[ajitpratap0]

Summary

Bug: since #331, gosqlx_version() has returned a sync.Once-cached C string. Its doc comment, gosqlx_free's doc, and pygosqlx.version() all treat that string as caller-owned and free it. As a result, the second pygosqlx.version() call in a process double-frees. macOS aborts with pointer being freed was not allocated; on glibc it is heap corruption. I reproduced it on main: pytest python/tests aborts in TestVersion::test_version_format.

Fix (C side follows the documented contract):

• gosqlx_version now returns a fresh C.CString(VERSION) on every call, and I removed the cache. The comments on gosqlx_version and gosqlx_free now state the ownership rule and that gosqlx_free(NULL) is a no-op.
• Export audit: every other export returns a fresh allocation via toJSON on both success and error paths, including the marshal-failure fallback. None returns NULL or a static pointer. NULL input goes through C.GoString(NULL) and becomes "". gosqlx_version was the only mismatch in the library.
• Second ownership bug, in the raw ctypes harness: pkg/cbinding/tests/test_ctypes.py declared results as c_char_p. That makes ctypes return a Python bytes copy, so gosqlx_free received a pointer it never allocated. It now uses c_void_p plus string_at, and it frees the version string too.

Tests

• Go: getVersion() now frees its result; previously the stress test's "cached singleton, just read it" comment covered a call that never freed. The new TestCGosqlxVersion_CallerOwned calls gosqlx_version twice, asserts distinct pointers and equal strings equal to VERSION, and frees both. I confirmed it fails on the old code. Also added TestCGosqlxFree_Nil. I renamed TestStressVersionSingleton to TestStressVersionConcurrent.
• Python: TestVersion::test_version_called_twice is the regression test. test_package_version.py asserts pygosqlx.__version__ and the setup.py version both equal the pyproject version.

CI: the new .github/workflows/python.yml runs on PRs and on pushes to main, filtered to pkg/**, python/**, go.mod, go.sum and the workflow file. It covers ubuntu-latest and macos-latest with Python 3.10 and 3.13, on Go 1.26. Steps: build with pkg/cbinding/build.sh, fail if the library is missing (so integration tests can't skip without anyone noticing), pip install -e "python[dev]", a two-call version() smoke test, pytest python/tests -v -rs, and the raw ctypes ABI tests. Actions are pinned to the same SHAs the repo already uses (checkout v4.4.0, setup-go v5.6.0, setup-python v5.6.0). Permissions are contents: read.

Design decisions I made (scoped worker, conservative choices):

1. pygosqlx.__version__ was "0.1.0" while pyproject and setup.py declare 0.2.0. The required __version__ == pyproject test can only pass if they agree, so I changed __version__ to "0.2.0". This is byte-identical to the same hunk in release PR docs(release): prepare v1.16.0 #845, so the two merge cleanly in either order. The package version itself is unchanged.
2. pip install -e "python[dev]" fails on current setuptools because setup.py passes long_description while pyproject's [project] table doesn't declare readme, which raises AttributeError in _long_description. I added readme = "README.md" to [project]. This changes metadata only, not the API or the version.
3. I reused the repo's existing setup-python pin (v5.6.0) instead of introducing v7, for consistency.

Test Plan

• Reproduced the abort on main with go build -buildmode=c-shared + GOSQLX_LIB_PATH=… pytest python/tests
• gofmt -l pkg/cbinding (clean), go vet ./pkg/cbinding/, golangci-lint run ./pkg/cbinding/... (0 issues)
• CGO_ENABLED=1 go test -race -count=1 ./pkg/cbinding/ passes; the new regression test fails against the old cbinding.go
• Simulated the workflow locally (macOS, Python 3.15): build.sh, pip install -e "python[dev]", smoke test, pytest python/tests (79 passed), test_ctypes.py (25 OK); pip wheel ./python builds pygosqlx-0.2.0
• Python Bindings workflow green on all 4 matrix legs

Breaking Changes

None for correct callers. A C caller that relied on the old (undocumented) behaviour and never freed the gosqlx_version() result now leaks a few bytes per call. The documented contract has always required gosqlx_free.

Release note

- **cbinding / pygosqlx:** fixed a crash where calling `pygosqlx.version()` more than once in a process aborted with "pointer being freed was not allocated" (double free). `gosqlx_version()` now returns a fresh, caller-owned string on every call, as documented; free it with `gosqlx_free`. `pygosqlx.__version__` now reports the packaged version (0.2.0), and editable installs work with current setuptools. A new Python Bindings CI workflow tests pygosqlx on Linux and macOS (Python 3.10 and 3.13).

🤖 Generated with Claude Code

https://claude.ai/code/session_01RkzVu3JY9Y6BRWfjesWts9

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
gosqlx-website	Ignored	Preview	Oct 21, 2026 6:15pm UTC

[github-actions]

Claude finished @ajitpratap0's task in 15s —— View job

---

Review of PR #847

Verdict: looks good. The fix is correct and minimal, and the tests target the real failure. I only read the diff and did not run the tests or CI. Nothing below blocks merging.

Fix (pkg/cbinding/cbinding.go)

• gosqlx_version now returns a fresh C.CString(VERSION) on every call. That matches the documented ownership contract and removes the double free.
• Dropping sync.Once and the sync import is clean. The cost is a few bytes allocated per call, which is acceptable.
• The doc updates on gosqlx_free (NULL is a no-op) and gosqlx_version are accurate.

Tests

• versionTwiceFreed skips the second free when the pointers match. That avoids crashing the test run on a regression, so the failure is reported as a clean assertion.
• TestCGosqlxVersion_CallerOwned is a good regression test. Distinct pointers are a reasonable proxy for "fresh allocation", though the allocator could in theory reuse an address. That can't happen here, because the first pointer is still live when the second call runs.
• test_ctypes.py: switching to c_void_p plus string_at is the correct way to keep the original pointer for gosqlx_free. Setting argtypes is a good hardening. In _call, the result_ptr is None check still works, because c_void_p returns None for NULL.
• test_version_called_twice and test_package_version.py are fine. Minor nit: the setup.py regex version\s*=\s*"..." takes the first match in the file, so it could pick up an unrelated python_requires-style or other version= string if one is added later. It is fine for now.

CI (.github/workflows/python.yml)

• The permissions, SHA pinning, concurrency group and fail-fast: false are all sensible.
• The "Locate shared library" step fails loudly if the library is missing, so integration tests can't skip silently. That's a good call.
• test -f and the uname case run under shell: bash, which is fine on both runners.
• Suggestion: add cache: pip to setup-python, and cache-dependency-path: python/pyproject.toml, to speed the matrix up. This is optional.
• The paths filter omits root-level files such as go.work, or a Taskfile that builds the library. Probably fine, but worth a thought.

Packaging

• __version__ = "0.2.0" and readme = "README.md" are sensible. The __version__ hunk is identical to docs(release): prepare v1.16.0 #845, so the PRs merge cleanly in either order.
• Confirm that python/README.md exists and is included in the sdist. A missing file would break pip install.

Remaining item: the checklist still shows the Python Bindings workflow as not yet green on all 4 matrix legs. Please confirm that before merging.

[github-actions]

Claude finished @ajitpratap0's task in 15s —— View job

---

Review of PR #847

Verdict: looks good. I read the diff and checked a few files in the checkout. I did not run the tests or CI.

• Read the diff and the surrounding files
• Check packaging (python/README.md)
• Post review

Fix (pkg/cbinding/cbinding.go)

• gosqlx_version returns a fresh C.CString(VERSION) on every call. That matches the documented ownership contract and removes the double free.
• Removing sync.Once and the sync import is clean. The cost is a few bytes per call.
• The doc updates on gosqlx_free (NULL is a no-op) and gosqlx_version are accurate.

Tests

• versionTwiceFreed skips the second free when the pointers are identical. A regression is then reported as a clean assertion instead of crashing the run.
• getVersion() now frees its result, so the stress test follows the contract. TestCGosqlxVersion_CallerOwned is a good regression test, and you confirmed it fails on the old code.
• test_ctypes.py: c_void_p plus string_at is the correct way to keep the original pointer for gosqlx_free.
• Nit: test_package_version.py reads the setup.py version with a regex that takes the first version = "..." match. It would pick up the wrong string if another version= literal is added earlier in the file. It is fine today.

CI (.github/workflows/python.yml)

• Permissions (contents: read), SHA pinning, the concurrency group and fail-fast: false are all sensible.
• The "Locate shared library" step fails loudly if the library is missing, so integration tests can't skip silently.
• Optional: add cache: pip and cache-dependency-path: python/pyproject.toml to setup-python to speed up the matrix.
• Optional: the paths filter omits go.work. It probably isn't needed, but check.

Packaging

• python/README.md exists, readme = "README.md" is now in pyproject.toml, and MANIFEST.in includes the README. The sdist concern is covered.
• The __version__ = "0.2.0" hunk is byte-identical to the one in docs(release): prepare v1.16.0 #845, so the two PRs merge cleanly in either order.

Before merging: the test plan still shows the Python Bindings workflow as not green on all 4 matrix legs. Please confirm that.