code

---

drafted by Claude on behalf of Daniel Stephenson

[dmccoystephenson]

Self-review rubric (scored against the diff and command output; one intrinsic pass):

• Scope: PASS. Three files changed, all for A 301, 302 or 303 answer sends the reporting key to the redirect target as an empty GET #31: trace_client.py (handler and opener), the test file (new test, handler gains Location/GET, one mock retargeted), and README.md (wire-format paragraph). 47 insertions, 2 deletions.
• Tests-new: PASS. No public name was added. _RedirectsNotFollowed/_OPENER are private and are exercised end-to-end by test_a_redirect_is_logged_and_dropped_never_followed_with_the_key.
• Tests-fix: PASS. Confirmed by stash-and-run on Python 3.8.10: with trace_client.py stashed, the new test FAILED (GET /elsewhere reached the other host carrying the key, as described in A 301, 302 or 303 answer sends the reporting key to the redirect target as an empty GET #31), and after restoring it the test PASSED.
• Sibling structure: PASS. No new files. The test follows the loop-over-statuses shape of test_a_success_status_other_than_201_is_logged_as_not_delivered.
• Sibling renames: PASS. Nothing was renamed.
• Docs: PASS. README "The wire format" now states the redirect behaviour. The class docstring's "down, slow, or rejecting" promise still holds, because a redirect is one more dropped report.
• Issue resolution: PASS. Both parts of the suggested fix in A 301, 302 or 303 answer sends the reporting key to the redirect target as an empty GET #31 are done: a non-following opener, and the redirect paragraph from the Java README.
• CI: PASS. test (3.8), test (3.10) and test (3.12) are green on the head. 3.12 covers the post-3.11 308 handling.
• Stdlib-only: PASS. The only added import is the package's own module in the tests. pyproject.toml is untouched.
• Py3.8 floor: PASS. test (3.8) is green, and the change was also run locally on 3.8.10.
• Never-raises / never-blocks: PASS. _OPENER.open sits in the same try as the urlopen it replaces. A declined redirect surfaces as HTTPError, which the existing branch handles. report() is unchanged.
• Log level: PASS. No _LOG call was added. The existing DEBUG line reports answered 3xx.
• Version agreement: PASS. No version string changed.
• Exports: PASS. No public names were added.
• Java parity noted: PASS. The PR body states that no Java change is needed and that the quotation from the Java README is taken from A 301, 302 or 303 answer sends the reporting key to the redirect target as an empty GET #31.

Judgment call for the owner:

• trace_client/trace_client.py, at the _OPENER = urllib.request.build_opener(...) line: the opener is built at import time, so its ProxyHandler reads the proxy environment variables (HTTPS_PROXY etc.) at import. Before this change, urlopen built its cached opener on the first report. A program that sets proxy variables after importing the client but before its first report would no longer have them honoured. Building the opener lazily in _send (or per request) would restore the old timing if that case matters.

Merge gate: wire-format change (_send), which is on the do-not-auto-merge list. Left open for owner review.

This review comment was drafted during a Gardener session (https://github.com/Stephenson-Software/gardener).

---

drafted by Claude on behalf of Daniel Stephenson